diff --git a/docs/guides/configuration.md b/docs/guides/configuration.md
index 8fbb467b..64130bb0 100644
--- a/docs/guides/configuration.md
+++ b/docs/guides/configuration.md
@@ -26,7 +26,6 @@ tool_timeout_sec: 20
generate_start_timeout_sec: 90
runtime:
- max_no_progress_streak: 5
max_repeat_cycle_streak: 3
max_turns: 90
hooks:
@@ -108,9 +107,8 @@ context:
| 字段 | 说明 |
|------|------|
-| `runtime.max_no_progress_streak` | 连续“无进展”轮次提醒阈值，默认 `5`；达到 `limit-1` 起会向模型注入纠偏提示，不会直接终止运行 |
-| `runtime.max_repeat_cycle_streak` | 连续“重复调用同一工具参数”提醒阈值，默认 `3`；达到阈值后触发重复循环提醒，不会直接终止运行 |
-| `runtime.max_turns` | 单次 Run 的最大推理轮数上限，默认 `40`；达到上限后直接终止并返回明确 stop reason |
+| `runtime.max_repeat_cycle_streak` | 连续“相同工具签名 + 相同结果指纹 + 相同子目标”阈值，默认 `3`；达到阈值后先注入重复循环提醒，提醒后仍重复则终止为 `repeat_cycle` |
+| `runtime.max_turns` | 单次 Run 的最大推理轮数上限，默认 `90`；达到上限后直接终止并返回明确 stop reason |
| `runtime.hooks.enabled` | hooks 总开关；关闭后不执行 runtime hooks |
| `runtime.hooks.user_hooks_enabled` | user hooks 开关；关闭后不加载 `runtime.hooks.items` |
| `runtime.hooks.default_timeout_sec` | user hook 默认超时秒数，需 `> 0` |
diff --git a/docs/stop-reason-and-decision-priority.md b/docs/stop-reason-and-decision-priority.md
index b637c676..03ce1ec0 100644
--- a/docs/stop-reason-and-decision-priority.md
+++ b/docs/stop-reason-and-decision-priority.md
@@ -7,9 +7,11 @@
- `max_turn_exceeded`
- `verification_failed`
- `accepted`
+- `missing_completion_signal`
+- `accept_check_failed`
- `todo_not_converged`
- `todo_waiting_external`
-- `no_progress_after_final_intercept`
+- `repeat_cycle`
- `max_turn_exceeded_with_unconverged_todos`
- `max_turn_exceeded_with_failed_verification`
- `verification_config_missing`
diff --git a/internal/cli/gateway_runtime_bridge.go b/internal/cli/gateway_runtime_bridge.go
index bd97e11a..a5dd41ae 100644
--- a/internal/cli/gateway_runtime_bridge.go
+++ b/internal/cli/gateway_runtime_bridge.go
@@ -1703,19 +1703,16 @@ func convertRuntimeSnapshot(snapshot agentruntime.RuntimeSnapshot) gateway.Runti
RunID: strings.TrimSpace(snapshot.RunID),
SessionID: strings.TrimSpace(snapshot.SessionID),
Phase: strings.TrimSpace(snapshot.Phase),
- TaskKind: strings.TrimSpace(snapshot.TaskKind),
UpdatedAt: snapshot.UpdatedAt,
Todos: convertRuntimeTodoSnapshot(snapshot.Todos),
Facts: map[string]any{
"runtime_facts": snapshot.Facts.RuntimeFacts,
},
Decision: map[string]any{
- "status": strings.TrimSpace(snapshot.Decision.Status),
- "stop_reason": strings.TrimSpace(snapshot.Decision.StopReason),
- "missing_facts": snapshot.Decision.MissingFacts,
- "required_next_actions": snapshot.Decision.RequiredNextActions,
- "user_visible_summary": strings.TrimSpace(snapshot.Decision.UserVisibleSummary),
- "internal_summary": strings.TrimSpace(snapshot.Decision.InternalSummary),
+ "status": strings.TrimSpace(snapshot.Decision.Status),
+ "stop_reason": strings.TrimSpace(snapshot.Decision.StopReason),
+ "summary": strings.TrimSpace(snapshot.Decision.Summary),
+ "details": append([]string(nil), snapshot.Decision.Details...),
},
SubAgents: map[string]any{
"started_count": snapshot.SubAgents.StartedCount,
diff --git a/internal/cli/gateway_runtime_bridge_test.go b/internal/cli/gateway_runtime_bridge_test.go
index 1bfaa619..10020c71 100644
--- a/internal/cli/gateway_runtime_bridge_test.go
+++ b/internal/cli/gateway_runtime_bridge_test.go
@@ -1009,7 +1009,6 @@ func TestGatewayRuntimePortBridgeListSessionTodosAndSnapshot(t *testing.T) {
RunID: "run-1",
SessionID: "session-2",
Phase: "acceptance",
- TaskKind: "workspace_write",
Decision: agentruntime.DecisionSnapshot{Status: "continue", StopReason: "unverified_write"},
SubAgents: agentruntime.SubAgentSnapshot{StartedCount: 1, CompletedCount: 1, FailedCount: 0},
},
diff --git a/internal/config/config_test.go b/internal/config/config_test.go
index 149307f9..c6a6cfd5 100644
--- a/internal/config/config_test.go
+++ b/internal/config/config_test.go
@@ -1721,7 +1721,6 @@ func TestValidateSnapshotPropagatesCompactError(t *testing.T) {
},
},
Runtime: RuntimeConfig{
- MaxNoProgressStreak: 3,
MaxRepeatCycleStreak: 3,
},
Context: ContextConfig{
@@ -1830,7 +1829,7 @@ func TestParseCurrentConfigRoundTripRuntimeConfig(t *testing.T) {
t.Parallel()
snapshot := testDefaultConfig().Clone()
- snapshot.Runtime.MaxNoProgressStreak = 5
+ snapshot.Runtime.MaxRepeatCycleStreak = 5
data, err := marshalPersistedConfig(snapshot)
if err != nil {
@@ -1841,8 +1840,8 @@ func TestParseCurrentConfigRoundTripRuntimeConfig(t *testing.T) {
if err != nil {
t.Fatalf("parseCurrentConfig() error = %v", err)
}
- if parsed.Runtime.MaxNoProgressStreak != 5 {
- t.Fatalf("expected max_no_progress_streak=5, got %d", parsed.Runtime.MaxNoProgressStreak)
+ if parsed.Runtime.MaxRepeatCycleStreak != 5 {
+ t.Fatalf("expected max_repeat_cycle_streak=5, got %d", parsed.Runtime.MaxRepeatCycleStreak)
}
}
@@ -1854,7 +1853,7 @@ selected_provider: openai
current_model: gpt-4.1
shell: bash
runtime:
- max_no_progress_streak: -2
+ max_repeat_cycle_streak: -2
`)
parsed, err := parseCurrentConfig(raw, StaticDefaults().Context, StaticDefaults().Memo)
@@ -1866,9 +1865,9 @@ runtime:
if err := parsed.ValidateSnapshot(); err != nil {
t.Fatalf("ValidateSnapshot() error = %v", err)
}
- if parsed.Runtime.MaxNoProgressStreak != DefaultMaxNoProgressStreak {
- t.Fatalf("expected default max_no_progress_streak=%d, got %d",
- DefaultMaxNoProgressStreak, parsed.Runtime.MaxNoProgressStreak)
+ if parsed.Runtime.MaxRepeatCycleStreak != DefaultMaxRepeatCycleStreak {
+ t.Fatalf("expected default max_repeat_cycle_streak=%d, got %d",
+ DefaultMaxRepeatCycleStreak, parsed.Runtime.MaxRepeatCycleStreak)
}
}
diff --git a/internal/config/context_budget_migration.go b/internal/config/context_budget_migration.go
index acf2e4bc..36fcbfed 100644
--- a/internal/config/context_budget_migration.go
+++ b/internal/config/context_budget_migration.go
@@ -134,6 +134,14 @@ func MigrateContextBudgetConfigContent(raw []byte) ([]byte, bool, []string, erro
if verificationChanged {
changed = true
}
+ legacyRuntimeChanged := migrateLegacyRuntimeConfigFields(doc)
+ if legacyRuntimeChanged {
+ changed = true
+ }
+ legacyMemoChanged := migrateLegacyMemoConfigFields(doc)
+ if legacyMemoChanged {
+ changed = true
+ }
if !changed {
return raw, false, nil, nil
@@ -146,6 +154,60 @@ func MigrateContextBudgetConfigContent(raw []byte) ([]byte, bool, []string, erro
return out, true, notes, nil
}
+// migrateLegacyRuntimeConfigFields 清理 runtime 下已废弃且会导致严格解析失败的历史字段。
+func migrateLegacyRuntimeConfigFields(doc map[string]any) bool {
+ runtimeValue, ok := doc["runtime"]
+ if !ok {
+ return false
+ }
+ runtimeMap, ok := migrationStringMap(runtimeValue)
+ if !ok {
+ return false
+ }
+
+ changed := false
+ for _, key := range []string{"max_no_progress_streak"} {
+ if _, exists := runtimeMap[key]; exists {
+ delete(runtimeMap, key)
+ changed = true
+ }
+ }
+ if !changed {
+ return false
+ }
+ doc["runtime"] = runtimeMap
+ return true
+}
+
+// migrateLegacyMemoConfigFields 清理 memo 下已移除且可安全丢弃的历史字段。
+func migrateLegacyMemoConfigFields(doc map[string]any) bool {
+ memoValue, ok := doc["memo"]
+ if !ok {
+ return false
+ }
+ memoMap, ok := migrationStringMap(memoValue)
+ if !ok {
+ return false
+ }
+
+ changed := false
+ for _, key := range []string{"extract_recent_messages"} {
+ if _, exists := memoMap[key]; exists {
+ delete(memoMap, key)
+ changed = true
+ }
+ }
+ if !changed {
+ return false
+ }
+ if len(memoMap) == 0 {
+ delete(doc, "memo")
+ } else {
+ doc["memo"] = memoMap
+ }
+ return true
+}
+
// migrateVerificationConfig 清理已废弃的 verification 字段，并将安全的旧 command string 收敛成 argv。
func migrateVerificationConfig(doc map[string]any) (bool, error) {
runtimeValue, ok := doc["runtime"]
@@ -166,7 +228,7 @@ func migrateVerificationConfig(doc map[string]any) (bool, error) {
}
changed := false
- for _, key := range []string{"enabled", "default_task_policy", "final_intercept", "max_retries", "hooks"} {
+ for _, key := range []string{"enabled", "default_task_policy", "final_intercept", "max_retries", "hooks", "max_no_progress"} {
if _, exists := verificationMap[key]; exists {
delete(verificationMap, key)
changed = true
diff --git a/internal/config/context_budget_migration_test.go b/internal/config/context_budget_migration_test.go
index a31deb68..51e648ce 100644
--- a/internal/config/context_budget_migration_test.go
+++ b/internal/config/context_budget_migration_test.go
@@ -206,6 +206,97 @@ runtime:
}
}
+func TestMigrateContextBudgetConfigContentRemovesLegacyRuntimeNoProgressField(t *testing.T) {
+ t.Parallel()
+
+ input := []byte(strings.TrimSpace(`
+runtime:
+ max_no_progress_streak: 5
+ max_repeat_cycle_streak: 3
+`) + "\n")
+
+ out, changed, notes, err := MigrateContextBudgetConfigContent(input)
+ if err != nil {
+ t.Fatalf("MigrateContextBudgetConfigContent() error = %v", err)
+ }
+ if !changed {
+ t.Fatal("expected migration change")
+ }
+ if len(notes) != 0 {
+ t.Fatalf("expected no migration notes, got %v", notes)
+ }
+
+ text := string(out)
+ if strings.Contains(text, "max_no_progress_streak") {
+ t.Fatalf("expected max_no_progress_streak removed, got:\n%s", text)
+ }
+ if !strings.Contains(text, "max_repeat_cycle_streak: 3") {
+ t.Fatalf("expected max_repeat_cycle_streak preserved, got:\n%s", text)
+ }
+}
+
+func TestMigrateContextBudgetConfigContentRemovesLegacyVerificationNoProgressField(t *testing.T) {
+ t.Parallel()
+
+ input := []byte(strings.TrimSpace(`
+runtime:
+ verification:
+ max_no_progress: 3
+ verifiers:
+ test:
+ timeout_sec: 30
+`) + "\n")
+
+ out, changed, notes, err := MigrateContextBudgetConfigContent(input)
+ if err != nil {
+ t.Fatalf("MigrateContextBudgetConfigContent() error = %v", err)
+ }
+ if !changed {
+ t.Fatal("expected migration change")
+ }
+ if len(notes) != 0 {
+ t.Fatalf("expected no migration notes, got %v", notes)
+ }
+
+ text := string(out)
+ if strings.Contains(text, "max_no_progress") {
+ t.Fatalf("expected max_no_progress removed, got:\n%s", text)
+ }
+ if !strings.Contains(text, "timeout_sec: 30") {
+ t.Fatalf("expected verifier config preserved, got:\n%s", text)
+ }
+}
+
+func TestMigrateContextBudgetConfigContentRemovesLegacyMemoExtractRecentMessagesField(t *testing.T) {
+ t.Parallel()
+
+ input := []byte(strings.TrimSpace(`
+memo:
+ auto_extract: true
+ extract_recent_messages: 4
+ extract_timeout_sec: 9
+`) + "\n")
+
+ out, changed, notes, err := MigrateContextBudgetConfigContent(input)
+ if err != nil {
+ t.Fatalf("MigrateContextBudgetConfigContent() error = %v", err)
+ }
+ if !changed {
+ t.Fatal("expected migration change")
+ }
+ if len(notes) != 0 {
+ t.Fatalf("expected no migration notes, got %v", notes)
+ }
+
+ text := string(out)
+ if strings.Contains(text, "extract_recent_messages") {
+ t.Fatalf("expected extract_recent_messages removed, got:\n%s", text)
+ }
+ if !strings.Contains(text, "auto_extract: true") || !strings.Contains(text, "extract_timeout_sec: 9") {
+ t.Fatalf("expected supported memo fields preserved, got:\n%s", text)
+ }
+}
+
func TestMigrateContextBudgetConfigFileCreatesBackup(t *testing.T) {
t.Parallel()
diff --git a/internal/config/runtime.go b/internal/config/runtime.go
index 57dd1375..36e035a2 100644
--- a/internal/config/runtime.go
+++ b/internal/config/runtime.go
@@ -8,14 +8,12 @@ import (
)
const (
- DefaultMaxNoProgressStreak = 5
DefaultMaxRepeatCycleStreak = 3
DefaultMaxTurns = 90
)
// RuntimeConfig 定义 runtime 层的可调参数。
type RuntimeConfig struct {
- MaxNoProgressStreak int `yaml:"max_no_progress_streak,omitempty"`
MaxRepeatCycleStreak int `yaml:"max_repeat_cycle_streak,omitempty"`
MaxTurns int `yaml:"max_turns,omitempty"`
Verification VerificationConfig `yaml:"verification,omitempty"`
@@ -32,7 +30,6 @@ type RuntimeAssetsConfig struct {
// defaultRuntimeConfig 返回 runtime 配置的静态默认值。
func defaultRuntimeConfig() RuntimeConfig {
return RuntimeConfig{
- MaxNoProgressStreak: DefaultMaxNoProgressStreak,
MaxRepeatCycleStreak: DefaultMaxRepeatCycleStreak,
MaxTurns: DefaultMaxTurns,
Verification: defaultVerificationConfig(),
@@ -52,7 +49,6 @@ func defaultRuntimeAssetsConfig() RuntimeAssetsConfig {
// Clone 复制 runtime 配置，避免调用方共享可变状态。
func (c RuntimeConfig) Clone() RuntimeConfig {
return RuntimeConfig{
- MaxNoProgressStreak: c.MaxNoProgressStreak,
MaxRepeatCycleStreak: c.MaxRepeatCycleStreak,
MaxTurns: c.MaxTurns,
Verification: c.Verification.Clone(),
@@ -66,9 +62,6 @@ func (c *RuntimeConfig) ApplyDefaults(defaults RuntimeConfig) {
if c == nil {
return
}
- if c.MaxNoProgressStreak <= 0 {
- c.MaxNoProgressStreak = defaults.MaxNoProgressStreak
- }
if c.MaxRepeatCycleStreak <= 0 {
c.MaxRepeatCycleStreak = defaults.MaxRepeatCycleStreak
}
@@ -82,9 +75,6 @@ func (c *RuntimeConfig) ApplyDefaults(defaults RuntimeConfig) {
// Validate 校验 runtime 配置是否满足最小约束。
func (c RuntimeConfig) Validate() error {
- if c.MaxNoProgressStreak <= 0 {
- return errors.New("max_no_progress_streak must be greater than 0")
- }
if c.MaxRepeatCycleStreak <= 0 {
return errors.New("max_repeat_cycle_streak must be greater than 0")
}
diff --git a/internal/config/runtime_test.go b/internal/config/runtime_test.go
index 0c67477a..e88fdfd3 100644
--- a/internal/config/runtime_test.go
+++ b/internal/config/runtime_test.go
@@ -5,21 +5,15 @@ import "testing"
func TestRuntimeConfigCloneAndDefaults(t *testing.T) {
t.Parallel()
- cfg := RuntimeConfig{MaxNoProgressStreak: 7, MaxRepeatCycleStreak: 4, MaxTurns: 21}
+ cfg := RuntimeConfig{MaxRepeatCycleStreak: 4, MaxTurns: 21}
cloned := cfg.Clone()
- if cloned.MaxNoProgressStreak != 7 || cloned.MaxRepeatCycleStreak != 4 || cloned.MaxTurns != 21 {
+ if cloned.MaxRepeatCycleStreak != 4 || cloned.MaxTurns != 21 {
t.Fatalf("Clone() mismatch: %+v", cloned)
}
defaults := defaultRuntimeConfig()
var zero RuntimeConfig
zero.ApplyDefaults(defaults)
- if zero.MaxNoProgressStreak != defaults.MaxNoProgressStreak {
- t.Fatalf("MaxNoProgressStreak = %d, want %d", zero.MaxNoProgressStreak, defaults.MaxNoProgressStreak)
- }
- if zero.Verification.MaxNoProgress != defaults.Verification.MaxNoProgress {
- t.Fatalf("Verification.MaxNoProgress = %d, want %d", zero.Verification.MaxNoProgress, defaults.Verification.MaxNoProgress)
- }
if len(zero.Verification.Verifiers) == 0 {
t.Fatal("expected default verifiers to be populated")
}
@@ -28,25 +22,20 @@ func TestRuntimeConfigCloneAndDefaults(t *testing.T) {
func TestRuntimeConfigValidate(t *testing.T) {
t.Parallel()
- if err := (RuntimeConfig{MaxNoProgressStreak: 1, MaxRepeatCycleStreak: 1, MaxTurns: 1}).Validate(); err != nil {
+ if err := (RuntimeConfig{MaxRepeatCycleStreak: 1, MaxTurns: 1}).Validate(); err != nil {
t.Fatalf("expected valid config, got %v", err)
}
- if err := (RuntimeConfig{MaxNoProgressStreak: 0, MaxRepeatCycleStreak: 1, MaxTurns: 1}).Validate(); err == nil {
- t.Fatal("expected max_no_progress_streak validation error")
- }
- if err := (RuntimeConfig{MaxNoProgressStreak: 1, MaxRepeatCycleStreak: 0, MaxTurns: 1}).Validate(); err == nil {
+ if err := (RuntimeConfig{MaxRepeatCycleStreak: 0, MaxTurns: 1}).Validate(); err == nil {
t.Fatal("expected max_repeat_cycle_streak validation error")
}
- if err := (RuntimeConfig{MaxNoProgressStreak: 1, MaxRepeatCycleStreak: 1, MaxTurns: -1}).Validate(); err == nil {
+ if err := (RuntimeConfig{MaxRepeatCycleStreak: 1, MaxTurns: -1}).Validate(); err == nil {
t.Fatal("expected max_turns validation error")
}
err := (RuntimeConfig{
- MaxNoProgressStreak: 1,
MaxRepeatCycleStreak: 1,
MaxTurns: 1,
Verification: VerificationConfig{
- MaxNoProgress: 1,
Verifiers: map[string]VerifierConfig{
"": {},
},
diff --git a/internal/config/verification.go b/internal/config/verification.go
index f578a58f..b6d46dc1 100644
--- a/internal/config/verification.go
+++ b/internal/config/verification.go
@@ -11,7 +11,6 @@ const (
verifierFileExists = "file_exists"
verifierContentMatch = "content_match"
verifierCommandSuccess = "command_success"
- verifierGitDiff = "git_diff"
verifierBuild = "build"
verifierTest = "test"
verifierLint = "lint"
@@ -51,7 +50,6 @@ var defaultVerificationDeniedCommands = []string{
// VerificationConfig 定义 runtime final 验收阶段的 verifier 执行配置。
type VerificationConfig struct {
- MaxNoProgress int `yaml:"max_no_progress,omitempty"`
Verifiers map[string]VerifierConfig `yaml:"verifiers,omitempty"`
ExecutionPolicy VerificationExecutionPolicyConfig `yaml:"execution_policy,omitempty"`
}
@@ -77,7 +75,6 @@ type VerificationExecutionPolicyConfig struct {
// defaultVerificationConfig 返回验证引擎默认策略。
func defaultVerificationConfig() VerificationConfig {
return VerificationConfig{
- MaxNoProgress: 2,
Verifiers: map[string]VerifierConfig{
verifierTodoConvergence: {
TimeoutSec: 5,
@@ -99,12 +96,6 @@ func defaultVerificationConfig() VerificationConfig {
OutputCapBytes: 128 * 1024,
Scope: verificationScopeProject,
},
- verifierGitDiff: {
- Command: []string{"git", "status", "--porcelain", "--untracked-files=normal"},
- TimeoutSec: 15,
- OutputCapBytes: 64 * 1024,
- Scope: verificationScopeProject,
- },
verifierBuild: {
TimeoutSec: 300,
OutputCapBytes: 256 * 1024,
@@ -145,7 +136,6 @@ func defaultVerificationExecutionPolicyConfig() VerificationExecutionPolicyConfi
// Clone 复制 verification 配置，避免 map/slice 共享底层数据。
func (c VerificationConfig) Clone() VerificationConfig {
cloned := VerificationConfig{
- MaxNoProgress: c.MaxNoProgress,
ExecutionPolicy: c.ExecutionPolicy.Clone(),
}
if len(c.Verifiers) > 0 {
@@ -162,9 +152,6 @@ func (c *VerificationConfig) ApplyDefaults(defaults VerificationConfig) {
if c == nil {
return
}
- if c.MaxNoProgress <= 0 {
- c.MaxNoProgress = defaults.MaxNoProgress
- }
if c.Verifiers == nil {
c.Verifiers = make(map[string]VerifierConfig, len(defaults.Verifiers))
}
@@ -182,9 +169,6 @@ func (c *VerificationConfig) ApplyDefaults(defaults VerificationConfig) {
// Validate 校验 verification 配置合法性。
func (c VerificationConfig) Validate() error {
- if c.MaxNoProgress <= 0 {
- return errors.New("runtime.verification.max_no_progress must be greater than 0")
- }
for name, verifier := range c.Verifiers {
if strings.TrimSpace(name) == "" {
return errors.New("runtime.verification.verifiers has empty name")
diff --git a/internal/config/verification_test.go b/internal/config/verification_test.go
index 4ce76a22..bd73b5e9 100644
--- a/internal/config/verification_test.go
+++ b/internal/config/verification_test.go
@@ -9,14 +9,11 @@ func TestVerificationConfigApplyDefaultsAndValidate(t *testing.T) {
cfg := VerificationConfig{}
cfg.ApplyDefaults(defaults)
- if cfg.MaxNoProgress != defaults.MaxNoProgress {
- t.Fatalf("MaxNoProgress = %d, want %d", cfg.MaxNoProgress, defaults.MaxNoProgress)
- }
if len(cfg.Verifiers) != len(defaults.Verifiers) {
t.Fatalf("verifier count = %d, want %d", len(cfg.Verifiers), len(defaults.Verifiers))
}
- if cfg.Verifiers[verifierGitDiff].Command[0] != "git" {
- t.Fatalf("expected git_diff default argv, got %#v", cfg.Verifiers[verifierGitDiff].Command)
+ if _, ok := cfg.Verifiers["git_diff"]; ok {
+ t.Fatal("git_diff verifier should not be configured by default")
}
if err := cfg.Validate(); err != nil {
t.Fatalf("Validate() error = %v", err)
@@ -27,12 +24,6 @@ func TestVerificationConfigValidateRejectsBadFields(t *testing.T) {
t.Parallel()
cfg := defaultVerificationConfig()
- cfg.MaxNoProgress = 0
- if err := cfg.Validate(); err == nil {
- t.Fatal("expected max_no_progress validation error")
- }
-
- cfg = defaultVerificationConfig()
cfg.Verifiers[" "] = VerifierConfig{}
if err := cfg.Validate(); err == nil {
t.Fatal("expected empty verifier name validation error")
diff --git a/internal/context/accept_checks_test.go b/internal/context/accept_checks_test.go
new file mode 100644
index 00000000..d3c88a4d
--- /dev/null
+++ b/internal/context/accept_checks_test.go
@@ -0,0 +1,11 @@
+package context
+
+import agentsession "neo-code/internal/session"
+
+func acceptText(items ...string) agentsession.AcceptChecks {
+ out := make(agentsession.AcceptChecks, 0, len(items))
+ for _, item := range items {
+ out = append(out, agentsession.AcceptCheck{Kind: agentsession.AcceptCheckOutputOnly, Target: item})
+ }
+ return out
+}
diff --git a/internal/context/builder_test.go b/internal/context/builder_test.go
index f0273319..5e45b0cc 100644
--- a/internal/context/builder_test.go
+++ b/internal/context/builder_test.go
@@ -159,13 +159,13 @@ func TestDefaultBuilderBuildIncludesPlanSections(t *testing.T) {
Goal: "引入 plan/build 模式",
Steps: []string{"扩展 session", "扩展 runtime"},
Constraints: []string{"保持 tools 边界"},
- Verify: []string{"go test ./internal/..."},
+ Verify: acceptText("go test ./internal/..."),
},
Summary: agentsession.SummaryView{
Goal: "引入 plan/build 模式",
KeySteps: []string{"扩展 session", "扩展 runtime"},
Constraints: []string{"保持 tools 边界"},
- Verify: []string{"go test ./internal/..."},
+ Verify: acceptText("go test ./internal/..."),
ActiveTodoIDs: []string{"todo-1"},
},
},
diff --git a/internal/context/source_plan_mode.go b/internal/context/source_plan_mode.go
index 5ca40fe9..910a5425 100644
--- a/internal/context/source_plan_mode.go
+++ b/internal/context/source_plan_mode.go
@@ -31,6 +31,13 @@ func (planModeContextSource) Sections(ctx context.Context, input BuildInput) ([]
}
if input.CurrentPlan == nil {
+ if stage == "plan" {
+ noPlanHint := promptSection{
+ Title: "Current Plan",
+ Content: "status: none\n\nNo current plan exists. You must create one by outputting a `plan_spec` + `summary_candidate` JSON before this turn ends.",
+ }
+ sections = append(sections, noPlanHint)
+ }
return sections, nil
}
planSection := renderCurrentPlanSection(input.CurrentPlan, input.InjectFullPlan)
@@ -80,7 +87,7 @@ func renderCurrentPlanSection(plan *agentsession.PlanArtifact, injectFull bool)
}
if len(plan.Summary.Verify) > 0 {
lines = append(lines, "verify:")
- for _, check := range plan.Summary.Verify {
+ for _, check := range plan.Summary.Verify.RenderLines() {
lines = append(lines, "- "+check)
}
}
diff --git a/internal/context/source_plan_mode_test.go b/internal/context/source_plan_mode_test.go
index ae0d7c12..39adf98e 100644
--- a/internal/context/source_plan_mode_test.go
+++ b/internal/context/source_plan_mode_test.go
@@ -111,7 +111,7 @@ func TestRenderCurrentPlanSectionInjectsFullPlan(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "完整计划",
Steps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
OpenQuestions: []string{"问题一"},
},
Summary: agentsession.SummaryView{
diff --git a/internal/context/source_todos.go b/internal/context/source_todos.go
index cb58cbf2..2fa8db1c 100644
--- a/internal/context/source_todos.go
+++ b/internal/context/source_todos.go
@@ -80,12 +80,6 @@ func (todosSource) Sections(ctx context.Context, input BuildInput) ([]promptSect
}
}
- lines = append(lines, "",
- "stale_todo_reminder: If any todo above is no longer relevant to the current task,",
- "or the user clearly switches to a different task, use todo_write to mark it completed",
- "only if the work is actually done; otherwise set_status=canceled before moving on.",
- )
-
return []promptSection{
{
Title: "Todo State",
diff --git a/internal/context/source_todos_test.go b/internal/context/source_todos_test.go
index 5276b9fe..a0bf7b66 100644
--- a/internal/context/source_todos_test.go
+++ b/internal/context/source_todos_test.go
@@ -61,11 +61,8 @@ func TestTodosSourceSections(t *testing.T) {
if len(lines) < 2 || !strings.Contains(lines[0], "in-progress") {
t.Fatalf("expected in_progress todo first, got %q", sections[0].Content)
}
- if !strings.Contains(sections[0].Content, "user clearly switches to a different task") {
- t.Fatalf("expected stale todo reminder to mention task switching, got %q", sections[0].Content)
- }
- if !strings.Contains(sections[0].Content, "only if the work is actually done") {
- t.Fatalf("expected stale todo reminder to distinguish completed from canceled, got %q", sections[0].Content)
+ if strings.Contains(sections[0].Content, "stale_todo_reminder") {
+ t.Fatalf("expected stale todo reminder to be removed, got %q", sections[0].Content)
}
}
diff --git a/internal/promptasset/assets.go b/internal/promptasset/assets.go
index 3a640f7b..642cb2c2 100644
--- a/internal/promptasset/assets.go
+++ b/internal/promptasset/assets.go
@@ -22,10 +22,12 @@ const (
var coreSections = loadCoreSections()
-var noProgressReminder = mustReadTemplate("templates/runtime/self_healing_no_progress.txt")
-
var repeatCycleReminder = mustReadTemplate("templates/runtime/self_healing_repeat_cycle.txt")
+var completionProtocolReminder = mustReadTemplate("templates/runtime/completion_protocol_reminder.md")
+
+var completionProtocolFinalReminder = mustReadTemplate("templates/runtime/completion_protocol_final_reminder.md")
+
var compactSystemPromptTemplate = mustReadTemplate("templates/context/compact_system_prompt.md")
var planModePlanPrompt = mustReadTemplate("templates/context/plan_mode_plan.md")
@@ -47,16 +49,21 @@ func CoreSections() []Section {
return append([]Section(nil), coreSections...)
}
-// NoProgressReminder 返回 runtime 无进展自愈提醒文案。
-func NoProgressReminder() string {
- return noProgressReminder
-}
-
// RepeatCycleReminder 返回 runtime 重复同参工具调用自愈提醒文案。
func RepeatCycleReminder() string {
return repeatCycleReminder
}
+// CompletionProtocolReminder 返回缺少 task_completion 时的普通协议提示。
+func CompletionProtocolReminder() string {
+ return completionProtocolReminder
+}
+
+// CompletionProtocolFinalReminder 返回缺少 task_completion 时的最终协议提示。
+func CompletionProtocolFinalReminder() string {
+ return completionProtocolFinalReminder
+}
+
// CompactSystemPrompt 返回 compact 场景使用的静态 system prompt。
func CompactSystemPrompt(taskStateContract string, summaryFormat string) string {
replacer := strings.NewReplacer(
diff --git a/internal/promptasset/assets_test.go b/internal/promptasset/assets_test.go
index 43c21456..169a09c8 100644
--- a/internal/promptasset/assets_test.go
+++ b/internal/promptasset/assets_test.go
@@ -58,12 +58,23 @@ func TestCorePromptContainsOperationalGuidance(t *testing.T) {
func TestRuntimeReminderTemplates(t *testing.T) {
t.Parallel()
- if !strings.Contains(NoProgressReminder(), "multiple consecutive attempts") {
- t.Fatalf("expected no-progress reminder guidance, got %q", NoProgressReminder())
- }
if !strings.Contains(RepeatCycleReminder(), "exact same arguments") {
t.Fatalf("expected repeat-cycle reminder guidance, got %q", RepeatCycleReminder())
}
+ for name, prompt := range map[string]string{
+ "completion": CompletionProtocolReminder(),
+ "final_completion": CompletionProtocolFinalReminder(),
+ } {
+ if !strings.Contains(prompt, "Completion retry rule") {
+ t.Fatalf("%s reminder should contain retry rule, got %q", name, prompt)
+ }
+ if !strings.Contains(prompt, "Do not repeat file lists") {
+ t.Fatalf("%s reminder should prevent repeated summaries, got %q", name, prompt)
+ }
+ if !strings.Contains(prompt, "at most one brief final sentence") {
+ t.Fatalf("%s reminder should keep final prose concise, got %q", name, prompt)
+ }
+ }
}
func TestPlanModePromptTemplates(t *testing.T) {
@@ -92,6 +103,12 @@ func TestPlanModePromptTemplates(t *testing.T) {
})
}
+ if !strings.Contains(PlanModePrompt("plan"), "summary_candidate.active_todo_ids") {
+ t.Fatalf("expected plan prompt to require active todo ownership")
+ }
+ if !strings.Contains(PlanModePrompt("build_execute"), "create current-run required todos") {
+ t.Fatalf("expected build prompt to require direct-build todo bootstrap")
+ }
if got := PlanModePrompt("unknown"); got != "" {
t.Fatalf("PlanModePrompt(unknown) = %q, want empty", got)
}
diff --git a/internal/promptasset/templates/context/plan_mode_build_execute.md b/internal/promptasset/templates/context/plan_mode_build_execute.md
index 25a024bc..7a964b5f 100644
--- a/internal/promptasset/templates/context/plan_mode_build_execute.md
+++ b/internal/promptasset/templates/context/plan_mode_build_execute.md
@@ -4,8 +4,12 @@ You are currently in build execution.
- If a current plan summary is attached, use it as guidance by default.
- If the summary is insufficient for the current task, consult the attached full plan view when available.
- If no current plan is attached, continue using task state, todos, and the conversation context.
+- If no current plan and no Todo State are attached, create current-run required todos with `todo_write` before the first substantive tool call for project analysis, documentation writing, code changes, multi-step debugging, or verification work.
+- Do not update or complete todo IDs that are not present in the current Todo State; create new current-run todos instead.
- Small necessary deviations are allowed, but explain why they are needed.
- Do not create or rewrite the current full plan in this stage.
- If the current plan appears outdated, explain the mismatch and continue, or recommend switching back to planning.
- Do not output `plan_spec` or `summary_candidate` in build execution.
-- When you believe the task tied to the current plan is complete, start your reply with a JSON object of the form `{"task_completion":{"completed":true}}`, then continue with the normal user-facing completion message.
+- When the task is complete, your final reply MUST start with `{"task_completion":{"completed":true}}` followed by your user-facing message. Without this signal, the runtime will issue up to two protocol reminders and then terminate the run.
+- Do NOT output `task_completion` while you still have tool calls to make. Tools always take priority over completion signals.
+- Acceptance is terminal: once you signal completion, the runtime performs a final yes/no check against the plan's verify criteria. If it fails, the run ends — there is no retry.
diff --git a/internal/promptasset/templates/context/plan_mode_plan.md b/internal/promptasset/templates/context/plan_mode_plan.md
index 7531cbce..0268dfdb 100644
--- a/internal/promptasset/templates/context/plan_mode_plan.md
+++ b/internal/promptasset/templates/context/plan_mode_plan.md
@@ -3,7 +3,15 @@ You are currently in the planning stage.
- You may research, analyze, ask clarifying questions, and produce a plan.
- Do not perform any write action in this stage.
- Do not rewrite the current full plan unless the conversation clearly requires creating or replacing the plan itself.
-- If you are only answering questions, comparing options, clarifying constraints, or refining details, do not output planning JSON.
+- **If no Current Plan section is attached, your first priority is to produce a plan.** The user has entered planning mode expecting a structured plan. Research the codebase as needed, then output a complete `plan_spec` + `summary_candidate` JSON. Do not end the turn with only a conversational answer when there is no existing plan.
+- If a Current Plan is already present, you may refine, replace, or discuss it. When the user asks a clarifying question or wants to explore options without committing to a new plan revision, you may answer conversationally without outputting planning JSON.
- Only output a JSON object containing `plan_spec` and `summary_candidate` when you are explicitly creating or rewriting the current full plan.
- `plan_spec` must include `goal`, `steps`, `constraints`, `verify`, `todos`, and `open_questions`.
+- `plan_spec.todos` **must not be empty**. Populate it with the major actionable items that the plan requires. Each todo must have a unique `id`, a descriptive `content`, and `status: "pending"`. Without todos the plan has no executable work items and the build stage cannot proceed.
- `summary_candidate` must include `goal`, `key_steps`, `constraints`, `verify`, and `active_todo_ids`.
+- If a Todo State section is attached, decide which non-terminal todos still belong to the current plan.
+- Todos that still belong to the current plan must appear in `plan_spec.todos` and their IDs must appear in `summary_candidate.active_todo_ids`.
+- Todos that do not belong to the current plan must not be copied into the new plan; create replacement plan-owned todos when ongoing work is still needed.
+- `verify` must be an array of structured check objects: `[{"kind":"...", "target":"...", "required":true}]`.
+- Supported `kind` values: `output_only` (chat/read-only), `workspace_change` (writes/edits), `command_success` (build/test/lint), `file_exists` (file artifacts), `content_contains` (content checks), `tool_fact` (named tool facts).
+- Examples: chat → `[{"kind":"output_only"}]`, fix → `[{"kind":"workspace_change"},{"kind":"command_success","target":"go test ./..."}]`, new file → `[{"kind":"workspace_change"},{"kind":"file_exists","target":"output.go"}]`.
diff --git a/internal/promptasset/templates/core/agent_identity.md b/internal/promptasset/templates/core/agent_identity.md
index c5c43e22..1547076c 100644
--- a/internal/promptasset/templates/core/agent_identity.md
+++ b/internal/promptasset/templates/core/agent_identity.md
@@ -43,7 +43,7 @@ Your final answer is only a completion candidate. It does not by itself prove th
Distinguish:
- `completion_gate`: whether it is reasonable to attempt finalization.
- `verification_gate`: whether the actual task requirements are satisfied.
-- `acceptance_decision`: the runtime's final accepted/continue/incomplete/failed decision.
+- `acceptance_decision`: the runtime's final accepted/failed decision. Acceptance is terminal — there is no "continue" or retry.
Do not finalize when any of these are true:
- Required todos are pending, in progress, blocked, or failed.
@@ -52,4 +52,4 @@ Do not finalize when any of these are true:
- Tool results indicate errors, truncation that affects confidence, or unresolved uncertainty.
- A subagent finished but the main task has not integrated and verified its result.
-If the runtime injects a reminder that completion was not accepted, continue execution and address the unmet condition. Do not argue with the reminder.
+If acceptance fails, the task is terminated. Do not try to continue — the run has ended.
diff --git a/internal/promptasset/templates/core/tool_usage.md b/internal/promptasset/templates/core/tool_usage.md
index c720174d..67598517 100644
--- a/internal/promptasset/templates/core/tool_usage.md
+++ b/internal/promptasset/templates/core/tool_usage.md
@@ -8,6 +8,10 @@
- Verify file content with `filesystem_read_file` + `expect_contains` and `verification_scope`; avoid `bash Get-Content` for routine checks.
- Use `filesystem_grep` to locate symbols, strings, and relevant code paths efficiently.
- Read tool results carefully before acting. Treat `status`, `ok`, `tool_call_id`, `truncated`, `meta.*`, exit codes, and `content` as the authoritative model-visible outcome of that call.
+- For explanation, Q&A, or concept-clarification requests, use tools only until you have enough evidence to answer.
+- After reading or searching the relevant files for an informational request, stop exploring and answer directly.
+- Do not restart the same interpretation loop after you already have enough context.
+- If two search/read passes do not change your conclusion, provide the answer and briefly state any remaining uncertainty.
## Repository exploration
When exploring the codebase, Git state, or current changes:
diff --git a/internal/promptasset/templates/runtime/completion_protocol_final_reminder.md b/internal/promptasset/templates/runtime/completion_protocol_final_reminder.md
new file mode 100644
index 00000000..bd9f374a
--- /dev/null
+++ b/internal/promptasset/templates/runtime/completion_protocol_final_reminder.md
@@ -0,0 +1,17 @@
+[Runtime Control]
+
+You again stopped calling tools without outputting `task_completion`.
+
+This is the final protocol reminder. Missing it again will terminate this run.
+
+Completion retry rule:
+Your previous prose may already be visible to the user. Do not duplicate, restate, expand, or re-list prior summaries.
+
+If the task is done:
+- Emit the required `task_completion` JSON exactly once.
+- After the JSON, write at most one brief final sentence.
+- Do not repeat file lists, completed steps, tool results, or previous summaries.
+
+If the task is not done:
+- Continue with the next necessary tool call.
+- Do not write another prose summary until the work is actually complete.
diff --git a/internal/promptasset/templates/runtime/completion_protocol_reminder.md b/internal/promptasset/templates/runtime/completion_protocol_reminder.md
new file mode 100644
index 00000000..f9deeada
--- /dev/null
+++ b/internal/promptasset/templates/runtime/completion_protocol_reminder.md
@@ -0,0 +1,15 @@
+[Runtime Control]
+
+You stopped calling tools without outputting `task_completion`.
+
+Completion retry rule:
+Your previous prose may already be visible to the user. Do not duplicate, restate, expand, or re-list prior summaries.
+
+If the task is done:
+- Emit the required `task_completion` JSON exactly once.
+- After the JSON, write at most one brief final sentence.
+- Do not repeat file lists, completed steps, tool results, or previous summaries.
+
+If the task is not done:
+- Continue with the next necessary tool call.
+- Do not write another prose summary until the work is actually complete.
diff --git a/internal/repository/path.go b/internal/repository/path.go
index b88d2fe3..83855dbc 100644
--- a/internal/repository/path.go
+++ b/internal/repository/path.go
@@ -185,7 +185,9 @@ func walkWorkspaceFiles(
func skipDirEntry(entry fs.DirEntry) bool {
name := strings.ToLower(strings.TrimSpace(entry.Name()))
switch name {
- case ".git", ".idea", ".vscode", "node_modules":
+ case ".git", ".idea", ".vscode", "node_modules",
+ ".cache", ".tmp", "tmp", "build", "dist", "out", "target", "coverage",
+ ".next", ".nuxt", ".turbo", ".parcel-cache", ".vite", "vendor", "bin", "obj":
return true
default:
return false
diff --git a/internal/repository/repository_coverage_test.go b/internal/repository/repository_coverage_test.go
index b025de3b..c0a30232 100644
--- a/internal/repository/repository_coverage_test.go
+++ b/internal/repository/repository_coverage_test.go
@@ -376,7 +376,14 @@ func TestRepositoryHelpersAndGitParsing(t *testing.T) {
workdir := t.TempDir()
mustWriteRepositoryFile(t, filepath.Join(workdir, "pkg", "a.go"), "package pkg\n\nconst Name = \"Widget\"\n")
mustWriteRepositoryFile(t, filepath.Join(workdir, "pkg", "b.txt"), "Widget appears twice\nWidget\n")
- mustWriteRepositoryFile(t, filepath.Join(workdir, "node_modules", "ignored.txt"), "ignored")
+ skippedDirs := []string{
+ ".cache", ".tmp", "tmp", "build", "dist", "out", "target", "coverage",
+ ".next", ".nuxt", ".turbo", ".parcel-cache", ".vite", "vendor", "bin", "obj",
+ "node_modules",
+ }
+ for _, dir := range skippedDirs {
+ mustWriteRepositoryFile(t, filepath.Join(workdir, dir, "ignored.txt"), "ignored")
+ }
if _, _, _, err := normalizeRetrievalQuery(workdir, RetrievalQuery{Mode: RetrievalModePath, Value: " "}); err == nil {
t.Fatal("expected empty query error")
@@ -417,14 +424,20 @@ func TestRepositoryHelpersAndGitParsing(t *testing.T) {
var visited []string
err = walkWorkspaceFiles(context.Background(), workdir, workdir, func(path string) error {
- visited = append(visited, filepath.Base(path))
+ rel, relErr := filepath.Rel(workdir, path)
+ if relErr != nil {
+ return relErr
+ }
+ visited = append(visited, filepath.Clean(rel))
return nil
})
if err != nil {
t.Fatalf("walkWorkspaceFiles() error = %v", err)
}
- if slices.Contains(visited, "ignored.txt") {
- t.Fatalf("expected node_modules to be skipped, got %v", visited)
+ for _, dir := range skippedDirs {
+ if slices.Contains(visited, filepath.Clean(filepath.Join(dir, "ignored.txt"))) {
+ t.Fatalf("expected %s to be skipped, got %v", dir, visited)
+ }
}
stopErr := errors.New("stop")
if err := walkWorkspaceFiles(context.Background(), workdir, workdir, func(path string) error { return stopErr }); !errors.Is(err, stopErr) {
@@ -508,6 +521,8 @@ func TestRepositoryReadSearchAndServiceEntrypoints(t *testing.T) {
mustWriteRepositoryFile(t, filepath.Join(workdir, "notes.py"), "def py_symbol():\n return 1\n")
mustWriteRepositoryFile(t, filepath.Join(workdir, ".env"), "SECRET=1\n")
mustWriteRepositoryFile(t, filepath.Join(workdir, "pkg", "bin.dat"), string([]byte{0x00, 0x01, 0x02}))
+ mustWriteRepositoryFile(t, filepath.Join(workdir, ".cache", "hidden.txt"), "hidden_alpha\n")
+ mustWriteRepositoryFile(t, filepath.Join(workdir, "build", "hidden.go"), "package build\n\nfunc HiddenWidget() {}\n")
service := NewService()
@@ -539,6 +554,13 @@ func TestRepositoryReadSearchAndServiceEntrypoints(t *testing.T) {
if len(textResult.Hits) != 1 || !textResult.Truncated || textResult.TotalCount == 0 {
t.Fatalf("unexpected text search result: %+v", textResult)
}
+ skippedTextResult, err := service.SearchText(context.Background(), workdir, "hidden_alpha", SearchOptions{Limit: 10})
+ if err != nil {
+ t.Fatalf("SearchText(skipped) error = %v", err)
+ }
+ if len(skippedTextResult.Hits) != 0 {
+ t.Fatalf("expected skipped directory text to be ignored, got %+v", skippedTextResult)
+ }
symbolResult, err := service.SearchSymbol(context.Background(), workdir, "BuildWidget", SearchOptions{Limit: 10})
if err != nil {
@@ -563,6 +585,13 @@ func TestRepositoryReadSearchAndServiceEntrypoints(t *testing.T) {
if len(fallbackResult.Hits) == 0 || fallbackResult.Hits[0].Kind != "reference" {
t.Fatalf("unexpected fallback symbol result: %+v", fallbackResult)
}
+ skippedSymbolResult, err := service.SearchSymbol(context.Background(), workdir, "HiddenWidget", SearchOptions{Limit: 10})
+ if err != nil {
+ t.Fatalf("SearchSymbol(skipped) error = %v", err)
+ }
+ if len(skippedSymbolResult.Hits) != 0 {
+ t.Fatalf("expected skipped directory symbol to be ignored, got %+v", skippedSymbolResult)
+ }
if got := extractGoSignature("func BuildWidget(\n\tname string,\n) string {\n\treturn name\n}\n", 1); !strings.Contains(got, "name string") {
t.Fatalf("extractGoSignature(multiline) = %q", got)
diff --git a/internal/runtime/accept_checks_test.go b/internal/runtime/accept_checks_test.go
new file mode 100644
index 00000000..2758e1be
--- /dev/null
+++ b/internal/runtime/accept_checks_test.go
@@ -0,0 +1,11 @@
+package runtime
+
+import agentsession "neo-code/internal/session"
+
+func acceptText(items ...string) agentsession.AcceptChecks {
+ out := make(agentsession.AcceptChecks, 0, len(items))
+ for _, item := range items {
+ out = append(out, agentsession.AcceptCheck{Kind: agentsession.AcceptCheckOutputOnly, Target: item})
+ }
+ return out
+}
diff --git a/internal/runtime/acceptance/decider.go b/internal/runtime/acceptance/decider.go
deleted file mode 100644
index 40dc295a..00000000
--- a/internal/runtime/acceptance/decider.go
+++ /dev/null
@@ -1,17 +0,0 @@
-package acceptance
-
-import "neo-code/internal/runtime/controlplane"
-
-// TerminalStatusFromAcceptance 将 acceptance 决策映射到 runtime 终态枚举。
-func TerminalStatusFromAcceptance(status AcceptanceStatus) controlplane.TerminalStatus {
- switch status {
- case AcceptanceAccepted:
- return controlplane.TerminalStatusCompleted
- case AcceptanceFailed:
- return controlplane.TerminalStatusFailed
- case AcceptanceIncomplete:
- return controlplane.TerminalStatusIncomplete
- default:
- return controlplane.TerminalStatusContinue
- }
-}
diff --git a/internal/runtime/acceptance/decider_test.go b/internal/runtime/acceptance/decider_test.go
deleted file mode 100644
index ebaa9b93..00000000
--- a/internal/runtime/acceptance/decider_test.go
+++ /dev/null
@@ -1,32 +0,0 @@
-package acceptance
-
-import (
- "testing"
-
- "neo-code/internal/runtime/controlplane"
-)
-
-func TestTerminalStatusFromAcceptance(t *testing.T) {
- t.Parallel()
-
- cases := []struct {
- status AcceptanceStatus
- want controlplane.TerminalStatus
- }{
- {status: AcceptanceAccepted, want: controlplane.TerminalStatusCompleted},
- {status: AcceptanceFailed, want: controlplane.TerminalStatusFailed},
- {status: AcceptanceIncomplete, want: controlplane.TerminalStatusIncomplete},
- {status: AcceptanceContinue, want: controlplane.TerminalStatusContinue},
- {status: AcceptanceStatus("other"), want: controlplane.TerminalStatusContinue},
- }
-
- for _, tc := range cases {
- tc := tc
- t.Run(string(tc.status), func(t *testing.T) {
- t.Parallel()
- if got := TerminalStatusFromAcceptance(tc.status); got != tc.want {
- t.Fatalf("TerminalStatusFromAcceptance(%q) = %q, want %q", tc.status, got, tc.want)
- }
- })
- }
-}
diff --git a/internal/runtime/acceptance/engine.go b/internal/runtime/acceptance/engine.go
deleted file mode 100644
index 809e03ba..00000000
--- a/internal/runtime/acceptance/engine.go
+++ /dev/null
@@ -1,140 +0,0 @@
-package acceptance
-
-import (
- "context"
- "fmt"
-
- "neo-code/internal/runtime/controlplane"
- "neo-code/internal/runtime/verify"
-)
-
-// Engine 负责聚合 completion gate 与 verifier gate，并输出唯一的收尾决策。
-type Engine struct {
- policy AcceptancePolicy
-}
-
-// NewEngine 创建 acceptance engine。
-func NewEngine(policy AcceptancePolicy) *Engine {
- if policy == nil {
- policy = DefaultPolicy{}
- }
- return &Engine{policy: policy}
-}
-
-// EvaluateFinal 执行 final acceptance 主链，输出结构化终态决策。
-func (e *Engine) EvaluateFinal(ctx context.Context, input FinalAcceptanceInput) (AcceptanceDecision, error) {
- decision := AcceptanceDecision{
- Status: AcceptanceContinue,
- StopReason: controlplane.StopReasonTodoNotConverged,
- CompletionBlockedReason: input.CompletionGate.Reason,
- UserVisibleSummary: "当前回合尚未达到可收尾条件，继续执行。",
- InternalSummary: "completion gate did not pass",
- ContinueHint: "There are unfinished required todos or unmet acceptance checks. Continue execution. Do not finalize yet.",
- }
- if input.CompletionGate.Passed {
- verifiers, err := e.policy.ResolveVerifiers(input.VerificationInput)
- if err != nil {
- return AcceptanceDecision{
- Status: AcceptanceFailed,
- StopReason: controlplane.StopReasonVerificationConfigMissing,
- ErrorClass: verify.ErrorClassEnvMissing,
- CompletionBlockedReason: input.CompletionGate.Reason,
- UserVisibleSummary: "验收配置无效，任务失败。",
- InternalSummary: fmt.Sprintf("verification profile resolution failed: %v", err),
- }, nil
- }
- orch := verify.Orchestrator{Verifiers: verifiers}
- gateDecision, err := orch.RunFinalVerification(ctx, input.VerificationInput)
- if err != nil {
- return AcceptanceDecision{}, err
- }
- decision = aggregateVerificationDecision(gateDecision)
- }
- decision.CompletionBlockedReason = input.CompletionGate.Reason
-
- if input.NoProgressExceeded && decision.Status == AcceptanceContinue {
- decision.Status = AcceptanceIncomplete
- decision.StopReason = controlplane.StopReasonNoProgressAfterFinalIntercept
- decision.UserVisibleSummary = "多次拦截 final 且无进展，已停止并标记为未完成。"
- decision.InternalSummary = "no-progress breaker triggered after repeated final interception"
- }
-
- if input.MaxTurnsReached && decision.Status == AcceptanceContinue {
- decision.Status = AcceptanceIncomplete
- if decision.StopReason == controlplane.StopReasonVerificationFailed {
- decision.StopReason = controlplane.StopReasonMaxTurnExceededWithFailedVerification
- } else {
- decision.StopReason = controlplane.StopReasonMaxTurnExceededWithUnconvergedTodos
- }
- decision.UserVisibleSummary = fmt.Sprintf("达到最大轮次限制（%d），任务未完成。", input.MaxTurnsLimit)
- decision.InternalSummary = "max turn reached while final was still intercepted"
- }
-
- return decision, nil
-}
-
-// aggregateVerificationDecision 将 verifier gate 的首个非 pass 结果映射为 acceptance 决策。
-func aggregateVerificationDecision(gate verify.VerificationGateDecision) AcceptanceDecision {
- first := firstNonPassResult(gate.Results)
- if first == nil {
- return AcceptanceDecision{
- Status: AcceptanceAccepted,
- StopReason: controlplane.StopReasonAccepted,
- UserVisibleSummary: "任务通过验收，已完成。",
- InternalSummary: "completion gate and verification gate both passed",
- VerifierResults: append([]verify.VerificationResult(nil), gate.Results...),
- HasProgress: true,
- }
- }
-
- switch first.Status {
- case verify.VerificationSoftBlock:
- return AcceptanceDecision{
- Status: AcceptanceContinue,
- StopReason: controlplane.StopReasonTodoNotConverged,
- UserVisibleSummary: "仍有未满足的验收条件，继续执行。",
- InternalSummary: "first verifier returned soft_block",
- ContinueHint: "There are unfinished required todos or unmet acceptance checks. Continue execution. Do not finalize yet.",
- VerifierResults: append([]verify.VerificationResult(nil), gate.Results...),
- }
- case verify.VerificationHardBlock:
- reason := controlplane.StopReasonTodoNotConverged
- if first.WaitingExternal {
- reason = controlplane.StopReasonTodoWaitingExternal
- }
- return AcceptanceDecision{
- Status: AcceptanceIncomplete,
- StopReason: reason,
- UserVisibleSummary: "任务仍依赖外部条件，当前以未完成状态结束。",
- InternalSummary: "first verifier returned hard_block",
- VerifierResults: append([]verify.VerificationResult(nil), gate.Results...),
- WaitingExternal: first.WaitingExternal,
- }
- default:
- stopReason := gate.Reason
- if stopReason == "" || stopReason == controlplane.StopReasonAccepted {
- stopReason = controlplane.StopReasonVerificationFailed
- }
- return AcceptanceDecision{
- Status: AcceptanceFailed,
- StopReason: stopReason,
- ErrorClass: first.ErrorClass,
- UserVisibleSummary: "验证未通过，任务失败。",
- InternalSummary: "first verifier returned fail",
- VerifierResults: append([]verify.VerificationResult(nil), gate.Results...),
- Retryable: first.Retryable,
- }
- }
-}
-
-// firstNonPassResult 返回首个非 pass 的 verifier 结果。
-func firstNonPassResult(results []verify.VerificationResult) *verify.VerificationResult {
- for _, result := range results {
- if result.Status == verify.VerificationPass {
- continue
- }
- cloned := result
- return &cloned
- }
- return nil
-}
diff --git a/internal/runtime/acceptance/engine_test.go b/internal/runtime/acceptance/engine_test.go
deleted file mode 100644
index 78a98e5d..00000000
--- a/internal/runtime/acceptance/engine_test.go
+++ /dev/null
@@ -1,146 +0,0 @@
-package acceptance
-
-import (
- "context"
- "testing"
-
- "neo-code/internal/runtime/controlplane"
- "neo-code/internal/runtime/verify"
- agentsession "neo-code/internal/session"
-)
-
-type staticPolicy struct {
- verifiers []verify.FinalVerifier
- err error
-}
-
-func (p staticPolicy) ResolveVerifiers(input verify.FinalVerifyInput) ([]verify.FinalVerifier, error) {
- _ = input
- return p.verifiers, p.err
-}
-
-type staticVerifier struct {
- name string
- result verify.VerificationResult
-}
-
-func (v staticVerifier) Name() string { return v.name }
-func (v staticVerifier) VerifyFinal(ctx context.Context, input verify.FinalVerifyInput) (verify.VerificationResult, error) {
- _ = ctx
- _ = input
- return v.result, nil
-}
-
-func TestEngineEvaluateFinal(t *testing.T) {
- t.Parallel()
-
- makeInput := func() FinalAcceptanceInput {
- return FinalAcceptanceInput{
- CompletionGate: CompletionGateDecision{Passed: true},
- VerificationInput: verify.FinalVerifyInput{
- TaskState: verify.TaskStateSnapshot{VerificationProfile: string(agentsession.VerificationProfileTaskOnly)},
- },
- }
- }
-
- t.Run("completion gate false returns continue", func(t *testing.T) {
- t.Parallel()
- decision, err := NewEngine(staticPolicy{}).EvaluateFinal(context.Background(), FinalAcceptanceInput{
- CompletionGate: CompletionGateDecision{Passed: false},
- })
- if err != nil {
- t.Fatalf("EvaluateFinal() error = %v", err)
- }
- if decision.Status != AcceptanceContinue {
- t.Fatalf("status = %q, want continue", decision.Status)
- }
- })
-
- t.Run("invalid profile becomes structured failed decision", func(t *testing.T) {
- t.Parallel()
- decision, err := NewEngine(staticPolicy{err: context.DeadlineExceeded}).EvaluateFinal(context.Background(), makeInput())
- if err != nil {
- t.Fatalf("EvaluateFinal() error = %v", err)
- }
- if decision.Status != AcceptanceFailed || decision.StopReason != controlplane.StopReasonVerificationConfigMissing {
- t.Fatalf("unexpected decision: %+v", decision)
- }
- })
-
- t.Run("soft block returns continue", func(t *testing.T) {
- t.Parallel()
- decision, err := NewEngine(staticPolicy{
- verifiers: []verify.FinalVerifier{
- staticVerifier{name: "todo", result: verify.VerificationResult{Name: "todo", Status: verify.VerificationSoftBlock}},
- },
- }).EvaluateFinal(context.Background(), makeInput())
- if err != nil {
- t.Fatalf("EvaluateFinal() error = %v", err)
- }
- if decision.Status != AcceptanceContinue {
- t.Fatalf("status = %q, want continue", decision.Status)
- }
- })
-
- t.Run("hard block returns incomplete", func(t *testing.T) {
- t.Parallel()
- decision, err := NewEngine(staticPolicy{
- verifiers: []verify.FinalVerifier{
- staticVerifier{name: "todo", result: verify.VerificationResult{Name: "todo", Status: verify.VerificationHardBlock, WaitingExternal: true}},
- },
- }).EvaluateFinal(context.Background(), makeInput())
- if err != nil {
- t.Fatalf("EvaluateFinal() error = %v", err)
- }
- if decision.Status != AcceptanceIncomplete || decision.StopReason != controlplane.StopReasonTodoWaitingExternal {
- t.Fatalf("unexpected decision: %+v", decision)
- }
- })
-
- t.Run("fail returns failed", func(t *testing.T) {
- t.Parallel()
- decision, err := NewEngine(staticPolicy{
- verifiers: []verify.FinalVerifier{
- staticVerifier{name: "build", result: verify.VerificationResult{Name: "build", Status: verify.VerificationFail, ErrorClass: verify.ErrorClassEnvMissing}},
- },
- }).EvaluateFinal(context.Background(), makeInput())
- if err != nil {
- t.Fatalf("EvaluateFinal() error = %v", err)
- }
- if decision.Status != AcceptanceFailed || decision.StopReason != controlplane.StopReasonVerificationConfigMissing {
- t.Fatalf("unexpected decision: %+v", decision)
- }
- })
-
- t.Run("all pass returns accepted", func(t *testing.T) {
- t.Parallel()
- decision, err := NewEngine(staticPolicy{
- verifiers: []verify.FinalVerifier{
- staticVerifier{name: "todo", result: verify.VerificationResult{Name: "todo", Status: verify.VerificationPass}},
- },
- }).EvaluateFinal(context.Background(), makeInput())
- if err != nil {
- t.Fatalf("EvaluateFinal() error = %v", err)
- }
- if decision.Status != AcceptanceAccepted {
- t.Fatalf("status = %q, want accepted", decision.Status)
- }
- })
-
- t.Run("retry exhausted no longer overrides final decision", func(t *testing.T) {
- t.Parallel()
- input := makeInput()
- input.VerificationInput.Todos = []verify.TodoSnapshot{{ID: "todo-1", Required: true, RetryCount: 1, RetryLimit: 1}}
- decision, err := NewEngine(staticPolicy{
- verifiers: []verify.FinalVerifier{
- staticVerifier{name: "todo", result: verify.VerificationResult{Name: "todo", Status: verify.VerificationPass}},
- },
- }).EvaluateFinal(context.Background(), input)
- if err != nil {
- t.Fatalf("EvaluateFinal() error = %v", err)
- }
- if decision.Status != AcceptanceAccepted || decision.StopReason != controlplane.StopReasonAccepted {
- t.Fatalf("unexpected decision: %+v", decision)
- }
- })
-}
diff --git a/internal/runtime/acceptance/error_class.go b/internal/runtime/acceptance/error_class.go
deleted file mode 100644
index edda8928..00000000
--- a/internal/runtime/acceptance/error_class.go
+++ /dev/null
@@ -1,6 +0,0 @@
-package acceptance
-
-import "neo-code/internal/runtime/verify"
-
-// ErrorClass 复用 verifier 层统一错误分类枚举。
-type ErrorClass = verify.ErrorClass
diff --git a/internal/runtime/acceptance/policy.go b/internal/runtime/acceptance/policy.go
deleted file mode 100644
index 784bada9..00000000
--- a/internal/runtime/acceptance/policy.go
+++ /dev/null
@@ -1,84 +0,0 @@
-package acceptance
-
-import (
- "fmt"
- "strings"
-
- "neo-code/internal/runtime/verify"
- agentsession "neo-code/internal/session"
-)
-
-// AcceptancePolicy 定义 final 验收时 verifier 选择策略。
-type AcceptancePolicy interface {
- ResolveVerifiers(input verify.FinalVerifyInput) ([]verify.FinalVerifier, error)
-}
-
-// DefaultPolicy 按 session-owned verification profile 解析 verifier 列表。
-type DefaultPolicy struct {
- Executor verify.CommandExecutor
-}
-
-// ResolveVerifiers 依据 verification profile 生成固定 verifier 执行列表。
-func (p DefaultPolicy) ResolveVerifiers(input verify.FinalVerifyInput) ([]verify.FinalVerifier, error) {
- profile := agentsession.VerificationProfile(strings.TrimSpace(input.TaskState.VerificationProfile))
- if !profile.Valid() {
- return nil, fmt.Errorf("invalid verification profile %q", input.TaskState.VerificationProfile)
- }
- names := mappedVerifierNames(profile)
- if len(names) == 0 {
- return nil, fmt.Errorf("verification profile %q has no verifier mapping", profile)
- }
- verifiers := make([]verify.FinalVerifier, 0, len(names))
- for _, name := range names {
- if verifier := p.buildVerifier(name); verifier != nil {
- verifiers = append(verifiers, verifier)
- }
- }
- return verifiers, nil
-}
-
-// buildVerifier 基于名称构建 verifier 实例。
-func (p DefaultPolicy) buildVerifier(name string) verify.FinalVerifier {
- switch strings.TrimSpace(name) {
- case "todo_convergence":
- return verify.TodoConvergenceVerifier{}
- case "file_exists":
- return verify.FileExistsVerifier{}
- case "content_match":
- return verify.ContentMatchVerifier{}
- case "command_success":
- return verify.CommandSuccessVerifier{VerifierName: "command_success", Executor: p.Executor}
- case "git_diff":
- return verify.GitDiffVerifier{Executor: p.Executor}
- case "build":
- return verify.NewBuildVerifier(p.Executor)
- case "test":
- return verify.NewTestVerifier(p.Executor)
- case "lint":
- return verify.NewLintVerifier(p.Executor)
- case "typecheck":
- return verify.NewTypecheckVerifier(p.Executor)
- default:
- return nil
- }
-}
-
-// mappedVerifierNames 返回 verification profile 对应的 verifier 名称集合。
-func mappedVerifierNames(profile agentsession.VerificationProfile) []string {
- switch profile {
- case agentsession.VerificationProfileTaskOnly:
- return []string{"todo_convergence"}
- case agentsession.VerificationProfileCreateFile, agentsession.VerificationProfileDocs:
- return []string{"todo_convergence", "file_exists", "content_match"}
- case agentsession.VerificationProfileConfig:
- return []string{"todo_convergence", "file_exists", "content_match", "command_success"}
- case agentsession.VerificationProfileEditCode:
- return []string{"todo_convergence", "git_diff", "build", "test", "typecheck"}
- case agentsession.VerificationProfileFixBug:
- return []string{"todo_convergence", "git_diff", "test", "build", "typecheck"}
- case agentsession.VerificationProfileRefactor:
- return []string{"todo_convergence", "git_diff", "build", "test", "lint", "typecheck"}
- default:
- return nil
- }
-}
diff --git a/internal/runtime/acceptance/policy_test.go b/internal/runtime/acceptance/policy_test.go
deleted file mode 100644
index 96c5cf1a..00000000
--- a/internal/runtime/acceptance/policy_test.go
+++ /dev/null
@@ -1,72 +0,0 @@
-package acceptance
-
-import (
- "testing"
-
- "neo-code/internal/runtime/verify"
- agentsession "neo-code/internal/session"
-)
-
-func TestMappedVerifierNames(t *testing.T) {
- t.Parallel()
-
- cases := []struct {
- profile agentsession.VerificationProfile
- want []string
- }{
- {profile: agentsession.VerificationProfileTaskOnly, want: []string{"todo_convergence"}},
- {profile: agentsession.VerificationProfileCreateFile, want: []string{"todo_convergence", "file_exists", "content_match"}},
- {profile: agentsession.VerificationProfileConfig, want: []string{"todo_convergence", "file_exists", "content_match", "command_success"}},
- {profile: agentsession.VerificationProfileEditCode, want: []string{"todo_convergence", "git_diff", "build", "test", "typecheck"}},
- {profile: agentsession.VerificationProfileRefactor, want: []string{"todo_convergence", "git_diff", "build", "test", "lint", "typecheck"}},
- }
-
- for _, tc := range cases {
- got := mappedVerifierNames(tc.profile)
- if len(got) != len(tc.want) {
- t.Fatalf("%s len = %d, want %d", tc.profile, len(got), len(tc.want))
- }
- for i := range tc.want {
- if got[i] != tc.want[i] {
- t.Fatalf("%s[%d] = %q, want %q", tc.profile, i, got[i], tc.want[i])
- }
- }
- }
-}
-
-func TestDefaultPolicyResolveVerifiers(t *testing.T) {
- t.Parallel()
-
- verifiers, err := (DefaultPolicy{}).ResolveVerifiers(verify.FinalVerifyInput{
- TaskState: verify.TaskStateSnapshot{VerificationProfile: string(agentsession.VerificationProfileEditCode)},
- })
- if err != nil {
- t.Fatalf("ResolveVerifiers() error = %v", err)
- }
- if len(verifiers) != 5 {
- t.Fatalf("ResolveVerifiers() len = %d, want 5", len(verifiers))
- }
- if verifiers[0].Name() != "todo_convergence" || verifiers[1].Name() != "git_diff" {
- t.Fatalf("unexpected verifier order: %s, %s", verifiers[0].Name(), verifiers[1].Name())
- }
-}
-
-func TestDefaultPolicyResolveVerifiersRejectsInvalidProfile(t *testing.T) {
- t.Parallel()
-
- _, err := (DefaultPolicy{}).ResolveVerifiers(verify.FinalVerifyInput{
- TaskState: verify.TaskStateSnapshot{VerificationProfile: "unknown"},
- })
- if err == nil {
- t.Fatal("expected invalid profile error")
- }
-}
-
-func TestDefaultPolicyResolveVerifiersRejectsMissingProfile(t *testing.T) {
- t.Parallel()
-
- _, err := (DefaultPolicy{}).ResolveVerifiers(verify.FinalVerifyInput{})
- if err == nil {
- t.Fatal("expected missing profile error")
- }
-}
diff --git a/internal/runtime/acceptance/stop_reason.go b/internal/runtime/acceptance/stop_reason.go
deleted file mode 100644
index 47892eb4..00000000
--- a/internal/runtime/acceptance/stop_reason.go
+++ /dev/null
@@ -1,6 +0,0 @@
-package acceptance
-
-import "neo-code/internal/runtime/controlplane"
-
-// StopReason 复用控制面统一停止原因枚举，避免 acceptance 层引入平行真源。
-type StopReason = controlplane.StopReason
diff --git a/internal/runtime/acceptance/types.go b/internal/runtime/acceptance/types.go
deleted file mode 100644
index b57db9a8..00000000
--- a/internal/runtime/acceptance/types.go
+++ /dev/null
@@ -1,54 +0,0 @@
-package acceptance
-
-import (
- "neo-code/internal/runtime/controlplane"
- "neo-code/internal/runtime/decider"
- "neo-code/internal/runtime/verify"
-)
-
-// CompletionGateDecision 表示 completion gate 评估结果。
-type CompletionGateDecision struct {
- Passed bool `json:"passed"`
- Reason string `json:"reason,omitempty"`
-}
-
-// AcceptanceStatus 表示 final 验收的统一决策状态。
-type AcceptanceStatus string
-
-const (
- AcceptanceAccepted AcceptanceStatus = "accepted"
- AcceptanceContinue AcceptanceStatus = "continue"
- AcceptanceIncomplete AcceptanceStatus = "incomplete"
- AcceptanceFailed AcceptanceStatus = "failed"
-)
-
-// AcceptanceDecision 表示 runtime beforeAcceptFinal 的结构化输出。
-type AcceptanceDecision struct {
- Status AcceptanceStatus `json:"status"`
- StopReason controlplane.StopReason `json:"stop_reason,omitempty"`
- ErrorClass verify.ErrorClass `json:"error_class,omitempty"`
- CompletionPassed bool `json:"completion_passed,omitempty"`
- VerificationPassed bool `json:"verification_passed,omitempty"`
- CompletionBlockedReason string `json:"completion_blocked_reason,omitempty"`
- MissingFacts []decider.MissingFact `json:"missing_facts,omitempty"`
- RequiredNextActions []decider.RequiredAction `json:"required_next_actions,omitempty"`
- RequiredInput *decider.RequiredInput `json:"required_input,omitempty"`
- IntentHint decider.TaskKind `json:"intent_hint,omitempty"`
- EffectiveTaskKind decider.TaskKind `json:"effective_task_kind,omitempty"`
- UserVisibleSummary string `json:"user_visible_summary,omitempty"`
- InternalSummary string `json:"internal_summary,omitempty"`
- ContinueHint string `json:"continue_hint,omitempty"`
- VerifierResults []verify.VerificationResult `json:"verifier_results,omitempty"`
- HasProgress bool `json:"has_progress,omitempty"`
- Retryable bool `json:"retryable,omitempty"`
- WaitingExternal bool `json:"waiting_external,omitempty"`
-}
-
-// FinalAcceptanceInput 表示 beforeAcceptFinal 需要的输入快照。
-type FinalAcceptanceInput struct {
- CompletionGate CompletionGateDecision `json:"completion_gate"`
- VerificationInput verify.FinalVerifyInput `json:"verification_input"`
- NoProgressExceeded bool `json:"no_progress_exceeded,omitempty"`
- MaxTurnsReached bool `json:"max_turns_reached,omitempty"`
- MaxTurnsLimit int `json:"max_turns_limit,omitempty"`
-}
diff --git a/internal/runtime/acceptance_events.go b/internal/runtime/acceptance_events.go
deleted file mode 100644
index 4b5e8297..00000000
--- a/internal/runtime/acceptance_events.go
+++ /dev/null
@@ -1,34 +0,0 @@
-package runtime
-
-import (
- "strings"
-
- "neo-code/internal/runtime/acceptance"
-)
-
-// emitAcceptanceDecisionEvents 将验收决策及其 verifier 轨迹统一转换为运行时事件，保证观测链路一致。
-func (s *Service) emitAcceptanceDecisionEvents(state *runState, decision acceptance.AcceptanceDecision) {
- for _, result := range decision.VerifierResults {
- s.emitRunScopedOptional(EventVerificationStageFinished, state, VerificationStageFinishedPayload{
- Name: result.Name,
- Status: result.Status,
- Summary: result.Summary,
- Reason: result.Reason,
- ErrorClass: result.ErrorClass,
- })
- }
- s.emitRunScopedOptional(EventVerificationFinished, state, VerificationFinishedPayload{
- AcceptanceStatus: decision.Status,
- StopReason: decision.StopReason,
- ErrorClass: decision.ErrorClass,
- })
- s.emitRunScopedOptional(EventAcceptanceDecided, state, AcceptanceDecidedPayload{
- Status: decision.Status,
- StopReason: decision.StopReason,
- ErrorClass: decision.ErrorClass,
- CompletionBlockedReason: strings.TrimSpace(decision.CompletionBlockedReason),
- UserVisibleSummary: decision.UserVisibleSummary,
- InternalSummary: decision.InternalSummary,
- ContinueHint: decision.ContinueHint,
- })
-}
diff --git a/internal/runtime/acceptance_service.go b/internal/runtime/acceptance_service.go
deleted file mode 100644
index b8ffe6ab..00000000
--- a/internal/runtime/acceptance_service.go
+++ /dev/null
@@ -1,366 +0,0 @@
-package runtime
-
-import (
- "context"
- "fmt"
- "strings"
-
- "neo-code/internal/config"
- "neo-code/internal/runtime/acceptance"
- "neo-code/internal/runtime/controlplane"
- "neo-code/internal/runtime/decider"
- runtimefacts "neo-code/internal/runtime/facts"
- "neo-code/internal/runtime/verify"
- agentsession "neo-code/internal/session"
-)
-
-// acceptanceServiceInput 收敛一次最终验收裁决所需的最小输入。
-type acceptanceServiceInput struct {
- RunID string
- SessionID string
- TaskKind decider.TaskKind
- UserGoal string
- CompletionPassed bool
- CompletionBlockedReason string
- Facts runtimefacts.RuntimeFacts
- Todos decider.TodoSnapshot
- Progress decider.ProgressSnapshot
- LastAssistantText string
- HookAnnotations []string
- HookGuards []decider.HookGuardSignal
- NoProgressStreak int
- MaxNoProgress int
- VerificationProfile agentsession.VerificationProfile
- VerificationInput verify.FinalVerifyInput
-}
-
-// acceptanceService 负责生成 runtime 唯一终态裁决输出。
-type acceptanceService struct{}
-
-// Decide 统一执行 completion/verification/decider 聚合，并输出 AcceptanceDecision。
-func (s *acceptanceService) Decide(ctx context.Context, input acceptanceServiceInput) (acceptance.AcceptanceDecision, error) {
- output := acceptance.AcceptanceDecision{
- Status: acceptance.AcceptanceContinue,
- StopReason: controlplane.StopReasonTodoNotConverged,
- CompletionPassed: input.CompletionPassed,
- VerificationPassed: false,
- CompletionBlockedReason: strings.TrimSpace(input.CompletionBlockedReason),
- }
- verificationGate, err := runVerificationGate(ctx, input)
- if err != nil {
- return acceptance.AcceptanceDecision{}, err
- }
- output.VerificationPassed = verificationGate.Passed
- output.VerifierResults = append([]verify.VerificationResult(nil), verificationGate.Results...)
-
- noProgressExceeded := input.MaxNoProgress > 0 && input.NoProgressStreak >= input.MaxNoProgress
- decision := decider.Decide(decider.DecisionInput{
- RunID: strings.TrimSpace(input.RunID),
- SessionID: strings.TrimSpace(input.SessionID),
- TaskKind: input.TaskKind,
- UserGoal: strings.TrimSpace(input.UserGoal),
- Facts: input.Facts,
- Todos: input.Todos,
- Progress: input.Progress,
- LastAssistantText: strings.TrimSpace(input.LastAssistantText),
- CompletionPassed: input.CompletionPassed,
- CompletionReason: strings.TrimSpace(input.CompletionBlockedReason),
- NoProgressExceeded: noProgressExceeded,
- HookAnnotations: append([]string(nil), input.HookAnnotations...),
- HookGuards: append([]decider.HookGuardSignal(nil), input.HookGuards...),
- })
-
- output.MissingFacts = append([]decider.MissingFact(nil), decision.MissingFacts...)
- output.RequiredNextActions = append([]decider.RequiredAction(nil), decision.RequiredNextActions...)
- if decision.RequiredInput != nil {
- cloned := *decision.RequiredInput
- if len(cloned.Details) > 0 {
- details := make(map[string]any, len(cloned.Details))
- for k, v := range cloned.Details {
- details[k] = v
- }
- cloned.Details = details
- }
- output.RequiredInput = &cloned
- }
- output.IntentHint = decision.IntentHint
- output.EffectiveTaskKind = decision.EffectiveTaskKind
- output.UserVisibleSummary = strings.TrimSpace(decision.UserVisibleSummary)
- output.InternalSummary = strings.TrimSpace(decision.InternalSummary)
- output.ContinueHint = strings.TrimSpace(buildDeciderContinueHint(decision))
- output.StopReason = toControlplaneStopReason(decision.StopReason)
- output.ErrorClass = ""
-
- if output.StopReason == "" {
- output.StopReason = controlplane.StopReasonTodoNotConverged
- }
- if noProgressExceeded && decision.Status == decider.DecisionIncomplete {
- output.StopReason = controlplane.StopReasonNoProgressAfterFinalIntercept
- }
-
- // accepted 必须同时通过 completion 与 verification gate。
- if input.CompletionPassed && verificationGate.Passed && decision.Status == decider.DecisionAccepted {
- output.Status = acceptance.AcceptanceAccepted
- output.StopReason = controlplane.StopReasonAccepted
- output.ContinueHint = ""
- output.CompletionPassed = true
- output.VerificationPassed = true
- return output, nil
- }
-
- // verification gate 全部通过时信任其结果：即使 decider 基于启发式返回 continue，
- // verification gate 已实际运行所有 profile 指定的 verifier 且全部 pass，应直接 accepted。
- // 避免 decider 与 verification gate 数据源不一致导致死循环。
- if input.CompletionPassed && verificationGate.Passed {
- output.Status = acceptance.AcceptanceAccepted
- output.StopReason = controlplane.StopReasonAccepted
- output.ContinueHint = ""
- output.CompletionPassed = true
- output.VerificationPassed = true
- return output, nil
- }
-
- if input.CompletionPassed && !verificationGate.Passed {
- return mergeVerificationFailure(output, verificationGate), nil
- }
-
- switch decision.Status {
- case decider.DecisionAccepted:
- // completion 不通过时即便 decider accepted，也必须继续。
- output.Status = acceptance.AcceptanceContinue
- case decider.DecisionFailed, decider.DecisionBlocked:
- output.Status = acceptance.AcceptanceFailed
- if output.StopReason == "" {
- output.StopReason = controlplane.StopReasonVerificationFailed
- }
- case decider.DecisionIncomplete:
- output.Status = acceptance.AcceptanceIncomplete
- if output.StopReason == "" {
- output.StopReason = controlplane.StopReasonNoProgressAfterFinalIntercept
- }
- default:
- output.Status = acceptance.AcceptanceContinue
- }
- if output.Status == acceptance.AcceptanceContinue && output.ContinueHint == "" {
- output.ContinueHint = finalContinueReminder
- }
- // 死循环兜底：多轮 final 被拦截且无进展 + 存在 open required todo → 追加强制清理指令
- if output.Status == acceptance.AcceptanceContinue && input.NoProgressStreak >= 2 && input.Todos.Summary.RequiredOpen > 0 {
- staleHint := buildStaleTodoResetHint(input.Todos.Summary.RequiredOpen, input.NoProgressStreak)
- if output.ContinueHint == "" {
- output.ContinueHint = staleHint
- } else {
- output.ContinueHint = output.ContinueHint + "\n\n" + staleHint
- }
- }
- if input.VerificationInput.RuntimeState.MaxTurnsReached && output.Status == acceptance.AcceptanceContinue {
- output.Status = acceptance.AcceptanceIncomplete
- if output.StopReason == controlplane.StopReasonVerificationFailed {
- output.StopReason = controlplane.StopReasonMaxTurnExceededWithFailedVerification
- } else {
- output.StopReason = controlplane.StopReasonMaxTurnExceededWithUnconvergedTodos
- }
- }
- output.ErrorClass = normalizeAcceptanceErrorClass(output.ErrorClass, input, output)
- return output, nil
-}
-
-// runVerificationGate 执行 verifier gate；completion 未通过时仅回填必要证据，不执行重 verifier。
-func runVerificationGate(ctx context.Context, input acceptanceServiceInput) (verify.VerificationGateDecision, error) {
- if !input.CompletionPassed {
- results := make([]verify.VerificationResult, 0, 1)
- if strings.EqualFold(strings.TrimSpace(input.CompletionBlockedReason), string(controlplane.CompletionBlockedReasonPendingTodo)) {
- if synthetic := synthesizeTodoConvergenceEvidence(toSessionTodos(input.Todos)); synthetic != nil {
- results = append(results, *synthetic)
- }
- }
- return verify.VerificationGateDecision{
- Passed: false,
- Reason: controlplane.StopReasonTodoNotConverged,
- Results: results,
- }, nil
- }
- if !input.VerificationProfile.Valid() {
- return verify.VerificationGateDecision{
- Passed: false,
- Reason: controlplane.StopReasonVerificationConfigMissing,
- Results: []verify.VerificationResult{{
- Name: "verification_profile",
- Status: verify.VerificationFail,
- Summary: "verification profile invalid",
- Reason: fmt.Sprintf("invalid verification profile %q", input.VerificationProfile),
- ErrorClass: verify.ErrorClassEnvMissing,
- }},
- }, nil
- }
-
- policy := acceptance.DefaultPolicy{Executor: verify.PolicyCommandExecutor{}}
- verifiers, err := policy.ResolveVerifiers(input.VerificationInput)
- if err != nil {
- return verify.VerificationGateDecision{
- Passed: false,
- Reason: controlplane.StopReasonVerificationConfigMissing,
- Results: []verify.VerificationResult{{
- Name: "verification_profile",
- Status: verify.VerificationFail,
- Summary: "verification profile resolution failed",
- Reason: err.Error(),
- ErrorClass: verify.ErrorClassEnvMissing,
- }},
- }, nil
- }
- orch := verify.Orchestrator{Verifiers: verifiers}
- return orch.RunFinalVerification(ctx, input.VerificationInput)
-}
-
-// mergeVerificationFailure 统一把 verification gate 非通过映射到终态决策。
-func mergeVerificationFailure(
- base acceptance.AcceptanceDecision,
- gate verify.VerificationGateDecision,
-) acceptance.AcceptanceDecision {
- out := base
- out.VerificationPassed = gate.Passed
- out.VerifierResults = append([]verify.VerificationResult(nil), gate.Results...)
- out.StopReason = gate.Reason
-
- first := firstNonPassVerifierResult(gate.Results)
- if gate.Passed || first == nil {
- return out
- }
- out.ErrorClass = first.ErrorClass
- switch first.Status {
- case verify.VerificationSoftBlock:
- out.Status = acceptance.AcceptanceContinue
- if out.StopReason == "" {
- out.StopReason = controlplane.StopReasonTodoNotConverged
- }
- if out.ContinueHint == "" {
- out.ContinueHint = finalContinueReminder
- }
- case verify.VerificationHardBlock:
- out.Status = acceptance.AcceptanceIncomplete
- if first.WaitingExternal {
- out.StopReason = controlplane.StopReasonTodoWaitingExternal
- }
- default:
- out.Status = acceptance.AcceptanceFailed
- if out.StopReason == "" || out.StopReason == controlplane.StopReasonAccepted {
- out.StopReason = controlplane.StopReasonVerificationFailed
- }
- if out.ErrorClass == "" {
- out.ErrorClass = verify.ErrorClassUnknown
- }
- }
- out.ErrorClass = normalizeAcceptanceErrorClass(out.ErrorClass, acceptanceServiceInput{}, out)
- return out
-}
-
-// normalizeAcceptanceErrorClass 统一补齐终态 error_class，避免 TUI/Gateway 出现 unknown/empty 推断歧义。
-func normalizeAcceptanceErrorClass(
- current verify.ErrorClass,
- input acceptanceServiceInput,
- decision acceptance.AcceptanceDecision,
-) verify.ErrorClass {
- if current != "" {
- return current
- }
- switch decision.StopReason {
- case controlplane.StopReasonVerificationConfigMissing:
- return verify.ErrorClassEnvMissing
- case controlplane.StopReasonVerificationExecutionDenied:
- return verify.ErrorClassPermissionDenied
- case controlplane.StopReasonVerificationExecutionError:
- return verify.ErrorClassUnknown
- case controlplane.StopReasonRequiredTodoFailed:
- return verify.ErrorClassUnknown
- case controlplane.StopReasonNoProgressAfterFinalIntercept:
- return verify.ErrorClassUnknown
- case controlplane.StopReasonVerificationFailed:
- if input.TaskKind == decider.TaskKindSubAgent && len(input.Facts.SubAgents.Failed) > 0 {
- return verify.ErrorClass("subagent_failed")
- }
- if input.TaskKind == decider.TaskKindWorkspaceWrite {
- if errClass := latestToolErrorClass(input.Facts.Errors.ToolErrors, "filesystem_write_file"); errClass != "" {
- return verify.ErrorClass(errClass)
- }
- }
- if errClass := latestToolErrorClass(input.Facts.Errors.ToolErrors, "spawn_subagent"); errClass != "" {
- return verify.ErrorClass(errClass)
- }
- if errClass := latestToolErrorClass(input.Facts.Errors.ToolErrors, "filesystem_write_file"); errClass != "" {
- return verify.ErrorClass(errClass)
- }
- }
- if decision.Status == acceptance.AcceptanceFailed || decision.Status == acceptance.AcceptanceIncomplete {
- return verify.ErrorClassUnknown
- }
- return ""
-}
-
-// latestToolErrorClass 返回目标工具最近一次非空错误分类。
-func latestToolErrorClass(errors []runtimefacts.ToolErrorFact, tool string) string {
- target := strings.TrimSpace(tool)
- for i := len(errors) - 1; i >= 0; i-- {
- entry := errors[i]
- if target != "" && !strings.EqualFold(strings.TrimSpace(entry.Tool), target) {
- continue
- }
- errClass := strings.TrimSpace(entry.ErrorClass)
- if errClass != "" {
- return errClass
- }
- }
- return ""
-}
-
-// buildStaleTodoResetHint 构造死循环兜底指令：当多轮 final 被拦截且无进展时，强制要求模型清理 stale todo。
-func buildStaleTodoResetHint(requiredOpen, noProgressStreak int) string {
- var b strings.Builder
- b.WriteString("\n")
- b.WriteString(fmt.Sprintf("CRITICAL: You have been blocked for %d consecutive final attempts with %d unfinished required todo(s).\n", noProgressStreak, requiredOpen))
- b.WriteString("If these todos are NO LONGER RELEVANT to the user's CURRENT request,\n")
- b.WriteString("you MUST mark them canceled using todo_write set_status=canceled RIGHT NOW.\n")
- b.WriteString("Do NOT attempt to complete stale todos that belong to a PREVIOUS task.\n")
- b.WriteString("After canceling irrelevant todos, proceed with the user's current request.\n")
- b.WriteString("")
- return b.String()
-}
-
-func firstNonPassVerifierResult(results []verify.VerificationResult) *verify.VerificationResult {
- for _, result := range results {
- if result.Status == verify.VerificationPass {
- continue
- }
- cloned := result
- return &cloned
- }
- return nil
-}
-
-func toSessionTodos(snapshot decider.TodoSnapshot) []agentsession.TodoItem {
- if len(snapshot.Items) == 0 {
- return nil
- }
- out := make([]agentsession.TodoItem, 0, len(snapshot.Items))
- for _, item := range snapshot.Items {
- required := item.Required
- status := agentsession.TodoStatus(strings.TrimSpace(item.Status))
- out = append(out, agentsession.TodoItem{
- ID: strings.TrimSpace(item.ID),
- Content: strings.TrimSpace(item.Content),
- Status: status,
- Required: &required,
- Artifacts: append([]string(nil), item.Artifacts...),
- FailureReason: strings.TrimSpace(item.FailureReason),
- })
- }
- return out
-}
-
-func resolveAcceptanceMaxNoProgress(cfg config.VerificationConfig) int {
- limit := cfg.MaxNoProgress
- if limit <= 0 {
- return 3
- }
- return limit
-}
diff --git a/internal/runtime/acceptance_service_test.go b/internal/runtime/acceptance_service_test.go
deleted file mode 100644
index 77d2b5d4..00000000
--- a/internal/runtime/acceptance_service_test.go
+++ /dev/null
@@ -1,513 +0,0 @@
-package runtime
-
-import (
- "context"
- "strings"
- "testing"
- "time"
-
- "neo-code/internal/config"
- providertypes "neo-code/internal/provider/types"
- "neo-code/internal/runtime/acceptance"
- "neo-code/internal/runtime/controlplane"
- "neo-code/internal/runtime/decider"
- runtimefacts "neo-code/internal/runtime/facts"
- runtimehooks "neo-code/internal/runtime/hooks"
- "neo-code/internal/runtime/verify"
- agentsession "neo-code/internal/session"
-)
-
-func TestBeforeCompletionDecisionAcceptanceHooksOnOffParity(t *testing.T) {
- t.Parallel()
-
- snapshot := TurnBudgetSnapshot{Config: config.StaticDefaults().Clone(), Workdir: t.TempDir()}
- assistant := providertypes.Message{
- Role: providertypes.RoleAssistant,
- Parts: []providertypes.ContentPart{providertypes.NewTextPart("done")},
- }
-
- offService := &Service{events: make(chan RuntimeEvent, 16)}
- offState := newRunState("run-hooks-off", agentsession.New("hooks-off"))
- offState.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- offDecision, err := offService.runBeforeCompletionDecisionAcceptance(
- context.Background(),
- &offState,
- snapshot,
- assistant,
- snapshot.Workdir,
- true,
- false,
- providertypes.RoleAssistant,
- )
- if err != nil {
- t.Fatalf("hooks-off decision error = %v", err)
- }
-
- onService := &Service{events: make(chan RuntimeEvent, 16)}
- baseRegistry := runtimehooks.NewRegistry()
- userRegistry := runtimehooks.NewRegistry()
- repoRegistry := runtimehooks.NewRegistry()
- if err := userRegistry.Register(runtimehooks.HookSpec{
- ID: "user-note",
- Point: runtimehooks.HookPointBeforeCompletionDecision,
- Scope: runtimehooks.HookScopeUser,
- Source: runtimehooks.HookSourceUser,
- Handler: func(_ context.Context, _ runtimehooks.HookContext) runtimehooks.HookResult {
- return runtimehooks.HookResult{Status: runtimehooks.HookResultPass, Message: "note"}
- },
- }); err != nil {
- t.Fatalf("register user hook: %v", err)
- }
- onService.SetHookExecutor(composeRuntimeHookExecutors(
- runtimehooks.NewExecutor(baseRegistry, nil, time.Second),
- runtimehooks.NewExecutor(userRegistry, nil, time.Second),
- runtimehooks.NewExecutor(repoRegistry, nil, time.Second),
- ))
- onState := newRunState("run-hooks-on", agentsession.New("hooks-on"))
- onState.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- onDecision, err := onService.runBeforeCompletionDecisionAcceptance(
- context.Background(),
- &onState,
- snapshot,
- assistant,
- snapshot.Workdir,
- true,
- false,
- providertypes.RoleAssistant,
- )
- if err != nil {
- t.Fatalf("hooks-on decision error = %v", err)
- }
-
- if offDecision.Status != onDecision.Status || offDecision.StopReason != onDecision.StopReason {
- t.Fatalf("hooks parity mismatch: off=%+v on=%+v", offDecision, onDecision)
- }
-
- offContinue, err := offService.runBeforeCompletionDecisionAcceptance(
- context.Background(),
- &offState,
- snapshot,
- assistant,
- snapshot.Workdir,
- false,
- false,
- providertypes.RoleAssistant,
- )
- if err != nil {
- t.Fatalf("hooks-off continue decision error = %v", err)
- }
- onContinue, err := onService.runBeforeCompletionDecisionAcceptance(
- context.Background(),
- &onState,
- snapshot,
- assistant,
- snapshot.Workdir,
- false,
- false,
- providertypes.RoleAssistant,
- )
- if err != nil {
- t.Fatalf("hooks-on continue decision error = %v", err)
- }
- if offContinue.Status != onContinue.Status || offContinue.StopReason != onContinue.StopReason {
- t.Fatalf("hooks continue parity mismatch: off=%+v on=%+v", offContinue, onContinue)
- }
-}
-
-func TestAcceptanceDecisionRequiresCompletionAndVerification(t *testing.T) {
- t.Parallel()
-
- service := &Service{events: make(chan RuntimeEvent, 16)}
- snapshot := TurnBudgetSnapshot{Config: config.StaticDefaults().Clone(), Workdir: t.TempDir()}
- assistant := providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("done")}}
-
- t.Run("completion_pass_but_verification_fail_not_accepted", func(t *testing.T) {
- state := newRunState("run-verify-fail", agentsession.New("verify-fail"))
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileCreateFile
- state.session.TaskState.KeyArtifacts = []string{"missing.txt"}
- decision, err := service.beforeAcceptFinal(context.Background(), &state, snapshot, assistant, true, beforeCompletionHookSignals{})
- if err != nil {
- t.Fatalf("beforeAcceptFinal error = %v", err)
- }
- if decision.Status == acceptance.AcceptanceAccepted {
- t.Fatalf("unexpected accepted decision: %+v", decision)
- }
- if !decision.CompletionPassed || decision.VerificationPassed {
- t.Fatalf("expected completion=true verification=false, got %+v", decision)
- }
- if len(decision.VerifierResults) == 0 {
- t.Fatalf("expected verification trace in decision")
- }
- })
-
- t.Run("completion_fail_not_accepted_even_if_task_only", func(t *testing.T) {
- state := newRunState("run-completion-fail", agentsession.New("completion-fail"))
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- decision, err := service.beforeAcceptFinal(context.Background(), &state, snapshot, assistant, false, beforeCompletionHookSignals{})
- if err != nil {
- t.Fatalf("beforeAcceptFinal error = %v", err)
- }
- if decision.Status == acceptance.AcceptanceAccepted {
- t.Fatalf("unexpected accepted decision: %+v", decision)
- }
- if decision.CompletionPassed {
- t.Fatalf("expected completion=false, got %+v", decision)
- }
- })
-
- t.Run("accepted_requires_both_true", func(t *testing.T) {
- state := newRunState("run-accepted", agentsession.New("accepted"))
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- decision, err := service.beforeAcceptFinal(context.Background(), &state, snapshot, assistant, true, beforeCompletionHookSignals{})
- if err != nil {
- t.Fatalf("beforeAcceptFinal error = %v", err)
- }
- if decision.Status != acceptance.AcceptanceAccepted {
- t.Fatalf("status=%q want accepted", decision.Status)
- }
- if !decision.CompletionPassed || !decision.VerificationPassed {
- t.Fatalf("accepted must satisfy completion+verification, got %+v", decision)
- }
- })
-
- t.Run("completed_required_todo_list_does_not_block_acceptance", func(t *testing.T) {
- state := newRunState("run-completed-todo", agentsession.New("completed-todo"))
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- required := true
- state.session.Todos = []agentsession.TodoItem{{
- ID: "todo-done",
- Content: "done",
- Status: agentsession.TodoStatusCompleted,
- Required: &required,
- }}
- decision, err := service.beforeAcceptFinal(context.Background(), &state, snapshot, assistant, true, beforeCompletionHookSignals{})
- if err != nil {
- t.Fatalf("beforeAcceptFinal error = %v", err)
- }
- if decision.Status != acceptance.AcceptanceAccepted {
- t.Fatalf("status=%q want accepted, decision=%+v", decision.Status, decision)
- }
- if !decision.CompletionPassed || !decision.VerificationPassed {
- t.Fatalf("completed todo list should pass completion+verification, got %+v", decision)
- }
- })
-}
-
-func TestBeforeCompletionDecisionUserRepoCannotDirectlyTerminal(t *testing.T) {
- t.Parallel()
-
- service := &Service{events: make(chan RuntimeEvent, 16)}
- baseRegistry := runtimehooks.NewRegistry()
- userRegistry := runtimehooks.NewRegistry()
- if err := userRegistry.Register(runtimehooks.HookSpec{
- ID: "user-guard",
- Point: runtimehooks.HookPointBeforeCompletionDecision,
- Scope: runtimehooks.HookScopeUser,
- Source: runtimehooks.HookSourceUser,
- Handler: func(_ context.Context, _ runtimehooks.HookContext) runtimehooks.HookResult {
- return runtimehooks.HookResult{Status: runtimehooks.HookResultFailed, Message: "guard"}
- },
- }); err != nil {
- t.Fatalf("register user guard hook: %v", err)
- }
- service.SetHookExecutor(composeRuntimeHookExecutors(
- runtimehooks.NewExecutor(baseRegistry, nil, time.Second),
- runtimehooks.NewExecutor(userRegistry, nil, time.Second),
- nil,
- ))
-
- state := newRunState("run-user-guard", agentsession.New("user-guard"))
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- snapshot := TurnBudgetSnapshot{Config: config.StaticDefaults().Clone(), Workdir: t.TempDir()}
- decision, err := service.runBeforeCompletionDecisionAcceptance(
- context.Background(),
- &state,
- snapshot,
- providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("done")}},
- snapshot.Workdir,
- true,
- false,
- providertypes.RoleAssistant,
- )
- if err != nil {
- t.Fatalf("runBeforeCompletionDecisionAcceptance error = %v", err)
- }
- if decision.Status != acceptance.AcceptanceAccepted {
- t.Fatalf("user guard should not directly terminal-block acceptance path, got %+v", decision)
- }
- if !strings.Contains(decision.InternalSummary, "hook signals consumed") {
- t.Fatalf("expected hook signal to be consumed by acceptance input, got %q", decision.InternalSummary)
- }
-}
-
-func TestVerificationTraceEmitsStageEvents(t *testing.T) {
- t.Parallel()
-
- service := &Service{events: make(chan RuntimeEvent, 32)}
- state := newRunState("run-verify-stage-events", agentsession.New("verify-stage-events"))
- decision := acceptance.AcceptanceDecision{
- Status: acceptance.AcceptanceContinue,
- StopReason: controlplane.StopReasonVerificationFailed,
- ErrorClass: "content_mismatch",
- VerifierResults: []verify.VerificationResult{
- {
- Name: "content_match",
- Status: verify.VerificationSoftBlock,
- Summary: "missing expected token",
- Reason: "content mismatch",
- ErrorClass: "content_mismatch",
- },
- },
- }
- service.emitAcceptanceDecisionEvents(&state, decision)
- events := collectRuntimeEvents(service.Events())
- stageCount := 0
- for _, evt := range events {
- if evt.Type == EventVerificationStageFinished {
- stageCount++
- }
- }
- if stageCount == 0 {
- t.Fatal("expected verification_stage_finished events from acceptance decision trace")
- }
-}
-
-func TestVerificationFailureProducesStopReasonAndErrorClass(t *testing.T) {
- t.Parallel()
-
- service := &Service{events: make(chan RuntimeEvent, 16)}
- state := newRunState("run-invalid-profile", agentsession.New("invalid-profile"))
- state.session.TaskState.VerificationProfile = "bad_profile"
- snapshot := TurnBudgetSnapshot{Config: config.StaticDefaults().Clone(), Workdir: t.TempDir()}
- decision, err := service.beforeAcceptFinal(
- context.Background(),
- &state,
- snapshot,
- providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("done")}},
- true,
- beforeCompletionHookSignals{},
- )
- if err != nil {
- t.Fatalf("beforeAcceptFinal error = %v", err)
- }
- if decision.Status != acceptance.AcceptanceFailed {
- t.Fatalf("status=%q want failed", decision.Status)
- }
- if decision.StopReason != controlplane.StopReasonVerificationConfigMissing {
- t.Fatalf("stop reason=%q want verification_config_missing", decision.StopReason)
- }
- if decision.ErrorClass == "" {
- t.Fatalf("verification failure must keep non-empty error_class: %+v", decision)
- }
-}
-
-func TestChatAnswerAcceptancePassesWithoutHeavyVerification(t *testing.T) {
- t.Parallel()
-
- service := &Service{events: make(chan RuntimeEvent, 16)}
- state := newRunState("run-chat-answer", agentsession.New("chat-answer"))
- state.taskKind = decider.TaskKindChatAnswer
- state.userGoal = "你好"
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
-
- snapshot := TurnBudgetSnapshot{Config: config.StaticDefaults().Clone(), Workdir: t.TempDir()}
- decision, err := service.beforeAcceptFinal(
- context.Background(),
- &state,
- snapshot,
- providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("你好")}},
- true,
- beforeCompletionHookSignals{},
- )
- if err != nil {
- t.Fatalf("beforeAcceptFinal error = %v", err)
- }
- if !decision.CompletionPassed || !decision.VerificationPassed {
- t.Fatalf("chat answer should pass completion+verification gate, got %+v", decision)
- }
- if decision.Status != acceptance.AcceptanceAccepted {
- t.Fatalf("status=%q want accepted", decision.Status)
- }
- if decision.StopReason != controlplane.StopReasonAccepted {
- t.Fatalf("stop reason=%q want accepted", decision.StopReason)
- }
-}
-
-func TestNormalizeAcceptanceErrorClassCoverage(t *testing.T) {
- t.Parallel()
-
- testInput := acceptanceServiceInput{
- TaskKind: decider.TaskKindWorkspaceWrite,
- Facts: runtimefacts.RuntimeFacts{
- Errors: runtimefacts.ErrorFacts{
- ToolErrors: []runtimefacts.ToolErrorFact{{
- Tool: "filesystem_write_file",
- ErrorClass: "permission_denied",
- }},
- },
- },
- }
- cases := []struct {
- name string
- input acceptanceServiceInput
- decision acceptance.AcceptanceDecision
- want verify.ErrorClass
- }{
- {
- name: "verification_config_missing",
- input: acceptanceServiceInput{},
- decision: acceptance.AcceptanceDecision{StopReason: controlplane.StopReasonVerificationConfigMissing, Status: acceptance.AcceptanceFailed},
- want: verify.ErrorClassEnvMissing,
- },
- {
- name: "verification_execution_denied",
- input: acceptanceServiceInput{},
- decision: acceptance.AcceptanceDecision{StopReason: controlplane.StopReasonVerificationExecutionDenied, Status: acceptance.AcceptanceFailed},
- want: verify.ErrorClassPermissionDenied,
- },
- {
- name: "verification_execution_error",
- input: acceptanceServiceInput{},
- decision: acceptance.AcceptanceDecision{StopReason: controlplane.StopReasonVerificationExecutionError, Status: acceptance.AcceptanceFailed},
- want: verify.ErrorClassUnknown,
- },
- {
- name: "required_todo_failed",
- input: acceptanceServiceInput{},
- decision: acceptance.AcceptanceDecision{StopReason: controlplane.StopReasonRequiredTodoFailed, Status: acceptance.AcceptanceFailed},
- want: verify.ErrorClassUnknown,
- },
- {
- name: "no_progress_after_final_intercept",
- input: acceptanceServiceInput{},
- decision: acceptance.AcceptanceDecision{StopReason: controlplane.StopReasonNoProgressAfterFinalIntercept, Status: acceptance.AcceptanceIncomplete},
- want: verify.ErrorClassUnknown,
- },
- {
- name: "subagent failed",
- input: acceptanceServiceInput{
- TaskKind: decider.TaskKindSubAgent,
- Facts: runtimefacts.RuntimeFacts{
- SubAgents: runtimefacts.SubAgentFacts{
- Failed: []runtimefacts.SubAgentFact{{TaskID: "sa-1"}},
- },
- },
- },
- decision: acceptance.AcceptanceDecision{StopReason: controlplane.StopReasonVerificationFailed, Status: acceptance.AcceptanceFailed},
- want: verify.ErrorClass("subagent_failed"),
- },
- {
- name: "workspace_write_hard_failure",
- input: testInput,
- decision: acceptance.AcceptanceDecision{StopReason: controlplane.StopReasonVerificationFailed, Status: acceptance.AcceptanceFailed},
- want: verify.ErrorClass("permission_denied"),
- },
- }
-
- for _, tc := range cases {
- t.Run(tc.name, func(t *testing.T) {
- got := normalizeAcceptanceErrorClass("", tc.input, tc.decision)
- if got != tc.want {
- t.Fatalf("normalizeAcceptanceErrorClass() = %q, want %q", got, tc.want)
- }
- })
- }
-}
-
-func TestNoProgressThresholdProducesIncomplete(t *testing.T) {
- t.Parallel()
-
- service := &Service{events: make(chan RuntimeEvent, 16)}
- state := newRunState("run-no-progress", agentsession.New("no-progress"))
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- state.finalInterceptStreak = 3
- state.mustUseToolAfterFinalContinue = true
- state.noToolAfterFinalContinueStreak = 3
-
- snapshot := TurnBudgetSnapshot{Config: config.StaticDefaults().Clone(), Workdir: t.TempDir()}
- decision, err := service.beforeAcceptFinal(
- context.Background(),
- &state,
- snapshot,
- providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("已完成")}},
- false,
- beforeCompletionHookSignals{},
- )
- if err != nil {
- t.Fatalf("beforeAcceptFinal error = %v", err)
- }
- if decision.Status != acceptance.AcceptanceIncomplete {
- t.Fatalf("status=%q want incomplete", decision.Status)
- }
- if decision.StopReason != controlplane.StopReasonNoProgressAfterFinalIntercept {
- t.Fatalf("stop_reason=%q want no_progress_after_final_intercept", decision.StopReason)
- }
-}
-
-func TestRunVerificationGateSkipsProfileValidationWhenCompletionBlocked(t *testing.T) {
- t.Parallel()
-
- gate, err := runVerificationGate(context.Background(), acceptanceServiceInput{
- CompletionPassed: false,
- CompletionBlockedReason: string(controlplane.CompletionBlockedReasonPendingTodo),
- VerificationProfile: "invalid_profile",
- Todos: decider.TodoSnapshot{
- Items: []decider.TodoViewItem{{
- ID: "todo-1",
- Content: "x",
- Status: "pending",
- Required: true,
- }},
- Summary: decider.TodoSummary{
- RequiredTotal: 1,
- RequiredOpen: 1,
- },
- },
- })
- if err != nil {
- t.Fatalf("runVerificationGate error = %v", err)
- }
- if gate.Reason != controlplane.StopReasonTodoNotConverged {
- t.Fatalf("reason=%q want todo_not_converged", gate.Reason)
- }
- if len(gate.Results) == 0 {
- t.Fatal("expected synthetic todo convergence evidence when completion is blocked")
- }
-}
-
-func TestBeforeAcceptFinalMarksIncompleteWhenFinalInterceptHitsMaxTurns(t *testing.T) {
- t.Parallel()
-
- service := &Service{events: make(chan RuntimeEvent, 16)}
- cfg := config.StaticDefaults().Clone()
- cfg.Runtime.MaxTurns = 1
- snapshot := TurnBudgetSnapshot{Config: cfg, Workdir: t.TempDir()}
-
- state := newRunState("run-max-turn-final-intercept", agentsession.New("max-turn-final-intercept"))
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- required := true
- state.session.Todos = []agentsession.TodoItem{{
- ID: "todo-1",
- Content: "pending",
- Status: agentsession.TodoStatusPending,
- Required: &required,
- }}
-
- decision, err := service.beforeAcceptFinal(
- context.Background(),
- &state,
- snapshot,
- providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("done")}},
- true,
- beforeCompletionHookSignals{},
- )
- if err != nil {
- t.Fatalf("beforeAcceptFinal error = %v", err)
- }
- if decision.Status != acceptance.AcceptanceIncomplete {
- t.Fatalf("status=%q want incomplete", decision.Status)
- }
- if decision.StopReason != controlplane.StopReasonMaxTurnExceededWithUnconvergedTodos {
- t.Fatalf("stop_reason=%q want max_turn_exceeded_with_unconverged_todos", decision.StopReason)
- }
-}
diff --git a/internal/runtime/acceptgate/checks.go b/internal/runtime/acceptgate/checks.go
new file mode 100644
index 00000000..b3f39002
--- /dev/null
+++ b/internal/runtime/acceptgate/checks.go
@@ -0,0 +1,248 @@
+package acceptgate
+
+import (
+ "path/filepath"
+ "strings"
+
+ agentsession "neo-code/internal/session"
+)
+
+func checkRequiredTodoFailures(todos []agentsession.TodoItem) CheckResult {
+ for _, todo := range todos {
+ if !todo.RequiredValue() {
+ continue
+ }
+ if todo.Status == agentsession.TodoStatusFailed {
+ return CheckResult{
+ Passed: false,
+ Name: "required_todo_failed",
+ Reason: "required todo failed: " + strings.TrimSpace(todo.ID),
+ }
+ }
+ }
+ return CheckResult{Passed: true, Name: "required_todo_failed"}
+}
+
+func checkRequiredTodoConvergence(todos []agentsession.TodoItem) CheckResult {
+ for _, todo := range todos {
+ if !todo.RequiredValue() {
+ continue
+ }
+ if !todo.Status.IsTerminal() {
+ return CheckResult{
+ Passed: false,
+ Name: "required_todo_convergence",
+ Reason: "required todo is not terminal: " + strings.TrimSpace(todo.ID),
+ }
+ }
+ }
+ return CheckResult{Passed: true, Name: "required_todo_convergence"}
+}
+
+func evaluateAcceptCheck(input Input, check agentsession.AcceptCheck) CheckResult {
+ check.Kind = strings.TrimSpace(check.Kind)
+ check.Target = strings.TrimSpace(check.Target)
+ switch check.Kind {
+ case agentsession.AcceptCheckOutputOnly:
+ return checkOutputOnly(input, check)
+ case agentsession.AcceptCheckWorkspaceChange:
+ return checkWorkspaceChange(input, check)
+ case agentsession.AcceptCheckCommandSuccess:
+ return checkCommandSuccess(input, check)
+ case agentsession.AcceptCheckFileExists:
+ return checkFileExists(input, check)
+ case agentsession.AcceptCheckContentContains:
+ return checkContentContains(input, check)
+ case agentsession.AcceptCheckToolFact:
+ return checkToolFact(input, check)
+ default:
+ return CheckResult{
+ Passed: false,
+ Name: checkName(check),
+ Kind: check.Kind,
+ Target: check.Target,
+ Reason: "unknown required accept check kind",
+ }
+ }
+}
+
+func checkOutputOnly(input Input, check agentsession.AcceptCheck) CheckResult {
+ if strings.TrimSpace(input.LastAssistantText) != "" {
+ return pass(check)
+ }
+ return fail(check, "assistant output is empty")
+}
+
+func checkWorkspaceChange(input Input, check agentsession.AcceptCheck) CheckResult {
+ if len(input.Facts.Files.Written) > 0 {
+ return pass(check)
+ }
+ for _, item := range input.Facts.Files.Exists {
+ switch strings.TrimSpace(item.Source) {
+ case "filesystem_write_file", "filesystem_edit", "bash", "workspace_write":
+ return pass(check)
+ }
+ }
+ return fail(check, "missing workspace change evidence")
+}
+
+func checkCommandSuccess(input Input, check agentsession.AcceptCheck) CheckResult {
+ target := normalizeCommand(check.Target)
+ if target == "" {
+ return fail(check, "command target is empty")
+ }
+ for _, fact := range input.Facts.Commands.Executed {
+ if !fact.Succeeded {
+ continue
+ }
+ if commandMatches(normalizeCommand(fact.Command), target, check.Match) {
+ return pass(check)
+ }
+ }
+ return fail(check, "missing successful command evidence")
+}
+
+func checkFileExists(input Input, check agentsession.AcceptCheck) CheckResult {
+ target := normalizePath(check.Target)
+ if target == "" {
+ return fail(check, "file target is empty")
+ }
+ for _, fact := range input.Facts.Files.Exists {
+ if normalizePath(fact.Path) == target {
+ return pass(check)
+ }
+ }
+ for _, fact := range input.Facts.Files.Written {
+ if normalizePath(fact.Path) == target {
+ return pass(check)
+ }
+ }
+ return fail(check, "missing file existence evidence")
+}
+
+func checkContentContains(input Input, check agentsession.AcceptCheck) CheckResult {
+ target := normalizePath(check.Target)
+ if target == "" {
+ return fail(check, "content target is empty")
+ }
+ for _, fact := range input.Facts.Files.ContentMatch {
+ if normalizePath(fact.Path) != target || !fact.VerificationPassed {
+ continue
+ }
+ if expected := strings.TrimSpace(check.Params["contains"]); expected != "" {
+ if !containsString(fact.ExpectedContains, expected) {
+ continue
+ }
+ }
+ return pass(check)
+ }
+ return fail(check, "missing content match evidence")
+}
+
+func checkToolFact(input Input, check agentsession.AcceptCheck) CheckResult {
+ scope := strings.TrimSpace(firstNonEmpty(check.Params["scope"], check.Target))
+ tool := strings.TrimSpace(check.Params["tool"])
+ for _, fact := range input.Facts.Verification.Passed {
+ if tool != "" && !strings.EqualFold(strings.TrimSpace(fact.Tool), tool) {
+ continue
+ }
+ if scope != "" && strings.TrimSpace(fact.Scope) != scope {
+ continue
+ }
+ return pass(check)
+ }
+ return fail(check, "missing tool verification fact")
+}
+
+func pass(check agentsession.AcceptCheck) CheckResult {
+ return CheckResult{Passed: true, Name: checkName(check), Kind: check.Kind, Target: check.Target}
+}
+
+func fail(check agentsession.AcceptCheck, reason string) CheckResult {
+ return CheckResult{Passed: false, Name: checkName(check), Kind: check.Kind, Target: check.Target, Reason: reason}
+}
+
+func checkName(check agentsession.AcceptCheck) string {
+ if id := strings.TrimSpace(check.ID); id != "" {
+ return id
+ }
+ if kind := strings.TrimSpace(check.Kind); kind != "" {
+ return kind
+ }
+ return "accept_check"
+}
+
+func commandMatches(actual, target, mode string) bool {
+ switch strings.TrimSpace(strings.ToLower(mode)) {
+ case "exact":
+ return actual == target
+ case "prefix":
+ return strings.HasPrefix(actual, target)
+ case "contains", "normalized_contains", "":
+ return actual == target || strings.Contains(actual, target)
+ default:
+ return actual == target || strings.Contains(actual, target)
+ }
+}
+
+func normalizeCommand(value string) string {
+ value = strings.TrimSpace(value)
+ if value == "" {
+ return ""
+ }
+ fields := strings.Fields(value)
+ out := make([]string, 0, len(fields))
+ for _, field := range fields {
+ if isEnvVarAssignment(field) {
+ continue
+ }
+ if strings.HasPrefix(strings.ToLower(field), "$env:") {
+ continue
+ }
+ out = append(out, field)
+ }
+ return strings.ToLower(strings.Join(out, " "))
+}
+
+// isEnvVarAssignment 识别裸环境变量赋值，避免把 CLI 的 -flag=value 当作环境变量剥离。
+func isEnvVarAssignment(field string) bool {
+ field = strings.TrimSpace(field)
+ if !strings.Contains(field, "=") {
+ return false
+ }
+ if strings.HasPrefix(field, "-") {
+ return false
+ }
+ if strings.Contains(field, "/") || strings.Contains(field, "\\") {
+ return false
+ }
+ return true
+}
+
+func normalizePath(value string) string {
+ value = strings.TrimSpace(value)
+ if value == "" {
+ return ""
+ }
+ cleaned := filepath.ToSlash(filepath.Clean(value))
+ cleaned = strings.TrimPrefix(cleaned, "./")
+ return strings.ToLower(cleaned)
+}
+
+func containsString(values []string, target string) bool {
+ for _, value := range values {
+ if strings.TrimSpace(value) == target {
+ return true
+ }
+ }
+ return false
+}
+
+func firstNonEmpty(values ...string) string {
+ for _, value := range values {
+ if strings.TrimSpace(value) != "" {
+ return value
+ }
+ }
+ return ""
+}
diff --git a/internal/runtime/acceptgate/checks_test.go b/internal/runtime/acceptgate/checks_test.go
new file mode 100644
index 00000000..32333756
--- /dev/null
+++ b/internal/runtime/acceptgate/checks_test.go
@@ -0,0 +1,56 @@
+package acceptgate
+
+import (
+ "context"
+ "testing"
+
+ runtimefacts "neo-code/internal/runtime/facts"
+ agentsession "neo-code/internal/session"
+)
+
+func TestNormalizeCommandKeepsCLIFlags(t *testing.T) {
+ t.Parallel()
+
+ got := normalizeCommand("go test ./... -run=TestFoo --filter=a=b -count=1")
+ want := "go test ./... -run=testfoo --filter=a=b -count=1"
+ if got != want {
+ t.Fatalf("normalizeCommand() = %q, want %q", got, want)
+ }
+}
+
+func TestNormalizeCommandStripsEnvVars(t *testing.T) {
+ t.Parallel()
+
+ got := normalizeCommand("CGO_ENABLED=0 GOFLAGS=-count=1 go test ./...")
+ if got != "go test ./..." {
+ t.Fatalf("normalizeCommand() = %q, want %q", got, "go test ./...")
+ }
+}
+
+func TestNormalizeCommandKeepsPathAssignments(t *testing.T) {
+ t.Parallel()
+
+ got := normalizeCommand("PKG=./cmd/... go test")
+ if got != "pkg=./cmd/... go test" {
+ t.Fatalf("normalizeCommand() = %q, want %q", got, "pkg=./cmd/... go test")
+ }
+}
+
+func TestEvaluateCommandSuccessKeepsFlagSpecificity(t *testing.T) {
+ t.Parallel()
+
+ report := Evaluate(context.Background(), Input{
+ PlanVerify: agentsession.AcceptChecks{
+ {Kind: agentsession.AcceptCheckCommandSuccess, Target: "go test ./... -run=TestFoo"},
+ },
+ Facts: runtimefacts.RuntimeFacts{
+ Commands: runtimefacts.CommandFacts{Executed: []runtimefacts.CommandFact{
+ {Tool: "bash", Command: "go test ./...", Succeeded: true},
+ }},
+ },
+ LastAssistantText: "done",
+ })
+ if report.Outcome != OutcomeFailed {
+ t.Fatalf("report = %+v, want failed because broad command must not satisfy -run-specific check", report)
+ }
+}
diff --git a/internal/runtime/acceptgate/gate.go b/internal/runtime/acceptgate/gate.go
new file mode 100644
index 00000000..b35d41ab
--- /dev/null
+++ b/internal/runtime/acceptgate/gate.go
@@ -0,0 +1,136 @@
+package acceptgate
+
+import (
+ "context"
+ "fmt"
+ "strings"
+
+ "neo-code/internal/runtime/controlplane"
+ runtimefacts "neo-code/internal/runtime/facts"
+ agentsession "neo-code/internal/session"
+)
+
+// Outcome 表示 Accept Gate 的二元终态结果。
+type Outcome string
+
+const (
+ // OutcomeAccepted 表示所有必需验收项均已满足。
+ OutcomeAccepted Outcome = "accepted"
+ // OutcomeFailed 表示至少一个必需验收项缺少运行期证据或状态未收敛。
+ OutcomeFailed Outcome = "failed"
+)
+
+// Input 汇总最终验收所需的运行期事实和 plan 状态。
+type Input struct {
+ PlanVerify agentsession.AcceptChecks
+ Facts runtimefacts.RuntimeFacts
+ Todos []agentsession.TodoItem
+ LastAssistantText string
+}
+
+// CheckResult 描述单个验收项的判定结果。
+type CheckResult struct {
+ Passed bool `json:"passed"`
+ Name string `json:"name"`
+ Kind string `json:"kind,omitempty"`
+ Target string `json:"target,omitempty"`
+ Reason string `json:"reason,omitempty"`
+}
+
+// Report 描述 Accept Gate 的完整判定报告。
+type Report struct {
+ Outcome Outcome `json:"status"`
+ StopReason controlplane.StopReason `json:"stop_reason,omitempty"`
+ Summary string `json:"summary,omitempty"`
+ Results []CheckResult `json:"results,omitempty"`
+}
+
+// Evaluate 按固定顺序检查 plan-owned todo 与 Plan.Verify 运行期证据。
+func Evaluate(ctx context.Context, input Input) Report {
+ if err := ctx.Err(); err != nil {
+ return Report{
+ Outcome: OutcomeFailed,
+ StopReason: controlplane.StopReasonFatalError,
+ Summary: err.Error(),
+ }
+ }
+
+ report := Report{
+ Outcome: OutcomeAccepted,
+ StopReason: controlplane.StopReasonAccepted,
+ }
+
+ report.add(checkRequiredTodoFailures(input.Todos))
+ report.add(checkRequiredTodoConvergence(input.Todos))
+
+ checks := input.PlanVerify.Normalize()
+ if len(checks) == 0 {
+ checks = agentsession.AcceptChecks{{Kind: agentsession.AcceptCheckOutputOnly}}
+ }
+ for _, check := range checks {
+ result := evaluateAcceptCheck(input, check)
+ if !check.RequiredValue() {
+ report.addOptional(result)
+ continue
+ }
+ report.add(result)
+ }
+ report.finalize()
+ return report
+}
+
+// add 记录必需验收项结果，并在失败时更新终态原因。
+func (r *Report) add(result CheckResult) {
+ if strings.TrimSpace(result.Name) == "" {
+ return
+ }
+ r.Results = append(r.Results, result)
+ if result.Passed {
+ return
+ }
+ r.Outcome = OutcomeFailed
+ switch result.Name {
+ case "required_todo_failed":
+ r.StopReason = controlplane.StopReasonRequiredTodoFailed
+ case "required_todo_convergence":
+ if r.StopReason != controlplane.StopReasonRequiredTodoFailed {
+ r.StopReason = controlplane.StopReasonTodoNotConverged
+ }
+ default:
+ if r.StopReason == "" || r.StopReason == controlplane.StopReasonAccepted {
+ r.StopReason = controlplane.StopReasonAcceptCheckFailed
+ }
+ }
+}
+
+// addOptional 保留可选验收项结果，但不让可选失败改变终态。
+func (r *Report) addOptional(result CheckResult) {
+ if strings.TrimSpace(result.Name) == "" {
+ return
+ }
+ r.Results = append(r.Results, result)
+}
+
+// finalize 汇总逐项失败原因，形成对上层展示稳定的终态摘要。
+func (r *Report) finalize() {
+ if r.Outcome == OutcomeAccepted {
+ r.StopReason = controlplane.StopReasonAccepted
+ r.Summary = "acceptance checks passed"
+ return
+ }
+ if r.StopReason == "" || r.StopReason == controlplane.StopReasonAccepted {
+ r.StopReason = controlplane.StopReasonAcceptCheckFailed
+ }
+ failures := make([]string, 0, len(r.Results))
+ for _, result := range r.Results {
+ if result.Passed {
+ continue
+ }
+ reason := strings.TrimSpace(result.Reason)
+ if reason == "" {
+ reason = "failed"
+ }
+ failures = append(failures, fmt.Sprintf("%s: %s", result.Name, reason))
+ }
+ r.Summary = strings.Join(failures, "; ")
+}
diff --git a/internal/runtime/acceptgate/gate_test.go b/internal/runtime/acceptgate/gate_test.go
new file mode 100644
index 00000000..15aefae0
--- /dev/null
+++ b/internal/runtime/acceptgate/gate_test.go
@@ -0,0 +1,177 @@
+package acceptgate
+
+import (
+ "context"
+ "encoding/json"
+ "testing"
+
+ "neo-code/internal/runtime/controlplane"
+ runtimefacts "neo-code/internal/runtime/facts"
+ agentsession "neo-code/internal/session"
+)
+
+func TestEvaluateFallbackOutputOnly(t *testing.T) {
+ t.Parallel()
+
+ report := Evaluate(context.Background(), Input{LastAssistantText: "done"})
+ if report.Outcome != OutcomeAccepted || report.StopReason != controlplane.StopReasonAccepted {
+ t.Fatalf("report = %+v, want accepted", report)
+ }
+
+ report = Evaluate(context.Background(), Input{})
+ if report.Outcome != OutcomeFailed || report.StopReason != controlplane.StopReasonAcceptCheckFailed {
+ t.Fatalf("report = %+v, want accept_check_failed", report)
+ }
+}
+
+func TestEvaluateCommandSuccess(t *testing.T) {
+ t.Parallel()
+
+ input := Input{
+ PlanVerify: agentsession.AcceptChecks{{Kind: agentsession.AcceptCheckCommandSuccess, Target: "go test ./..."}},
+ Facts: runtimefacts.RuntimeFacts{
+ Commands: runtimefacts.CommandFacts{Executed: []runtimefacts.CommandFact{
+ {Tool: "bash", Command: "GOFLAGS=-count=1 go test ./...", Succeeded: true},
+ }},
+ },
+ LastAssistantText: "done",
+ }
+ if report := Evaluate(context.Background(), input); report.Outcome != OutcomeAccepted {
+ t.Fatalf("report = %+v, want accepted", report)
+ }
+
+ input.Facts.Commands.Executed[0].Succeeded = false
+ if report := Evaluate(context.Background(), input); report.Outcome != OutcomeFailed {
+ t.Fatalf("report = %+v, want failed", report)
+ }
+}
+
+func TestEvaluateWorkspaceChangeUsesRuntimeFactsOnly(t *testing.T) {
+ t.Parallel()
+
+ input := Input{
+ PlanVerify: agentsession.AcceptChecks{{Kind: agentsession.AcceptCheckWorkspaceChange}},
+ Facts: runtimefacts.RuntimeFacts{
+ Files: runtimefacts.FileFacts{Written: []runtimefacts.FileWriteFact{{Path: "internal/foo.go"}}},
+ },
+ LastAssistantText: "done",
+ }
+ if report := Evaluate(context.Background(), input); report.Outcome != OutcomeAccepted {
+ t.Fatalf("written fact report = %+v, want accepted", report)
+ }
+
+ input.Facts.Files.Written = nil
+ input.Facts.Files.Exists = []runtimefacts.FileExistFact{{Path: "internal/foo.go", Source: "filesystem_write_file"}}
+ if report := Evaluate(context.Background(), input); report.Outcome != OutcomeAccepted {
+ t.Fatalf("write-source exists report = %+v, want accepted", report)
+ }
+
+ input.Facts.Files.Exists = []runtimefacts.FileExistFact{{Path: "internal/foo.go", Source: "filesystem_read_file"}}
+ if report := Evaluate(context.Background(), input); report.Outcome != OutcomeFailed {
+ t.Fatalf("read-only fact report = %+v, want failed", report)
+ }
+
+ input.Facts.Files.Exists = []runtimefacts.FileExistFact{{Path: "internal/foo.go", Source: "filesystem_write_file_noop"}}
+ if report := Evaluate(context.Background(), input); report.Outcome != OutcomeFailed {
+ t.Fatalf("noop write fact report = %+v, want failed", report)
+ }
+}
+
+func TestEvaluateFileAndContentFacts(t *testing.T) {
+ t.Parallel()
+
+ input := Input{
+ PlanVerify: agentsession.AcceptChecks{
+ {Kind: agentsession.AcceptCheckFileExists, Target: "./README.md"},
+ {Kind: agentsession.AcceptCheckContentContains, Target: "README.md", Params: map[string]string{"contains": "NeoCode"}},
+ },
+ Facts: runtimefacts.RuntimeFacts{
+ Files: runtimefacts.FileFacts{
+ Exists: []runtimefacts.FileExistFact{{Path: "README.md", Source: "filesystem_read_file"}},
+ ContentMatch: []runtimefacts.FileContentMatchFact{{
+ Path: "README.md",
+ ExpectedContains: []string{"NeoCode"},
+ VerificationPassed: true,
+ }},
+ },
+ },
+ LastAssistantText: "done",
+ }
+ if report := Evaluate(context.Background(), input); report.Outcome != OutcomeAccepted {
+ t.Fatalf("report = %+v, want accepted", report)
+ }
+
+ input.Facts.Files.ContentMatch[0].VerificationPassed = false
+ report := Evaluate(context.Background(), input)
+ if report.Outcome != OutcomeFailed || len(report.Results) != 4 {
+ t.Fatalf("report = %+v, want failed with all results", report)
+ }
+}
+
+func TestEvaluateToolFactAndUnknownKind(t *testing.T) {
+ t.Parallel()
+
+ input := Input{
+ PlanVerify: agentsession.AcceptChecks{
+ {Kind: agentsession.AcceptCheckToolFact, Params: map[string]string{"tool": "bash", "scope": "test"}},
+ {Kind: "future_check"},
+ },
+ Facts: runtimefacts.RuntimeFacts{
+ Verification: runtimefacts.VerificationFacts{
+ Passed: []runtimefacts.VerificationFact{{Tool: "bash", Scope: "test"}},
+ },
+ },
+ LastAssistantText: "done",
+ }
+ report := Evaluate(context.Background(), input)
+ if report.Outcome != OutcomeFailed || report.StopReason != controlplane.StopReasonAcceptCheckFailed {
+ t.Fatalf("report = %+v, want unknown kind failure", report)
+ }
+ if report.Results[len(report.Results)-1].Reason != "unknown required accept check kind" {
+ t.Fatalf("last result = %+v, want unknown kind reason", report.Results[len(report.Results)-1])
+ }
+}
+
+func TestEvaluateOptionalUnknownKindDoesNotFail(t *testing.T) {
+ t.Parallel()
+
+ optional := false
+ var checks agentsession.AcceptChecks
+ if err := json.Unmarshal([]byte(`[{"kind":"output_only"},{"kind":"future_check","required":false}]`), &checks); err != nil {
+ t.Fatalf("Unmarshal() error = %v", err)
+ }
+ checks = append(checks, agentsession.AcceptCheck{Kind: "go_literal_optional", Required: &optional})
+ report := Evaluate(context.Background(), Input{
+ PlanVerify: checks,
+ LastAssistantText: "done",
+ })
+ if report.Outcome != OutcomeAccepted {
+ t.Fatalf("report = %+v, want accepted", report)
+ }
+ if len(report.Results) != 5 {
+ t.Fatalf("results len = %d, want 5", len(report.Results))
+ }
+}
+
+func TestEvaluateTodoPriority(t *testing.T) {
+ t.Parallel()
+
+ required := true
+ input := Input{
+ PlanVerify: agentsession.AcceptChecks{{Kind: agentsession.AcceptCheckOutputOnly}},
+ LastAssistantText: "done",
+ Todos: []agentsession.TodoItem{
+ {ID: "todo-1", Status: agentsession.TodoStatusFailed, Required: &required},
+ },
+ }
+ report := Evaluate(context.Background(), input)
+ if report.Outcome != OutcomeFailed || report.StopReason != controlplane.StopReasonRequiredTodoFailed {
+ t.Fatalf("failed todo report = %+v, want required_todo_failed", report)
+ }
+
+ input.Todos[0].Status = agentsession.TodoStatusPending
+ report = Evaluate(context.Background(), input)
+ if report.Outcome != OutcomeFailed || report.StopReason != controlplane.StopReasonTodoNotConverged {
+ t.Fatalf("pending todo report = %+v, want todo_not_converged", report)
+ }
+}
diff --git a/internal/runtime/acceptgate_runtime.go b/internal/runtime/acceptgate_runtime.go
new file mode 100644
index 00000000..a8c890f4
--- /dev/null
+++ b/internal/runtime/acceptgate_runtime.go
@@ -0,0 +1,131 @@
+package runtime
+
+import (
+ "context"
+ "strings"
+
+ "neo-code/internal/partsrender"
+ "neo-code/internal/promptasset"
+ providertypes "neo-code/internal/provider/types"
+ "neo-code/internal/runtime/acceptgate"
+ runtimefacts "neo-code/internal/runtime/facts"
+ agentsession "neo-code/internal/session"
+)
+
+const missingCompletionSignalLimit = 6
+
+// completionProtocolReminderForStreak 根据连续缺失完成信号的次数返回对应协议提示。
+func completionProtocolReminderForStreak(streak int) string {
+ if streak >= missingCompletionSignalLimit-1 {
+ return promptasset.CompletionProtocolFinalReminder()
+ }
+ return promptasset.CompletionProtocolReminder()
+}
+
+// evaluateAcceptGate 从运行态提取事实快照，并执行最终 Accept Gate。
+func (s *Service) evaluateAcceptGate(ctx context.Context, state *runState, assistantMessage providertypes.Message) acceptgate.Report {
+ if state == nil {
+ return acceptgate.Evaluate(ctx, acceptgate.Input{})
+ }
+ state.mu.Lock()
+ var planVerify agentsession.AcceptChecks
+ var currentPlan *agentsession.PlanArtifact
+ if state.session.CurrentPlan != nil {
+ currentPlan = state.session.CurrentPlan.Clone()
+ planVerify = currentPlan.Summary.Verify.Clone()
+ if len(planVerify) == 0 {
+ planVerify = currentPlan.Spec.Verify.Clone()
+ }
+ }
+ todos := selectPlanOwnedTodos(currentPlan, cloneTodosForPersistence(state.session.Todos))
+ factsSnapshot := runtimefacts.RuntimeFacts{}
+ if state.factsCollector != nil {
+ factsSnapshot = state.factsCollector.Snapshot()
+ }
+ state.mu.Unlock()
+
+ return acceptgate.Evaluate(ctx, acceptgate.Input{
+ PlanVerify: planVerify,
+ Facts: factsSnapshot,
+ Todos: todos,
+ LastAssistantText: renderAssistantTextWithoutCompletion(assistantMessage),
+ })
+}
+
+// selectPlanOwnedTodos 只把当前计划拥有的 todo 交给终态验收，避免无 plan 的 chat/read-only 被旧 todo 污染。
+func selectPlanOwnedTodos(plan *agentsession.PlanArtifact, todos []agentsession.TodoItem) []agentsession.TodoItem {
+ if plan == nil || len(todos) == 0 {
+ return nil
+ }
+ owned := make(map[string]struct{})
+ for _, id := range plan.Summary.ActiveTodoIDs {
+ id = strings.TrimSpace(id)
+ if id != "" {
+ owned[id] = struct{}{}
+ }
+ }
+ for _, todo := range plan.Spec.Todos {
+ id := strings.TrimSpace(todo.ID)
+ if id != "" {
+ owned[id] = struct{}{}
+ }
+ }
+ selected := make([]agentsession.TodoItem, 0, len(todos))
+ for _, todo := range todos {
+ if _, ok := owned[strings.TrimSpace(todo.ID)]; ok {
+ selected = append(selected, todo)
+ continue
+ }
+ if isPostPlanRequiredTodo(plan, todo) {
+ selected = append(selected, todo)
+ }
+ }
+ return selected
+}
+
+// isPostPlanRequiredTodo 判断计划执行期新增的必需 todo 是否应纳入当前计划验收。
+func isPostPlanRequiredTodo(plan *agentsession.PlanArtifact, todo agentsession.TodoItem) bool {
+ if plan == nil || !todo.RequiredValue() || todo.Status.IsTerminal() {
+ return false
+ }
+ if plan.CreatedAt.IsZero() || todo.CreatedAt.IsZero() {
+ return false
+ }
+ return !todo.CreatedAt.Before(plan.CreatedAt)
+}
+
+// emitAcceptGateReport 将 Accept Gate 报告发布为统一 acceptance_decided 事件。
+func (s *Service) emitAcceptGateReport(state *runState, report acceptgate.Report) {
+ status := string(acceptgate.OutcomeFailed)
+ if report.Outcome == acceptgate.OutcomeAccepted {
+ status = string(acceptgate.OutcomeAccepted)
+ }
+ s.emitRunScopedOptional(EventAcceptanceDecided, state, AcceptanceDecidedPayload{
+ Status: status,
+ StopReason: report.StopReason,
+ Summary: report.Summary,
+ Results: append([]acceptgate.CheckResult(nil), report.Results...),
+ })
+}
+
+func renderAssistantTextWithoutCompletion(message providertypes.Message) string {
+ text := strings.TrimSpace(partsrender.RenderDisplayParts(message.Parts))
+ if text == "" {
+ return ""
+ }
+ candidate, ok := extractPlanningJSONObjectIfPresent(text, "task_completion")
+ if !ok {
+ return text
+ }
+ return strings.TrimSpace(stripPlanningJSONObjectText(text, candidate))
+}
+
+// stripCompletionSignalFromAssistantMessage 移除仅供 runtime 控制使用的 task_completion JSON，保留用户可见回复。
+func stripCompletionSignalFromAssistantMessage(message providertypes.Message) providertypes.Message {
+ text := renderAssistantTextWithoutCompletion(message)
+ if strings.TrimSpace(text) == strings.TrimSpace(partsrender.RenderDisplayParts(message.Parts)) {
+ return message
+ }
+ message.Parts = []providertypes.ContentPart{providertypes.NewTextPart(text)}
+ return message
+}
diff --git a/internal/runtime/acceptgate_runtime_test.go b/internal/runtime/acceptgate_runtime_test.go
new file mode 100644
index 00000000..32bab233
--- /dev/null
+++ b/internal/runtime/acceptgate_runtime_test.go
@@ -0,0 +1,81 @@
+package runtime
+
+import (
+ "testing"
+ "time"
+
+ agentsession "neo-code/internal/session"
+)
+
+func TestSelectPlanOwnedTodosIncludesPostPlanRequired(t *testing.T) {
+ t.Parallel()
+
+ required := true
+ optional := false
+ createdAt := time.Date(2026, 5, 9, 10, 0, 0, 0, time.UTC)
+ plan := &agentsession.PlanArtifact{
+ Status: agentsession.PlanStatusApproved,
+ CreatedAt: createdAt,
+ Spec: agentsession.PlanSpec{Todos: []agentsession.TodoItem{
+ {ID: "plan-owned", Status: agentsession.TodoStatusPending},
+ }},
+ }
+ todos := []agentsession.TodoItem{
+ {
+ ID: "plan-owned",
+ Status: agentsession.TodoStatusPending,
+ Required: &required,
+ CreatedAt: createdAt.Add(-time.Hour),
+ },
+ {
+ ID: "post-required",
+ Status: agentsession.TodoStatusPending,
+ Required: &required,
+ CreatedAt: createdAt.Add(time.Minute),
+ },
+ {
+ ID: "old-required",
+ Status: agentsession.TodoStatusPending,
+ Required: &required,
+ CreatedAt: createdAt.Add(-time.Minute),
+ },
+ {
+ ID: "post-optional",
+ Status: agentsession.TodoStatusPending,
+ Required: &optional,
+ CreatedAt: createdAt.Add(time.Minute),
+ },
+ {
+ ID: "post-completed",
+ Status: agentsession.TodoStatusCompleted,
+ Required: &required,
+ CreatedAt: createdAt.Add(time.Minute),
+ },
+ }
+
+ selected := selectPlanOwnedTodos(plan, todos)
+ if len(selected) != 2 {
+ t.Fatalf("selected length = %d, want 2: %+v", len(selected), selected)
+ }
+ if selected[0].ID != "plan-owned" || selected[1].ID != "post-required" {
+ t.Fatalf("selected IDs = [%s %s], want [plan-owned post-required]", selected[0].ID, selected[1].ID)
+ }
+}
+
+func TestSelectPlanOwnedTodosRequiresPlanForPostPlanRequired(t *testing.T) {
+ t.Parallel()
+
+ required := true
+ todos := []agentsession.TodoItem{
+ {
+ ID: "post-required",
+ Status: agentsession.TodoStatusPending,
+ Required: &required,
+ CreatedAt: time.Now(),
+ },
+ }
+
+ if selected := selectPlanOwnedTodos(nil, todos); selected != nil {
+ t.Fatalf("selectPlanOwnedTodos(nil) = %+v, want nil", selected)
+ }
+}
diff --git a/internal/runtime/ask.go b/internal/runtime/ask.go
index 76041989..b1ee47a5 100644
--- a/internal/runtime/ask.go
+++ b/internal/runtime/ask.go
@@ -11,6 +11,7 @@ import (
"neo-code/internal/config"
agentcontext "neo-code/internal/context"
+ "neo-code/internal/partsrender"
"neo-code/internal/provider"
providertypes "neo-code/internal/provider/types"
"neo-code/internal/runtime/streaming"
@@ -145,7 +146,7 @@ func (s *Service) Ask(ctx context.Context, input AskInput) error {
return failAsk(streamOutcome.err)
}
- reply := strings.TrimSpace(renderPartsForVerification(streamOutcome.message.Parts))
+ reply := strings.TrimSpace(partsrender.RenderDisplayParts(streamOutcome.message.Parts))
session = appendAskMessage(session, "assistant", reply)
if err := s.askStore.Save(ctx, session); err != nil {
return failAsk(err)
diff --git a/internal/runtime/before_completion_orchestrator.go b/internal/runtime/before_completion_orchestrator.go
deleted file mode 100644
index 77974edf..00000000
--- a/internal/runtime/before_completion_orchestrator.go
+++ /dev/null
@@ -1,160 +0,0 @@
-package runtime
-
-import (
- "context"
- "strings"
-
- providertypes "neo-code/internal/provider/types"
- "neo-code/internal/runtime/acceptance"
- "neo-code/internal/runtime/decider"
- runtimehooks "neo-code/internal/runtime/hooks"
-)
-
-// beforeCompletionHookSignals 收敛 before_completion_decision 阶段 user/repo hook 的可消费信号。
-type beforeCompletionHookSignals struct {
- Annotations []string
- Guards []decider.HookGuardSignal
-}
-
-// runBeforeCompletionDecisionAcceptance 执行 before_completion_decision 专用编排：
-// 1) 先执行 user/repo hooks 收集 annotation/guard signal；
-// 2) 再执行普通 internal hooks 用于观测；
-// 3) 最后由 runtime 内部 AcceptanceService 作为 before_completion_decision 的收口裁决阶段，生成唯一 AcceptanceDecision。
-// AcceptanceDecision 走强类型 runtime 内部路径，不通过通用 HookResult metadata 承载。
-func (s *Service) runBeforeCompletionDecisionAcceptance(
- ctx context.Context,
- state *runState,
- snapshot TurnBudgetSnapshot,
- assistant providertypes.Message,
- workdir string,
- completionPassed bool,
- hasToolCalls bool,
- assistantRole string,
-) (acceptance.AcceptanceDecision, error) {
- if s == nil {
- return acceptance.AcceptanceDecision{}, nil
- }
-
- point := runtimehooks.HookPointBeforeCompletionDecision
- hookInput := s.buildRunHookContext(
- state,
- runtimehooks.HookContext{
- Metadata: map[string]any{
- "completion_passed": completionPassed,
- "has_tool_calls": hasToolCalls,
- "assistant_role": strings.TrimSpace(assistantRole),
- "workdir": strings.TrimSpace(workdir),
- },
- },
- )
- scopedCtx := withRuntimeHookEnvelope(ctx, hookRuntimeEnvelope{
- RunID: firstNonBlank(hookRunIDFromState(state), hookInput.RunID),
- SessionID: firstNonBlank(hookSessionIDFromState(state), hookInput.SessionID),
- Turn: hookTurnFromState(state),
- Phase: hookPhaseFromState(state),
- })
-
- signals := beforeCompletionHookSignals{}
- if s.hookExecutor != nil {
- baseExecutor, userExecutor, repoExecutor := splitHookExecutors(s.hookExecutor)
-
- for _, item := range []struct {
- executor HookExecutor
- source runtimehooks.HookSource
- }{
- {executor: userExecutor, source: runtimehooks.HookSourceUser},
- {executor: repoExecutor, source: runtimehooks.HookSourceRepo},
- } {
- if item.executor == nil {
- continue
- }
- output := item.executor.Run(scopedCtx, point, hookInput.Clone())
- annotations, guards := collectBeforeCompletionSignals(output, item.source)
- signals.Annotations = append(signals.Annotations, annotations...)
- signals.Guards = append(signals.Guards, guards...)
- s.recordUserHookAnnotations(state, output)
- }
-
- // internal hooks 在该点位最后执行；其结果仅用于观测，不参与 user/repo signal 收集。
- if baseExecutor != nil {
- output := baseExecutor.Run(scopedCtx, point, hookInput.Clone())
- s.recordUserHookAnnotations(state, output)
- }
- }
- s.emitRunScopedOptional(EventVerificationStarted, state, VerificationStartedPayload{
- CompletionPassed: completionPassed,
- CompletionBlockedReason: strings.TrimSpace(string(state.completion.CompletionBlockedReason)),
- })
- // 收口裁决阶段：消费 completion/facts/todo/verification/user-repo signals，生成唯一终态裁决。
- return s.beforeAcceptFinal(ctx, state, snapshot, assistant, completionPassed, signals)
-}
-
-// buildRunHookContext 构造带 run/session 元数据的 hook 输入。
-func (s *Service) buildRunHookContext(state *runState, input runtimehooks.HookContext) runtimehooks.HookContext {
- runID := firstNonBlank(hookRunIDFromState(state), input.RunID)
- sessionID := firstNonBlank(hookSessionIDFromState(state), input.SessionID)
- input.RunID = firstNonBlank(input.RunID, runID)
- input.SessionID = firstNonBlank(input.SessionID, sessionID)
- if input.Metadata == nil {
- input.Metadata = make(map[string]any, 8)
- }
- input.Metadata["run_id"] = input.RunID
- input.Metadata["session_id"] = input.SessionID
- if state != nil {
- input.Metadata["runtime_run_token"] = state.runToken
- if _, exists := input.Metadata["phase"]; !exists {
- input.Metadata["phase"] = hookPhaseFromState(state)
- }
- input.Metadata["turn"] = hookTurnFromState(state)
- }
- return input
-}
-
-// splitHookExecutors 拆解 composeRuntimeHookExecutors 形成的链，恢复 internal/user/repo 三段执行器。
-func splitHookExecutors(executor HookExecutor) (base HookExecutor, user HookExecutor, repo HookExecutor) {
- switch typed := executor.(type) {
- case *repoComposedHookExecutor:
- subBase, subUser, _ := splitHookExecutors(typed.base)
- return subBase, subUser, typed.repo
- case *userComposedHookExecutor:
- subBase, _, subRepo := splitHookExecutors(typed.base)
- return subBase, typed.user, subRepo
- default:
- return executor, nil, nil
- }
-}
-
-// collectBeforeCompletionSignals 从 user/repo hook 结果提取 annotation 与 guard 信号。
-func collectBeforeCompletionSignals(
- output runtimehooks.RunOutput,
- defaultSource runtimehooks.HookSource,
-) ([]string, []decider.HookGuardSignal) {
- if len(output.Results) == 0 {
- return nil, nil
- }
- annotations := make([]string, 0, len(output.Results))
- guards := make([]decider.HookGuardSignal, 0, len(output.Results))
- for _, result := range output.Results {
- source := strings.TrimSpace(string(result.Source))
- if source == "" {
- source = strings.TrimSpace(string(defaultSource))
- }
- message := strings.TrimSpace(result.Message)
- errText := strings.TrimSpace(result.Error)
- if message != "" {
- annotations = append(annotations, message)
- }
-
- isGuard := result.Status == runtimehooks.HookResultFailed || result.Metadata.GuardSignal
- if !isGuard {
- continue
- }
- guard := decider.HookGuardSignal{
- HookID: strings.TrimSpace(result.HookID),
- Source: source,
- Message: firstNonBlank(message, errText),
- }
- guards = append(guards, guard)
- }
- return annotations, guards
-}
diff --git a/internal/runtime/budget_models.go b/internal/runtime/budget_models.go
index e56d9768..12c8ec94 100644
--- a/internal/runtime/budget_models.go
+++ b/internal/runtime/budget_models.go
@@ -20,7 +20,6 @@ type TurnBudgetSnapshot struct {
PromptBudget int
BudgetSource string
CompactCount int
- NoProgressStreakLimit int
RepeatCycleStreakLimit int
InjectFullPlan bool
ContextWindow int
@@ -62,7 +61,6 @@ func newTurnBudgetSnapshot(
promptBudget int,
budgetSource string,
compactCount int,
- noProgressStreakLimit int,
repeatCycleStreakLimit int,
injectFullPlan bool,
contextWindow int,
@@ -84,7 +82,6 @@ func newTurnBudgetSnapshot(
PromptBudget: promptBudget,
BudgetSource: budgetSource,
CompactCount: compactCount,
- NoProgressStreakLimit: noProgressStreakLimit,
RepeatCycleStreakLimit: repeatCycleStreakLimit,
InjectFullPlan: injectFullPlan,
ContextWindow: contextWindow,
diff --git a/internal/runtime/controlplane/progress.go b/internal/runtime/controlplane/progress.go
index 7a74438a..f185ac43 100644
--- a/internal/runtime/controlplane/progress.go
+++ b/internal/runtime/controlplane/progress.go
@@ -1,21 +1,5 @@
package controlplane
-// ProgressEvidenceKind 标识 runtime 聚合得到的结构化进展证据。
-type ProgressEvidenceKind string
-
-const (
- // EvidenceTaskStateChanged 表示任务状态发生合法迁移。
- EvidenceTaskStateChanged ProgressEvidenceKind = "TASK_STATE_CHANGED"
- // EvidenceTodoStateChanged 表示 todo 列表发生结构化变化。
- EvidenceTodoStateChanged ProgressEvidenceKind = "TODO_STATE_CHANGED"
- // EvidenceWriteApplied 表示本轮产生了有效文件改动。
- EvidenceWriteApplied ProgressEvidenceKind = "WRITE_APPLIED"
- // EvidenceVerifyPassed 表示本轮存在明确的验证成功信号（仅与写入证据组合后算业务推进）。
- EvidenceVerifyPassed ProgressEvidenceKind = "VERIFY_PASSED"
- // EvidenceNewInfoNonDup 表示本轮引入了去重后的新信息。
- EvidenceNewInfoNonDup ProgressEvidenceKind = "NEW_INFO_NON_DUP"
-)
-
// SubgoalRelation 表示当前轮子目标与上一轮的关系。
type SubgoalRelation string
@@ -28,7 +12,7 @@ const (
SubgoalRelationUnknown SubgoalRelation = "unknown"
)
-// StalledProgressState 表示当前进展是否已进入软卡住状态。
+// StalledProgressState 表示当前重复循环检测是否已进入软卡住状态。
type StalledProgressState string
const (
@@ -44,38 +28,23 @@ type ReminderKind string
const (
// ReminderKindNone 表示当前轮无需注入提醒。
ReminderKindNone ReminderKind = ""
- // ReminderKindNoProgress 表示应注入无进展提醒。
- ReminderKindNoProgress ReminderKind = "REMINDER_NO_PROGRESS"
// ReminderKindRepeatCycle 表示应注入重复循环提醒。
ReminderKindRepeatCycle ReminderKind = "REMINDER_REPEAT_CYCLE"
- // ReminderKindGenericStalled 表示应注入通用 stalled 提醒。
- ReminderKindGenericStalled ReminderKind = "REMINDER_GENERIC_STALLED"
)
-// ProgressEvidenceRecord 描述一条结构化进展证据。
-type ProgressEvidenceRecord struct {
- Kind ProgressEvidenceKind `json:"kind"`
- Detail string `json:"detail,omitempty"`
-}
-
-// ProgressScore 表示一次 progress 评估后的完整快照。
+// ProgressScore 表示一次重复循环检测后的快照。
type ProgressScore struct {
- HasBusinessProgress bool `json:"has_business_progress"`
- HasExplorationProgress bool `json:"has_exploration_progress"`
- StrongEvidenceCount int `json:"strong_evidence_count"`
- MediumEvidenceCount int `json:"medium_evidence_count"`
- WeakEvidenceCount int `json:"weak_evidence_count"`
- ExplorationStreak int `json:"exploration_streak"`
- NoProgressStreak int `json:"no_progress_streak"`
- RepeatCycleStreak int `json:"repeat_cycle_streak"`
- SameToolSignature bool `json:"same_tool_signature"`
- SameResultFingerprint bool `json:"same_result_fingerprint"`
- SameSubgoal SubgoalRelation `json:"same_subgoal"`
- StalledProgressState StalledProgressState `json:"stalled_progress_state"`
- ReminderKind ReminderKind `json:"reminder_kind,omitempty"`
+ RepeatCycleStreak int `json:"repeat_cycle_streak"`
+ SameToolSignature bool `json:"same_tool_signature"`
+ SameResultFingerprint bool `json:"same_result_fingerprint"`
+ SameSubgoal SubgoalRelation `json:"same_subgoal"`
+ StalledProgressState StalledProgressState `json:"stalled_progress_state"`
+ ReminderKind ReminderKind `json:"reminder_kind,omitempty"`
+ ShouldTerminate bool `json:"should_terminate"`
+ TerminateReason StopReason `json:"terminate_reason,omitempty"`
}
-// ProgressState 保存跨轮 progress 判定所需的历史快照。
+// ProgressState 保存跨轮重复循环检测所需的历史快照。
type ProgressState struct {
LastScore ProgressScore `json:"last_score"`
LastToolSignature string `json:"last_tool_signature,omitempty"`
@@ -83,27 +52,17 @@ type ProgressState struct {
LastSubgoalFingerprint string `json:"last_subgoal_fingerprint,omitempty"`
}
-// ProgressInput 描述一次 progress 评估所需的事实输入。
+// ProgressInput 描述一次重复循环检测所需的指纹输入。
type ProgressInput struct {
- RunState RunState
- Evidence []ProgressEvidenceRecord
CurrentToolSignature string
ResultFingerprint string
SubgoalFingerprint string
- NoProgressLimit int
RepeatCycleLimit int
}
-// EvaluateProgress 基于上一轮状态和本轮事实生成新的 progress 快照。
+// EvaluateProgress 基于上一轮指纹和本轮指纹检测 agent 是否陷入重复循环。
func EvaluateProgress(state ProgressState, input ProgressInput) ProgressState {
next := ProgressScore{}
- flags := summarizeEvidence(input.Evidence)
-
- next.StrongEvidenceCount = flags.strongCount
- next.MediumEvidenceCount = flags.mediumCount
- next.WeakEvidenceCount = flags.weakCount
- next.HasBusinessProgress = flags.strongCount > 0 || (flags.hasWrite && flags.hasVerify)
- next.HasExplorationProgress = !next.HasBusinessProgress && isExplorationProgress(input.RunState, flags)
next.SameToolSignature = input.CurrentToolSignature != "" &&
state.LastToolSignature != "" &&
input.CurrentToolSignature == state.LastToolSignature
@@ -112,35 +71,23 @@ func EvaluateProgress(state ProgressState, input ProgressInput) ProgressState {
input.ResultFingerprint == state.LastResultFingerprint
next.SameSubgoal = compareSubgoalFingerprint(state.LastSubgoalFingerprint, input.SubgoalFingerprint)
- if next.HasBusinessProgress {
- next.ExplorationStreak = 0
- next.NoProgressStreak = 0
- } else if next.HasExplorationProgress {
- next.ExplorationStreak = state.LastScore.ExplorationStreak + 1
- next.NoProgressStreak = state.LastScore.NoProgressStreak
- if next.ExplorationStreak > explorationWindowForPhase(input.RunState) {
- next.NoProgressStreak++
- }
- } else {
- next.ExplorationStreak = 0
- next.NoProgressStreak = state.LastScore.NoProgressStreak + 1
- }
-
- if next.HasBusinessProgress {
- next.RepeatCycleStreak = 0
- } else if next.SameToolSignature && next.SameResultFingerprint && next.SameSubgoal == SubgoalRelationSame {
+ if next.SameToolSignature && next.SameResultFingerprint && next.SameSubgoal == SubgoalRelationSame {
next.RepeatCycleStreak = state.LastScore.RepeatCycleStreak + 1
} else {
next.RepeatCycleStreak = 0
}
- if shouldStall(next, input.NoProgressLimit, input.RepeatCycleLimit) {
+ if shouldStall(next, input.RepeatCycleLimit) {
next.StalledProgressState = StalledProgressStalled
- next.ReminderKind = selectReminderKind(next)
+ next.ReminderKind = ReminderKindRepeatCycle
} else {
next.StalledProgressState = StalledProgressHealthy
next.ReminderKind = ReminderKindNone
}
+ if shouldTerminateAfterStalledReminder(state.LastScore, next) {
+ next.ShouldTerminate = true
+ next.TerminateReason = StopReasonRepeatCycle
+ }
return ProgressState{
LastScore: next,
@@ -150,55 +97,13 @@ func EvaluateProgress(state ProgressState, input ProgressInput) ProgressState {
}
}
-type evidenceFlags struct {
- strongCount int
- mediumCount int
- weakCount int
- hasWrite bool
- hasVerify bool
-}
-
-// summarizeEvidence 汇总本轮 evidence 的强中弱计数与关键标记。
-func summarizeEvidence(records []ProgressEvidenceRecord) evidenceFlags {
- var flags evidenceFlags
- for _, record := range records {
- switch record.Kind {
- case EvidenceTaskStateChanged, EvidenceTodoStateChanged:
- flags.strongCount++
- case EvidenceWriteApplied, EvidenceVerifyPassed:
- flags.mediumCount++
- case EvidenceNewInfoNonDup:
- flags.weakCount++
- }
-
- switch record.Kind {
- case EvidenceWriteApplied:
- flags.hasWrite = true
- case EvidenceVerifyPassed:
- flags.hasVerify = true
- }
- }
- return flags
-}
-
-// isExplorationProgress 判断本轮是否属于可被宽容窗口吸收的探索型推进。
-func isExplorationProgress(runState RunState, flags evidenceFlags) bool {
- if runState != RunStatePlan && runState != RunStateExecute {
+// shouldTerminateAfterStalledReminder 只在 repeat stalled 已提醒过一轮后才允许硬终止。
+func shouldTerminateAfterStalledReminder(previous ProgressScore, current ProgressScore) bool {
+ if current.StalledProgressState != StalledProgressStalled || current.ReminderKind == ReminderKindNone {
return false
}
- return flags.weakCount > 0
-}
-
-// explorationWindowForPhase 返回不同阶段允许的 exploration 宽容窗口。
-func explorationWindowForPhase(runState RunState) int {
- switch runState {
- case RunStatePlan:
- return 4
- case RunStateExecute:
- return 2
- default:
- return 0
- }
+ return previous.StalledProgressState == StalledProgressStalled &&
+ previous.ReminderKind == current.ReminderKind
}
// compareSubgoalFingerprint 判断当前轮与上一轮的子目标关系。
@@ -215,24 +120,7 @@ func compareSubgoalFingerprint(previous string, current string) SubgoalRelation
return SubgoalRelationDifferent
}
-// shouldStall 判断当前快照是否应进入 stalled。
-func shouldStall(score ProgressScore, noProgressLimit int, repeatLimit int) bool {
- if repeatLimit > 0 && score.RepeatCycleStreak >= repeatLimit {
- return true
- }
- if noProgressLimit > 0 && score.NoProgressStreak >= noProgressLimit {
- return true
- }
- return false
-}
-
-// selectReminderKind 选择 stalled 场景下应注入的提醒类型。
-func selectReminderKind(score ProgressScore) ReminderKind {
- if score.RepeatCycleStreak > 0 && score.SameToolSignature && score.SameResultFingerprint {
- return ReminderKindRepeatCycle
- }
- if score.NoProgressStreak > 0 {
- return ReminderKindNoProgress
- }
- return ReminderKindGenericStalled
+// shouldStall 判断当前快照是否应进入 repeat stalled。
+func shouldStall(score ProgressScore, repeatLimit int) bool {
+ return repeatLimit > 0 && score.RepeatCycleStreak >= repeatLimit
}
diff --git a/internal/runtime/controlplane/progress_test.go b/internal/runtime/controlplane/progress_test.go
index fe450eda..288ebd05 100644
--- a/internal/runtime/controlplane/progress_test.go
+++ b/internal/runtime/controlplane/progress_test.go
@@ -2,91 +2,6 @@ package controlplane
import "testing"
-func TestEvaluateProgressBusinessProgressResetsStreaks(t *testing.T) {
- t.Parallel()
-
- state := ProgressState{
- LastScore: ProgressScore{
- ExplorationStreak: 2,
- NoProgressStreak: 3,
- RepeatCycleStreak: 1,
- },
- }
-
- got := EvaluateProgress(state, ProgressInput{
- RunState: RunStateExecute,
- Evidence: []ProgressEvidenceRecord{
- {Kind: EvidenceTodoStateChanged},
- },
- NoProgressLimit: 3,
- RepeatCycleLimit: 3,
- })
-
- if !got.LastScore.HasBusinessProgress {
- t.Fatalf("expected business progress")
- }
- if got.LastScore.NoProgressStreak != 0 {
- t.Fatalf("no-progress streak = %d, want 0", got.LastScore.NoProgressStreak)
- }
- if got.LastScore.RepeatCycleStreak != 0 {
- t.Fatalf("repeat streak = %d, want 0", got.LastScore.RepeatCycleStreak)
- }
-}
-
-func TestEvaluateProgressExplorationUsesWindow(t *testing.T) {
- t.Parallel()
-
- state := ProgressState{
- LastScore: ProgressScore{
- ExplorationStreak: 3,
- NoProgressStreak: 1,
- },
- }
-
- got := EvaluateProgress(state, ProgressInput{
- RunState: RunStatePlan,
- Evidence: []ProgressEvidenceRecord{
- {Kind: EvidenceNewInfoNonDup},
- },
- NoProgressLimit: 3,
- RepeatCycleLimit: 3,
- })
-
- if !got.LastScore.HasExplorationProgress {
- t.Fatalf("expected exploration progress")
- }
- if got.LastScore.ExplorationStreak != 4 {
- t.Fatalf("exploration streak = %d, want 4", got.LastScore.ExplorationStreak)
- }
- if got.LastScore.NoProgressStreak != 1 {
- t.Fatalf("no-progress streak = %d, want unchanged 1", got.LastScore.NoProgressStreak)
- }
-}
-
-func TestEvaluateProgressExplorationExhaustionStartsNoProgress(t *testing.T) {
- t.Parallel()
-
- state := ProgressState{
- LastScore: ProgressScore{
- ExplorationStreak: 4,
- NoProgressStreak: 1,
- },
- }
-
- got := EvaluateProgress(state, ProgressInput{
- RunState: RunStatePlan,
- Evidence: []ProgressEvidenceRecord{
- {Kind: EvidenceNewInfoNonDup},
- },
- NoProgressLimit: 3,
- RepeatCycleLimit: 3,
- })
-
- if got.LastScore.NoProgressStreak != 2 {
- t.Fatalf("no-progress streak = %d, want 2", got.LastScore.NoProgressStreak)
- }
-}
-
func TestEvaluateProgressRepeatCycleRequiresSameResultAndSubgoal(t *testing.T) {
t.Parallel()
@@ -98,11 +13,9 @@ func TestEvaluateProgressRepeatCycleRequiresSameResultAndSubgoal(t *testing.T) {
}
got := EvaluateProgress(state, ProgressInput{
- RunState: RunStateExecute,
CurrentToolSignature: "sig",
ResultFingerprint: "result",
SubgoalFingerprint: "subgoal",
- NoProgressLimit: 3,
RepeatCycleLimit: 3,
})
@@ -115,53 +28,108 @@ func TestEvaluateProgressRepeatCycleRequiresSameResultAndSubgoal(t *testing.T) {
if got.LastScore.ReminderKind != ReminderKindRepeatCycle {
t.Fatalf("reminder = %q, want %q", got.LastScore.ReminderKind, ReminderKindRepeatCycle)
}
+ if got.LastScore.ShouldTerminate {
+ t.Fatal("first stalled repeat-cycle should warn before hard terminate")
+ }
}
-func TestEvaluateProgressUnknownSubgoalDoesNotAdvanceRepeat(t *testing.T) {
+func TestEvaluateProgressRepeatCycleTerminatesAfterReminder(t *testing.T) {
t.Parallel()
state := ProgressState{
- LastScore: ProgressScore{RepeatCycleStreak: 1},
+ LastScore: ProgressScore{
+ RepeatCycleStreak: 3,
+ StalledProgressState: StalledProgressStalled,
+ ReminderKind: ReminderKindRepeatCycle,
+ },
LastToolSignature: "sig",
LastResultFingerprint: "result",
LastSubgoalFingerprint: "subgoal",
}
got := EvaluateProgress(state, ProgressInput{
- RunState: RunStateExecute,
CurrentToolSignature: "sig",
ResultFingerprint: "result",
- SubgoalFingerprint: "",
- NoProgressLimit: 3,
+ SubgoalFingerprint: "subgoal",
RepeatCycleLimit: 3,
})
- if got.LastScore.SameSubgoal != SubgoalRelationUnknown {
- t.Fatalf("same subgoal = %q, want %q", got.LastScore.SameSubgoal, SubgoalRelationUnknown)
- }
- if got.LastScore.RepeatCycleStreak != 0 {
- t.Fatalf("repeat streak = %d, want 0", got.LastScore.RepeatCycleStreak)
+ if !got.LastScore.ShouldTerminate || got.LastScore.TerminateReason != StopReasonRepeatCycle {
+ t.Fatalf("score = %+v, want repeat-cycle hard terminate", got.LastScore)
}
}
-func TestEvaluateProgressVerifyPassedAloneIsNotBusinessProgress(t *testing.T) {
+func TestEvaluateProgressDifferentToolResultOrSubgoalResetsRepeat(t *testing.T) {
t.Parallel()
- got := EvaluateProgress(ProgressState{}, ProgressInput{
- RunState: RunStateVerify,
- Evidence: []ProgressEvidenceRecord{
- {Kind: EvidenceVerifyPassed},
+ tests := []struct {
+ name string
+ input ProgressInput
+ same SubgoalRelation
+ }{
+ {
+ name: "different tool",
+ input: ProgressInput{
+ CurrentToolSignature: "other",
+ ResultFingerprint: "result",
+ SubgoalFingerprint: "subgoal",
+ RepeatCycleLimit: 3,
+ },
+ same: SubgoalRelationSame,
+ },
+ {
+ name: "different result",
+ input: ProgressInput{
+ CurrentToolSignature: "sig",
+ ResultFingerprint: "other",
+ SubgoalFingerprint: "subgoal",
+ RepeatCycleLimit: 3,
+ },
+ same: SubgoalRelationSame,
+ },
+ {
+ name: "different subgoal",
+ input: ProgressInput{
+ CurrentToolSignature: "sig",
+ ResultFingerprint: "result",
+ SubgoalFingerprint: "other",
+ RepeatCycleLimit: 3,
+ },
+ same: SubgoalRelationDifferent,
+ },
+ {
+ name: "unknown subgoal",
+ input: ProgressInput{
+ CurrentToolSignature: "sig",
+ ResultFingerprint: "result",
+ SubgoalFingerprint: "",
+ RepeatCycleLimit: 3,
+ },
+ same: SubgoalRelationUnknown,
},
- NoProgressLimit: 3,
- RepeatCycleLimit: 3,
- })
- if got.LastScore.HasBusinessProgress {
- t.Fatalf("expected verify-passed alone to not count as business progress")
}
- if got.LastScore.StrongEvidenceCount != 0 {
- t.Fatalf("strong evidence = %d, want 0", got.LastScore.StrongEvidenceCount)
+
+ state := ProgressState{
+ LastScore: ProgressScore{RepeatCycleStreak: 2},
+ LastToolSignature: "sig",
+ LastResultFingerprint: "result",
+ LastSubgoalFingerprint: "subgoal",
}
- if got.LastScore.MediumEvidenceCount != 1 {
- t.Fatalf("medium evidence = %d, want 1", got.LastScore.MediumEvidenceCount)
+ for _, tt := range tests {
+ tt := tt
+ t.Run(tt.name, func(t *testing.T) {
+ t.Parallel()
+
+ got := EvaluateProgress(state, tt.input)
+ if got.LastScore.SameSubgoal != tt.same {
+ t.Fatalf("same subgoal = %q, want %q", got.LastScore.SameSubgoal, tt.same)
+ }
+ if got.LastScore.RepeatCycleStreak != 0 {
+ t.Fatalf("repeat streak = %d, want 0", got.LastScore.RepeatCycleStreak)
+ }
+ if got.LastScore.StalledProgressState != StalledProgressHealthy {
+ t.Fatalf("stalled state = %q, want healthy", got.LastScore.StalledProgressState)
+ }
+ })
}
}
diff --git a/internal/runtime/controlplane/stop_reason.go b/internal/runtime/controlplane/stop_reason.go
index 737ae17e..c7ff736e 100644
--- a/internal/runtime/controlplane/stop_reason.go
+++ b/internal/runtime/controlplane/stop_reason.go
@@ -16,12 +16,16 @@ const (
StopReasonVerificationFailed StopReason = "verification_failed"
// StopReasonAccepted 表示 completion gate 与 verifier gate 均通过并完成收尾。
StopReasonAccepted StopReason = "accepted"
+ // StopReasonMissingCompletionSignal 表示模型停止调用工具但没有输出结构化完成信号。
+ StopReasonMissingCompletionSignal StopReason = "missing_completion_signal"
+ // StopReasonAcceptCheckFailed 表示最终 Accept Gate 的验收项失败。
+ StopReasonAcceptCheckFailed StopReason = "accept_check_failed"
// StopReasonTodoNotConverged 表示 required todo 尚未收敛。
StopReasonTodoNotConverged StopReason = "todo_not_converged"
// StopReasonTodoWaitingExternal 表示 required todo 仍在等待外部条件。
StopReasonTodoWaitingExternal StopReason = "todo_waiting_external"
- // StopReasonNoProgressAfterFinalIntercept 表示 final 连续被拦截且没有新进展。
- StopReasonNoProgressAfterFinalIntercept StopReason = "no_progress_after_final_intercept"
+ // StopReasonRepeatCycle 表示运行重复相同动作/结果并触发硬终止。
+ StopReasonRepeatCycle StopReason = "repeat_cycle"
// StopReasonMaxTurnExceededWithUnconvergedTodos 表示达到最大轮次时 todo 仍未收敛。
StopReasonMaxTurnExceededWithUnconvergedTodos StopReason = "max_turn_exceeded_with_unconverged_todos"
// StopReasonMaxTurnExceededWithFailedVerification 表示达到最大轮次时 verifier 已失败。
diff --git a/internal/runtime/decider/decide.go b/internal/runtime/decider/decide.go
deleted file mode 100644
index 38577bdb..00000000
--- a/internal/runtime/decider/decide.go
+++ /dev/null
@@ -1,723 +0,0 @@
-package decider
-
-import (
- "fmt"
- "path/filepath"
- "regexp"
- "strings"
-
- "neo-code/internal/runtime/facts"
-)
-
-// TODO(runtime-control-plane): FinalDecider 当前同时承担了意图推断、有效任务类型推导、事实解释、
-// 终态裁决与下一步动作合成。待验收循环稳定后拆分为 IntentInferer、EffectiveTaskKindDeriver、
-// AcceptanceProfile 与 NextActionPlanner，降低单模块职责耦合。
-//
-// Decide 执行最终终态裁决，作为 runtime 的唯一决策入口。
-func Decide(input DecisionInput) Decision {
- intent := InferTaskIntent(input.UserGoal)
- hint := input.TaskKind
- if strings.TrimSpace(string(hint)) == "" {
- hint = intent.Hint
- }
- effectiveTaskKind := DeriveEffectiveTaskKind(hint, input.Facts, input.Todos)
-
- var decision Decision
- if input.Todos.Summary.RequiredFailed > 0 {
- decision = Decision{
- Status: DecisionFailed,
- StopReason: "required_todo_failed",
- UserVisibleSummary: "存在 required todo 失败，任务已终止。",
- InternalSummary: "required todo entered failed terminal state",
- }
- return finalizeDecision(decision, hint, effectiveTaskKind, input)
- }
- if input.NoProgressExceeded {
- decision = Decision{
- Status: DecisionIncomplete,
- StopReason: "no_progress_after_final_intercept",
- UserVisibleSummary: "连续多轮缺少新事实，任务以未完成结束。",
- InternalSummary: "no progress exceeded while final intercepted",
- }
- return finalizeDecision(decision, hint, effectiveTaskKind, input)
- }
- if !input.CompletionPassed {
- decision = continueWithCompletionReason(input)
- return finalizeDecision(decision, hint, effectiveTaskKind, input)
- }
-
- switch effectiveTaskKind {
- case TaskKindTodoState:
- decision = decideTodoState(input)
- case TaskKindWorkspaceWrite:
- decision = decideWorkspaceWrite(input)
- case TaskKindSubAgent:
- decision = decideSubAgent(input)
- case TaskKindReadOnly:
- decision = decideReadOnly(input)
- case TaskKindMixed:
- decision = decideMixed(input)
- case TaskKindChatAnswer:
- fallthrough
- default:
- decision = Decision{
- Status: DecisionAccepted,
- StopReason: "accepted",
- UserVisibleSummary: "任务完成。",
- InternalSummary: "chat answer accepted by completion gate",
- }
- }
- return finalizeDecision(decision, hint, effectiveTaskKind, input)
-}
-
-// continueWithCompletionReason 把 completion gate 阻塞转成可执行缺失事实提示。
-func continueWithCompletionReason(input DecisionInput) Decision {
- reason := strings.TrimSpace(input.CompletionReason)
- switch reason {
- case "pending_todo":
- openTodos := collectOpenRequiredTodos(input.Todos.Items)
- if len(openTodos) == 0 {
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{Kind: "required_todo_terminal"}},
- RequiredInput: &RequiredInput{
- Kind: "missing_required_todo_id",
- Message: "缺少 required todo 标识，无法推进状态收敛。",
- },
- UserVisibleSummary: "仍有 required todo 未收敛，但当前无法确定待推进项。",
- InternalSummary: "completion blocked by pending_todo without resolvable open todo id",
- }
- }
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{
- Kind: "required_todo_terminal",
- Target: strings.Join(openTodos, ","),
- Details: map[string]any{"open_required_ids": openTodos},
- }},
- RequiredNextActions: []RequiredAction{{
- Tool: "todo_write",
- ArgsHint: map[string]any{
- "action": "set_status",
- "id": firstOrEmpty(openTodos),
- "status": "completed",
- },
- }},
- UserVisibleSummary: "仍有 required todo 未收敛，需要继续推进 todo 状态。",
- InternalSummary: "completion blocked by pending_todo",
- }
- case "unverified_write":
- target, expectedContent, ok := selectVerificationTarget(input)
- if !ok {
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{
- Kind: "file_written",
- Details: map[string]any{
- "reason": "cannot infer target path/content from user goal",
- },
- }},
- RequiredInput: &RequiredInput{
- Kind: "missing_file_target_or_content",
- Message: "无法从当前任务中确定要验证的文件路径或内容，需要用户补充。",
- },
- UserVisibleSummary: "写入事实尚未完成验证，但当前缺少可执行验证目标。",
- InternalSummary: "completion blocked by unverified_write without resolvable verification target",
- }
- }
- return buildWriteVerificationDecision(
- target,
- expectedContent,
- "写入事实尚未完成验证，需要补充 verification facts。",
- "completion blocked by unverified_write",
- )
- case "post_execute_closure_required":
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{
- Kind: "post_execute_closure",
- Target: "latest_tool_results",
- }},
- RequiredInput: &RequiredInput{
- Kind: "missing_post_execute_closure",
- Message: "需要基于最新工具结果补充闭环信息后再尝试完成。",
- },
- UserVisibleSummary: "请先基于最新工具结果完成闭环，再尝试最终收尾。",
- InternalSummary: "completion blocked by post_execute_closure_required",
- }
- default:
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- UserVisibleSummary: "仍缺少可验证事实，请继续调用工具推进任务。",
- InternalSummary: "completion gate blocked without classified reason",
- }
- }
-}
-
-// decideTodoState 依据 todo 快照判定状态类任务。
-func decideTodoState(input DecisionInput) Decision {
- if input.Todos.Summary.Total == 0 && len(input.Facts.Todos.CreatedIDs) == 0 {
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{
- Kind: "todo_created",
- }},
- RequiredInput: &RequiredInput{
- Kind: "missing_todo_content",
- Message: "尚未提供 Todo 内容，需要用户补充待办事项。",
- },
- UserVisibleSummary: "尚未创建目标 Todo，请先调用 todo_write。",
- InternalSummary: "todo_state task missing created todo facts",
- }
- }
- if input.Todos.Summary.RequiredOpen > 0 {
- openIDs := collectOpenRequiredTodos(input.Todos.Items)
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{
- Kind: "required_todo_terminal",
- Target: strings.Join(openIDs, ","),
- Details: map[string]any{"open_required_ids": openIDs},
- }},
- RequiredNextActions: []RequiredAction{{
- Tool: "todo_write",
- ArgsHint: map[string]any{
- "action": "set_status",
- "id": firstOrEmpty(openIDs),
- "status": "completed",
- },
- }},
- UserVisibleSummary: "Todo 已创建但 required 项仍未完成。",
- InternalSummary: "todo_state task still has open required todos",
- }
- }
- return Decision{
- Status: DecisionAccepted,
- StopReason: "accepted",
- UserVisibleSummary: "Todo 状态已满足任务目标。",
- InternalSummary: "todo_state facts satisfied",
- }
-}
-
-// decideWorkspaceWrite 依据写入与验证事实判定文件任务。
-func decideWorkspaceWrite(input DecisionInput) Decision {
- if len(input.Facts.Files.Written) == 0 {
- if hasSatisfiedWorkspaceWriteWithoutNewWrite(input) {
- return Decision{
- Status: DecisionAccepted,
- StopReason: "accepted",
- UserVisibleSummary: "目标文件状态已满足，无需重复写入。",
- InternalSummary: "workspace_write satisfied by noop_write verification facts",
- }
- }
- if !hasExplicitFileTarget(input.UserGoal) {
- return Decision{
- Status: DecisionAccepted,
- StopReason: "accepted",
- UserVisibleSummary: "任务未声明明确文件目标，已按通用编辑任务收尾。",
- InternalSummary: "workspace_write downgraded to generic edit due missing explicit file target",
- }
- }
- errorDetail := latestToolErrorDetail(input.Facts.Errors.ToolErrors, "filesystem_write_file")
- details := map[string]any{}
- if errorDetail != "" {
- details["last_write_error"] = errorDetail
- }
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{
- Kind: "file_written",
- Details: details,
- }},
- RequiredInput: &RequiredInput{
- Kind: "missing_file_target_or_content",
- Message: "无法从当前任务中确定要写入的文件路径或内容，需要用户补充。",
- },
- UserVisibleSummary: "还没有写入事实，请先执行文件写入。",
- InternalSummary: "workspace_write task missing file_written fact",
- }
- }
- target, expectedContent, ok := selectVerificationTarget(input)
- if !ok {
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{
- Kind: "verification_passed",
- Details: map[string]any{
- "reason": "missing resolvable verification target",
- },
- }},
- RequiredInput: &RequiredInput{
- Kind: "missing_file_target_or_content",
- Message: "无法确定当前需要验收的写入目标，请补充文件路径或重试写入。",
- },
- UserVisibleSummary: "已检测到写入事实，但无法确定验收目标。",
- InternalSummary: "workspace_write has writes but cannot resolve verification target",
- }
- }
- if hasWorkspaceWriteHardFailure(input.Facts.Errors.ToolErrors, target) {
- return Decision{
- Status: DecisionFailed,
- StopReason: "verification_failed",
- UserVisibleSummary: "文件写入出现持续失败，任务终止。请检查路径权限或写入策略。",
- InternalSummary: "workspace_write hard failure detected from tool error facts",
- }
- }
- if !hasVerificationForTarget(input.Facts, target) {
- return buildWriteVerificationDecision(
- target,
- expectedContent,
- "已写入文件但尚未形成通过的验证事实。",
- "workspace_write task missing",
- )
- }
- return Decision{
- Status: DecisionAccepted,
- StopReason: "accepted",
- UserVisibleSummary: "文件写入与验证事实已满足。",
- InternalSummary: "workspace_write facts satisfied",
- }
-}
-
-// decideSubAgent 依据子代理启动/完成事实判定子代理任务。
-func decideSubAgent(input DecisionInput) Decision {
- if len(input.Facts.SubAgents.Started) == 0 {
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{
- Kind: "subagent_started",
- }},
- RequiredInput: &RequiredInput{
- Kind: "missing_subagent_instruction",
- Message: "需要明确的子代理任务指令后才能执行 spawn_subagent。",
- },
- UserVisibleSummary: "尚未产生子代理启动事实，请显式调用 spawn_subagent。",
- InternalSummary: "subagent task missing start fact",
- }
- }
- if len(input.Facts.SubAgents.Failed) > 0 && len(input.Facts.SubAgents.Completed) == 0 {
- return Decision{
- Status: DecisionFailed,
- StopReason: "verification_failed",
- UserVisibleSummary: "子代理执行失败，任务终止。",
- InternalSummary: "subagent task failed without completion fact",
- }
- }
- if len(input.Facts.SubAgents.Completed) == 0 {
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- UserVisibleSummary: "子代理已启动但尚未完成。",
- InternalSummary: "subagent task started but no completed fact",
- }
- }
- if isWriteIntentGoal(input.UserGoal) && !hasSubAgentArtifactEvidence(input.Facts) {
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{
- Kind: "subagent_artifact_or_file_fact",
- Target: "workspace_artifact",
- }},
- RequiredInput: &RequiredInput{
- Kind: "missing_subagent_artifact_path",
- Message: "需要提供子代理产物路径或可验证的文件目标。",
- },
- UserVisibleSummary: "子代理已完成，但缺少可验证的产物事实。",
- InternalSummary: "subagent completed without artifact/file evidence for write-intent goal",
- }
- }
- return Decision{
- Status: DecisionAccepted,
- StopReason: "accepted",
- UserVisibleSummary: "子代理完成事实已满足。",
- InternalSummary: "subagent task completed facts satisfied",
- }
-}
-
-// decideReadOnly 判定只读任务是否可结束。
-func decideReadOnly(input DecisionInput) Decision {
- if len(input.Facts.Files.Exists) == 0 && len(input.Facts.Commands.Executed) == 0 && len(input.LastAssistantText) == 0 {
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- UserVisibleSummary: "尚无可验证读取事实，请先执行只读工具。",
- InternalSummary: "read_only task has no read/search facts",
- }
- }
- return Decision{
- Status: DecisionAccepted,
- StopReason: "accepted",
- UserVisibleSummary: "只读分析任务已完成。",
- InternalSummary: "read_only facts satisfied",
- }
-}
-
-// decideMixed 对混合任务采用保守策略：必须同时具备状态推进与至少一个验证事实。
-func decideMixed(input DecisionInput) Decision {
- if len(input.Facts.Verification.Passed) == 0 {
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- UserVisibleSummary: "混合任务尚未形成验证通过事实。",
- InternalSummary: "mixed task missing verification passed facts",
- }
- }
- if input.Todos.Summary.RequiredOpen > 0 {
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- UserVisibleSummary: "混合任务 required todo 尚未收敛。",
- InternalSummary: "mixed task has open required todos",
- }
- }
- return Decision{
- Status: DecisionAccepted,
- StopReason: "accepted",
- UserVisibleSummary: "混合任务事实已满足。",
- InternalSummary: "mixed task satisfied by verification + todo closure",
- }
-}
-
-// buildWriteVerificationDecision 统一构造写入后缺少验证事实时的继续决策。
-func buildWriteVerificationDecision(target string, expectedContent string, userSummary string, internalPrefix string) Decision {
- scope := fmt.Sprintf("artifact:%s", target)
- if expectedContent != "" {
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{
- Kind: "verification_passed",
- Target: target,
- Expected: expectedContent,
- }},
- RequiredNextActions: []RequiredAction{{
- Tool: "filesystem_read_file",
- ArgsHint: map[string]any{
- "path": target,
- "expect_contains": []string{expectedContent},
- "verification_scope": scope,
- },
- }},
- UserVisibleSummary: userSummary,
- InternalSummary: internalPrefix + " content verification facts bound to target artifact",
- }
- }
- return Decision{
- Status: DecisionContinue,
- StopReason: "todo_not_converged",
- MissingFacts: []MissingFact{{
- Kind: "file_exists",
- Target: target,
- }},
- RequiredNextActions: []RequiredAction{{
- Tool: "filesystem_glob",
- ArgsHint: map[string]any{
- "pattern": target,
- "expect_min_matches": 1,
- "verification_scope": scope,
- },
- }},
- UserVisibleSummary: userSummary,
- InternalSummary: internalPrefix + " existence verification facts bound to target artifact",
- }
-}
-
-// collectOpenRequiredTodos 收集 required 且未终态的 todo id。
-func collectOpenRequiredTodos(items []TodoViewItem) []string {
- ids := make([]string, 0)
- for _, item := range items {
- if !item.Required {
- continue
- }
- switch strings.ToLower(strings.TrimSpace(item.Status)) {
- case "completed", "failed", "canceled":
- continue
- default:
- if id := strings.TrimSpace(item.ID); id != "" {
- ids = append(ids, id)
- }
- }
- }
- return ids
-}
-
-// firstOrEmpty 返回首个元素，不存在时返回空串。
-func firstOrEmpty(values []string) string {
- if len(values) == 0 {
- return ""
- }
- return values[0]
-}
-
-// hasVerificationForTarget 判断目标文件是否已经有通过的验证事实，避免跨文件误判 accepted。
-func hasVerificationForTarget(allFacts facts.RuntimeFacts, targetPath string) bool {
- target := strings.TrimSpace(targetPath)
- if target == "" {
- return false
- }
- targetArtifactScope := "artifact:" + target
- normalizedTarget := strings.ToLower(filepath.Clean(target))
-
- for _, fact := range allFacts.Verification.Passed {
- scope := strings.TrimSpace(fact.Scope)
- if scope == "" {
- continue
- }
- normalizedScope := strings.ToLower(filepath.Clean(scope))
- if strings.EqualFold(scope, target) || strings.EqualFold(scope, targetArtifactScope) || normalizedScope == normalizedTarget {
- return true
- }
- if strings.HasPrefix(strings.ToLower(scope), "artifact:") {
- normalized := strings.TrimPrefix(scope, "artifact:")
- cleaned := strings.TrimSpace(normalized)
- if strings.EqualFold(cleaned, target) || strings.ToLower(filepath.Clean(cleaned)) == normalizedTarget {
- return true
- }
- }
- }
- for _, fact := range allFacts.Files.ContentMatch {
- if !fact.VerificationPassed {
- continue
- }
- if strings.EqualFold(strings.TrimSpace(fact.Path), target) ||
- strings.ToLower(filepath.Clean(strings.TrimSpace(fact.Path))) == normalizedTarget {
- return true
- }
- }
- return false
-}
-
-// latestToolErrorDetail 返回指定工具的最新错误摘要，便于构造可执行 continue 提示。
-func latestToolErrorDetail(errors []facts.ToolErrorFact, toolName string) string {
- targetTool := strings.TrimSpace(toolName)
- for i := len(errors) - 1; i >= 0; i-- {
- fact := errors[i]
- if !strings.EqualFold(strings.TrimSpace(fact.Tool), targetTool) {
- continue
- }
- content := strings.TrimSpace(fact.Content)
- if content == "" {
- content = strings.TrimSpace(fact.ErrorClass)
- }
- if content != "" {
- return content
- }
- }
- return ""
-}
-
-// hasWorkspaceWriteHardFailure 判断写入目标是否出现高置信不可恢复错误，防止无意义循环重试。
-func hasWorkspaceWriteHardFailure(errors []facts.ToolErrorFact, targetPath string) bool {
- target := strings.TrimSpace(targetPath)
- if target == "" {
- return false
- }
- targetLower := strings.ToLower(target)
- targetBaseLower := strings.ToLower(strings.TrimSpace(filepath.Base(target)))
- errorCount := 0
- for _, fact := range errors {
- if !strings.EqualFold(strings.TrimSpace(fact.Tool), "filesystem_write_file") {
- continue
- }
- content := strings.ToLower(strings.TrimSpace(fact.Content))
- if content == "" {
- content = strings.ToLower(strings.TrimSpace(fact.ErrorClass))
- }
- mentionsTarget := strings.Contains(content, targetLower) ||
- (targetBaseLower != "" && strings.Contains(content, targetBaseLower))
- if !mentionsTarget {
- continue
- }
- if strings.Contains(content, "permission denied") ||
- strings.Contains(content, "path not allowed") || strings.Contains(content, "no such file") {
- errorCount++
- }
- }
- return errorCount >= 2
-}
-
-// isWriteIntentGoal 判断用户目标是否显式要求产物写入。
-func isWriteIntentGoal(goal string) bool {
- return containsAny(strings.ToLower(strings.TrimSpace(goal)),
- "创建文件", "写入", "修改文件", "新增文件", "create file", "write file", "edit file", "update file", ".txt", ".go", ".md", ".json")
-}
-
-// hasExplicitFileTarget 判断用户目标是否包含可定位文件目标，避免对泛化“编辑一下”任务过度拦截。
-func hasExplicitFileTarget(goal string) bool {
- normalized := strings.ToLower(strings.TrimSpace(goal))
- return containsAny(
- normalized,
- ".txt", ".go", ".md", ".json", ".yaml", ".yml", ".ts", ".tsx", ".py", "/",
- "readme", "package.json",
- )
-}
-
-// hasSatisfiedWorkspaceWriteWithoutNewWrite 判断未产生新写入时是否已有可验收写入结果。
-func hasSatisfiedWorkspaceWriteWithoutNewWrite(input DecisionInput) bool {
- goalPaths := extractGoalPaths(input.UserGoal)
- for _, goalPath := range goalPaths {
- if hasVerificationForTarget(input.Facts, goalPath) {
- return true
- }
- }
- if len(goalPaths) > 0 {
- return false
- }
- target, _, ok := selectVerificationTarget(input)
- return ok && hasVerificationForTarget(input.Facts, target)
-}
-
-// hasSubAgentArtifactEvidence 判断子代理任务是否已有可验证产物事实。
-func hasSubAgentArtifactEvidence(allFacts facts.RuntimeFacts) bool {
- for _, fact := range allFacts.SubAgents.Completed {
- if len(fact.Artifacts) > 0 {
- return true
- }
- }
- if len(allFacts.Files.Written) > 0 || len(allFacts.Files.Exists) > 0 || len(allFacts.Files.ContentMatch) > 0 {
- return true
- }
- return false
-}
-
-// latestWriteVerificationHint 返回最适合下一步验证动作的写入目标与期望内容（若可用）。
-func latestWriteVerificationHint(allFacts facts.RuntimeFacts, preferredPath string) (string, string) {
- normalizedPreferred := strings.TrimSpace(preferredPath)
- writes := allFacts.Files.Written
- for i := len(writes) - 1; i >= 0; i-- {
- fact := writes[i]
- path := strings.TrimSpace(fact.Path)
- if path == "" {
- continue
- }
- if normalizedPreferred != "" && !strings.EqualFold(path, normalizedPreferred) {
- continue
- }
- return path, strings.TrimSpace(fact.ExpectedContent)
- }
- if normalizedPreferred != "" {
- return normalizedPreferred, ""
- }
- return "", ""
-}
-
-// finalizeDecision 统一补全决策元信息，确保快照可观测 hint 与 effective kind。
-func finalizeDecision(decision Decision, intentHint TaskKind, effective TaskKind, input DecisionInput) Decision {
- if len(input.HookAnnotations) > 0 || len(input.HookGuards) > 0 {
- detail := fmt.Sprintf(
- "hook signals consumed (annotations=%d guards=%d)",
- len(input.HookAnnotations),
- len(input.HookGuards),
- )
- if strings.TrimSpace(decision.InternalSummary) == "" {
- decision.InternalSummary = detail
- } else {
- decision.InternalSummary = strings.TrimSpace(decision.InternalSummary) + "; " + detail
- }
- }
- decision.IntentHint = intentHint
- decision.EffectiveTaskKind = effective
- return decision
-}
-
-var filePathPattern = regexp.MustCompile(`(?i)(?:^|[\s"'` + "`" + `])([a-z0-9_\-./]+\.[a-z0-9]{1,8})(?:$|[\s"'` + "`" + `,;:])`)
-
-// selectVerificationTarget 选择当前回合应验证的写入目标，避免回退到历史首条写入。
-func selectVerificationTarget(input DecisionInput) (path string, expectedContent string, ok bool) {
- goalPaths := extractGoalPaths(input.UserGoal)
- for i := len(goalPaths) - 1; i >= 0; i-- {
- goalPath := goalPaths[i]
- for j := len(input.Facts.Files.Written) - 1; j >= 0; j-- {
- writeFact := input.Facts.Files.Written[j]
- if strings.EqualFold(strings.TrimSpace(writeFact.Path), goalPath) {
- return strings.TrimSpace(writeFact.Path), strings.TrimSpace(writeFact.ExpectedContent), true
- }
- }
- }
- for i := len(input.Facts.Files.Written) - 1; i >= 0; i-- {
- writeFact := input.Facts.Files.Written[i]
- target := strings.TrimSpace(writeFact.Path)
- if target == "" {
- continue
- }
- if !hasVerificationForTarget(input.Facts, target) {
- return target, strings.TrimSpace(writeFact.ExpectedContent), true
- }
- }
- for i := len(input.Facts.Files.Written) - 1; i >= 0; i-- {
- writeFact := input.Facts.Files.Written[i]
- target := strings.TrimSpace(writeFact.Path)
- if target != "" {
- return target, strings.TrimSpace(writeFact.ExpectedContent), true
- }
- }
- for i := len(goalPaths) - 1; i >= 0; i-- {
- goalPath := strings.TrimSpace(goalPaths[i])
- if goalPath != "" && hasVerificationForTarget(input.Facts, goalPath) {
- return goalPath, "", true
- }
- }
- for i := len(input.Facts.Files.ContentMatch) - 1; i >= 0; i-- {
- matchFact := input.Facts.Files.ContentMatch[i]
- if !matchFact.VerificationPassed {
- continue
- }
- target := strings.TrimSpace(matchFact.Path)
- if target == "" {
- continue
- }
- return target, firstOrEmpty(matchFact.ExpectedContains), true
- }
- for i := len(input.Facts.Verification.Passed) - 1; i >= 0; i-- {
- verifyFact := input.Facts.Verification.Passed[i]
- scope := strings.TrimSpace(verifyFact.Scope)
- if scope == "" {
- continue
- }
- target := scope
- if strings.HasPrefix(strings.ToLower(target), "artifact:") {
- target = strings.TrimSpace(target[len("artifact:"):])
- }
- if target == "" {
- continue
- }
- return target, "", true
- }
- return "", "", false
-}
-
-// extractGoalPaths 从用户目标文本提取可能的文件路径。
-func extractGoalPaths(goal string) []string {
- matches := filePathPattern.FindAllStringSubmatch(strings.TrimSpace(goal), -1)
- if len(matches) == 0 {
- return nil
- }
- seen := make(map[string]struct{}, len(matches))
- out := make([]string, 0, len(matches))
- for _, match := range matches {
- if len(match) < 2 {
- continue
- }
- value := strings.TrimSpace(match[1])
- if value == "" {
- continue
- }
- if _, exists := seen[strings.ToLower(value)]; exists {
- continue
- }
- seen[strings.ToLower(value)] = struct{}{}
- out = append(out, value)
- }
- return out
-}
diff --git a/internal/runtime/decider/decide_additional_test.go b/internal/runtime/decider/decide_additional_test.go
deleted file mode 100644
index 017421e6..00000000
--- a/internal/runtime/decider/decide_additional_test.go
+++ /dev/null
@@ -1,155 +0,0 @@
-package decider
-
-import (
- "testing"
-
- runtimefacts "neo-code/internal/runtime/facts"
-)
-
-func TestContinueWithCompletionReasonBranches(t *testing.T) {
- t.Parallel()
-
- t.Run("pending_todo without open ids requires input", func(t *testing.T) {
- t.Parallel()
- decision := continueWithCompletionReason(DecisionInput{
- CompletionReason: "pending_todo",
- Todos: TodoSnapshot{
- Items: []TodoViewItem{{ID: "x", Required: true, Status: "completed"}},
- },
- })
- if decision.RequiredInput == nil || decision.RequiredInput.Kind != "missing_required_todo_id" {
- t.Fatalf("required input = %+v", decision.RequiredInput)
- }
- })
-
- t.Run("unverified_write without target requires input", func(t *testing.T) {
- t.Parallel()
- decision := continueWithCompletionReason(DecisionInput{
- CompletionReason: "unverified_write",
- })
- if decision.RequiredInput == nil || decision.RequiredInput.Kind != "missing_file_target_or_content" {
- t.Fatalf("required input = %+v", decision.RequiredInput)
- }
- })
-
- t.Run("post_execute_closure_required maps to closure missing fact", func(t *testing.T) {
- t.Parallel()
- decision := continueWithCompletionReason(DecisionInput{
- CompletionReason: "post_execute_closure_required",
- })
- if len(decision.MissingFacts) == 0 || decision.MissingFacts[0].Kind != "post_execute_closure" {
- t.Fatalf("missing facts = %+v", decision.MissingFacts)
- }
- })
-}
-
-func TestDecideTaskSpecificBranches(t *testing.T) {
- t.Parallel()
-
- t.Run("todo_state without creation facts requests todo_write", func(t *testing.T) {
- t.Parallel()
- decision := decideTodoState(DecisionInput{})
- if decision.RequiredInput == nil || decision.RequiredInput.Kind != "missing_todo_content" {
- t.Fatalf("required input = %+v", decision.RequiredInput)
- }
- })
-
- t.Run("mixed accepts when verification passed and required todos closed", func(t *testing.T) {
- t.Parallel()
- decision := decideMixed(DecisionInput{
- Facts: runtimefacts.RuntimeFacts{
- Verification: runtimefacts.VerificationFacts{
- Passed: []runtimefacts.VerificationFact{{Tool: "filesystem_read_file", Scope: "artifact:a.txt"}},
- },
- },
- Todos: TodoSnapshot{Summary: TodoSummary{RequiredOpen: 0}},
- })
- if decision.Status != DecisionAccepted {
- t.Fatalf("status = %q, want accepted", decision.Status)
- }
- })
-
- t.Run("subagent failed without completion returns failed", func(t *testing.T) {
- t.Parallel()
- decision := decideSubAgent(DecisionInput{
- Facts: runtimefacts.RuntimeFacts{
- SubAgents: runtimefacts.SubAgentFacts{
- Started: []runtimefacts.SubAgentFact{{TaskID: "sa-1"}},
- Failed: []runtimefacts.SubAgentFact{{TaskID: "sa-1"}},
- },
- },
- })
- if decision.Status != DecisionFailed {
- t.Fatalf("status = %q, want failed", decision.Status)
- }
- })
-}
-
-func TestHelperFunctionsBranches(t *testing.T) {
- t.Parallel()
-
- if got := latestToolErrorDetail([]runtimefacts.ToolErrorFact{{Tool: "filesystem_write_file", ErrorClass: "timeout"}}, "filesystem_write_file"); got != "timeout" {
- t.Fatalf("latestToolErrorDetail = %q, want timeout", got)
- }
-
- if !hasVerificationForTarget(runtimefacts.RuntimeFacts{
- Verification: runtimefacts.VerificationFacts{
- Passed: []runtimefacts.VerificationFact{{Scope: "artifact:./docs/../docs/readme.md"}},
- },
- }, "docs/readme.md") {
- t.Fatal("expected normalized artifact scope match")
- }
-
- if got := extractGoalPaths(`edit "A.TXT", then update a.txt and b.md`); len(got) != 2 {
- t.Fatalf("extractGoalPaths len = %d, want 2", len(got))
- }
-}
-
-func TestHelperFunctionsAdditionalBranches(t *testing.T) {
- t.Parallel()
-
- if got := firstOrEmpty(nil); got != "" {
- t.Fatalf("firstOrEmpty(nil) = %q, want empty", got)
- }
- if got := firstOrEmpty([]string{"a", "b"}); got != "a" {
- t.Fatalf("firstOrEmpty = %q, want a", got)
- }
-
- open := collectOpenRequiredTodos([]TodoViewItem{
- {ID: "todo-1", Required: true, Status: "pending"},
- {ID: "todo-2", Required: true, Status: "in_progress"},
- {ID: "todo-3", Required: true, Status: "completed"},
- {ID: "", Required: true, Status: "pending"},
- })
- if len(open) != 2 || open[0] != "todo-1" || open[1] != "todo-2" {
- t.Fatalf("collectOpenRequiredTodos = %#v, want [todo-1 todo-2]", open)
- }
-
- if got := latestToolErrorDetail([]runtimefacts.ToolErrorFact{
- {Tool: "filesystem_write_file", Content: " last content "},
- }, "filesystem_write_file"); got != "last content" {
- t.Fatalf("latestToolErrorDetail(content fallback) = %q, want trimmed content", got)
- }
-
- if hasWorkspaceWriteHardFailure([]runtimefacts.ToolErrorFact{
- {Tool: "filesystem_write_file", ErrorClass: "permission_denied", Content: "permission denied for /tmp/other.txt"},
- {Tool: "filesystem_write_file", ErrorClass: "permission_denied", Content: "permission denied for /tmp/other2.txt"},
- }, "/tmp/target.txt") {
- t.Fatal("hasWorkspaceWriteHardFailure should require target correlation")
- }
-
- target, expected, ok := selectVerificationTarget(DecisionInput{
- UserGoal: "please edit ./docs/readme.md and include hello",
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{
- {Path: "a.md", WorkspaceWrite: true},
- {Path: "docs/readme.md", WorkspaceWrite: true, ExpectedContent: "hello"},
- },
- },
- },
- })
- if !ok || target != "docs/readme.md" || expected != "hello" {
- t.Fatalf("selectVerificationTarget = target=%q expected=%q ok=%v", target, expected, ok)
- }
-}
diff --git a/internal/runtime/decider/decide_test.go b/internal/runtime/decider/decide_test.go
deleted file mode 100644
index c7a72177..00000000
--- a/internal/runtime/decider/decide_test.go
+++ /dev/null
@@ -1,624 +0,0 @@
-package decider
-
-import (
- "encoding/json"
- "strings"
- "testing"
-
- runtimefacts "neo-code/internal/runtime/facts"
-)
-
-func assertDecisionStatus(t *testing.T, decision Decision, want DecisionStatus) {
- t.Helper()
- if decision.Status != want {
- t.Fatalf("status = %q, want %q", decision.Status, want)
- }
-}
-
-func TestDecideRequiredTodoFailedStopsImmediately(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindTodoState,
- Todos: TodoSnapshot{
- Summary: TodoSummary{RequiredFailed: 1},
- },
- })
-
- if decision.Status != DecisionFailed {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionFailed)
- }
- if decision.StopReason != "required_todo_failed" {
- t.Fatalf("stop_reason = %q, want required_todo_failed", decision.StopReason)
- }
-}
-
-func TestDecideUsesEffectiveTaskKindFromFacts(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindChatAnswer,
- CompletionPassed: true,
- UserGoal: "你好",
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{{Path: "test.txt", WorkspaceWrite: true}},
- },
- },
- })
- if decision.EffectiveTaskKind != TaskKindWorkspaceWrite {
- t.Fatalf("effective kind = %q, want %q", decision.EffectiveTaskKind, TaskKindWorkspaceWrite)
- }
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
-}
-
-func TestDecideNoProgressExceededReturnsIncomplete(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- NoProgressExceeded: true,
- })
-
- if decision.Status != DecisionIncomplete {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionIncomplete)
- }
- if decision.StopReason != "no_progress_after_final_intercept" {
- t.Fatalf("stop_reason = %q, want no_progress_after_final_intercept", decision.StopReason)
- }
-}
-
-func TestDecideCompletionBlockedReasonPendingTodo(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindTodoState,
- CompletionPassed: false,
- CompletionReason: "pending_todo",
- Todos: TodoSnapshot{
- Items: []TodoViewItem{
- {ID: "todo-1", Required: true, Status: "pending"},
- },
- },
- })
-
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
- if len(decision.MissingFacts) == 0 || decision.MissingFacts[0].Kind != "required_todo_terminal" {
- t.Fatalf("missing facts = %+v", decision.MissingFacts)
- }
- if len(decision.RequiredNextActions) == 0 || decision.RequiredNextActions[0].Tool != "todo_write" {
- t.Fatalf("required actions = %+v", decision.RequiredNextActions)
- }
-}
-
-func TestDecideWorkspaceWriteNeedsVerificationThenAccepts(t *testing.T) {
- continueDecision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{{Path: "test.txt", Bytes: 1, WorkspaceWrite: true}},
- },
- },
- })
- if continueDecision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", continueDecision.Status, DecisionContinue)
- }
- if len(continueDecision.MissingFacts) == 0 || continueDecision.MissingFacts[0].Kind != "file_exists" {
- t.Fatalf("missing facts = %+v, want file_exists", continueDecision.MissingFacts)
- }
- if len(continueDecision.RequiredNextActions) == 0 || continueDecision.RequiredNextActions[0].Tool != "filesystem_glob" {
- t.Fatalf("required actions = %+v, want filesystem_glob", continueDecision.RequiredNextActions)
- }
-
- acceptedDecision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{{Path: "test.txt", Bytes: 1, WorkspaceWrite: true}},
- },
- Verification: runtimefacts.VerificationFacts{
- Passed: []runtimefacts.VerificationFact{{Tool: "filesystem_read_file", Scope: "artifact:test.txt"}},
- },
- },
- })
- if acceptedDecision.Status != DecisionAccepted {
- t.Fatalf("status = %q, want %q", acceptedDecision.Status, DecisionAccepted)
- }
-}
-
-func TestDecideWorkspaceWriteNoopSatisfiedByVerificationFacts(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- UserGoal: "创建 2.txt 内容为 2",
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Exists: []runtimefacts.FileExistFact{{Path: "2.txt", Source: "filesystem_write_file_noop"}},
- ContentMatch: []runtimefacts.FileContentMatchFact{{
- Path: "2.txt",
- Scope: "artifact:2.txt",
- ExpectedContains: []string{"2"},
- VerificationPassed: true,
- }},
- },
- Verification: runtimefacts.VerificationFacts{
- Passed: []runtimefacts.VerificationFact{{Tool: "filesystem_write_file", Scope: "artifact:2.txt"}},
- },
- },
- })
- if decision.Status != DecisionAccepted {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionAccepted)
- }
-}
-
-func TestDecideWorkspaceWriteRepeatedNoopShouldStayAccepted(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- UserGoal: "创建 2.txt 内容为 2",
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{
- {Path: "2.txt", Bytes: 1, WorkspaceWrite: true, ExpectedContent: "2"},
- },
- Exists: []runtimefacts.FileExistFact{
- {Path: "2.txt", Source: "filesystem_write_file"},
- {Path: "2.txt", Source: "filesystem_write_file_noop"},
- },
- ContentMatch: []runtimefacts.FileContentMatchFact{{
- Path: "2.txt",
- Scope: "artifact:2.txt",
- ExpectedContains: []string{"2"},
- VerificationPassed: true,
- }},
- },
- Verification: runtimefacts.VerificationFacts{
- Performed: []runtimefacts.VerificationFact{{Tool: "filesystem_write_file", Scope: "artifact:2.txt"}},
- Passed: []runtimefacts.VerificationFact{{Tool: "filesystem_write_file", Scope: "artifact:2.txt"}},
- },
- },
- })
- if decision.Status != DecisionAccepted {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionAccepted)
- }
-}
-
-func TestDecideWorkspaceWriteVerificationMustBindTarget(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{{Path: "test.txt", Bytes: 1, WorkspaceWrite: true}},
- },
- Verification: runtimefacts.VerificationFacts{
- Passed: []runtimefacts.VerificationFact{{Tool: "filesystem_read_file", Scope: "artifact:other.txt"}},
- },
- },
- })
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
- if len(decision.MissingFacts) == 0 || decision.MissingFacts[0].Target != "test.txt" {
- t.Fatalf("missing facts = %+v, want target test.txt", decision.MissingFacts)
- }
-}
-
-func TestDecideWorkspaceWriteVerificationShouldNotMatchByBasenameOnly(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{{Path: "src/readme.md", Bytes: 1, WorkspaceWrite: true}},
- },
- Verification: runtimefacts.VerificationFacts{
- Passed: []runtimefacts.VerificationFact{{Tool: "filesystem_read_file", Scope: "artifact:docs/readme.md"}},
- },
- },
- })
- assertDecisionStatus(t, decision, DecisionContinue)
- if len(decision.MissingFacts) == 0 || decision.MissingFacts[0].Target != "src/readme.md" {
- t.Fatalf("missing facts = %+v, want target src/readme.md", decision.MissingFacts)
- }
-}
-
-func TestDecideCompletionBlockedUnverifiedWriteUsesExpectedContentWhenAvailable(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: false,
- CompletionReason: "unverified_write",
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{
- {Path: "2.txt", Bytes: 1, WorkspaceWrite: true, ExpectedContent: "2"},
- },
- },
- },
- })
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
- if len(decision.RequiredNextActions) == 0 || decision.RequiredNextActions[0].Tool != "filesystem_read_file" {
- t.Fatalf("required actions = %+v, want filesystem_read_file", decision.RequiredNextActions)
- }
- expectContains, _ := decision.RequiredNextActions[0].ArgsHint["expect_contains"].([]string)
- if len(expectContains) != 1 || expectContains[0] != "2" {
- t.Fatalf("expect_contains = %#v, want [\"2\"]", decision.RequiredNextActions[0].ArgsHint["expect_contains"])
- }
-}
-
-func TestDecideCompletionBlockedUnverifiedWriteFallsBackToExistsWhenContentUnknown(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: false,
- CompletionReason: "unverified_write",
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{
- {Path: "2.txt", Bytes: 1, WorkspaceWrite: true},
- },
- },
- },
- })
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
- if len(decision.MissingFacts) == 0 || decision.MissingFacts[0].Kind != "file_exists" {
- t.Fatalf("missing facts = %+v, want file_exists", decision.MissingFacts)
- }
- if len(decision.RequiredNextActions) == 0 || decision.RequiredNextActions[0].Tool != "filesystem_glob" {
- t.Fatalf("required actions = %+v, want filesystem_glob", decision.RequiredNextActions)
- }
-}
-
-func TestDecideWorkspaceWriteHardFailureStops(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{{Path: "Z:/not-exist/test.txt", Bytes: 1, WorkspaceWrite: true}},
- },
- Errors: runtimefacts.ErrorFacts{
- ToolErrors: []runtimefacts.ToolErrorFact{
- {Tool: "filesystem_write_file", ErrorClass: "permission_denied", Content: "permission denied for Z:/not-exist/test.txt"},
- {Tool: "filesystem_write_file", ErrorClass: "permission_denied", Content: "permission denied for Z:/not-exist/test.txt"},
- },
- },
- },
- })
- if decision.Status != DecisionFailed {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionFailed)
- }
-}
-
-func TestDecideWorkspaceWriteHardFailureRequiresTargetCorrelation(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{{Path: "target/test.txt", Bytes: 1, WorkspaceWrite: true}},
- },
- Errors: runtimefacts.ErrorFacts{
- ToolErrors: []runtimefacts.ToolErrorFact{
- {Tool: "filesystem_write_file", ErrorClass: "permission_denied", Content: "permission denied for other/path.txt"},
- {Tool: "filesystem_write_file", ErrorClass: "permission_denied", Content: "permission denied for another/path.txt"},
- },
- },
- },
- })
- assertDecisionStatus(t, decision, DecisionContinue)
-}
-
-func TestDecideWorkspaceWriteWithoutExplicitTargetFallsBackToAccepted(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- UserGoal: "edit file",
- })
- if decision.Status != DecisionAccepted {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionAccepted)
- }
-}
-
-func TestLatestWriteVerificationHintBranches(t *testing.T) {
- facts := runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{
- {Path: "a.txt", ExpectedContent: "A"},
- {Path: "b.txt", ExpectedContent: "B"},
- },
- },
- }
- path, expected := latestWriteVerificationHint(facts, "b.txt")
- if path != "b.txt" || expected != "B" {
- t.Fatalf("hint for preferred path = (%q,%q), want (b.txt,B)", path, expected)
- }
-
- path, expected = latestWriteVerificationHint(facts, "missing.txt")
- if path != "missing.txt" || expected != "" {
- t.Fatalf("fallback preferred hint = (%q,%q), want (missing.txt,\"\")", path, expected)
- }
-
- path, expected = latestWriteVerificationHint(facts, "")
- if path != "b.txt" || expected != "B" {
- t.Fatalf("latest hint = (%q,%q), want (b.txt,B)", path, expected)
- }
-}
-
-func TestDecideWorkspaceWriteMissingFactsShouldRequestInputNotPlaceholderAction(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- UserGoal: "请帮我修一下",
- })
- if decision.Status != DecisionAccepted {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionAccepted)
- }
-
- decision = Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: false,
- CompletionReason: "unverified_write",
- UserGoal: "请帮我修一下",
- })
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
- if decision.RequiredInput == nil {
- t.Fatalf("required_input is nil")
- }
- if len(decision.RequiredNextActions) != 0 {
- t.Fatalf("required actions = %+v, want empty", decision.RequiredNextActions)
- }
-}
-
-func TestDecideWorkspaceWriteSelectsLatestMentionedTarget(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- UserGoal: "创建 2.txt 内容为 2",
- Facts: runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{
- {Path: "1.txt", WorkspaceWrite: true, ExpectedContent: "1"},
- {Path: "2.txt", WorkspaceWrite: true, ExpectedContent: "2"},
- },
- },
- },
- })
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
- if len(decision.MissingFacts) == 0 || decision.MissingFacts[0].Target != "2.txt" {
- t.Fatalf("missing facts = %+v, want target 2.txt", decision.MissingFacts)
- }
-}
-
-func TestDecideRequiredNextActionsShouldNotContainPlaceholders(t *testing.T) {
- decisions := []Decision{
- Decide(DecisionInput{
- TaskKind: TaskKindSubAgent,
- CompletionPassed: true,
- UserGoal: "用 subagent 创建 test1.txt 内容为 1",
- }),
- Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: false,
- CompletionReason: "unverified_write",
- UserGoal: "请继续",
- }),
- Decide(DecisionInput{
- TaskKind: TaskKindTodoState,
- CompletionPassed: true,
- UserGoal: "创建 todo",
- }),
- }
- for i, decision := range decisions {
- payload, err := json.Marshal(decision.RequiredNextActions)
- if err != nil {
- t.Fatalf("marshal required_next_actions[%d] failed: %v", i, err)
- }
- serialized := string(payload)
- if strings.Contains(serialized, "<") || strings.Contains(serialized, ">") {
- t.Fatalf("required_next_actions[%d] contains placeholder: %s", i, serialized)
- }
- }
-}
-
-func TestDecideSubAgentRequiresCompletedFact(t *testing.T) {
- continueDecision := Decide(DecisionInput{
- TaskKind: TaskKindSubAgent,
- CompletionPassed: true,
- Facts: runtimefacts.RuntimeFacts{
- SubAgents: runtimefacts.SubAgentFacts{
- Started: []runtimefacts.SubAgentFact{{TaskID: "sa-1", Role: "reviewer"}},
- },
- },
- })
- if continueDecision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", continueDecision.Status, DecisionContinue)
- }
-
- failedDecision := Decide(DecisionInput{
- TaskKind: TaskKindSubAgent,
- CompletionPassed: true,
- Facts: runtimefacts.RuntimeFacts{
- SubAgents: runtimefacts.SubAgentFacts{
- Started: []runtimefacts.SubAgentFact{{TaskID: "sa-1", Role: "reviewer"}},
- Failed: []runtimefacts.SubAgentFact{{TaskID: "sa-1", Role: "reviewer"}},
- },
- },
- })
- if failedDecision.Status != DecisionFailed {
- t.Fatalf("status = %q, want %q", failedDecision.Status, DecisionFailed)
- }
-
- acceptedDecision := Decide(DecisionInput{
- TaskKind: TaskKindSubAgent,
- CompletionPassed: true,
- Facts: runtimefacts.RuntimeFacts{
- SubAgents: runtimefacts.SubAgentFacts{
- Started: []runtimefacts.SubAgentFact{{TaskID: "sa-1", Role: "reviewer"}},
- Completed: []runtimefacts.SubAgentFact{{TaskID: "sa-1", Role: "reviewer"}},
- },
- },
- })
- if acceptedDecision.Status != DecisionAccepted {
- t.Fatalf("status = %q, want %q", acceptedDecision.Status, DecisionAccepted)
- }
-}
-
-func TestDecideSubAgentWriteIntentNeedsArtifactEvidence(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindSubAgent,
- CompletionPassed: true,
- UserGoal: "用 subagent 创建 test1.txt，内容为 1",
- Facts: runtimefacts.RuntimeFacts{
- SubAgents: runtimefacts.SubAgentFacts{
- Started: []runtimefacts.SubAgentFact{{TaskID: "sa-1", Role: "coder"}},
- Completed: []runtimefacts.SubAgentFact{{TaskID: "sa-1", Role: "coder"}},
- },
- },
- })
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
- if len(decision.MissingFacts) == 0 || decision.MissingFacts[0].Kind != "subagent_artifact_or_file_fact" {
- t.Fatalf("missing facts = %+v", decision.MissingFacts)
- }
-
- accepted := Decide(DecisionInput{
- TaskKind: TaskKindSubAgent,
- CompletionPassed: true,
- UserGoal: "用 subagent 创建 test1.txt，内容为 1",
- Facts: runtimefacts.RuntimeFacts{
- SubAgents: runtimefacts.SubAgentFacts{
- Started: []runtimefacts.SubAgentFact{{TaskID: "sa-1", Role: "coder"}},
- Completed: []runtimefacts.SubAgentFact{{
- TaskID: "sa-1", Role: "coder", Artifacts: []string{"test1.txt"},
- }},
- },
- },
- })
- if accepted.Status != DecisionAccepted {
- t.Fatalf("status = %q, want %q", accepted.Status, DecisionAccepted)
- }
-}
-
-func TestDecideTodoStateAcceptsWithoutFileVerification(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindTodoState,
- CompletionPassed: true,
- Todos: TodoSnapshot{
- Items: []TodoViewItem{
- {ID: "todo-1", Content: "创建 Todo", Status: "pending", Required: false},
- },
- Summary: TodoSummary{Total: 1, RequiredTotal: 0},
- },
- Facts: runtimefacts.RuntimeFacts{
- Todos: runtimefacts.TodoFacts{CreatedIDs: []string{"todo-1"}},
- },
- })
- if decision.Status != DecisionAccepted {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionAccepted)
- }
-}
-
-func TestDecideReadOnlyBranches(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindReadOnly,
- CompletionPassed: true,
- })
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
-
- decision = Decide(DecisionInput{
- TaskKind: TaskKindReadOnly,
- CompletionPassed: true,
- LastAssistantText: "analysis done",
- })
- if decision.Status != DecisionAccepted {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionAccepted)
- }
-
- decision = Decide(DecisionInput{
- TaskKind: TaskKindReadOnly,
- CompletionPassed: true,
- Facts: runtimefacts.RuntimeFacts{
- Commands: runtimefacts.CommandFacts{
- Executed: []runtimefacts.CommandFact{{Tool: "bash", Command: "ls", ExitCode: 0, Succeeded: true}},
- },
- },
- })
- if decision.Status != DecisionAccepted {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionAccepted)
- }
-}
-
-func TestDecideMixedBranches(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindMixed,
- CompletionPassed: true,
- })
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
-
- decision = Decide(DecisionInput{
- TaskKind: TaskKindMixed,
- CompletionPassed: true,
- Todos: TodoSnapshot{
- Summary: TodoSummary{RequiredOpen: 1},
- },
- Facts: runtimefacts.RuntimeFacts{
- Verification: runtimefacts.VerificationFacts{
- Passed: []runtimefacts.VerificationFact{{Tool: "filesystem_glob", Scope: "artifact:test.txt"}},
- },
- },
- })
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
-
- decision = Decide(DecisionInput{
- TaskKind: TaskKindMixed,
- CompletionPassed: true,
- LastAssistantText: "analysis done",
- Facts: runtimefacts.RuntimeFacts{
- Verification: runtimefacts.VerificationFacts{
- Passed: []runtimefacts.VerificationFact{{Tool: "filesystem_glob", Scope: "artifact:test.txt"}},
- },
- },
- })
- if decision.Status != DecisionAccepted {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionAccepted)
- }
-}
-
-func TestDecideWorkspaceWriteInjectsLatestToolErrorIntoMissingFact(t *testing.T) {
- decision := Decide(DecisionInput{
- TaskKind: TaskKindWorkspaceWrite,
- CompletionPassed: true,
- UserGoal: "please update README.md",
- Facts: runtimefacts.RuntimeFacts{
- Errors: runtimefacts.ErrorFacts{
- ToolErrors: []runtimefacts.ToolErrorFact{
- {Tool: "filesystem_write_file", ErrorClass: "permission_denied", Content: "first error"},
- {Tool: "filesystem_write_file", ErrorClass: "generic_error", Content: ""},
- },
- },
- },
- })
- if decision.Status != DecisionContinue {
- t.Fatalf("status = %q, want %q", decision.Status, DecisionContinue)
- }
- if len(decision.MissingFacts) == 0 {
- t.Fatalf("missing facts = %+v", decision.MissingFacts)
- }
- details := decision.MissingFacts[0].Details
- if details["last_write_error"] != "generic_error" {
- t.Fatalf("last_write_error = %#v, want generic_error", details["last_write_error"])
- }
-}
diff --git a/internal/runtime/decider/infer.go b/internal/runtime/decider/infer.go
deleted file mode 100644
index c9a24adc..00000000
--- a/internal/runtime/decider/infer.go
+++ /dev/null
@@ -1,125 +0,0 @@
-package decider
-
-import (
- "strings"
-
- runtimefacts "neo-code/internal/runtime/facts"
-)
-
-// InferTaskKind 通过规则推断任务类型，避免依赖模型分类。
-func InferTaskKind(goal string) TaskKind {
- return InferTaskIntent(goal).Hint
-}
-
-// InferTaskIntent 基于用户文本推断弱意图，仅作 hint，不作为强验收依据。
-func InferTaskIntent(goal string) TaskIntent {
- text := strings.ToLower(strings.TrimSpace(goal))
- if text == "" {
- return TaskIntent{Hint: TaskKindChatAnswer, Confidence: 0.2, Reasons: []string{"empty_goal"}}
- }
-
- hasTodo := containsAny(text, "todo", "待办")
- hasSubAgent := containsAny(text, "subagent", "子代理")
- hasWriteVerb := containsAny(
- text,
- "创建文件", "写入", "修改文件", "编辑文件", "新增文件", "补丁", "修复代码",
- "create file", "write file", "edit file", "update file", "apply patch",
- )
- hasFileTarget := containsAny(text, ".txt", ".go", ".md", ".json", ".yaml", ".yml", ".ts", ".tsx")
- hasNamedWriteTarget := containsAny(text, "readme", "package.json")
- hasWriteIntentToken := containsAny(text, "创建", "写", "改", "补", "edit", "write", "update", "create", "modify")
- hasWrite := hasWriteVerb || ((hasFileTarget || hasNamedWriteTarget) && hasWriteIntentToken)
- hasRead := containsAny(
- text,
- "读取", "查看", "看看", "总结", "分析", "检索", "搜索", "审查", "review", "verify", "验证", "校验", "怎么修",
- "read", "grep", "glob", "list", "inspect", "analyze", "summarize",
- )
- hasReviewIntent := containsAny(text, "review", "审查", "总结", "分析", "analyze", "summarize")
- hasPlan := containsAny(text, "计划", "规划", "plan", "todo 列表", "todo list")
- hasTodoAction := containsAny(text, "创建 todo", "更新 todo", "完成 todo", "标记 todo", "todo")
-
- intent := TaskIntent{Hint: TaskKindChatAnswer, Confidence: 0.35, Reasons: []string{"fallback_chat"}}
- switch {
- case hasSubAgent && hasWrite:
- intent.Hint = TaskKindSubAgent
- intent.Confidence = 0.9
- intent.Reasons = []string{"subagent_keyword", "write_intent"}
- case hasSubAgent:
- intent.Hint = TaskKindSubAgent
- intent.Confidence = 0.82
- intent.Reasons = []string{"subagent_keyword"}
- case hasTodo && hasTodoAction:
- intent.Hint = TaskKindTodoState
- intent.Confidence = 0.78
- intent.Reasons = []string{"todo_action_keyword"}
- case hasPlan && hasTodo && !hasWrite:
- intent.Hint = TaskKindTodoState
- intent.Confidence = 0.84
- intent.Reasons = []string{"todo_keyword", "plan_keyword"}
- case hasWrite && hasReviewIntent:
- intent.Hint = TaskKindMixed
- intent.Confidence = 0.75
- intent.Reasons = []string{"write_intent", "read_intent"}
- case hasWrite:
- intent.Hint = TaskKindWorkspaceWrite
- intent.Confidence = 0.72
- intent.Reasons = []string{"write_intent"}
- case hasRead:
- intent.Hint = TaskKindReadOnly
- intent.Confidence = 0.7
- intent.Reasons = []string{"read_intent"}
- }
- return intent
-}
-
-// DeriveEffectiveTaskKind 基于运行事实修正任务类型；文本 hint 仅作回退。
-func DeriveEffectiveTaskKind(hint TaskKind, allFacts runtimefacts.RuntimeFacts, todos TodoSnapshot) TaskKind {
- hasWrite := len(allFacts.Files.Written) > 0 || len(allFacts.Files.ContentMatch) > 0
- if !hasWrite && hint == TaskKindWorkspaceWrite && hasArtifactVerificationPassed(allFacts) {
- hasWrite = true
- }
- hasVerification := len(allFacts.Verification.Passed) > 0
- hasSubAgent := len(allFacts.SubAgents.Started) > 0 || len(allFacts.SubAgents.Completed) > 0 || len(allFacts.SubAgents.Failed) > 0
- hasTodo := todos.Summary.Total > 0 || len(allFacts.Todos.CreatedIDs) > 0 || len(allFacts.Todos.CompletedIDs) > 0 || len(allFacts.Todos.FailedIDs) > 0
- hasRead := len(allFacts.Files.Exists) > 0 || len(allFacts.Commands.Executed) > 0
-
- switch {
- case hasSubAgent && (hasWrite || hasTodo || hasVerification):
- return TaskKindMixed
- case hasSubAgent:
- return TaskKindSubAgent
- case hasWrite && (hasTodo || hasVerification || hasRead):
- return TaskKindWorkspaceWrite
- case hasWrite:
- return TaskKindWorkspaceWrite
- case hasTodo && !hasWrite:
- return TaskKindTodoState
- case hasRead || hasVerification:
- return TaskKindReadOnly
- case strings.TrimSpace(string(hint)) != "":
- return hint
- default:
- return TaskKindChatAnswer
- }
-}
-
-// hasArtifactVerificationPassed 判断是否存在与产物路径绑定的通过验证事实。
-func hasArtifactVerificationPassed(allFacts runtimefacts.RuntimeFacts) bool {
- for _, fact := range allFacts.Verification.Passed {
- scope := strings.TrimSpace(fact.Scope)
- if strings.HasPrefix(strings.ToLower(scope), "artifact:") {
- return true
- }
- }
- return false
-}
-
-// containsAny 判断文本是否包含任一关键词。
-func containsAny(text string, keywords ...string) bool {
- for _, keyword := range keywords {
- if strings.Contains(text, strings.ToLower(strings.TrimSpace(keyword))) {
- return true
- }
- }
- return false
-}
diff --git a/internal/runtime/decider/infer_test.go b/internal/runtime/decider/infer_test.go
deleted file mode 100644
index b6c8fe05..00000000
--- a/internal/runtime/decider/infer_test.go
+++ /dev/null
@@ -1,117 +0,0 @@
-package decider
-
-import (
- "testing"
-
- runtimefacts "neo-code/internal/runtime/facts"
-)
-
-func TestInferTaskKind(t *testing.T) {
- t.Parallel()
-
- tests := []struct {
- name string
- goal string
- want TaskKind
- }{
- {
- name: "todo plan",
- goal: "请创建 todo 列表并规划后续任务",
- want: TaskKindTodoState,
- },
- {
- name: "workspace write",
- goal: "创建文件 test.txt 并写入 1",
- want: TaskKindWorkspaceWrite,
- },
- {
- name: "review read only",
- goal: "review README.md 并总结风险",
- want: TaskKindReadOnly,
- },
- {
- name: "mixed write and review",
- goal: "edit main.go then review changes",
- want: TaskKindMixed,
- },
- {
- name: "subagent explicit",
- goal: "用 subagent 创建 test1.txt，内容为 1",
- want: TaskKindSubAgent,
- },
- {
- name: "chat answer fallback",
- goal: "什么是 NeoCode",
- want: TaskKindChatAnswer,
- },
- {
- name: "greeting chat answer",
- goal: "你好",
- want: TaskKindChatAnswer,
- },
- {
- name: "bug fix discussion should not be workspace write",
- goal: "帮我看看这个 bug 怎么修",
- want: TaskKindReadOnly,
- },
- {
- name: "review implementation should be read only",
- goal: "review this implementation and suggest fixes",
- want: TaskKindReadOnly,
- },
- {
- name: "update readme should be workspace write",
- goal: "把 README 补一下",
- want: TaskKindWorkspaceWrite,
- },
- {
- name: "todo creation should be todo state",
- goal: "创建一个 Todo，内容是 1",
- want: TaskKindTodoState,
- },
- {
- name: "todo content contains write target should still be todo state hint",
- goal: "创建一个 Todo，内容是创建 test.txt 内容为 1",
- want: TaskKindTodoState,
- },
- }
-
- for _, tt := range tests {
- tt := tt
- t.Run(tt.name, func(t *testing.T) {
- t.Parallel()
- got := InferTaskKind(tt.goal)
- if got != tt.want {
- t.Fatalf("InferTaskKind(%q) = %q, want %q", tt.goal, got, tt.want)
- }
- })
- }
-}
-
-func TestDeriveEffectiveTaskKindFactsOverrideHint(t *testing.T) {
- t.Parallel()
-
- if got := DeriveEffectiveTaskKind(TaskKindChatAnswer, runtimefacts.RuntimeFacts{
- Files: runtimefacts.FileFacts{
- Written: []runtimefacts.FileWriteFact{{Path: "a.txt", WorkspaceWrite: true}},
- },
- }, TodoSnapshot{}); got != TaskKindWorkspaceWrite {
- t.Fatalf("effective kind = %q, want %q", got, TaskKindWorkspaceWrite)
- }
-
- if got := DeriveEffectiveTaskKind(TaskKindWorkspaceWrite, runtimefacts.RuntimeFacts{
- Commands: runtimefacts.CommandFacts{
- Executed: []runtimefacts.CommandFact{{Tool: "bash", Command: "ls", Succeeded: true}},
- },
- }, TodoSnapshot{}); got != TaskKindReadOnly {
- t.Fatalf("effective kind = %q, want %q", got, TaskKindReadOnly)
- }
-
- if got := DeriveEffectiveTaskKind(TaskKindWorkspaceWrite, runtimefacts.RuntimeFacts{
- Verification: runtimefacts.VerificationFacts{
- Passed: []runtimefacts.VerificationFact{{Tool: "filesystem_write_file", Scope: "artifact:2.txt"}},
- },
- }, TodoSnapshot{}); got != TaskKindWorkspaceWrite {
- t.Fatalf("effective kind = %q, want %q when artifact verification exists", got, TaskKindWorkspaceWrite)
- }
-}
diff --git a/internal/runtime/decider/types.go b/internal/runtime/decider/types.go
deleted file mode 100644
index d45eba92..00000000
--- a/internal/runtime/decider/types.go
+++ /dev/null
@@ -1,134 +0,0 @@
-package decider
-
-import "neo-code/internal/runtime/facts"
-
-// TaskKind 描述任务验收的主类型。
-type TaskKind string
-
-const (
- // TaskKindChatAnswer 表示普通问答任务。
- TaskKindChatAnswer TaskKind = "chat_answer"
- // TaskKindTodoState 表示 todo 状态任务。
- TaskKindTodoState TaskKind = "todo_state"
- // TaskKindWorkspaceWrite 表示工作区写入任务。
- TaskKindWorkspaceWrite TaskKind = "workspace_write"
- // TaskKindSubAgent 表示显式子代理任务。
- TaskKindSubAgent TaskKind = "subagent"
- // TaskKindReadOnly 表示只读分析任务。
- TaskKindReadOnly TaskKind = "read_only"
- // TaskKindMixed 表示混合任务。
- TaskKindMixed TaskKind = "mixed"
-)
-
-// DecisionStatus 表示终态决策状态。
-type DecisionStatus string
-
-const (
- // DecisionAccepted 表示满足收尾条件。
- DecisionAccepted DecisionStatus = "accepted"
- // DecisionContinue 表示仍需继续执行。
- DecisionContinue DecisionStatus = "continue"
- // DecisionFailed 表示任务失败终止。
- DecisionFailed DecisionStatus = "failed"
- // DecisionBlocked 表示被外部条件阻塞。
- DecisionBlocked DecisionStatus = "blocked"
- // DecisionIncomplete 表示长时间无进展后未完成终止。
- DecisionIncomplete DecisionStatus = "incomplete"
-)
-
-// MissingFact 描述 continue 场景下缺失的客观事实。
-type MissingFact struct {
- Kind string `json:"kind"`
- Target string `json:"target,omitempty"`
- Expected string `json:"expected,omitempty"`
- Details map[string]any `json:"details,omitempty"`
-}
-
-// RequiredAction 描述下一轮建议工具动作。
-type RequiredAction struct {
- Tool string `json:"tool"`
- ArgsHint map[string]any `json:"args_hint,omitempty"`
-}
-
-// RequiredInput 描述继续执行前必须补充的人类输入。
-type RequiredInput struct {
- Kind string `json:"kind"`
- Message string `json:"message"`
- Details map[string]any `json:"details,omitempty"`
-}
-
-// HookGuardSignal 描述 before_completion_decision user/repo hook 产生的守卫信号。
-type HookGuardSignal struct {
- HookID string `json:"hook_id,omitempty"`
- Source string `json:"source,omitempty"`
- Message string `json:"message,omitempty"`
-}
-
-// TaskIntent 描述由用户文本推断出的弱意图线索。
-type TaskIntent struct {
- Hint TaskKind `json:"hint,omitempty"`
- Confidence float64 `json:"confidence,omitempty"`
- Reasons []string `json:"reasons,omitempty"`
-}
-
-// Decision 描述最终裁决结果。
-type Decision struct {
- Status DecisionStatus `json:"status"`
- StopReason string `json:"stop_reason,omitempty"`
- MissingFacts []MissingFact `json:"missing_facts,omitempty"`
- RequiredNextActions []RequiredAction `json:"required_next_actions,omitempty"`
- RequiredInput *RequiredInput `json:"required_input,omitempty"`
- IntentHint TaskKind `json:"intent_hint,omitempty"`
- EffectiveTaskKind TaskKind `json:"effective_task_kind,omitempty"`
- UserVisibleSummary string `json:"user_visible_summary,omitempty"`
- InternalSummary string `json:"internal_summary,omitempty"`
-}
-
-// TodoViewItem 描述决策所需 todo 快照条目。
-type TodoViewItem struct {
- ID string
- Content string
- Status string
- Required bool
- Artifacts []string
- FailureReason string
- BlockedReason string
- Revision int64
-}
-
-// TodoSummary 描述决策所需 todo 汇总。
-type TodoSummary struct {
- Total int
- RequiredTotal int
- RequiredCompleted int
- RequiredFailed int
- RequiredOpen int
-}
-
-// TodoSnapshot 描述决策所需 todo 快照。
-type TodoSnapshot struct {
- Items []TodoViewItem
- Summary TodoSummary
-}
-
-// ProgressSnapshot 描述 no-progress 判定所需信息。
-type ProgressSnapshot struct {
- FactCount int
-}
-
-// DecisionInput 描述终态裁决输入。
-type DecisionInput struct {
- RunID string
- SessionID string
- TaskKind TaskKind
- UserGoal string
- Facts facts.RuntimeFacts
- Todos TodoSnapshot
- Progress ProgressSnapshot
- LastAssistantText string
- CompletionPassed bool
- CompletionReason string
- NoProgressExceeded bool
- HookAnnotations []string
- HookGuards []HookGuardSignal
-}
diff --git a/internal/runtime/events.go b/internal/runtime/events.go
index 39a767a8..d4b3e1ee 100644
--- a/internal/runtime/events.go
+++ b/internal/runtime/events.go
@@ -3,7 +3,7 @@ package runtime
import (
"time"
- "neo-code/internal/runtime/acceptance"
+ "neo-code/internal/runtime/acceptgate"
"neo-code/internal/runtime/controlplane"
"neo-code/internal/runtime/verify"
)
@@ -77,9 +77,9 @@ type VerificationStageFinishedPayload struct {
// VerificationFinishedPayload 描述整体验证流程结束事件。
type VerificationFinishedPayload struct {
- AcceptanceStatus acceptance.AcceptanceStatus `json:"acceptance_status"`
- StopReason controlplane.StopReason `json:"stop_reason,omitempty"`
- ErrorClass verify.ErrorClass `json:"error_class,omitempty"`
+ AcceptanceStatus string `json:"acceptance_status"`
+ StopReason controlplane.StopReason `json:"stop_reason,omitempty"`
+ ErrorClass verify.ErrorClass `json:"error_class,omitempty"`
}
// VerificationCompletedPayload 描述验证通过并可完成的事件。
@@ -95,13 +95,10 @@ type VerificationFailedPayload struct {
// AcceptanceDecidedPayload 描述 acceptance engine 决议结果。
type AcceptanceDecidedPayload struct {
- Status acceptance.AcceptanceStatus `json:"status"`
- StopReason controlplane.StopReason `json:"stop_reason,omitempty"`
- ErrorClass verify.ErrorClass `json:"error_class,omitempty"`
- CompletionBlockedReason string `json:"completion_blocked_reason,omitempty"`
- UserVisibleSummary string `json:"user_visible_summary,omitempty"`
- InternalSummary string `json:"internal_summary,omitempty"`
- ContinueHint string `json:"continue_hint,omitempty"`
+ Status string `json:"status"`
+ StopReason controlplane.StopReason `json:"stop_reason,omitempty"`
+ Summary string `json:"summary,omitempty"`
+ Results []acceptgate.CheckResult `json:"results,omitempty"`
}
// LedgerReconciledPayload 为账本对账预留负载。
@@ -521,16 +518,16 @@ type RunDiffSummaryPayload struct {
// UserQuestionRequestedPayload 描述 ask_user 提问事件负载。
type UserQuestionRequestedPayload struct {
- RequestID string `json:"request_id"`
- QuestionID string `json:"question_id"`
- Title string `json:"title"`
- Description string `json:"description"`
- Kind string `json:"kind"`
- Options []any `json:"options,omitempty"`
- Required bool `json:"required"`
- AllowSkip bool `json:"allow_skip"`
- MaxChoices int `json:"max_choices,omitempty"`
- TimeoutSec int `json:"timeout_sec,omitempty"`
+ RequestID string `json:"request_id"`
+ QuestionID string `json:"question_id"`
+ Title string `json:"title"`
+ Description string `json:"description"`
+ Kind string `json:"kind"`
+ Options []any `json:"options,omitempty"`
+ Required bool `json:"required"`
+ AllowSkip bool `json:"allow_skip"`
+ MaxChoices int `json:"max_choices,omitempty"`
+ TimeoutSec int `json:"timeout_sec,omitempty"`
}
// UserQuestionResolvedPayload 描述 ask_user 回答/跳过/超时事件负载。
diff --git a/internal/runtime/facts/collector.go b/internal/runtime/facts/collector.go
index fa1ba230..a4a78f60 100644
--- a/internal/runtime/facts/collector.go
+++ b/internal/runtime/facts/collector.go
@@ -110,7 +110,7 @@ func (c *Collector) ApplyToolResult(toolName string, result tools.ToolResult) {
c.applyVerificationFacts(name, result)
}
-// applyWorkspaceWritePathFacts 将工具 metadata 中声明的写入路径转成可被 decider 验收的文件事实。
+// applyWorkspaceWritePathFacts 将工具 metadata 中声明的写入路径转成可被 Accept Gate 消费的文件事实。
func (c *Collector) applyWorkspaceWritePathFacts(result tools.ToolResult, source string) {
if !result.Facts.WorkspaceWrite {
return
diff --git a/internal/runtime/final_acceptance.go b/internal/runtime/final_acceptance.go
deleted file mode 100644
index 28e89b0f..00000000
--- a/internal/runtime/final_acceptance.go
+++ /dev/null
@@ -1,815 +0,0 @@
-package runtime
-
-import (
- "context"
- "encoding/json"
- "fmt"
- "slices"
- "strings"
-
- providertypes "neo-code/internal/provider/types"
- "neo-code/internal/runtime/acceptance"
- "neo-code/internal/runtime/controlplane"
- "neo-code/internal/runtime/decider"
- runtimefacts "neo-code/internal/runtime/facts"
- "neo-code/internal/runtime/verify"
- agentsession "neo-code/internal/session"
-)
-
-const finalContinueReminder = "There are unfinished required todos or unmet acceptance checks. Continue execution. Do not finalize yet."
-
-// beforeAcceptFinal 在 runtime 接受模型 final 前执行唯一的 completion/verifier/acceptance 闭环。
-func (s *Service) beforeAcceptFinal(
- ctx context.Context,
- state *runState,
- snapshot TurnBudgetSnapshot,
- assistant providertypes.Message,
- completionPassed bool,
- signals beforeCompletionHookSignals,
-) (acceptance.AcceptanceDecision, error) {
- if state == nil {
- return acceptance.AcceptanceDecision{}, nil
- }
-
- maxNoProgress := resolveAcceptanceMaxNoProgress(snapshot.Config.Runtime.Verification)
- noProgressStreak := state.finalInterceptStreak
- if noProgressStreak < 0 {
- noProgressStreak = 0
- }
- if state.mustUseToolAfterFinalContinue && state.noToolAfterFinalContinueStreak > noProgressStreak {
- noProgressStreak = state.noToolAfterFinalContinueStreak
- }
-
- input := s.buildAcceptanceServiceInput(
- state,
- snapshot,
- assistant,
- completionPassed,
- signals,
- noProgressStreak,
- maxNoProgress,
- )
- service := &acceptanceService{}
- acceptanceDecision, err := service.Decide(ctx, input)
- if err != nil {
- return acceptance.AcceptanceDecision{}, err
- }
- deciderDecision := toDeciderDecisionFromAcceptance(acceptanceDecision)
- state.mu.Lock()
- state.lastDeciderDecision = deciderDecision
- pendingFinalProgress := state.pendingFinalProgress
- state.mu.Unlock()
- s.emitRunScopedOptional(EventDecisionMade, state, deciderDecision)
- s.emitRuntimeSnapshotUpdated(ctx, state, "decision_made")
- if acceptanceDecision.Status == acceptance.AcceptanceContinue && pendingFinalProgress {
- acceptanceDecision.HasProgress = true
- }
- return acceptanceDecision, nil
-}
-
-// buildAcceptanceServiceInput 从当前运行态抽取 AcceptanceService 所需输入。
-func (s *Service) buildAcceptanceServiceInput(
- state *runState,
- snapshot TurnBudgetSnapshot,
- assistant providertypes.Message,
- completionPassed bool,
- signals beforeCompletionHookSignals,
- noProgressStreak int,
- maxNoProgress int,
-) acceptanceServiceInput {
- state.mu.Lock()
- taskKind := state.taskKind
- userGoal := state.userGoal
- completionReason := strings.TrimSpace(string(state.completion.CompletionBlockedReason))
- verificationProfile := state.session.TaskState.VerificationProfile
- sessionMessages := append([]providertypes.Message(nil), state.session.Messages...)
- sessionTodos := cloneTodosForPersistence(state.session.Todos)
- sessionTaskState := state.session.TaskState
- todoSnapshot := buildTodoSnapshotFromItems(sessionTodos)
- factsSnapshot := runtimefacts.RuntimeFacts{}
- if state.factsCollector != nil {
- factsSnapshot = state.factsCollector.Snapshot()
- }
- taskID := strings.TrimSpace(state.taskID)
- runID := strings.TrimSpace(state.runID)
- sessionID := strings.TrimSpace(state.session.ID)
- turn := state.turn
- maxTurnsReached := state.maxTurnsReached
- maxTurns := resolveRuntimeMaxTurns(snapshot.Config.Runtime)
- state.mu.Unlock()
-
- if strings.TrimSpace(userGoal) == "" {
- userGoal = renderPartsForVerification(assistant.Parts)
- }
- if strings.TrimSpace(string(taskKind)) == "" {
- taskKind = decider.InferTaskKind(userGoal)
- }
- if todoSnapshot.Summary.RequiredOpen > 0 {
- completionPassed = false
- if completionReason == "" {
- completionReason = string(controlplane.CompletionBlockedReasonPendingTodo)
- }
- }
- if !maxTurnsReached && maxTurns > 0 && turn+1 >= maxTurns {
- maxTurnsReached = true
- }
- verifyInput := verify.FinalVerifyInput{
- SessionID: sessionID,
- RunID: runID,
- TaskID: taskID,
- Workdir: snapshot.Workdir,
- Messages: buildVerifyMessages(sessionMessages),
- Todos: buildVerifyTodos(sessionTodos),
- LastAssistantFinal: renderPartsForVerification(assistant.Parts),
- TaskState: buildVerifyTaskState(sessionTaskState),
- RuntimeState: verify.RuntimeStateSnapshot{
- Turn: turn,
- MaxTurns: maxTurns,
- MaxTurnsReached: maxTurnsReached,
- FinalInterceptStreak: noProgressStreak,
- },
- VerificationConfig: snapshot.Config.Runtime.Verification.Clone(),
- }
- return acceptanceServiceInput{
- RunID: runID,
- SessionID: sessionID,
- TaskKind: taskKind,
- UserGoal: userGoal,
- CompletionPassed: completionPassed,
- CompletionBlockedReason: completionReason,
- Facts: factsSnapshot,
- Todos: toDeciderTodoSnapshot(todoSnapshot),
- Progress: toDeciderProgress(factsSnapshot),
- LastAssistantText: renderPartsForVerification(assistant.Parts),
- HookAnnotations: append([]string(nil), signals.Annotations...),
- HookGuards: append([]decider.HookGuardSignal(nil), signals.Guards...),
- NoProgressStreak: noProgressStreak,
- MaxNoProgress: maxNoProgress,
- VerificationProfile: verificationProfile,
- VerificationInput: verifyInput,
- }
-}
-
-// toDeciderDecisionFromAcceptance 将统一 acceptance 决策投影为 runtime snapshot 兼容的 decider 视图。
-func toDeciderDecisionFromAcceptance(decision acceptance.AcceptanceDecision) decider.Decision {
- status := decider.DecisionContinue
- switch decision.Status {
- case acceptance.AcceptanceAccepted:
- status = decider.DecisionAccepted
- case acceptance.AcceptanceFailed:
- status = decider.DecisionFailed
- case acceptance.AcceptanceIncomplete:
- status = decider.DecisionIncomplete
- }
- return decider.Decision{
- Status: status,
- StopReason: strings.TrimSpace(string(decision.StopReason)),
- MissingFacts: append([]decider.MissingFact(nil), decision.MissingFacts...),
- RequiredNextActions: append([]decider.RequiredAction(nil), decision.RequiredNextActions...),
- RequiredInput: cloneRequiredInput(decision.RequiredInput),
- IntentHint: decision.IntentHint,
- EffectiveTaskKind: decision.EffectiveTaskKind,
- UserVisibleSummary: strings.TrimSpace(decision.UserVisibleSummary),
- InternalSummary: strings.TrimSpace(decision.InternalSummary),
- }
-}
-
-func cloneRequiredInput(in *decider.RequiredInput) *decider.RequiredInput {
- if in == nil {
- return nil
- }
- cloned := *in
- if len(in.Details) == 0 {
- return &cloned
- }
- cloned.Details = make(map[string]any, len(in.Details))
- for k, v := range in.Details {
- cloned.Details[k] = v
- }
- return &cloned
-}
-
-// mapDeciderDecisionToAcceptance 把 FinalDecider 裁决映射到 acceptance 协议。
-// Deprecated: 仅保留给 legacy 回滚对照与测试使用；P7 主链直接消费 AcceptanceService.Decide 产物。
-func mapDeciderDecisionToAcceptance(decision decider.Decision) acceptance.AcceptanceDecision {
- out := acceptance.AcceptanceDecision{
- StopReason: toControlplaneStopReason(decision.StopReason),
- RequiredInput: cloneRequiredInput(decision.RequiredInput),
- IntentHint: decision.IntentHint,
- EffectiveTaskKind: decision.EffectiveTaskKind,
- UserVisibleSummary: strings.TrimSpace(decision.UserVisibleSummary),
- InternalSummary: strings.TrimSpace(decision.InternalSummary),
- ContinueHint: buildDeciderContinueHint(decision),
- }
- switch decision.Status {
- case decider.DecisionAccepted:
- out.Status = acceptance.AcceptanceAccepted
- if out.StopReason == "" {
- out.StopReason = controlplane.StopReasonAccepted
- }
- case decider.DecisionFailed, decider.DecisionBlocked:
- out.Status = acceptance.AcceptanceFailed
- if out.StopReason == "" {
- out.StopReason = controlplane.StopReasonVerificationFailed
- }
- case decider.DecisionIncomplete:
- out.Status = acceptance.AcceptanceIncomplete
- if out.StopReason == "" {
- out.StopReason = controlplane.StopReasonNoProgressAfterFinalIntercept
- }
- default:
- out.Status = acceptance.AcceptanceContinue
- if out.StopReason == "" {
- out.StopReason = controlplane.StopReasonTodoNotConverged
- }
- }
- return out
-}
-
-// toDeciderTodoSnapshot 转换 runtime todo 快照到 decider 输入结构。
-func toDeciderTodoSnapshot(snapshot TodoSnapshot) decider.TodoSnapshot {
- out := decider.TodoSnapshot{
- Summary: decider.TodoSummary{
- Total: snapshot.Summary.Total,
- RequiredTotal: snapshot.Summary.RequiredTotal,
- RequiredCompleted: snapshot.Summary.RequiredCompleted,
- RequiredFailed: snapshot.Summary.RequiredFailed,
- RequiredOpen: snapshot.Summary.RequiredOpen,
- },
- }
- if len(snapshot.Items) == 0 {
- return out
- }
- out.Items = make([]decider.TodoViewItem, 0, len(snapshot.Items))
- for _, item := range snapshot.Items {
- out.Items = append(out.Items, decider.TodoViewItem{
- ID: strings.TrimSpace(item.ID),
- Content: strings.TrimSpace(item.Content),
- Status: strings.TrimSpace(item.Status),
- Required: item.Required,
- Artifacts: append([]string(nil), item.Artifacts...),
- FailureReason: strings.TrimSpace(item.FailureReason),
- BlockedReason: strings.TrimSpace(item.BlockedReason),
- Revision: item.Revision,
- })
- }
- return out
-}
-
-// toDeciderProgress 构建 decider 所需的最小进度快照。
-func toDeciderProgress(factsSnapshot runtimefacts.RuntimeFacts) decider.ProgressSnapshot {
- return decider.ProgressSnapshot{
- FactCount: max(0, factsSnapshot.Progress.ObservedFactCount),
- }
-}
-
-// toControlplaneStopReason 把 decider stop reason 映射为 controlplane 枚举。
-func toControlplaneStopReason(reason string) controlplane.StopReason {
- normalized := strings.TrimSpace(reason)
- switch normalized {
- case string(controlplane.StopReasonAccepted):
- return controlplane.StopReasonAccepted
- case string(controlplane.StopReasonTodoNotConverged):
- return controlplane.StopReasonTodoNotConverged
- case string(controlplane.StopReasonNoProgressAfterFinalIntercept):
- return controlplane.StopReasonNoProgressAfterFinalIntercept
- case string(controlplane.StopReasonRequiredTodoFailed):
- return controlplane.StopReasonRequiredTodoFailed
- case string(controlplane.StopReasonVerificationFailed):
- return controlplane.StopReasonVerificationFailed
- case string(controlplane.StopReasonTodoWaitingExternal):
- return controlplane.StopReasonTodoWaitingExternal
- case string(controlplane.StopReasonVerificationConfigMissing):
- return controlplane.StopReasonVerificationConfigMissing
- case string(controlplane.StopReasonVerificationExecutionDenied):
- return controlplane.StopReasonVerificationExecutionDenied
- case string(controlplane.StopReasonVerificationExecutionError):
- return controlplane.StopReasonVerificationExecutionError
- default:
- return ""
- }
-}
-
-// buildDeciderContinueHint 生成 FinalDecider continue 场景下的结构化执行提示。
-func buildDeciderContinueHint(decision decider.Decision) string {
- if decision.Status != decider.DecisionContinue {
- return ""
- }
- var builder strings.Builder
- builder.WriteString("\n")
- if summary := strings.TrimSpace(decision.UserVisibleSummary); summary != "" {
- builder.WriteString("")
- builder.WriteString(xmlEscape(summary))
- builder.WriteString("\n")
- }
- if len(decision.MissingFacts) > 0 {
- builder.WriteString("\n")
- for _, fact := range decision.MissingFacts {
- builder.WriteString(fmt.Sprintf(
- "%s\n",
- xmlEscape(strings.TrimSpace(fact.Kind)),
- xmlEscape(strings.TrimSpace(fact.Target)),
- xmlEscape(evidenceJSONPreview(fact.Details)),
- ))
- }
- builder.WriteString("\n")
- }
- if len(decision.RequiredNextActions) > 0 {
- builder.WriteString("\n")
- for _, action := range decision.RequiredNextActions {
- builder.WriteString(fmt.Sprintf(
- "%s\n",
- xmlEscape(strings.TrimSpace(action.Tool)),
- xmlEscape(evidenceJSONPreview(action.ArgsHint)),
- ))
- }
- builder.WriteString("\n")
- }
- builder.WriteString("Do not claim completion with plain text. Call tools to produce objective facts before final response.\n")
- builder.WriteString("")
- return strings.TrimSpace(builder.String())
-}
-
-// beforeAcceptFinalLegacy 是历史 acceptance/verify 实现，仅用于回滚对照与测试覆盖。
-// Deprecated: P7 主链不再调用该路径，最终裁决统一走 beforeAcceptFinal -> AcceptanceService。
-func (s *Service) beforeAcceptFinalLegacy(
- ctx context.Context,
- state *runState,
- snapshot TurnBudgetSnapshot,
- assistant providertypes.Message,
- completionPassed bool,
-) (acceptance.AcceptanceDecision, error) {
- if state == nil {
- return acceptance.AcceptanceDecision{}, nil
- }
-
- verificationCfg := snapshot.Config.Runtime.Verification.Clone()
- policy := acceptance.DefaultPolicy{
- Executor: verify.PolicyCommandExecutor{},
- }
- engine := acceptance.NewEngine(policy)
-
- maxNoProgress := verificationCfg.MaxNoProgress
- if maxNoProgress <= 0 {
- maxNoProgress = 3
- }
- noProgressStreak := state.finalInterceptStreak
- if noProgressStreak < 0 {
- noProgressStreak = 0
- }
- if state.mustUseToolAfterFinalContinue && state.noToolAfterFinalContinueStreak > noProgressStreak {
- noProgressStreak = state.noToolAfterFinalContinueStreak
- }
- maxTurnsLimit := state.maxTurnsLimit
- maxTurnsReached := state.maxTurnsReached
- if !maxTurnsReached {
- resolvedMaxTurns := resolveRuntimeMaxTurns(snapshot.Config.Runtime)
- if resolvedMaxTurns > 0 && state.turn+1 >= resolvedMaxTurns {
- maxTurnsReached = true
- maxTurnsLimit = resolvedMaxTurns
- }
- }
-
- input := acceptance.FinalAcceptanceInput{
- CompletionGate: acceptance.CompletionGateDecision{
- Passed: completionPassed,
- Reason: string(state.completion.CompletionBlockedReason),
- },
- VerificationInput: verify.FinalVerifyInput{
- SessionID: state.session.ID,
- RunID: state.runID,
- TaskID: state.taskID,
- Workdir: snapshot.Workdir,
- Messages: buildVerifyMessages(state.session.Messages),
- Todos: buildVerifyTodos(state.session.Todos),
- LastAssistantFinal: renderPartsForVerification(assistant.Parts),
- ToolResults: nil,
- TaskState: buildVerifyTaskState(state.session.TaskState),
- RuntimeState: verify.RuntimeStateSnapshot{
- Turn: state.turn,
- MaxTurns: resolveRuntimeMaxTurns(snapshot.Config.Runtime),
- MaxTurnsReached: maxTurnsReached,
- FinalInterceptStreak: noProgressStreak,
- },
- VerificationConfig: verificationCfg,
- },
- NoProgressExceeded: noProgressStreak >= maxNoProgress,
- MaxTurnsReached: maxTurnsReached,
- MaxTurnsLimit: maxTurnsLimit,
- }
-
- decision, err := engine.EvaluateFinal(ctx, input)
- if err != nil {
- return acceptance.AcceptanceDecision{}, err
- }
- if decision.Status == acceptance.AcceptanceContinue && len(decision.VerifierResults) == 0 {
- if synthetic := synthesizeTodoConvergenceEvidence(state.session.Todos); synthetic != nil {
- decision.VerifierResults = append(decision.VerifierResults, *synthetic)
- }
- }
- if decision.Status == acceptance.AcceptanceContinue && state.pendingFinalProgress {
- decision.HasProgress = true
- }
- if strings.TrimSpace(decision.CompletionBlockedReason) == "" {
- decision.CompletionBlockedReason = strings.TrimSpace(string(state.completion.CompletionBlockedReason))
- }
- if decision.Status == acceptance.AcceptanceContinue {
- decision.ContinueHint = buildAcceptanceContinueHint(decision)
- }
- return decision, nil
-}
-
-// synthesizeTodoConvergenceEvidence 在 completion gate 拦截且 verifier 未运行时，回填 todo 证据供 continue hint 使用。
-func synthesizeTodoConvergenceEvidence(todos []agentsession.TodoItem) *verify.VerificationResult {
- if len(todos) == 0 {
- return nil
- }
- pendingIDs := make([]string, 0)
- inProgressIDs := make([]string, 0)
- blockedIDs := make([]string, 0)
- statusByID := make(map[string]string)
- artifactsByID := make(map[string][]string)
- checksByID := make(map[string][]verify.TodoContentCheckSnapshot)
-
- for _, todo := range todos {
- if !todo.RequiredValue() {
- continue
- }
- id := strings.TrimSpace(todo.ID)
- if id == "" {
- continue
- }
- status := strings.TrimSpace(string(todo.Status))
- statusByID[id] = status
- switch status {
- case string(agentsession.TodoStatusPending):
- pendingIDs = append(pendingIDs, id)
- case string(agentsession.TodoStatusInProgress):
- inProgressIDs = append(inProgressIDs, id)
- case string(agentsession.TodoStatusBlocked):
- blockedIDs = append(blockedIDs, id)
- }
- if len(todo.Artifacts) > 0 {
- artifactsByID[id] = append([]string(nil), todo.Artifacts...)
- }
- if len(todo.ContentChecks) > 0 {
- checksByID[id] = buildVerifyTodoContentChecks(todo.ContentChecks)
- }
- }
-
- if len(pendingIDs) == 0 && len(inProgressIDs) == 0 && len(blockedIDs) == 0 {
- return nil
- }
- slices.Sort(pendingIDs)
- slices.Sort(inProgressIDs)
- slices.Sort(blockedIDs)
-
- return &verify.VerificationResult{
- Name: "todo_convergence",
- Status: verify.VerificationSoftBlock,
- Summary: "required todos are not converged",
- Reason: "required todos are still pending, in progress, or blocked",
- Evidence: map[string]any{
- "pending_ids": pendingIDs,
- "in_progress_ids": inProgressIDs,
- "blocked_ids": blockedIDs,
- "todo_statuses": statusByID,
- "todo_artifacts": artifactsByID,
- "todo_checks": checksByID,
- },
- }
-}
-
-// buildAcceptanceContinueHint 构造带 verifier 证据的 continue 提示，强制下一轮先补工具事实再尝试 final。
-func buildAcceptanceContinueHint(decision acceptance.AcceptanceDecision) string {
- const actionDirective = "Do not claim completion with plain text. Next turn MUST call todo_write and/or verification tools to add objective facts before any final response."
- blockedReason := strings.TrimSpace(decision.CompletionBlockedReason)
- if len(decision.VerifierResults) == 0 && blockedReason == "" {
- if base := strings.TrimSpace(decision.ContinueHint); base != "" {
- return strings.TrimSpace(base + "\n" + actionDirective)
- }
- return strings.TrimSpace(finalContinueReminder + "\n" + actionDirective)
- }
-
- var builder strings.Builder
- builder.WriteString("\n")
- if blockedReason != "" {
- builder.WriteString(fmt.Sprintf("%s\n", xmlEscape(blockedReason)))
- }
- builder.WriteString("")
- builder.WriteString(actionDirective)
- builder.WriteString("\n")
-
- if section := renderCompletionBlockedReasonHintSection(blockedReason, decision.VerifierResults); section != "" {
- builder.WriteString(section)
- }
- if section := renderTodoConvergenceHintSection(decision.VerifierResults); section != "" {
- builder.WriteString(section)
- }
- if section := renderVerifierFailureHintSection(decision.VerifierResults); section != "" {
- builder.WriteString(section)
- }
- builder.WriteString("")
- return strings.TrimSpace(builder.String())
-}
-
-// renderCompletionBlockedReasonHintSection 根据 completion gate 阻塞原因输出结构化执行指令。
-func renderCompletionBlockedReasonHintSection(
- blockedReason string,
- results []verify.VerificationResult,
-) string {
- switch strings.TrimSpace(blockedReason) {
- case string(controlplane.CompletionBlockedReasonPendingTodo):
- pending := extractPendingTodoIDs(results)
- if len(pending) == 0 {
- return "Use todo_write to move required todos to terminal states, then retry acceptance.\n"
- }
- return fmt.Sprintf(
- "%sUse todo_write to close these required todos before final response.\n",
- strings.Join(pending, ","),
- )
- case string(controlplane.CompletionBlockedReasonUnverifiedWrite):
- return "Produce VerificationPerformed and VerificationPassed facts via verification tools before final response.\n"
- case string(controlplane.CompletionBlockedReasonPostExecuteClosureRequired):
- return "First close loop from latest tool results (todo updates/artifact checks), then retry final acceptance.\n"
- default:
- return ""
- }
-}
-
-// extractPendingTodoIDs 从 verifier 证据提取 required 未收敛 todo 列表。
-func extractPendingTodoIDs(results []verify.VerificationResult) []string {
- for _, result := range results {
- if strings.TrimSpace(result.Name) != "todo_convergence" {
- continue
- }
- evidence := result.Evidence
- if len(evidence) == 0 {
- return nil
- }
- ids := append([]string{}, evidenceStringList(evidence["pending_ids"])...)
- ids = append(ids, evidenceStringList(evidence["in_progress_ids"])...)
- ids = append(ids, evidenceStringList(evidence["blocked_ids"])...)
- return normalizeEvidenceList(ids)
- }
- return nil
-}
-
-// renderTodoConvergenceHintSection 渲染 todo_convergence 证据，明确 pending/in_progress/blocked 清单。
-func renderTodoConvergenceHintSection(results []verify.VerificationResult) string {
- for _, result := range results {
- if strings.TrimSpace(result.Name) != "todo_convergence" {
- continue
- }
- evidence := result.Evidence
- if len(evidence) == 0 {
- return ""
- }
- pending := evidenceStringList(evidence["pending_ids"])
- inProgress := evidenceStringList(evidence["in_progress_ids"])
- blocked := evidenceStringList(evidence["blocked_ids"])
- waitingExternal := evidenceStringList(evidence["waiting_external_ids"])
- statuses := evidenceJSONPreview(evidence["todo_statuses"])
- artifacts := evidenceJSONPreview(evidence["todo_artifacts"])
- checks := evidenceJSONPreview(evidence["todo_checks"])
-
- var builder strings.Builder
- builder.WriteString("\n")
- builder.WriteString(fmt.Sprintf("%s\n", strings.Join(pending, ",")))
- builder.WriteString(fmt.Sprintf("%s\n", strings.Join(inProgress, ",")))
- builder.WriteString(fmt.Sprintf("%s\n", strings.Join(blocked, ",")))
- if len(waitingExternal) > 0 {
- builder.WriteString(fmt.Sprintf("%s\n", strings.Join(waitingExternal, ",")))
- }
- if statuses != "" {
- builder.WriteString(fmt.Sprintf("%s\n", xmlEscape(statuses)))
- }
- if artifacts != "" {
- builder.WriteString(fmt.Sprintf("%s\n", xmlEscape(artifacts)))
- }
- if checks != "" {
- builder.WriteString(fmt.Sprintf("%s\n", xmlEscape(checks)))
- }
- builder.WriteString("For each listed todo, use todo_write status transitions and attach artifacts/check facts via tools. Do not finalize yet.\n")
- builder.WriteString("\n")
- return builder.String()
- }
- return ""
-}
-
-// renderVerifierFailureHintSection 渲染非通过 verifier 的摘要，避免 continue 只有泛化提醒。
-func renderVerifierFailureHintSection(results []verify.VerificationResult) string {
- nonPass := make([]verify.VerificationResult, 0, len(results))
- for _, result := range results {
- if result.Status == verify.VerificationPass {
- continue
- }
- nonPass = append(nonPass, result)
- }
- if len(nonPass) == 0 {
- return ""
- }
- sortVerificationResults(nonPass)
-
- var builder strings.Builder
- builder.WriteString("\n")
- for _, result := range nonPass {
- builder.WriteString(fmt.Sprintf(
- "%s%s\n",
- xmlEscape(strings.TrimSpace(result.Name)),
- xmlEscape(string(result.Status)),
- xmlEscape(strings.TrimSpace(result.Summary)),
- xmlEscape(strings.TrimSpace(result.Reason)),
- ))
- }
- builder.WriteString("\n")
- return builder.String()
-}
-
-// evidenceStringList 将 verifier evidence 中的字符串列表统一提取为去重、去空白后的有序值。
-func evidenceStringList(value any) []string {
- switch typed := value.(type) {
- case []string:
- return normalizeEvidenceList(typed)
- case []any:
- values := make([]string, 0, len(typed))
- for _, item := range typed {
- switch raw := item.(type) {
- case string:
- values = append(values, raw)
- default:
- if encoded, err := json.Marshal(raw); err == nil {
- values = append(values, string(encoded))
- }
- }
- }
- return normalizeEvidenceList(values)
- default:
- return nil
- }
-}
-
-// evidenceJSONPreview 将 evidence 任意结构转成紧凑 JSON 文本，便于作为提示中的可执行事实。
-func evidenceJSONPreview(value any) string {
- if value == nil {
- return ""
- }
- encoded, err := json.Marshal(value)
- if err != nil {
- return ""
- }
- return strings.TrimSpace(string(encoded))
-}
-
-// normalizeEvidenceList 对 evidence 文本列表做去重与排序，保证提示稳定可测。
-func normalizeEvidenceList(values []string) []string {
- if len(values) == 0 {
- return nil
- }
- seen := make(map[string]struct{}, len(values))
- normalized := make([]string, 0, len(values))
- for _, value := range values {
- trimmed := strings.TrimSpace(value)
- if trimmed == "" {
- continue
- }
- if _, exists := seen[trimmed]; exists {
- continue
- }
- seen[trimmed] = struct{}{}
- normalized = append(normalized, trimmed)
- }
- if len(normalized) == 0 {
- return nil
- }
- slices.Sort(normalized)
- return normalized
-}
-
-// sortVerificationResults 保证 verifier 输出顺序稳定，减少提示抖动。
-func sortVerificationResults(results []verify.VerificationResult) {
- slices.SortFunc(results, func(a verify.VerificationResult, b verify.VerificationResult) int {
- return strings.Compare(strings.TrimSpace(a.Name), strings.TrimSpace(b.Name))
- })
-}
-
-// xmlEscape 对可见提示中的 verifier 文本做最小转义，避免破坏 XML 结构。
-func xmlEscape(value string) string {
- replacer := strings.NewReplacer(
- "&", "&",
- "<", "<",
- ">", ">",
- "\"", """,
- "'", "'",
- )
- return replacer.Replace(value)
-}
-
-// recordAcceptanceTerminal 将 acceptance 输出映射为 runtime 唯一终态记录。
-func recordAcceptanceTerminal(state *runState, decision acceptance.AcceptanceDecision) {
- if state == nil {
- return
- }
- status := acceptance.TerminalStatusFromAcceptance(decision.Status)
- state.markTerminalDecision(status, decision.StopReason, strings.TrimSpace(decision.InternalSummary))
-}
-
-// buildVerifyTodos 将 session todo 转换为 verifier 快照。
-func buildVerifyTodos(items []agentsession.TodoItem) []verify.TodoSnapshot {
- if len(items) == 0 {
- return nil
- }
- todos := make([]verify.TodoSnapshot, 0, len(items))
- for _, item := range items {
- todos = append(todos, verify.TodoSnapshot{
- ID: strings.TrimSpace(item.ID),
- Content: strings.TrimSpace(item.Content),
- Status: strings.TrimSpace(string(item.Status)),
- Required: item.RequiredValue(),
- BlockedReason: strings.TrimSpace(string(item.BlockedReason)),
- Acceptance: append([]string(nil), item.Acceptance...),
- Artifacts: append([]string(nil), item.Artifacts...),
- Supersedes: append([]string(nil), item.Supersedes...),
- ContentChecks: buildVerifyTodoContentChecks(item.ContentChecks),
- RetryCount: item.RetryCount,
- RetryLimit: item.RetryLimit,
- FailureReason: strings.TrimSpace(item.FailureReason),
- })
- }
- return todos
-}
-
-// buildVerifyTodoContentChecks 将 session 内容校验规则转换为 verifier 快照。
-func buildVerifyTodoContentChecks(items []agentsession.TodoContentCheck) []verify.TodoContentCheckSnapshot {
- if len(items) == 0 {
- return nil
- }
- checks := make([]verify.TodoContentCheckSnapshot, 0, len(items))
- for _, item := range items {
- checks = append(checks, verify.TodoContentCheckSnapshot{
- Artifact: strings.TrimSpace(item.Artifact),
- Contains: append([]string(nil), item.Contains...),
- })
- }
- return checks
-}
-
-// buildVerifyTaskState 将 task_state 中与验收相关的结构化字段投影给 verifier。
-func buildVerifyTaskState(state agentsession.TaskState) verify.TaskStateSnapshot {
- return verify.TaskStateSnapshot{
- VerificationProfile: string(state.VerificationProfile),
- KeyArtifacts: append([]string(nil), state.KeyArtifacts...),
- }
-}
-
-// buildVerifyMessages 将会话消息压缩为 verifier 所需的最小快照。
-func buildVerifyMessages(messages []providertypes.Message) []verify.MessageLike {
- if len(messages) == 0 {
- return nil
- }
- out := make([]verify.MessageLike, 0, len(messages))
- for _, message := range messages {
- out = append(out, verify.MessageLike{
- Role: strings.TrimSpace(message.Role),
- Content: renderPartsForVerification(message.Parts),
- })
- }
- return out
-}
-
-// renderPartsForVerification 将消息分片合并为 verifier 侧可读文本。
-func renderPartsForVerification(parts []providertypes.ContentPart) string {
- if len(parts) == 0 {
- return ""
- }
- segments := make([]string, 0, len(parts))
- for _, part := range parts {
- if part.Kind != providertypes.ContentPartText {
- continue
- }
- text := strings.TrimSpace(part.Text)
- if text == "" {
- continue
- }
- segments = append(segments, text)
- }
- return strings.Join(segments, "\n")
-}
-
-// applyAcceptanceResultProgress 根据 acceptance 输出更新 final 拦截计数唯一真相源。
-func applyAcceptanceResultProgress(state *runState, decision acceptance.AcceptanceDecision) {
- if state == nil {
- return
- }
- switch decision.Status {
- case acceptance.AcceptanceContinue:
- if state.pendingFinalProgress {
- state.finalInterceptStreak = 0
- } else {
- state.finalInterceptStreak++
- }
- default:
- state.finalInterceptStreak = 0
- }
- state.pendingFinalProgress = false
-}
diff --git a/internal/runtime/final_acceptance_additional_test.go b/internal/runtime/final_acceptance_additional_test.go
deleted file mode 100644
index b706002b..00000000
--- a/internal/runtime/final_acceptance_additional_test.go
+++ /dev/null
@@ -1,107 +0,0 @@
-package runtime
-
-import (
- "context"
- "strings"
- "testing"
-
- "neo-code/internal/config"
- providertypes "neo-code/internal/provider/types"
- "neo-code/internal/runtime/acceptance"
- "neo-code/internal/runtime/controlplane"
- "neo-code/internal/runtime/decider"
- "neo-code/internal/runtime/verify"
- agentsession "neo-code/internal/session"
-)
-
-func TestFinalAcceptanceMappingAndLegacyPaths(t *testing.T) {
- t.Parallel()
-
- t.Run("map decider statuses", func(t *testing.T) {
- t.Parallel()
- got := mapDeciderDecisionToAcceptance(decider.Decision{Status: decider.DecisionBlocked})
- if got.Status != acceptance.AcceptanceFailed || got.StopReason != controlplane.StopReasonVerificationFailed {
- t.Fatalf("blocked mapping = %+v", got)
- }
- got = mapDeciderDecisionToAcceptance(decider.Decision{Status: decider.DecisionContinue})
- if got.Status != acceptance.AcceptanceContinue || got.StopReason != controlplane.StopReasonTodoNotConverged {
- t.Fatalf("continue mapping = %+v", got)
- }
- })
-
- t.Run("projection keeps required input and task kinds", func(t *testing.T) {
- t.Parallel()
- required := &decider.RequiredInput{
- Kind: "missing_file_target_or_content",
- Message: "need target path",
- Details: map[string]any{"path": "test.txt"},
- }
- projected := toDeciderDecisionFromAcceptance(acceptance.AcceptanceDecision{
- Status: acceptance.AcceptanceContinue,
- StopReason: controlplane.StopReasonTodoNotConverged,
- RequiredInput: required,
- IntentHint: decider.TaskKindWorkspaceWrite,
- EffectiveTaskKind: decider.TaskKindWorkspaceWrite,
- })
- if projected.RequiredInput == nil || projected.RequiredInput.Kind != "missing_file_target_or_content" {
- t.Fatalf("required input lost in projection: %+v", projected)
- }
- if projected.IntentHint != decider.TaskKindWorkspaceWrite || projected.EffectiveTaskKind != decider.TaskKindWorkspaceWrite {
- t.Fatalf("task kind hints lost in projection: %+v", projected)
- }
- })
-
- t.Run("legacy path adds continue hint", func(t *testing.T) {
- t.Parallel()
- service := &Service{}
- state := newRunState("run-legacy", agentsession.New("legacy"))
- required := true
- state.session.Todos = []agentsession.TodoItem{
- {ID: "todo-1", Status: agentsession.TodoStatusPending, Required: &required},
- }
- snapshot := TurnBudgetSnapshot{Config: config.StaticDefaults().Clone(), Workdir: t.TempDir()}
- decision, err := service.beforeAcceptFinalLegacy(context.Background(), &state, snapshot, providertypes.Message{}, false)
- if err != nil {
- t.Fatalf("beforeAcceptFinalLegacy() error = %v", err)
- }
- if decision.Status != acceptance.AcceptanceContinue {
- t.Fatalf("legacy status = %q, want continue", decision.Status)
- }
- if !strings.Contains(decision.ContinueHint, "") {
- t.Fatalf("continue hint = %q", decision.ContinueHint)
- }
- })
-}
-
-func TestFinalAcceptanceHelperBranches(t *testing.T) {
- t.Parallel()
-
- if got := buildAcceptanceContinueHint(acceptance.AcceptanceDecision{
- Status: acceptance.AcceptanceContinue,
- ContinueHint: "base",
- VerifierResults: nil,
- }); !strings.Contains(got, "base") {
- t.Fatalf("continue hint fallback = %q", got)
- }
-
- if got := renderCompletionBlockedReasonHintSection("pending_todo", nil); !strings.Contains(got, "required_action") {
- t.Fatalf("pending_todo fallback hint = %q", got)
- }
- if got := renderCompletionBlockedReasonHintSection("unverified_write", nil); !strings.Contains(got, "VerificationPerformed") {
- t.Fatalf("unverified_write hint = %q", got)
- }
-
- results := []verify.VerificationResult{
- {Name: "z", Status: verify.VerificationSoftBlock},
- {Name: "a", Status: verify.VerificationHardBlock},
- {Name: "ok", Status: verify.VerificationPass},
- }
- section := renderVerifierFailureHintSection(results)
- if !strings.Contains(section, "name=\"a\"") || !strings.Contains(section, "name=\"z\"") {
- t.Fatalf("verifier section = %q", section)
- }
-
- if xmlEscape(``) == `` {
- t.Fatal("xmlEscape should escape special chars")
- }
-}
diff --git a/internal/runtime/final_acceptance_test.go b/internal/runtime/final_acceptance_test.go
deleted file mode 100644
index e08b1096..00000000
--- a/internal/runtime/final_acceptance_test.go
+++ /dev/null
@@ -1,273 +0,0 @@
-package runtime
-
-import (
- "context"
- "encoding/json"
- "strings"
- "testing"
-
- "neo-code/internal/config"
- providertypes "neo-code/internal/provider/types"
- "neo-code/internal/runtime/acceptance"
- "neo-code/internal/runtime/verify"
- agentsession "neo-code/internal/session"
-)
-
-func TestBeforeAcceptFinalDecisionPaths(t *testing.T) {
- t.Parallel()
-
- service := &Service{}
- baseCfg := config.StaticDefaults().Clone()
- snapshot := TurnBudgetSnapshot{
- Config: baseCfg,
- Workdir: t.TempDir(),
- }
-
- t.Run("pending required todo -> continue", func(t *testing.T) {
- t.Parallel()
- state := newRunState("run-continue", agentsession.New("continue"))
- required := true
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- state.session.Todos = []agentsession.TodoItem{
- {ID: "todo-1", Content: "do work", Status: agentsession.TodoStatusPending, Required: &required},
- }
- decision, err := service.beforeAcceptFinal(context.Background(), &state, snapshot, providertypes.Message{
- Role: providertypes.RoleAssistant,
- Parts: []providertypes.ContentPart{providertypes.NewTextPart("done")},
- }, true, beforeCompletionHookSignals{})
- if err != nil {
- t.Fatalf("beforeAcceptFinal() error = %v", err)
- }
- if decision.Status != acceptance.AcceptanceContinue {
- t.Fatalf("status = %q, want continue", decision.Status)
- }
- })
-
- t.Run("invalid profile -> failed", func(t *testing.T) {
- t.Parallel()
- state := newRunState("run-invalid-profile", agentsession.New("invalid-profile"))
- state.session.TaskState.VerificationProfile = "bad"
- decision, err := service.beforeAcceptFinal(context.Background(), &state, snapshot, providertypes.Message{}, true, beforeCompletionHookSignals{})
- if err != nil {
- t.Fatalf("beforeAcceptFinal() error = %v", err)
- }
- if decision.Status != acceptance.AcceptanceFailed {
- t.Fatalf("status = %q, want failed", decision.Status)
- }
- })
-
- t.Run("continue carries pending final progress signal", func(t *testing.T) {
- t.Parallel()
- state := newRunState("run-progress", agentsession.New("progress"))
- required := true
- state.pendingFinalProgress = true
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- state.session.Todos = []agentsession.TodoItem{
- {ID: "todo-1", Content: "do work", Status: agentsession.TodoStatusPending, Required: &required},
- }
- decision, err := service.beforeAcceptFinal(context.Background(), &state, snapshot, providertypes.Message{}, true, beforeCompletionHookSignals{})
- if err != nil {
- t.Fatalf("beforeAcceptFinal() error = %v", err)
- }
- if !decision.HasProgress {
- t.Fatal("expected continue decision to carry pending final progress")
- }
- })
-
- t.Run("all converged -> accepted", func(t *testing.T) {
- t.Parallel()
- state := newRunState("run-accepted", agentsession.New("accepted"))
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- decision, err := service.beforeAcceptFinal(context.Background(), &state, snapshot, providertypes.Message{}, true, beforeCompletionHookSignals{})
- if err != nil {
- t.Fatalf("beforeAcceptFinal() error = %v", err)
- }
- if decision.Status != acceptance.AcceptanceAccepted {
- t.Fatalf("status = %q, want accepted", decision.Status)
- }
- })
-
- t.Run("final intercept streak drives no-progress breaker", func(t *testing.T) {
- t.Parallel()
- state := newRunState("run-incomplete", agentsession.New("incomplete"))
- required := true
- state.finalInterceptStreak = snapshot.Config.Runtime.Verification.MaxNoProgress
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- state.session.Todos = []agentsession.TodoItem{
- {ID: "todo-1", Content: "do work", Status: agentsession.TodoStatusPending, Required: &required},
- }
- decision, err := service.beforeAcceptFinal(context.Background(), &state, snapshot, providertypes.Message{}, true, beforeCompletionHookSignals{})
- if err != nil {
- t.Fatalf("beforeAcceptFinal() error = %v", err)
- }
- if decision.Status != acceptance.AcceptanceIncomplete {
- t.Fatalf("status = %q, want incomplete", decision.Status)
- }
- })
-}
-
-func TestFinalAcceptanceHelpers(t *testing.T) {
- t.Parallel()
-
- t.Run("buildVerifyTaskState includes profile", func(t *testing.T) {
- t.Parallel()
- got := buildVerifyTaskState(agentsession.TaskState{
- VerificationProfile: agentsession.VerificationProfileDocs,
- KeyArtifacts: []string{"README.md"},
- })
- if got.VerificationProfile != "docs" || len(got.KeyArtifacts) != 1 {
- t.Fatalf("unexpected task state snapshot: %+v", got)
- }
- })
-
- t.Run("applyAcceptanceResultProgress uses pending final progress", func(t *testing.T) {
- t.Parallel()
- state := newRunState("run-progress", agentsession.New("progress"))
- state.finalInterceptStreak = 2
- state.pendingFinalProgress = true
- applyAcceptanceResultProgress(&state, acceptance.AcceptanceDecision{Status: acceptance.AcceptanceContinue})
- if state.finalInterceptStreak != 0 || state.pendingFinalProgress {
- t.Fatalf("unexpected state after progress reset: streak=%d, pending=%v", state.finalInterceptStreak, state.pendingFinalProgress)
- }
-
- applyAcceptanceResultProgress(&state, acceptance.AcceptanceDecision{Status: acceptance.AcceptanceContinue})
- if state.finalInterceptStreak != 1 {
- t.Fatalf("streak = %d, want 1", state.finalInterceptStreak)
- }
- })
-
- t.Run("buildAcceptanceContinueHint includes actionable evidence and tool requirement", func(t *testing.T) {
- t.Parallel()
- decision := acceptance.AcceptanceDecision{
- Status: acceptance.AcceptanceContinue,
- CompletionBlockedReason: "pending_todo",
- VerifierResults: []verify.VerificationResult{
- {
- Name: "todo_convergence",
- Status: verify.VerificationSoftBlock,
- Summary: "required todos are not converged",
- Reason: "required todos are still pending, in progress, or internally blocked",
- Evidence: map[string]any{
- "pending_ids": []string{"todo-2", "todo-1"},
- "in_progress_ids": []string{"todo-3"},
- "blocked_ids": []string{"todo-4"},
- },
- },
- },
- }
- hint := buildAcceptanceContinueHint(decision)
- if !strings.Contains(hint, "") {
- t.Fatalf("hint should contain acceptance xml envelope, got %q", hint)
- }
- if !strings.Contains(hint, "MUST call todo_write") {
- t.Fatalf("hint should force tool-based facts, got %q", hint)
- }
- if !strings.Contains(hint, "todo-1,todo-2") {
- t.Fatalf("hint should include sorted pending ids, got %q", hint)
- }
- if !strings.Contains(hint, "pending_todo") {
- t.Fatalf("hint should include completion blocked reason, got %q", hint)
- }
- })
-
- t.Run("buildAcceptanceContinueHint emits unverified_write guidance", func(t *testing.T) {
- t.Parallel()
- hint := buildAcceptanceContinueHint(acceptance.AcceptanceDecision{
- Status: acceptance.AcceptanceContinue,
- CompletionBlockedReason: "unverified_write",
- })
- if !strings.Contains(hint, "unverified_write") {
- t.Fatalf("hint should include unverified_write reason, got %q", hint)
- }
- if !strings.Contains(hint, "VerificationPerformed") || !strings.Contains(hint, "VerificationPassed") {
- t.Fatalf("hint should require verification facts, got %q", hint)
- }
- })
-
- t.Run("synthesizeTodoConvergenceEvidence projects required todos", func(t *testing.T) {
- t.Parallel()
- required := true
- result := synthesizeTodoConvergenceEvidence([]agentsession.TodoItem{
- {ID: "todo-1", Content: "a", Status: agentsession.TodoStatusPending, Required: &required},
- {ID: "todo-2", Content: "b", Status: agentsession.TodoStatusInProgress, Required: &required},
- {ID: "todo-3", Content: "c", Status: agentsession.TodoStatusCompleted, Required: &required},
- })
- if result == nil {
- t.Fatal("expected synthetic verifier result")
- }
- if result.Name != "todo_convergence" || result.Status != verify.VerificationSoftBlock {
- t.Fatalf("unexpected synthetic result: %+v", *result)
- }
- pending, _ := result.Evidence["pending_ids"].([]string)
- if len(pending) != 1 || pending[0] != "todo-1" {
- t.Fatalf("pending ids = %+v, want [todo-1]", pending)
- }
- })
-
- t.Run("buildVerifyTodos and content checks keep normalized values", func(t *testing.T) {
- t.Parallel()
- required := true
- todos := buildVerifyTodos([]agentsession.TodoItem{
- {
- ID: " todo-1 ",
- Content: " do work ",
- Status: agentsession.TodoStatusInProgress,
- Required: &required,
- Acceptance: []string{"a"},
- Artifacts: []string{"x.md"},
- Supersedes: []string{"todo-0"},
- ContentChecks: []agentsession.TodoContentCheck{
- {Artifact: " README.md ", Contains: []string{"done"}},
- },
- RetryCount: 1,
- RetryLimit: 2,
- FailureReason: " none ",
- },
- })
- if len(todos) != 1 {
- t.Fatalf("todos len = %d, want 1", len(todos))
- }
- if todos[0].ID != "todo-1" || todos[0].Content != "do work" || todos[0].Status != "in_progress" {
- t.Fatalf("unexpected todo snapshot: %+v", todos[0])
- }
- if len(todos[0].ContentChecks) != 1 || todos[0].ContentChecks[0].Artifact != "README.md" {
- t.Fatalf("unexpected content checks: %+v", todos[0].ContentChecks)
- }
- })
-
- t.Run("buildVerifyMessages ignores non-text and trims content", func(t *testing.T) {
- t.Parallel()
- messages := buildVerifyMessages([]providertypes.Message{
- {
- Role: " assistant ",
- Parts: []providertypes.ContentPart{
- providertypes.NewTextPart(" first "),
- {Kind: "tool_call", Text: "ignored"},
- providertypes.NewTextPart("second"),
- },
- },
- })
- if len(messages) != 1 {
- t.Fatalf("messages len = %d, want 1", len(messages))
- }
- if messages[0].Role != "assistant" || messages[0].Content != "first\nsecond" {
- t.Fatalf("unexpected message snapshot: %+v", messages[0])
- }
- })
-
- t.Run("evidence helpers normalize and serialize", func(t *testing.T) {
- t.Parallel()
- items := evidenceStringList([]any{" b ", map[string]any{"k": 1}, "a", "a"})
- if len(items) != 3 || items[0] != "a" || items[1] != "b" {
- t.Fatalf("unexpected evidence list: %+v", items)
- }
- var m map[string]any
- if err := json.Unmarshal([]byte(items[2]), &m); err != nil {
- t.Fatalf("expected JSON encoded item, got %q", items[2])
- }
- preview := evidenceJSONPreview(map[string]any{"x": 1})
- if preview == "" || !strings.Contains(preview, "\"x\":1") {
- t.Fatalf("unexpected preview: %q", preview)
- }
- })
-}
diff --git a/internal/runtime/hooks_integration_test.go b/internal/runtime/hooks_integration_test.go
index 3b98eba4..aa820bd4 100644
--- a/internal/runtime/hooks_integration_test.go
+++ b/internal/runtime/hooks_integration_test.go
@@ -15,7 +15,6 @@ import (
approvalflow "neo-code/internal/runtime/approval"
"neo-code/internal/runtime/controlplane"
runtimehooks "neo-code/internal/runtime/hooks"
- agentsession "neo-code/internal/session"
"neo-code/internal/subagent"
"neo-code/internal/tools"
)
@@ -266,103 +265,9 @@ func TestRunBeforeCompletionDecisionHookBlockIsObservedOnly(t *testing.T) {
if eventIndex(events, EventHookBlocked) >= 0 {
t.Fatalf("before_completion_decision should not emit hook_blocked when point is observe-only")
}
- assertEventContains(t, events, EventHookFinished)
assertEventContains(t, events, EventAgentDone)
- if eventIndex(events, EventHookFinished) > eventIndex(events, EventVerificationStarted) {
- t.Fatalf("before_completion_decision hook should finish before verification_started")
- }
- if capturedWorkdir == "" {
- t.Fatalf("expected before_completion_decision hook metadata to include workdir")
- }
-}
-
-func TestBeforeCompletionDecisionOrchestratorRunsUserRepoBeforeInternalAndFeedsDecision(t *testing.T) {
- t.Parallel()
-
- service := &Service{events: make(chan RuntimeEvent, 16)}
-
- var (
- mu sync.Mutex
- callFlow []string
- )
- appendCall := func(value string) {
- mu.Lock()
- callFlow = append(callFlow, value)
- mu.Unlock()
- }
-
- baseRegistry := runtimehooks.NewRegistry()
- if err := baseRegistry.Register(runtimehooks.HookSpec{
- ID: "internal-before-completion",
- Point: runtimehooks.HookPointBeforeCompletionDecision,
- Scope: runtimehooks.HookScopeInternal,
- Source: runtimehooks.HookSourceInternal,
- Handler: func(_ context.Context, _ runtimehooks.HookContext) runtimehooks.HookResult {
- appendCall("internal")
- return runtimehooks.HookResult{Status: runtimehooks.HookResultPass}
- },
- }); err != nil {
- t.Fatalf("register internal hook: %v", err)
- }
-
- userRegistry := runtimehooks.NewRegistry()
- if err := userRegistry.Register(runtimehooks.HookSpec{
- ID: "user-before-completion",
- Point: runtimehooks.HookPointBeforeCompletionDecision,
- Scope: runtimehooks.HookScopeUser,
- Source: runtimehooks.HookSourceUser,
- Handler: func(_ context.Context, _ runtimehooks.HookContext) runtimehooks.HookResult {
- appendCall("user")
- return runtimehooks.HookResult{Status: runtimehooks.HookResultFailed, Message: "user guard signal"}
- },
- }); err != nil {
- t.Fatalf("register user hook: %v", err)
- }
-
- repoRegistry := runtimehooks.NewRegistry()
- if err := repoRegistry.Register(runtimehooks.HookSpec{
- ID: "repo-before-completion",
- Point: runtimehooks.HookPointBeforeCompletionDecision,
- Scope: runtimehooks.HookScopeRepo,
- Source: runtimehooks.HookSourceRepo,
- Handler: func(_ context.Context, _ runtimehooks.HookContext) runtimehooks.HookResult {
- appendCall("repo")
- return runtimehooks.HookResult{Status: runtimehooks.HookResultPass, Message: "repo annotation"}
- },
- }); err != nil {
- t.Fatalf("register repo hook: %v", err)
- }
-
- baseExecutor := runtimehooks.NewExecutor(baseRegistry, newHookRuntimeEventEmitter(service), time.Second)
- userExecutor := runtimehooks.NewExecutor(userRegistry, newHookRuntimeEventEmitter(service), time.Second)
- repoExecutor := runtimehooks.NewExecutor(repoRegistry, newHookRuntimeEventEmitter(service), time.Second)
- service.SetHookExecutor(composeRuntimeHookExecutors(baseExecutor, userExecutor, repoExecutor))
-
- session := newRuntimeSession("session-before-completion-orchestrator")
- state := newRunState("run-before-completion-orchestrator", session)
- snapshotCfg := TurnBudgetSnapshot{
- Config: config.StaticDefaults().Clone(),
- Workdir: t.TempDir(),
- }
- state.session.TaskState.VerificationProfile = agentsession.VerificationProfileTaskOnly
- decision, err := service.runBeforeCompletionDecisionAcceptance(
- context.Background(),
- &state,
- snapshotCfg,
- providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("done")}},
- t.TempDir(),
- true,
- false,
- providertypes.RoleAssistant,
- )
- if err != nil {
- t.Fatalf("runBeforeCompletionDecisionAcceptance() error = %v", err)
- }
- if got := strings.Join(callFlow, ","); got != "user,repo,internal" {
- t.Fatalf("before_completion_decision hook order = %q, want %q", got, "user,repo,internal")
- }
- if !strings.Contains(decision.InternalSummary, "hook signals consumed") {
- t.Fatalf("decision internal summary should include hook signal context, got %q", decision.InternalSummary)
+ if capturedWorkdir != "" {
+ t.Fatalf("before_completion_decision hook should not run as an authoritative terminal gate")
}
}
diff --git a/internal/runtime/permission.go b/internal/runtime/permission.go
index b61bb253..79e8e3dd 100644
--- a/internal/runtime/permission.go
+++ b/internal/runtime/permission.go
@@ -57,8 +57,10 @@ const (
minInlineSubAgentToolTimeout = 30 * time.Second
defaultDiagnoseToolTimeout = 60 * time.Second
defaultPermissionToolTimeout = 20 * time.Second
+ defaultWorkspaceScanToolTimeout = 60 * time.Second
defaultAskUserToolTimeout = 5 * time.Minute
maxAskUserToolTimeout = time.Hour
+ maxAdaptiveToolTimeout = 160 * time.Second
)
// permissionExecutionInput 汇总一次工具执行与审批协作所需的上下文。
@@ -168,24 +170,28 @@ func (s *Service) executeToolCallWithPermission(ctx context.Context, input permi
}
callInput.SubAgentInvoker = newRuntimeSubAgentInvoker(s, input.RunID, input.SessionID, input.AgentID, input.Workdir)
- effectiveTimeout := resolveToolExecutionTimeout(input.Call, input.ToolTimeout)
+ baseTimeout := resolveToolExecutionTimeout(input.Call, input.ToolTimeout)
+ effectiveTimeout := resolveAdaptiveToolExecutionTimeout(input.State, input.Call, baseTimeout)
runCtx, cancel := context.WithTimeout(ctx, effectiveTimeout)
defer cancel()
if s.runnerToolDispatcher != nil {
result, handled, dispatchErr := s.runnerToolDispatcher.TryDispatch(runCtx, input.SessionID, input.RunID, callInput)
if handled {
+ recordAdaptiveToolTimeoutResult(input.State, input.Call, result, dispatchErr)
return result, dispatchErr
}
}
result, execErr := s.toolManager.Execute(runCtx, callInput)
if execErr == nil {
+ recordAdaptiveToolTimeoutResult(input.State, input.Call, result, nil)
return result, nil
}
var permissionErr *tools.PermissionDecisionError
if !errors.As(execErr, &permissionErr) {
+ recordAdaptiveToolTimeoutResult(input.State, input.Call, result, execErr)
return result, execErr
}
@@ -220,6 +226,7 @@ func (s *Service) executeToolCallWithPermission(ctx context.Context, input permi
Reason: reason,
Enforced: true,
})
+ recordAdaptiveToolTimeoutResult(input.State, input.Call, blockedResult, errors.New(reason))
return blockedResult, errors.New(reason)
}
@@ -233,6 +240,7 @@ func (s *Service) executeToolCallWithPermission(ctx context.Context, input permi
permissionResolutionStatus(permissionErr.Decision()),
permissionErr.RememberScope(),
)
+ recordAdaptiveToolTimeoutResult(input.State, input.Call, result, execErr)
return result, execErr
}
@@ -296,6 +304,7 @@ func (s *Service) executeToolCallWithPermission(ctx context.Context, input permi
retryCtx, retryCancel := context.WithTimeout(ctx, effectiveTimeout)
retryResult, retryErr := s.toolManager.Execute(retryCtx, callInput)
retryCancel()
+ recordAdaptiveToolTimeoutResult(input.State, input.Call, retryResult, retryErr)
return retryResult, retryErr
}
@@ -312,6 +321,12 @@ func resolveToolExecutionTimeout(call providertypes.ToolCall, fallback time.Dura
}
return base
}
+ if isWorkspaceScanTool(name) {
+ if base < defaultWorkspaceScanToolTimeout {
+ return defaultWorkspaceScanToolTimeout
+ }
+ return base
+ }
if strings.EqualFold(name, tools.ToolNameAskUser) {
requested := parseAskUserTimeoutFromArguments(call.Arguments)
if requested <= 0 {
@@ -341,6 +356,136 @@ func resolveToolExecutionTimeout(call providertypes.ToolCall, fallback time.Dura
return base
}
+// isWorkspaceScanTool 识别会遍历工作区的搜索工具，用于给首轮执行预留更合理的时间。
+func isWorkspaceScanTool(name string) bool {
+ return strings.EqualFold(name, tools.ToolNameCodebaseSearchText) ||
+ strings.EqualFold(name, tools.ToolNameCodebaseSearchSymbol) ||
+ strings.EqualFold(name, tools.ToolNameFilesystemGrep) ||
+ strings.EqualFold(name, tools.ToolNameFilesystemGlob)
+}
+
+// resolveAdaptiveToolExecutionTimeout 根据同一 Run 内同签名工具的 timeout 次数指数放大超时。
+func resolveAdaptiveToolExecutionTimeout(state *runState, call providertypes.ToolCall, base time.Duration) time.Duration {
+ if state == nil || !supportsAdaptiveToolTimeout(call.Name) {
+ return base
+ }
+ key := toolTimeoutBackoffKey(call)
+ if key == "" {
+ return base
+ }
+ state.mu.Lock()
+ attempts := state.toolTimeoutBackoff[key]
+ state.mu.Unlock()
+ timeout := base
+ for attempts > 0 && timeout < maxAdaptiveToolTimeout {
+ timeout *= 2
+ if timeout > maxAdaptiveToolTimeout {
+ timeout = maxAdaptiveToolTimeout
+ }
+ attempts--
+ }
+ return timeout
+}
+
+// recordAdaptiveToolTimeoutResult 记录工具 timeout 结果；成功或非 timeout 错误会清除该签名的倍增状态。
+func recordAdaptiveToolTimeoutResult(state *runState, call providertypes.ToolCall, result tools.ToolResult, err error) {
+ if state == nil || !supportsAdaptiveToolTimeout(call.Name) {
+ return
+ }
+ key := toolTimeoutBackoffKey(call)
+ if key == "" {
+ return
+ }
+ state.mu.Lock()
+ defer state.mu.Unlock()
+ if toolExecutionTimedOut(result, err) {
+ if state.toolTimeoutBackoff == nil {
+ state.toolTimeoutBackoff = make(map[string]int)
+ }
+ state.toolTimeoutBackoff[key]++
+ return
+ }
+ delete(state.toolTimeoutBackoff, key)
+}
+
+// supportsAdaptiveToolTimeout 仅对普通工具调用启用倍增，避免覆盖交互/子代理等自带超时语义。
+func supportsAdaptiveToolTimeout(name string) bool {
+ normalized := strings.TrimSpace(name)
+ if normalized == "" {
+ return false
+ }
+ switch {
+ case strings.EqualFold(normalized, tools.ToolNameAskUser),
+ strings.EqualFold(normalized, tools.ToolNameSpawnSubAgent),
+ strings.EqualFold(normalized, tools.ToolNameDiagnose):
+ return false
+ default:
+ return true
+ }
+}
+
+// toolTimeoutBackoffKey 将工具名和规范化参数组合为本轮 timeout 倍增键。
+func toolTimeoutBackoffKey(call providertypes.ToolCall) string {
+ if isWorkspaceScanTool(call.Name) {
+ return workspaceScanToolTimeoutBackoffKey(call)
+ }
+ signature := computeToolSignature([]providertypes.ToolCall{call})
+ if strings.TrimSpace(signature) == "" {
+ return ""
+ }
+ return strings.ToLower(strings.TrimSpace(call.Name)) + "\x00" + signature
+}
+
+// workspaceScanToolTimeoutBackoffKey 仅按扫描工具和范围聚合 timeout，避免换关键词后丢失退避状态。
+func workspaceScanToolTimeoutBackoffKey(call providertypes.ToolCall) string {
+ name := strings.ToLower(strings.TrimSpace(call.Name))
+ if name == "" {
+ return ""
+ }
+ return name + "\x00" + workspaceScanScopeFromArguments(call.Arguments)
+}
+
+// workspaceScanScopeFromArguments 从搜索工具参数中抽取扫描范围，解析失败时回落到全工作区范围。
+func workspaceScanScopeFromArguments(raw string) string {
+ if strings.TrimSpace(raw) == "" {
+ return "."
+ }
+ var payload struct {
+ Dir string `json:"dir"`
+ ScopeDir string `json:"scope_dir"`
+ Workdir string `json:"workdir"`
+ }
+ if err := json.Unmarshal([]byte(raw), &payload); err != nil {
+ return "."
+ }
+ scope := strings.TrimSpace(payload.ScopeDir)
+ if scope == "" {
+ scope = strings.TrimSpace(payload.Dir)
+ }
+ if scope == "" {
+ scope = "."
+ }
+ workdir := strings.TrimSpace(payload.Workdir)
+ if workdir == "" {
+ return scope
+ }
+ return workdir + "/" + scope
+}
+
+// toolExecutionTimedOut 判断工具结果是否代表执行超时。
+func toolExecutionTimedOut(result tools.ToolResult, err error) bool {
+ if errors.Is(err, context.DeadlineExceeded) {
+ return true
+ }
+ if strings.EqualFold(strings.TrimSpace(result.ErrorClass), "timeout") {
+ return true
+ }
+ content := strings.ToLower(strings.TrimSpace(result.Content))
+ return strings.Contains(content, "context deadline exceeded") ||
+ strings.Contains(content, "timed out") ||
+ strings.Contains(content, "timeout")
+}
+
// parseAskUserTimeoutFromArguments 解析 ask_user 的 timeout_sec，并返回持续时间。
func parseAskUserTimeoutFromArguments(raw string) time.Duration {
if strings.TrimSpace(raw) == "" {
diff --git a/internal/runtime/permission_test.go b/internal/runtime/permission_test.go
index 7a1ae74b..269a373b 100644
--- a/internal/runtime/permission_test.go
+++ b/internal/runtime/permission_test.go
@@ -11,6 +11,7 @@ import (
providertypes "neo-code/internal/provider/types"
approvalflow "neo-code/internal/runtime/approval"
"neo-code/internal/security"
+ agentsession "neo-code/internal/session"
"neo-code/internal/tools"
"neo-code/internal/tools/mcp"
)
@@ -1332,6 +1333,139 @@ func TestResolveToolExecutionTimeoutForAskUser(t *testing.T) {
}
}
+func TestResolveToolExecutionTimeoutForWorkspaceScanTools(t *testing.T) {
+ t.Parallel()
+
+ base := 20 * time.Second
+ for _, name := range []string{
+ tools.ToolNameCodebaseSearchText,
+ tools.ToolNameCodebaseSearchSymbol,
+ tools.ToolNameFilesystemGrep,
+ tools.ToolNameFilesystemGlob,
+ } {
+ got := resolveToolExecutionTimeout(providertypes.ToolCall{
+ Name: name,
+ Arguments: `{"query":"plan mode"}`,
+ }, base)
+ if got != defaultWorkspaceScanToolTimeout {
+ t.Fatalf("%s timeout = %v, want %v", name, got, defaultWorkspaceScanToolTimeout)
+ }
+ }
+
+ largerBase := 90 * time.Second
+ got := resolveToolExecutionTimeout(providertypes.ToolCall{
+ Name: tools.ToolNameCodebaseSearchText,
+ Arguments: `{"query":"plan mode"}`,
+ }, largerBase)
+ if got != largerBase {
+ t.Fatalf("expected larger base timeout %v to win, got %v", largerBase, got)
+ }
+}
+
+func TestResolveAdaptiveToolExecutionTimeoutBackoff(t *testing.T) {
+ t.Parallel()
+
+ state := newRunState("run-timeout-backoff", agentsession.New("timeout-backoff"))
+ call := providertypes.ToolCall{
+ Name: tools.ToolNameBash,
+ Arguments: `{"command":"go test ./..."}`,
+ }
+ base := 20 * time.Second
+
+ if got := resolveAdaptiveToolExecutionTimeout(&state, call, base); got != 20*time.Second {
+ t.Fatalf("first timeout = %v, want 20s", got)
+ }
+ recordAdaptiveToolTimeoutResult(&state, call, tools.ToolResult{}, context.DeadlineExceeded)
+ if got := resolveAdaptiveToolExecutionTimeout(&state, call, base); got != 40*time.Second {
+ t.Fatalf("second timeout = %v, want 40s", got)
+ }
+ recordAdaptiveToolTimeoutResult(&state, call, tools.ToolResult{ErrorClass: "timeout"}, nil)
+ if got := resolveAdaptiveToolExecutionTimeout(&state, call, base); got != 80*time.Second {
+ t.Fatalf("third timeout = %v, want 80s", got)
+ }
+ recordAdaptiveToolTimeoutResult(&state, call, tools.ToolResult{Content: "command timed out"}, nil)
+ if got := resolveAdaptiveToolExecutionTimeout(&state, call, base); got != 160*time.Second {
+ t.Fatalf("fourth timeout = %v, want 160s", got)
+ }
+ recordAdaptiveToolTimeoutResult(&state, call, tools.ToolResult{Content: "timeout"}, nil)
+ if got := resolveAdaptiveToolExecutionTimeout(&state, call, base); got != 160*time.Second {
+ t.Fatalf("capped timeout = %v, want 160s", got)
+ }
+
+ recordAdaptiveToolTimeoutResult(&state, call, tools.ToolResult{Name: tools.ToolNameBash, Content: "ok"}, nil)
+ if got := resolveAdaptiveToolExecutionTimeout(&state, call, base); got != base {
+ t.Fatalf("reset timeout = %v, want %v", got, base)
+ }
+}
+
+func TestResolveAdaptiveToolExecutionTimeoutForWorkspaceScanTools(t *testing.T) {
+ t.Parallel()
+
+ state := newRunState("run-codebase-timeout-backoff", agentsession.New("codebase-timeout-backoff"))
+ call := providertypes.ToolCall{
+ Name: tools.ToolNameCodebaseSearchText,
+ Arguments: `{"query":"plan mode","scope_dir":"internal/runtime"}`,
+ }
+ sameScopeDifferentQuery := providertypes.ToolCall{
+ Name: tools.ToolNameCodebaseSearchText,
+ Arguments: `{"query":"todo","scope_dir":"internal/runtime"}`,
+ }
+ differentScope := providertypes.ToolCall{
+ Name: tools.ToolNameCodebaseSearchText,
+ Arguments: `{"query":"todo","scope_dir":"internal/session"}`,
+ }
+ base := defaultWorkspaceScanToolTimeout
+
+ if got := resolveAdaptiveToolExecutionTimeout(&state, call, base); got != 60*time.Second {
+ t.Fatalf("first codebase timeout = %v, want 60s", got)
+ }
+ recordAdaptiveToolTimeoutResult(&state, call, tools.ToolResult{}, context.DeadlineExceeded)
+ if got := resolveAdaptiveToolExecutionTimeout(&state, sameScopeDifferentQuery, base); got != 120*time.Second {
+ t.Fatalf("same-scope codebase timeout = %v, want 120s", got)
+ }
+ if got := resolveAdaptiveToolExecutionTimeout(&state, differentScope, base); got != base {
+ t.Fatalf("different-scope codebase timeout = %v, want %v", got, base)
+ }
+ recordAdaptiveToolTimeoutResult(&state, call, tools.ToolResult{ErrorClass: "timeout"}, nil)
+ if got := resolveAdaptiveToolExecutionTimeout(&state, call, base); got != 160*time.Second {
+ t.Fatalf("third codebase timeout = %v, want 160s", got)
+ }
+ recordAdaptiveToolTimeoutResult(&state, call, tools.ToolResult{Name: tools.ToolNameCodebaseSearchText, Content: "ok"}, nil)
+ if got := resolveAdaptiveToolExecutionTimeout(&state, call, base); got != base {
+ t.Fatalf("reset codebase timeout = %v, want %v", got, base)
+ }
+
+ grepState := newRunState("run-grep-timeout-backoff", agentsession.New("grep-timeout-backoff"))
+ grepCall := providertypes.ToolCall{
+ Name: tools.ToolNameFilesystemGrep,
+ Arguments: `{"pattern":"CurrentPlan","dir":"internal/runtime"}`,
+ }
+ grepSameDir := providertypes.ToolCall{
+ Name: tools.ToolNameFilesystemGrep,
+ Arguments: `{"pattern":"PlanArtifact","dir":"internal/runtime"}`,
+ }
+ recordAdaptiveToolTimeoutResult(&grepState, grepCall, tools.ToolResult{}, context.DeadlineExceeded)
+ if got := resolveAdaptiveToolExecutionTimeout(&grepState, grepSameDir, base); got != 120*time.Second {
+ t.Fatalf("same-dir grep timeout = %v, want 120s", got)
+ }
+}
+
+func TestResolveAdaptiveToolExecutionTimeoutSkipsInteractiveTools(t *testing.T) {
+ t.Parallel()
+
+ state := newRunState("run-timeout-no-backoff", agentsession.New("timeout-no-backoff"))
+ call := providertypes.ToolCall{
+ Name: tools.ToolNameAskUser,
+ Arguments: `{"question_id":"q1","title":"T","kind":"text"}`,
+ }
+ base := 20 * time.Second
+
+ recordAdaptiveToolTimeoutResult(&state, call, tools.ToolResult{ErrorClass: "timeout"}, nil)
+ if got := resolveAdaptiveToolExecutionTimeout(&state, call, base); got != base {
+ t.Fatalf("ask_user adaptive timeout = %v, want unchanged %v", got, base)
+ }
+}
+
func TestResolveToolExecutionTimeoutFallbackAndHelpers(t *testing.T) {
t.Parallel()
diff --git a/internal/runtime/planning.go b/internal/runtime/planning.go
index a906951a..fcdc3827 100644
--- a/internal/runtime/planning.go
+++ b/internal/runtime/planning.go
@@ -18,11 +18,11 @@ const (
)
type summaryCandidate struct {
- Goal string `json:"goal"`
- KeySteps []string `json:"key_steps"`
- Constraints []string `json:"constraints"`
- Verify []string `json:"verify"`
- ActiveTodoIDs []string `json:"active_todo_ids"`
+ Goal string `json:"goal"`
+ KeySteps []string `json:"key_steps"`
+ Constraints []string `json:"constraints"`
+ Verify agentsession.AcceptChecks `json:"verify"`
+ ActiveTodoIDs []string `json:"active_todo_ids"`
}
type planTurnOutput struct {
@@ -119,7 +119,7 @@ func normalizeSummaryCandidate(candidate summaryCandidate) agentsession.SummaryV
Goal: strings.TrimSpace(candidate.Goal),
KeySteps: append([]string(nil), candidate.KeySteps...),
Constraints: append([]string(nil), candidate.Constraints...),
- Verify: append([]string(nil), candidate.Verify...),
+ Verify: candidate.Verify.Clone(),
ActiveTodoIDs: append([]string(nil), candidate.ActiveTodoIDs...),
}
}
@@ -305,6 +305,20 @@ func applyCurrentPlanRevision(session *agentsession.Session, plan *agentsession.
if oldPlan := session.CurrentPlan; oldPlan != nil && oldPlan.Revision < plan.Revision {
agentsession.CancelTodosByIDs(session.Todos, oldPlan.Summary.ActiveTodoIDs)
}
+ // 将 PlanSpec.Todos 中尚不存在于 session.Todos 的条目补入，
+ // 避免 plan 模式下模型后续通过 todo_write 引用这些 ID 时找不到。
+ for _, planTodo := range plan.Spec.Todos {
+ id := strings.TrimSpace(planTodo.ID)
+ if id == "" {
+ continue
+ }
+ if _, exists := session.FindTodo(id); exists {
+ continue
+ }
+ if err := session.AddTodo(planTodo); err != nil {
+ return false
+ }
+ }
session.CurrentPlan = plan
session.PlanApprovalPendingFullAlign = false
session.PlanCompletionPendingFullReview = false
diff --git a/internal/runtime/planning_test.go b/internal/runtime/planning_test.go
index 5bcd873e..bd7e9b63 100644
--- a/internal/runtime/planning_test.go
+++ b/internal/runtime/planning_test.go
@@ -267,14 +267,14 @@ func TestBuildPlanArtifact(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "旧计划",
Steps: []string{"旧步骤"},
- Verify: []string{"旧验证"},
+ Verify: acceptText("旧验证"),
},
}
output := planTurnOutput{
PlanSpec: agentsession.PlanSpec{
Goal: "新计划",
Steps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
Todos: []agentsession.TodoItem{
{ID: "todo-1", Content: "待办", Status: agentsession.TodoStatusPending},
},
@@ -282,7 +282,7 @@ func TestBuildPlanArtifact(t *testing.T) {
SummaryCandidate: summaryCandidate{
Goal: "新计划",
KeySteps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
ActiveTodoIDs: []string{"todo-1"},
},
}
@@ -319,7 +319,7 @@ func TestMarkCurrentPlanCompleted(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "执行当前计划",
Steps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
},
}
if !markCurrentPlanCompleted(&session, true) {
@@ -342,7 +342,7 @@ func TestMarkCurrentPlanCompleted(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "草案计划",
Steps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
},
}
if markCurrentPlanCompleted(&session, false) {
@@ -361,7 +361,7 @@ func TestPlanningNeedsFullPlan(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "Use full plan when alignment is pending",
Steps: []string{"align plan"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
Todos: []agentsession.TodoItem{
{ID: "todo-1", Content: "align plan", Status: agentsession.TodoStatusPending},
},
@@ -369,7 +369,7 @@ func TestPlanningNeedsFullPlan(t *testing.T) {
Summary: agentsession.SummaryView{
Goal: "Use full plan when alignment is pending",
KeySteps: []string{"align plan"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
ActiveTodoIDs: []string{"todo-1"},
},
}
@@ -416,7 +416,7 @@ func TestApproveCurrentPlan(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "批准当前计划",
Steps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
},
}
if err := approveCurrentPlan(&session, "plan-approve", 3); err != nil {
@@ -441,7 +441,7 @@ func TestRememberFullPlanRevisionClearsAlignmentFlags(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "完成全文对齐",
Steps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
},
}
session.PlanApprovalPendingFullAlign = true
@@ -479,7 +479,7 @@ func TestMarkCurrentPlanRestorePendingAndContextDirty(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "restore full plan",
Steps: []string{"step one"},
- Verify: []string{"verify one"},
+ Verify: acceptText("verify one"),
},
}
if !markCurrentPlanRestorePending(&session) {
@@ -535,7 +535,7 @@ func TestApproveCurrentPlanValidationErrors(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "审批校验",
Steps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
},
}
diff --git a/internal/runtime/repository_context_additional_test.go b/internal/runtime/repository_context_additional_test.go
index d2a3bf37..1ceab281 100644
--- a/internal/runtime/repository_context_additional_test.go
+++ b/internal/runtime/repository_context_additional_test.go
@@ -5,8 +5,8 @@ import (
"errors"
"testing"
- "neo-code/internal/repository"
providertypes "neo-code/internal/provider/types"
+ "neo-code/internal/repository"
agentsession "neo-code/internal/session"
)
diff --git a/internal/runtime/repository_context_test.go b/internal/runtime/repository_context_test.go
index cf7379c3..6ed20824 100644
--- a/internal/runtime/repository_context_test.go
+++ b/internal/runtime/repository_context_test.go
@@ -6,8 +6,8 @@ import (
"testing"
agentcontext "neo-code/internal/context"
- "neo-code/internal/repository"
providertypes "neo-code/internal/provider/types"
+ "neo-code/internal/repository"
agentsession "neo-code/internal/session"
"neo-code/internal/tools"
)
diff --git a/internal/runtime/run.go b/internal/runtime/run.go
index 6823053f..fa42daf0 100644
--- a/internal/runtime/run.go
+++ b/internal/runtime/run.go
@@ -14,10 +14,11 @@ import (
"neo-code/internal/config"
agentcontext "neo-code/internal/context"
contextcompact "neo-code/internal/context/compact"
+ "neo-code/internal/partsrender"
"neo-code/internal/promptasset"
"neo-code/internal/provider"
providertypes "neo-code/internal/provider/types"
- "neo-code/internal/runtime/acceptance"
+ "neo-code/internal/runtime/acceptgate"
"neo-code/internal/runtime/controlplane"
runtimehooks "neo-code/internal/runtime/hooks"
"neo-code/internal/runtime/streaming"
@@ -25,8 +26,6 @@ import (
"neo-code/internal/tools"
)
-var selfHealingReminder = promptasset.NoProgressReminder()
-
var selfHealingRepeatReminder = promptasset.RepeatCycleReminder()
const (
@@ -63,20 +62,6 @@ func computeToolSignature(calls []providertypes.ToolCall) string {
return hex.EncodeToString(hash[:])
}
-// computeTodoStateSignature 计算当前 Todo 列表的状态签名，用于识别 dispatch 是否产生了真实状态变化。
-func computeTodoStateSignature(items []agentsession.TodoItem) string {
- normalized := cloneTodosForPersistence(items)
- if len(normalized) == 0 {
- return ""
- }
- encoded, err := json.Marshal(normalized)
- if err != nil {
- return ""
- }
- hash := sha256.Sum256(encoded)
- return hex.EncodeToString(hash[:])
-}
-
// Run 执行一次完整的 ReAct 闭环：保存用户输入、驱动模型、执行工具并发出事件。
// 已有会话会先加锁再加载/更新，确保同一会话并发 Run 不会出现状态覆盖；
// 新会话在创建后再绑定会话锁，不同会话可并行执行。
@@ -183,8 +168,7 @@ func (s *Service) Run(ctx context.Context, input UserInput) (err error) {
agentsession.NormalizeAgentMode(session.AgentMode) == agentsession.AgentModePlan
state.taskID = strings.TrimSpace(input.TaskID)
state.agentID = strings.TrimSpace(input.AgentID)
- state.taskKind = inferTaskKindFromInput(input.Parts)
- state.userGoal = strings.TrimSpace(renderPartsForVerification(input.Parts))
+ state.userGoal = strings.TrimSpace(partsrender.RenderDisplayParts(input.Parts))
if input.CapabilityToken != nil {
token := input.CapabilityToken.Normalize()
state.capabilityToken = &token
@@ -223,6 +207,12 @@ func (s *Service) Run(ctx context.Context, input UserInput) (err error) {
if err := s.appendUserMessageAndSave(ctx, &state, input.Parts); err != nil {
return s.handleRunError(err)
}
+ if err := s.maybeAppendTodoBootstrapReminder(ctx, &state); err != nil {
+ return s.handleRunError(err)
+ }
+ if err := s.maybeAppendPlanBootstrapReminder(ctx, &state); err != nil {
+ return s.handleRunError(err)
+ }
s.emitRuntimeSnapshotUpdated(ctx, &state, "session_start")
s.updateResumeCheckpoint(ctx, &state, "plan", "")
@@ -314,16 +304,10 @@ func (s *Service) Run(ctx context.Context, input UserInput) (err error) {
return s.handleRunError(err)
}
hasToolCalls := len(turnOutput.assistant.ToolCalls) > 0
- state.mu.Lock()
- if hasToolCalls {
- state.mustUseToolAfterFinalContinue = false
- state.noToolAfterFinalContinueStreak = 0
- } else if state.mustUseToolAfterFinalContinue {
- state.pendingFinalProgress = false
- state.noToolAfterFinalContinueStreak++
- }
- state.mu.Unlock()
if hasToolCalls {
+ state.mu.Lock()
+ state.missingCompletionSignalStreak = 0
+ state.mu.Unlock()
if err := s.appendAssistantMessageAndSave(
ctx,
&state,
@@ -360,7 +344,7 @@ func (s *Service) Run(ctx context.Context, input UserInput) (err error) {
turnOutput.assistant,
hasToolCalls,
)
- completionState, completed := controlplane.EvaluateCompletion(
+ completionState, _ := controlplane.EvaluateCompletion(
state.completion,
hasToolCalls,
)
@@ -396,106 +380,84 @@ func (s *Service) Run(ctx context.Context, input UserInput) (err error) {
s.emitRunScoped(ctx, EventAgentDone, &state, planMessage)
return nil
}
+ if strings.TrimSpace(partsrender.RenderDisplayParts(turnOutput.assistant.Parts)) != "" {
+ if err := s.appendAssistantMessageOnlyAndSave(ctx, &state, turnOutput.assistant); err != nil {
+ return s.handleRunError(err)
+ }
+ s.emitRunScoped(ctx, EventAgentDone, &state, turnOutput.assistant)
+ return nil
+ }
}
completionSignaled, err := maybeParseCompletionTurnOutput(turnOutput.assistant)
if err != nil {
return s.handleRunError(err)
}
+ hasThinking := len(turnOutput.assistant.ThinkingMetadata) > 0
+ if !completionSignaled {
+ if hasThinking {
+ break turnAttempt
+ }
+ state.mu.Lock()
+ state.missingCompletionSignalStreak++
+ missingCompletionSignalStreak := state.missingCompletionSignalStreak
+ state.mu.Unlock()
+ if missingCompletionSignalStreak < missingCompletionSignalLimit {
+ reminder := completionProtocolReminderForStreak(missingCompletionSignalStreak)
+ setPendingSystemReminder(&state, reminder)
+ break turnAttempt
+ }
+ state.markTerminalDecision(
+ controlplane.TerminalStatusIncomplete,
+ controlplane.StopReasonMissingCompletionSignal,
+ "assistant stopped without task_completion",
+ )
+ s.emitRunScoped(ctx, EventAgentDone, &state, turnOutput.assistant)
+ return nil
+ }
+
if err := s.setBaseRunState(ctx, &state, controlplane.RunStateVerify); err != nil {
return s.handleRunError(err)
}
s.updateResumeCheckpoint(ctx, &state, "verify", "completed")
- acceptanceDecision, err := s.runBeforeCompletionDecisionAcceptance(
- ctx,
- &state,
- snapshot,
- turnOutput.assistant,
- snapshot.Workdir,
- completed,
- hasToolCalls,
- turnOutput.assistant.Role,
- )
- if err != nil {
- return s.handleRunError(err)
- }
- s.emitAcceptanceDecisionEvents(&state, acceptanceDecision)
- applyAcceptanceResultProgress(&state, acceptanceDecision)
-
- switch acceptanceDecision.Status {
- case acceptance.AcceptanceAccepted:
- state.lastAcceptanceBlockedReason = ""
- state.mustUseToolAfterFinalContinue = false
- state.noToolAfterFinalContinueStreak = 0
+ report := s.evaluateAcceptGate(ctx, &state, turnOutput.assistant)
+ s.emitAcceptGateReport(&state, report)
+ assistantForFinal := stripCompletionSignalFromAssistantMessage(turnOutput.assistant)
+ if report.Outcome == acceptgate.OutcomeAccepted {
+ state.mu.Lock()
+ state.missingCompletionSignalStreak = 0
+ state.mu.Unlock()
if markCurrentPlanCompleted(&state.session, completionSignaled) {
state.touchSession()
if err := s.sessionStore.UpdateSessionState(ctx, sessionStateInputFromSession(state.session)); err != nil {
return s.handleRunError(err)
}
}
- if err := s.appendAssistantMessageOnlyAndSave(ctx, &state, turnOutput.assistant); err != nil {
+ if err := s.appendAssistantMessageOnlyAndSave(ctx, &state, assistantForFinal); err != nil {
return s.handleRunError(err)
}
s.emitRunScopedOptional(EventVerificationCompleted, &state, VerificationCompletedPayload{
- StopReason: acceptanceDecision.StopReason,
+ StopReason: report.StopReason,
})
- recordAcceptanceTerminal(&state, acceptanceDecision)
- s.emitRunScoped(ctx, EventAgentDone, &state, turnOutput.assistant)
+ state.markTerminalDecision(controlplane.TerminalStatusCompleted, report.StopReason, report.Summary)
+ s.emitRunScoped(ctx, EventAgentDone, &state, assistantForFinal)
s.triggerMemoExtraction(state.session.ID, runBoundaryMessagesForMemo(&state), state.rememberedThisRun)
+
return nil
- case acceptance.AcceptanceContinue:
- state.lastAcceptanceBlockedReason = strings.TrimSpace(acceptanceDecision.CompletionBlockedReason)
- state.mustUseToolAfterFinalContinue = true
- if state.noToolAfterFinalContinueStreak == 0 {
- state.noToolAfterFinalContinueStreak = 1
- }
- reminder := strings.TrimSpace(buildAcceptanceContinueHint(acceptanceDecision))
- if reminder == "" {
- reminder = finalContinueReminder
- }
- if err := s.appendSystemMessageAndSave(ctx, &state, reminder); err != nil {
- return s.handleRunError(err)
- }
- break turnAttempt
- case acceptance.AcceptanceIncomplete:
- state.lastAcceptanceBlockedReason = ""
- state.mustUseToolAfterFinalContinue = false
- state.noToolAfterFinalContinueStreak = 0
- if err := s.appendAssistantMessageOnlyAndSave(ctx, &state, turnOutput.assistant); err != nil {
- return s.handleRunError(err)
- }
- recordAcceptanceTerminal(&state, acceptanceDecision)
- s.emitRunScoped(ctx, EventAgentDone, &state, turnOutput.assistant)
- return nil
- case acceptance.AcceptanceFailed:
- state.lastAcceptanceBlockedReason = ""
- state.mustUseToolAfterFinalContinue = false
- state.noToolAfterFinalContinueStreak = 0
- if err := s.appendAssistantMessageOnlyAndSave(ctx, &state, turnOutput.assistant); err != nil {
- return s.handleRunError(err)
- }
- s.emitRunScopedOptional(EventVerificationFailed, &state, VerificationFailedPayload{
- StopReason: acceptanceDecision.StopReason,
- ErrorClass: acceptanceDecision.ErrorClass,
- })
- recordAcceptanceTerminal(&state, acceptanceDecision)
- s.emitRunScoped(ctx, EventAgentDone, &state, turnOutput.assistant)
- return nil
- default:
- state.lastAcceptanceBlockedReason = ""
- state.mustUseToolAfterFinalContinue = false
- state.noToolAfterFinalContinueStreak = 0
- if err := s.appendAssistantMessageOnlyAndSave(ctx, &state, turnOutput.assistant); err != nil {
- return s.handleRunError(err)
- }
- recordAcceptanceTerminal(&state, acceptanceDecision)
- s.emitRunScoped(ctx, EventAgentDone, &state, turnOutput.assistant)
- return nil
}
+ state.mu.Lock()
+ state.missingCompletionSignalStreak = 0
+ state.mu.Unlock()
+ if err := s.appendAssistantMessageOnlyAndSave(ctx, &state, assistantForFinal); err != nil {
+ return s.handleRunError(err)
+ }
+ s.emitRunScopedOptional(EventVerificationFailed, &state, VerificationFailedPayload{
+ StopReason: report.StopReason,
+ })
+ state.markTerminalDecision(controlplane.TerminalStatusFailed, report.StopReason, report.Summary)
+ s.emitRunScoped(ctx, EventAgentDone, &state, assistantForFinal)
+ return nil
}
- beforeTask := state.session.TaskState.Clone()
- beforeTodos := cloneTodosForPersistence(state.session.Todos)
-
if err := s.setBaseRunState(ctx, &state, controlplane.RunStateExecute); err != nil {
return s.handleRunError(err)
}
@@ -519,33 +481,26 @@ func (s *Service) Run(ctx context.Context, input UserInput) (err error) {
state.completion = applyToolExecutionCompletion(state.completion, summary)
afterTask := state.session.TaskState.Clone()
afterTodos := cloneTodosForPersistence(state.session.Todos)
- progressRunState := controlplane.RunStateExecute
- if resolvePlanningStageForState(&state) == planStagePlan {
- progressRunState = controlplane.RunStatePlan
- }
progressInput := collectProgressInput(
- progressRunState,
- beforeTask,
afterTask,
- beforeTodos,
afterTodos,
summary,
- snapshot.NoProgressStreakLimit,
snapshot.RepeatCycleStreakLimit,
)
state.progress = controlplane.EvaluateProgress(state.progress, progressInput)
currentScore := state.progress.LastScore
- if shouldPromotePendingFinalProgress(
- currentScore,
- summary,
- state.completion,
- state.lastAcceptanceBlockedReason,
- ) {
- state.pendingFinalProgress = true
- }
state.mu.Unlock()
s.emitRunScoped(ctx, EventProgressEvaluated, &state, ProgressEvaluatedPayload{Score: currentScore})
+ if currentScore.ShouldTerminate {
+ reason := currentScore.TerminateReason
+ if reason == "" {
+ reason = controlplane.StopReasonRepeatCycle
+ }
+ state.markTerminalDecision(controlplane.TerminalStatusIncomplete, reason, "progress hard stop")
+ s.emitRunScoped(ctx, EventAgentDone, &state, providertypes.Message{Role: providertypes.RoleAssistant})
+ return nil
+ }
if err := s.setBaseRunState(ctx, &state, controlplane.RunStateVerify); err != nil {
return s.handleRunError(err)
}
@@ -633,9 +588,11 @@ func (s *Service) prepareTurnBudgetSnapshot(ctx context.Context, state *runState
score := state.progress.LastScore
state.mu.Unlock()
- limit := resolveNoProgressStreakLimit(cfg.Runtime)
repeatLimit := resolveRepeatCycleStreakLimit(cfg.Runtime)
systemPrompt := withProgressReminder(builtContext.SystemPrompt, score)
+ if pendingReminder := drainPendingSystemReminder(state); pendingReminder != "" {
+ systemPrompt = mergeEphemeralHookNotificationIntoSystemPrompt(systemPrompt, pendingReminder)
+ }
if notificationHint := strings.TrimSpace(s.drainHookNotificationsForTurn(state)); notificationHint != "" {
systemPrompt = mergeEphemeralHookNotificationIntoSystemPrompt(systemPrompt, notificationHint)
}
@@ -675,7 +632,6 @@ func (s *Service) prepareTurnBudgetSnapshot(ctx context.Context, state *runState
promptBudget,
budgetSource,
state.compactCount,
- limit,
repeatLimit,
injectFullPlan,
contextWindow,
@@ -683,14 +639,6 @@ func (s *Service) prepareTurnBudgetSnapshot(ctx context.Context, state *runState
), false, nil
}
-// resolveNoProgressStreakLimit 统一解析熔断阈值，避免运行期出现无效值导致分支行为不一致。
-func resolveNoProgressStreakLimit(rc config.RuntimeConfig) int {
- if rc.MaxNoProgressStreak <= 0 {
- return config.DefaultMaxNoProgressStreak
- }
- return rc.MaxNoProgressStreak
-}
-
// resolveRepeatCycleStreakLimit 统一解析重复调用循环阈值。
func resolveRepeatCycleStreakLimit(rc config.RuntimeConfig) int {
if rc.MaxRepeatCycleStreak <= 0 {
@@ -707,6 +655,32 @@ func resolveRuntimeMaxTurns(rc config.RuntimeConfig) int {
return rc.MaxTurns
}
+// setPendingSystemReminder 暂存只用于下一轮 provider 请求的系统提醒，避免写入会话历史。
+func setPendingSystemReminder(state *runState, reminder string) {
+ if state == nil {
+ return
+ }
+ reminder = strings.TrimSpace(reminder)
+ if reminder == "" {
+ return
+ }
+ state.mu.Lock()
+ state.pendingSystemReminder = reminder
+ state.mu.Unlock()
+}
+
+// drainPendingSystemReminder 读取并清空本轮待注入的系统提醒，保证提醒只进入一次 provider 请求。
+func drainPendingSystemReminder(state *runState) string {
+ if state == nil {
+ return ""
+ }
+ state.mu.Lock()
+ defer state.mu.Unlock()
+ reminder := strings.TrimSpace(state.pendingSystemReminder)
+ state.pendingSystemReminder = ""
+ return reminder
+}
+
// callProvider 使用冻结后的 TurnBudgetSnapshot 执行单次 provider 调用。
func (s *Service) callProvider(
ctx context.Context,
@@ -1154,8 +1128,6 @@ func withProgressReminder(systemPrompt string, score controlplane.ProgressScore)
switch score.ReminderKind {
case controlplane.ReminderKindRepeatCycle:
reminder = selfHealingRepeatReminder
- case controlplane.ReminderKindNoProgress, controlplane.ReminderKindGenericStalled:
- reminder = selfHealingReminder
default:
return systemPrompt
}
diff --git a/internal/runtime/runtime_progress_test.go b/internal/runtime/runtime_progress_test.go
index df2d56f2..16b1e4c8 100644
--- a/internal/runtime/runtime_progress_test.go
+++ b/internal/runtime/runtime_progress_test.go
@@ -2,7 +2,6 @@ package runtime
import (
"context"
- "strconv"
"strings"
"sync/atomic"
"testing"
@@ -16,52 +15,45 @@ import (
todotool "neo-code/internal/tools/todo"
)
-func TestProgressStreakNoLongerStopsRun(t *testing.T) {
+func TestRepeatCycleStreakNoLongerStopsRunAndInjectsReminder(t *testing.T) {
t.Setenv("TEST_KEY", "dummy")
cfg := config.Config{
- Providers: []config.ProviderConfig{{Name: "test-progress", Driver: "test", BaseURL: "http://localhost", Model: "test", APIKeyEnv: "TEST_KEY"}},
- SelectedProvider: "test-progress",
+ Providers: []config.ProviderConfig{{Name: "test-repeat", Driver: "test", BaseURL: "http://localhost", Model: "test", APIKeyEnv: "TEST_KEY"}},
+ SelectedProvider: "test-repeat",
Workdir: t.TempDir(),
Runtime: config.RuntimeConfig{
- MaxNoProgressStreak: 3,
- MaxRepeatCycleStreak: 6,
+ MaxRepeatCycleStreak: 3,
},
}
+ var executeCalls int32
+ var providerCalls int32
toolManager := &stubToolManager{
specs: []providertypes.ToolSpec{
- {Name: "tool_error"},
+ {Name: "tool_repeat"},
},
executeFn: func(ctx context.Context, input tools.ToolCallInput) (tools.ToolResult, error) {
- // Always return error to avoid generating progress
- return tools.ToolResult{Content: "error occurred", IsError: true}, nil
+ atomic.AddInt32(&executeCalls, 1)
+ return tools.ToolResult{Name: input.Name, Content: "ok", IsError: false}, nil
},
}
var promptInjected bool
- var providerCalls int32
- var signatureSeq int32
providerFactory := &scriptedProviderFactory{
provider: &scriptedProvider{
chatFn: func(ctx context.Context, req providertypes.GenerateRequest, events chan<- providertypes.StreamEvent) error {
call := atomic.AddInt32(&providerCalls, 1)
- seq := atomic.AddInt32(&signatureSeq, 1)
- if strings.Contains(req.SystemPrompt, selfHealingReminder) {
+ if strings.Contains(req.SystemPrompt, selfHealingRepeatReminder) {
promptInjected = true
}
if call >= 5 {
- events <- providertypes.NewTextDeltaStreamEvent("done")
+ events <- providertypes.NewTextDeltaStreamEvent("{\"task_completion\":{\"completed\":true}}\ndone")
events <- providertypes.NewMessageDoneStreamEvent("stop", nil)
return nil
}
- // the model always decides to call the tool
- events <- providertypes.NewToolCallStartStreamEvent(0, "call_err", "tool_error")
- events <- providertypes.NewToolCallDeltaStreamEvent(
- 0,
- "call_err",
- `{"seq":`+strconv.FormatInt(int64(seq), 10)+`}`,
- )
+ events <- providertypes.NewToolCallStartStreamEvent(0, "call_repeat", "tool_repeat")
+ events <- providertypes.NewToolCallDeltaStreamEvent(0, "call_repeat", `{"path":"x"}`)
events <- providertypes.NewMessageDoneStreamEvent("tool_calls", nil)
return nil
},
@@ -69,7 +61,6 @@ func TestProgressStreakNoLongerStopsRun(t *testing.T) {
}
manager := config.NewManager(config.NewLoader(t.TempDir(), &cfg))
-
service := NewWithFactory(
manager,
toolManager,
@@ -78,110 +69,36 @@ func TestProgressStreakNoLongerStopsRun(t *testing.T) {
nil,
)
- input := UserInput{
- RunID: "run-progress",
- Parts: []providertypes.ContentPart{providertypes.NewTextPart("trigger error loop")},
- }
-
- if err := service.Run(context.Background(), input); err != nil {
- t.Fatalf("expected run success without no-progress hard stop, got %v", err)
+ err := service.Run(context.Background(), UserInput{
+ RunID: "run-repeat-streak",
+ Parts: []providertypes.ContentPart{providertypes.NewTextPart("trigger repeat loop")},
+ })
+ if err != nil {
+ t.Fatalf("expected run to recover after repeat-cycle reminder, got %v", err)
}
events := collectRuntimeEvents(service.Events())
assertStopReasonDecided(t, events, controlplane.StopReasonAccepted, "")
if !promptInjected {
- t.Error("expected self-healing prompt to be injected before repetitive no-progress turns")
- }
- if providerCalls != 5 {
- t.Fatalf("expected 5 provider turns (4 tool cycles + done), got %d", providerCalls)
- }
-}
-
-func TestProgressEvidenceResetsNoProgressStreak(t *testing.T) {
- t.Setenv("TEST_KEY", "dummy")
-
- cfg := config.Config{
- Providers: []config.ProviderConfig{{Name: "test-progress", Driver: "test", BaseURL: "http://localhost", Model: "test", APIKeyEnv: "TEST_KEY"}},
- SelectedProvider: "test-progress",
- Workdir: t.TempDir(),
+ t.Fatal("expected repeat self-healing prompt injection before recovery")
}
-
- var executeCalls int32
- toolManager := &stubToolManager{
- specs: []providertypes.ToolSpec{
- {Name: "tool_mixed"},
- },
- executeFn: func(ctx context.Context, input tools.ToolCallInput) (tools.ToolResult, error) {
- call := int(atomic.AddInt32(&executeCalls, 1))
- if call == 3 {
- return tools.ToolResult{Name: input.Name, Content: "ok", IsError: false}, nil
- }
- return tools.ToolResult{Name: input.Name, Content: "error occurred", IsError: true}, nil
- },
- }
-
- var providerCalls int32
- var signatureSeq int32
- providerFactory := &scriptedProviderFactory{
- provider: &scriptedProvider{
- chatFn: func(ctx context.Context, req providertypes.GenerateRequest, events chan<- providertypes.StreamEvent) error {
- call := int(atomic.AddInt32(&providerCalls, 1))
- if call <= 4 {
- seq := atomic.AddInt32(&signatureSeq, 1)
- events <- providertypes.NewToolCallStartStreamEvent(0, "call_mixed", "tool_mixed")
- events <- providertypes.NewToolCallDeltaStreamEvent(
- 0,
- "call_mixed",
- `{"seq":`+strconv.FormatInt(int64(seq), 10)+`}`,
- )
- events <- providertypes.NewMessageDoneStreamEvent("tool_calls", nil)
- return nil
- }
- events <- providertypes.NewTextDeltaStreamEvent("done")
- events <- providertypes.NewMessageDoneStreamEvent("stop", nil)
- return nil
- },
- },
- }
-
- manager := config.NewManager(config.NewLoader(t.TempDir(), &cfg))
- service := NewWithFactory(
- manager,
- toolManager,
- newMemoryStore(),
- providerFactory,
- nil,
- )
-
- err := service.Run(context.Background(), UserInput{
- RunID: "run-progress-reset",
- Parts: []providertypes.ContentPart{providertypes.NewTextPart("trigger mixed progress loop")},
- })
- if err != nil {
- t.Fatalf("expected run to finish successfully, got %v", err)
- }
-
if executeCalls != 4 {
- t.Fatalf("expected 4 tool executions, got %d", executeCalls)
+ t.Fatalf("expected 4 repeated tool executions before repeat reminder recovery, got %d", executeCalls)
}
if providerCalls != 5 {
- t.Fatalf("expected 5 provider calls (4 tool turns + 1 done), got %d", providerCalls)
+ t.Fatalf("expected 5 provider turns including recovery response, got %d", providerCalls)
}
-
- events := collectRuntimeEvents(service.Events())
- assertStopReasonDecided(t, events, controlplane.StopReasonAccepted, "")
}
-func TestRepeatCycleStreakNoLongerStopsRunAndInjectsReminder(t *testing.T) {
+func TestRepeatCycleTerminatesAfterReminderIfStillStalled(t *testing.T) {
t.Setenv("TEST_KEY", "dummy")
cfg := config.Config{
- Providers: []config.ProviderConfig{{Name: "test-repeat", Driver: "test", BaseURL: "http://localhost", Model: "test", APIKeyEnv: "TEST_KEY"}},
- SelectedProvider: "test-repeat",
+ Providers: []config.ProviderConfig{{Name: "test-repeat-hard-stop", Driver: "test", BaseURL: "http://localhost", Model: "test", APIKeyEnv: "TEST_KEY"}},
+ SelectedProvider: "test-repeat-hard-stop",
Workdir: t.TempDir(),
Runtime: config.RuntimeConfig{
- MaxNoProgressStreak: 10,
MaxRepeatCycleStreak: 3,
},
}
@@ -189,9 +106,7 @@ func TestRepeatCycleStreakNoLongerStopsRunAndInjectsReminder(t *testing.T) {
var executeCalls int32
var providerCalls int32
toolManager := &stubToolManager{
- specs: []providertypes.ToolSpec{
- {Name: "tool_repeat"},
- },
+ specs: []providertypes.ToolSpec{{Name: "tool_repeat"}},
executeFn: func(ctx context.Context, input tools.ToolCallInput) (tools.ToolResult, error) {
atomic.AddInt32(&executeCalls, 1)
return tools.ToolResult{Name: input.Name, Content: "ok", IsError: false}, nil
@@ -202,15 +117,10 @@ func TestRepeatCycleStreakNoLongerStopsRunAndInjectsReminder(t *testing.T) {
providerFactory := &scriptedProviderFactory{
provider: &scriptedProvider{
chatFn: func(ctx context.Context, req providertypes.GenerateRequest, events chan<- providertypes.StreamEvent) error {
- call := atomic.AddInt32(&providerCalls, 1)
+ atomic.AddInt32(&providerCalls, 1)
if strings.Contains(req.SystemPrompt, selfHealingRepeatReminder) {
promptInjected = true
}
- if call >= 5 {
- events <- providertypes.NewTextDeltaStreamEvent("done")
- events <- providertypes.NewMessageDoneStreamEvent("stop", nil)
- return nil
- }
events <- providertypes.NewToolCallStartStreamEvent(0, "call_repeat", "tool_repeat")
events <- providertypes.NewToolCallDeltaStreamEvent(0, "call_repeat", `{"path":"x"}`)
events <- providertypes.NewMessageDoneStreamEvent("tool_calls", nil)
@@ -220,33 +130,26 @@ func TestRepeatCycleStreakNoLongerStopsRunAndInjectsReminder(t *testing.T) {
}
manager := config.NewManager(config.NewLoader(t.TempDir(), &cfg))
- service := NewWithFactory(
- manager,
- toolManager,
- newMemoryStore(),
- providerFactory,
- nil,
- )
+ service := NewWithFactory(manager, toolManager, newMemoryStore(), providerFactory, nil)
- err := service.Run(context.Background(), UserInput{
- RunID: "run-repeat-streak",
- Parts: []providertypes.ContentPart{providertypes.NewTextPart("trigger repeat loop")},
- })
- if err != nil {
- t.Fatalf("expected run success without repeat hard stop, got %v", err)
+ if err := service.Run(context.Background(), UserInput{
+ RunID: "run-repeat-hard-stop",
+ Parts: []providertypes.ContentPart{providertypes.NewTextPart("trigger unrecovered repeat loop")},
+ }); err != nil {
+ t.Fatalf("expected run to stop cleanly on repeat-cycle, got %v", err)
}
events := collectRuntimeEvents(service.Events())
- assertStopReasonDecided(t, events, controlplane.StopReasonAccepted, "")
+ assertStopReasonDecided(t, events, controlplane.StopReasonRepeatCycle, "")
if !promptInjected {
- t.Fatal("expected repeat self-healing prompt to be injected before repeat limit is reached")
+ t.Fatal("expected repeat self-healing prompt injection before hard repeat-cycle termination")
}
- if executeCalls != 4 {
- t.Fatalf("expected repeated tool executions to continue until model stops, got %d", executeCalls)
+ if executeCalls != 5 {
+ t.Fatalf("expected 5 repeated tool executions before repeat-cycle termination, got %d", executeCalls)
}
if providerCalls != 5 {
- t.Fatalf("expected 5 provider turns (4 tool cycles + done), got %d", providerCalls)
+ t.Fatalf("expected 5 provider turns before repeat-cycle termination, got %d", providerCalls)
}
}
@@ -258,7 +161,6 @@ func TestRepeatCycleFailedCallsNoLongerHardStop(t *testing.T) {
SelectedProvider: "test-repeat-fail",
Workdir: t.TempDir(),
Runtime: config.RuntimeConfig{
- MaxNoProgressStreak: 10,
MaxRepeatCycleStreak: 3,
},
}
@@ -280,7 +182,7 @@ func TestRepeatCycleFailedCallsNoLongerHardStop(t *testing.T) {
chatFn: func(ctx context.Context, req providertypes.GenerateRequest, events chan<- providertypes.StreamEvent) error {
call := atomic.AddInt32(&providerCalls, 1)
if call >= 5 {
- events <- providertypes.NewTextDeltaStreamEvent("done")
+ events <- providertypes.NewTextDeltaStreamEvent("{\"task_completion\":{\"completed\":true}}\ndone")
events <- providertypes.NewMessageDoneStreamEvent("stop", nil)
return nil
}
@@ -306,14 +208,16 @@ func TestRepeatCycleFailedCallsNoLongerHardStop(t *testing.T) {
Parts: []providertypes.ContentPart{providertypes.NewTextPart("trigger repeat fail loop")},
})
if err != nil {
- t.Fatalf("expected run success without repeat hard stop, got %v", err)
+ t.Fatalf("expected run to recover after repeat-cycle reminder, got %v", err)
}
if executeCalls != 4 {
- t.Fatalf("expected failed repeated calls to continue until model stops, got %d", executeCalls)
+ t.Fatalf("expected 4 failed repeated calls before recovery, got %d", executeCalls)
}
if providerCalls != 5 {
- t.Fatalf("expected 5 provider turns (4 tool cycles + done), got %d", providerCalls)
+ t.Fatalf("expected 5 provider turns including recovery response, got %d", providerCalls)
}
+ events := collectRuntimeEvents(service.Events())
+ assertStopReasonDecided(t, events, controlplane.StopReasonAccepted, "")
}
func TestRunStopsWhenMaxTurnsReached(t *testing.T) {
@@ -444,7 +348,6 @@ func TestPrepareTurnSnapshotInjectRepeatReminderWithEmptyPrompt(t *testing.T) {
func TestPrepareTurnBudgetSnapshotRepeatReminderTakesPriority(t *testing.T) {
manager := newRuntimeConfigManager(t)
if err := manager.Update(context.Background(), func(cfg *config.Config) error {
- cfg.Runtime.MaxNoProgressStreak = 3
cfg.Runtime.MaxRepeatCycleStreak = 3
return nil
}); err != nil {
@@ -461,7 +364,6 @@ func TestPrepareTurnBudgetSnapshotRepeatReminderTakesPriority(t *testing.T) {
toolManager: &stubToolManager{},
}
state := newRunState("run-reminder-priority", newRuntimeSession("session-reminder-priority"))
- state.progress.LastScore.NoProgressStreak = 2
state.progress.LastScore.RepeatCycleStreak = 2
state.progress.LastScore.StalledProgressState = controlplane.StalledProgressStalled
state.progress.LastScore.ReminderKind = controlplane.ReminderKindRepeatCycle
@@ -476,19 +378,9 @@ func TestPrepareTurnBudgetSnapshotRepeatReminderTakesPriority(t *testing.T) {
if !strings.Contains(snapshot.Request.SystemPrompt, selfHealingRepeatReminder) {
t.Fatalf("expected prompt to contain repeat reminder, got %q", snapshot.Request.SystemPrompt)
}
- if strings.Contains(snapshot.Request.SystemPrompt, selfHealingReminder) {
- t.Fatalf("expected no-progress reminder to be skipped when repeat reminder is injected, got %q", snapshot.Request.SystemPrompt)
- }
}
func TestResolveStreakLimitDefaults(t *testing.T) {
- if got := resolveNoProgressStreakLimit(config.RuntimeConfig{MaxNoProgressStreak: 0}); got != config.DefaultMaxNoProgressStreak {
- t.Fatalf("expected default no-progress limit %d, got %d", config.DefaultMaxNoProgressStreak, got)
- }
- if got := resolveNoProgressStreakLimit(config.RuntimeConfig{MaxNoProgressStreak: 8}); got != 8 {
- t.Fatalf("expected explicit no-progress limit 8, got %d", got)
- }
-
if got := resolveRepeatCycleStreakLimit(config.RuntimeConfig{MaxRepeatCycleStreak: -1}); got != config.DefaultMaxRepeatCycleStreak {
t.Fatalf("expected default repeat limit %d, got %d", config.DefaultMaxRepeatCycleStreak, got)
}
@@ -504,59 +396,11 @@ func TestResolveStreakLimitDefaults(t *testing.T) {
}
}
-func TestComputeTodoStateSignature(t *testing.T) {
- t.Parallel()
-
- if got := computeTodoStateSignature(nil); got != "" {
- t.Fatalf("computeTodoStateSignature(nil) = %q", got)
- }
-
- base := []agentsession.TodoItem{
- {
- ID: "t1",
- Content: "task",
- Status: agentsession.TodoStatusPending,
- Executor: agentsession.TodoExecutorAgent,
- },
- }
- sig1 := computeTodoStateSignature(base)
- if strings.TrimSpace(sig1) == "" {
- t.Fatal("expected non-empty signature")
- }
-
- same := []agentsession.TodoItem{
- {
- ID: "t1",
- Content: "task",
- Status: agentsession.TodoStatusPending,
- Executor: agentsession.TodoExecutorAgent,
- },
- }
- sig2 := computeTodoStateSignature(same)
- if sig1 != sig2 {
- t.Fatalf("expected stable signature, got %q vs %q", sig1, sig2)
- }
-
- changed := []agentsession.TodoItem{
- {
- ID: "t1",
- Content: "task",
- Status: agentsession.TodoStatusCompleted,
- Executor: agentsession.TodoExecutorAgent,
- },
- }
- sig3 := computeTodoStateSignature(changed)
- if sig3 == sig1 {
- t.Fatalf("expected changed signature when todo state changes")
- }
-}
-
func TestNoToolIncompleteTurnStillEvaluatesProgressAndInjectsReminder(t *testing.T) {
t.Parallel()
manager := newRuntimeConfigManager(t)
if err := manager.Update(context.Background(), func(cfg *config.Config) error {
- cfg.Runtime.MaxNoProgressStreak = 1
return nil
}); err != nil {
t.Fatalf("update config: %v", err)
@@ -579,11 +423,12 @@ func TestNoToolIncompleteTurnStillEvaluatesProgressAndInjectsReminder(t *testing
registry.Register(todotool.New())
providerImpl := &scriptedProvider{
+ requireExplicitCompletion: true,
responses: []scriptedResponse{
{
Message: providertypes.Message{
Role: providertypes.RoleAssistant,
- Parts: []providertypes.ContentPart{providertypes.NewTextPart("done")},
+ Parts: []providertypes.ContentPart{providertypes.NewTextPart("{\"task_completion\":{\"completed\":false}}\ndone")},
},
FinishReason: "stop",
},
@@ -616,7 +461,7 @@ func TestNoToolIncompleteTurnStillEvaluatesProgressAndInjectsReminder(t *testing
{
Message: providertypes.Message{
Role: providertypes.RoleAssistant,
- Parts: []providertypes.ContentPart{providertypes.NewTextPart("done")},
+ Parts: []providertypes.ContentPart{providertypes.NewTextPart("{\"task_completion\":{\"completed\":true}}\ndone")},
},
FinishReason: "stop",
},
@@ -642,19 +487,28 @@ func TestNoToolIncompleteTurnStillEvaluatesProgressAndInjectsReminder(t *testing
if len(providerImpl.requests) < 2 {
t.Fatalf("expected at least 2 provider requests, got %d", len(providerImpl.requests))
}
- foundReminder := false
- for _, message := range providerImpl.requests[1].Messages {
+ secondSystemPrompt := providerImpl.requests[1].SystemPrompt
+ if !strings.Contains(secondSystemPrompt, "[Runtime Control]") ||
+ !strings.Contains(secondSystemPrompt, "task_completion") {
+ t.Fatalf("expected runtime protocol note in second provider request system prompt, got %q", secondSystemPrompt)
+ }
+ if len(providerImpl.requests) > 2 {
+ thirdSystemPrompt := providerImpl.requests[2].SystemPrompt
+ if strings.Contains(thirdSystemPrompt, "[Runtime Control]") &&
+ strings.Contains(thirdSystemPrompt, "task_completion") {
+ t.Fatalf("expected runtime protocol note to be injected once, got third system prompt %q", thirdSystemPrompt)
+ }
+ }
+
+ savedSession := store.sessions[session.ID]
+ for _, message := range savedSession.Messages {
content := renderPartsForTest(message.Parts)
if message.Role == providertypes.RoleSystem &&
- strings.Contains(content, "") &&
- strings.Contains(content, "MUST call todo_write") {
- foundReminder = true
- break
+ strings.Contains(content, "[Runtime Control]") &&
+ strings.Contains(content, "task_completion") {
+ t.Fatalf("expected completion reminder to stay out of session transcript, found %q", content)
}
}
- if !foundReminder {
- t.Fatalf("expected continue reminder in second provider request messages, got %+v", providerImpl.requests[1].Messages)
- }
events := collectRuntimeEvents(service.Events())
assertEventContains(t, events, EventProgressEvaluated)
@@ -666,7 +520,6 @@ func TestAcceptanceContinueWithoutToolCallStopsAsIncomplete(t *testing.T) {
manager := newRuntimeConfigManager(t)
if err := manager.Update(context.Background(), func(cfg *config.Config) error {
- cfg.Runtime.Verification.MaxNoProgress = 2
return nil
}); err != nil {
t.Fatalf("update config: %v", err)
@@ -687,28 +540,14 @@ func TestAcceptanceContinueWithoutToolCallStopsAsIncomplete(t *testing.T) {
store.sessions[session.ID] = cloneSession(session)
providerImpl := &scriptedProvider{
+ requireExplicitCompletion: true,
responses: []scriptedResponse{
- {
- Message: providertypes.Message{
- Role: providertypes.RoleAssistant,
- Parts: []providertypes.ContentPart{providertypes.NewTextPart("我已经完成了")},
- },
- FinishReason: "stop",
- },
- {
- Message: providertypes.Message{
- Role: providertypes.RoleAssistant,
- Parts: []providertypes.ContentPart{providertypes.NewTextPart("任务已完成")},
- },
- FinishReason: "stop",
- },
- {
- Message: providertypes.Message{
- Role: providertypes.RoleAssistant,
- Parts: []providertypes.ContentPart{providertypes.NewTextPart("不应再到这里")},
- },
- FinishReason: "stop",
- },
+ {Message: providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("{\"task_completion\":{\"completed\":false}}\n1")}}, FinishReason: "stop"},
+ {Message: providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("{\"task_completion\":{\"completed\":false}}\n2")}}, FinishReason: "stop"},
+ {Message: providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("{\"task_completion\":{\"completed\":false}}\n3")}}, FinishReason: "stop"},
+ {Message: providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("{\"task_completion\":{\"completed\":false}}\n4")}}, FinishReason: "stop"},
+ {Message: providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("{\"task_completion\":{\"completed\":false}}\n5")}}, FinishReason: "stop"},
+ {Message: providertypes.Message{Role: providertypes.RoleAssistant, Parts: []providertypes.ContentPart{providertypes.NewTextPart("不应再到这里")}}, FinishReason: "stop"},
},
}
@@ -728,49 +567,28 @@ func TestAcceptanceContinueWithoutToolCallStopsAsIncomplete(t *testing.T) {
t.Fatalf("Run() error = %v", err)
}
- if len(providerImpl.requests) != 2 {
- t.Fatalf("expected runtime to stop after two no-tool continues, got %d requests", len(providerImpl.requests))
+ if len(providerImpl.requests) != 6 {
+ t.Fatalf("expected runtime to stop after six missing completion signals, got %d requests", len(providerImpl.requests))
}
- secondRequestMessages := providerImpl.requests[1].Messages
- foundHint := false
- for _, message := range secondRequestMessages {
- if message.Role != providertypes.RoleSystem {
- continue
- }
+ // 第 6 个请求（streak=5 时注入最终提醒后）应包含最终协议提醒
+ fifthSystemPrompt := providerImpl.requests[5].SystemPrompt
+ if !strings.Contains(fifthSystemPrompt, "[Runtime Control]") ||
+ !strings.Contains(fifthSystemPrompt, "final protocol reminder") ||
+ !strings.Contains(fifthSystemPrompt, "task_completion") {
+ t.Fatalf("expected final runtime protocol note in request 5 system prompt, got %q", fifthSystemPrompt)
+ }
+ savedSession := store.sessions[session.ID]
+ for _, message := range savedSession.Messages {
content := renderPartsForTest(message.Parts)
- if strings.Contains(content, "") && strings.Contains(content, "MUST call todo_write") {
- foundHint = true
- break
+ if message.Role == providertypes.RoleSystem &&
+ strings.Contains(content, "[Runtime Control]") &&
+ strings.Contains(content, "task_completion") {
+ t.Fatalf("expected completion reminder to stay out of session transcript, found %q", content)
}
}
- if !foundHint {
- t.Fatalf("expected actionable acceptance continue hint, got messages: %+v", secondRequestMessages)
- }
events := collectRuntimeEvents(service.Events())
- assertStopReasonDecided(t, events, controlplane.StopReasonNoProgressAfterFinalIntercept, "")
- foundVerificationReason := false
- foundAcceptanceReason := false
- for _, event := range events {
- switch event.Type {
- case EventVerificationStarted:
- payload, ok := event.Payload.(VerificationStartedPayload)
- if ok && strings.TrimSpace(payload.CompletionBlockedReason) == string(controlplane.CompletionBlockedReasonPendingTodo) {
- foundVerificationReason = true
- }
- case EventAcceptanceDecided:
- payload, ok := event.Payload.(AcceptanceDecidedPayload)
- if ok && strings.TrimSpace(payload.CompletionBlockedReason) == string(controlplane.CompletionBlockedReasonPendingTodo) {
- foundAcceptanceReason = true
- }
- }
- }
- if !foundVerificationReason {
- t.Fatal("expected verification_started payload to include completion_blocked_reason=pending_todo")
- }
- if !foundAcceptanceReason {
- t.Fatal("expected acceptance_decided payload to include completion_blocked_reason=pending_todo")
- }
+ assertStopReasonDecided(t, events, controlplane.StopReasonMissingCompletionSignal, "")
}
func assertStopReasonDecided(t *testing.T, events []RuntimeEvent, wantReason controlplane.StopReason, wantDetail string) {
diff --git a/internal/runtime/runtime_snapshot.go b/internal/runtime/runtime_snapshot.go
index 4dae4b90..e76a7c69 100644
--- a/internal/runtime/runtime_snapshot.go
+++ b/internal/runtime/runtime_snapshot.go
@@ -5,7 +5,6 @@ import (
"strings"
"time"
- "neo-code/internal/runtime/decider"
runtimefacts "neo-code/internal/runtime/facts"
agentsession "neo-code/internal/session"
)
@@ -15,7 +14,6 @@ type RuntimeSnapshot struct {
RunID string `json:"run_id"`
SessionID string `json:"session_id"`
Phase string `json:"phase,omitempty"`
- TaskKind string `json:"task_kind,omitempty"`
UpdatedAt time.Time `json:"updated_at"`
Todos TodoSnapshot `json:"todos"`
Facts FactsSnapshot `json:"facts"`
@@ -32,15 +30,10 @@ type FactsSnapshot struct {
// DecisionSnapshot 是终态裁决快照。
type DecisionSnapshot struct {
- Status string `json:"status,omitempty"`
- StopReason string `json:"stop_reason,omitempty"`
- MissingFacts []decider.MissingFact `json:"missing_facts,omitempty"`
- RequiredNextActions []decider.RequiredAction `json:"required_next_actions,omitempty"`
- RequiredInput *decider.RequiredInput `json:"required_input,omitempty"`
- IntentHint string `json:"intent_hint,omitempty"`
- EffectiveTaskKind string `json:"effective_task_kind,omitempty"`
- UserVisibleSummary string `json:"user_visible_summary,omitempty"`
- InternalSummary string `json:"internal_summary,omitempty"`
+ Status string `json:"status,omitempty"`
+ StopReason string `json:"stop_reason,omitempty"`
+ Summary string `json:"summary,omitempty"`
+ Details []string `json:"details,omitempty"`
}
// SubAgentSnapshot 汇总子代理事实状态，避免客户端自行遍历事实结构。
@@ -84,16 +77,13 @@ func buildRuntimeSnapshot(state *runState) RuntimeSnapshot {
factsSnapshot = state.factsCollector.Snapshot()
}
- decisionSnapshot := DecisionSnapshot{
- Status: strings.TrimSpace(string(state.lastDeciderDecision.Status)),
- StopReason: strings.TrimSpace(state.lastDeciderDecision.StopReason),
- MissingFacts: append([]decider.MissingFact(nil), state.lastDeciderDecision.MissingFacts...),
- RequiredNextActions: append([]decider.RequiredAction(nil), state.lastDeciderDecision.RequiredNextActions...),
- RequiredInput: state.lastDeciderDecision.RequiredInput,
- IntentHint: strings.TrimSpace(string(state.lastDeciderDecision.IntentHint)),
- EffectiveTaskKind: strings.TrimSpace(string(state.lastDeciderDecision.EffectiveTaskKind)),
- UserVisibleSummary: strings.TrimSpace(state.lastDeciderDecision.UserVisibleSummary),
- InternalSummary: strings.TrimSpace(state.lastDeciderDecision.InternalSummary),
+ decisionSnapshot := DecisionSnapshot{}
+ if state.terminalSet || state.terminalStatus != "" || state.terminalStopReason != "" {
+ decisionSnapshot = DecisionSnapshot{
+ Status: strings.TrimSpace(string(state.terminalStatus)),
+ StopReason: strings.TrimSpace(string(state.terminalStopReason)),
+ Summary: strings.TrimSpace(state.terminalStopDetail),
+ }
}
pendingUserQuestion := clonePendingUserQuestion(state.pendingUserQuestion)
@@ -101,7 +91,6 @@ func buildRuntimeSnapshot(state *runState) RuntimeSnapshot {
RunID: strings.TrimSpace(state.runID),
SessionID: strings.TrimSpace(state.session.ID),
Phase: strings.TrimSpace(string(state.lifecycle)),
- TaskKind: strings.TrimSpace(string(state.taskKind)),
UpdatedAt: time.Now(),
Todos: todoSnapshot,
Facts: FactsSnapshot{
@@ -208,7 +197,6 @@ func (s *Service) GetRuntimeSnapshot(ctx context.Context, sessionID string) (Run
snapshot := RuntimeSnapshot{
SessionID: normalizedSessionID,
Phase: "",
- TaskKind: string(decider.TaskKindMixed),
UpdatedAt: session.UpdatedAt,
Todos: buildTodoSnapshotFromItems(session.ListTodos()),
Facts: FactsSnapshot{
diff --git a/internal/runtime/runtime_snapshot_test.go b/internal/runtime/runtime_snapshot_test.go
index e297db56..390b8851 100644
--- a/internal/runtime/runtime_snapshot_test.go
+++ b/internal/runtime/runtime_snapshot_test.go
@@ -6,7 +6,6 @@ import (
"testing"
"time"
- "neo-code/internal/runtime/decider"
runtimefacts "neo-code/internal/runtime/facts"
agentsession "neo-code/internal/session"
)
@@ -80,9 +79,6 @@ func TestGetRuntimeSnapshotBranches(t *testing.T) {
if got.SessionID != session.ID {
t.Fatalf("session id = %q, want %q", got.SessionID, session.ID)
}
- if got.TaskKind != string(decider.TaskKindMixed) {
- t.Fatalf("task kind = %q, want %q", got.TaskKind, decider.TaskKindMixed)
- }
if got.Facts.RuntimeFacts.Progress.ObservedFactCount != 0 {
t.Fatalf("unexpected facts snapshot: %+v", got.Facts.RuntimeFacts)
}
diff --git a/internal/runtime/runtime_test.go b/internal/runtime/runtime_test.go
index 5e1ca56d..f87bede5 100644
--- a/internal/runtime/runtime_test.go
+++ b/internal/runtime/runtime_test.go
@@ -618,13 +618,14 @@ func (s *blockingLoadStore) CleanupExpiredSessions(ctx context.Context, maxAge t
}
type scriptedProvider struct {
- name string
- streams [][]providertypes.StreamEvent
- responses []scriptedResponse
- requests []providertypes.GenerateRequest
- callCount int
- estimateFn func(ctx context.Context, req providertypes.GenerateRequest) (providertypes.BudgetEstimate, error)
- chatFn func(ctx context.Context, req providertypes.GenerateRequest, events chan<- providertypes.StreamEvent) error
+ name string
+ streams [][]providertypes.StreamEvent
+ responses []scriptedResponse
+ requests []providertypes.GenerateRequest
+ callCount int
+ requireExplicitCompletion bool
+ estimateFn func(ctx context.Context, req providertypes.GenerateRequest) (providertypes.BudgetEstimate, error)
+ chatFn func(ctx context.Context, req providertypes.GenerateRequest, events chan<- providertypes.StreamEvent) error
}
func (p *scriptedProvider) EstimateInputTokens(
@@ -654,10 +655,32 @@ func (p *scriptedProvider) Generate(ctx context.Context, req providertypes.Gener
p.callCount++
if p.chatFn != nil {
- return p.chatFn(ctx, req, events)
+ stream, err := p.collectChatFnStream(ctx, req)
+ if p.shouldInjectDefaultCompletionForStream(stream) {
+ select {
+ case events <- providertypes.NewTextDeltaStreamEvent("{\"task_completion\":{\"completed\":true}}\n"):
+ case <-ctx.Done():
+ return ctx.Err()
+ }
+ }
+ for _, event := range stream {
+ select {
+ case events <- event:
+ case <-ctx.Done():
+ return ctx.Err()
+ }
+ }
+ return err
}
if callIndex < len(p.streams) {
+ if p.shouldInjectDefaultCompletionForStream(p.streams[callIndex]) {
+ select {
+ case events <- providertypes.NewTextDeltaStreamEvent("{\"task_completion\":{\"completed\":true}}\n"):
+ case <-ctx.Done():
+ return ctx.Err()
+ }
+ }
for _, event := range p.streams[callIndex] {
select {
case events <- event:
@@ -675,6 +698,7 @@ func (p *scriptedProvider) Generate(ctx context.Context, req providertypes.Gener
}
if callIndex < len(p.responses) {
response := p.responses[callIndex]
+ response.Message = p.withDefaultCompletionSignal(response.Message)
for index, toolCall := range response.Message.ToolCalls {
select {
case events <- providertypes.NewToolCallStartStreamEvent(index, toolCall.ID, toolCall.Name):
@@ -704,6 +728,68 @@ func (p *scriptedProvider) Generate(ctx context.Context, req providertypes.Gener
return nil
}
+// withDefaultCompletionSignal 让旧测试脚本中的普通最终回复满足新的 task_completion 协议。
+func (p *scriptedProvider) withDefaultCompletionSignal(message providertypes.Message) providertypes.Message {
+ if p.requireExplicitCompletion || len(message.ToolCalls) > 0 {
+ return message
+ }
+ text := renderPartsForTest(message.Parts)
+ if strings.TrimSpace(text) == "" ||
+ strings.Contains(text, `"task_completion"`) ||
+ strings.Contains(text, `"plan_spec"`) {
+ return message
+ }
+ message.Parts = []providertypes.ContentPart{
+ providertypes.NewTextPart("{\"task_completion\":{\"completed\":true}}\n" + text),
+ }
+ return message
+}
+
+// collectChatFnStream 收集自定义测试 provider 的流事件，便于统一补齐完成信号。
+func (p *scriptedProvider) collectChatFnStream(
+ ctx context.Context,
+ req providertypes.GenerateRequest,
+) ([]providertypes.StreamEvent, error) {
+ proxy := make(chan providertypes.StreamEvent, 1024)
+ errCh := make(chan error, 1)
+ go func() {
+ errCh <- p.chatFn(ctx, req, proxy)
+ close(proxy)
+ }()
+ stream := make([]providertypes.StreamEvent, 0)
+ for event := range proxy {
+ stream = append(stream, event)
+ }
+ return stream, <-errCh
+}
+
+// shouldInjectDefaultCompletionForStream 为旧测试中的纯文本流补齐完成信号，工具调用流和显式完成流保持原样。
+func (p *scriptedProvider) shouldInjectDefaultCompletionForStream(stream []providertypes.StreamEvent) bool {
+ if p.requireExplicitCompletion {
+ return false
+ }
+ hasText := false
+ for _, event := range stream {
+ switch event.Type {
+ case providertypes.StreamEventToolCallStart, providertypes.StreamEventToolCallDelta:
+ return false
+ case providertypes.StreamEventTextDelta:
+ if event.TextDelta == nil {
+ continue
+ }
+ text := strings.TrimSpace(event.TextDelta.Text)
+ if text == "" {
+ continue
+ }
+ hasText = true
+ if strings.Contains(text, `"task_completion"`) || strings.Contains(text, `"plan_spec"`) {
+ return false
+ }
+ }
+ }
+ return hasText
+}
+
// streamContainsMessageDone 判断测试流中是否已显式包含结束事件，避免辅助 provider 重复补发 message_done。
func streamContainsMessageDone(events []providertypes.StreamEvent) bool {
for _, event := range events {
@@ -1220,8 +1306,8 @@ func TestServiceRunSchedulesMemoExtractionFromCurrentRunBoundary(t *testing.T) {
if len(messages) != 2 {
t.Fatalf("scheduled messages = %#v, want current run user+assistant only", messages)
}
- if renderPartsForVerification(messages[0].Parts) != "new user" ||
- renderPartsForVerification(messages[1].Parts) != "new final" {
+ if renderPartsForTest(messages[0].Parts) != "new user" ||
+ renderPartsForTest(messages[1].Parts) != "new final" {
t.Fatalf("scheduled messages crossed run boundary: %#v", messages)
}
}
@@ -2479,7 +2565,7 @@ func TestServiceRunErrorPaths(t *testing.T) {
})
}
responses = append(responses, scriptedResponse{
- Message: providertypes.Message{Parts: []providertypes.ContentPart{providertypes.NewTextPart("done after many cycles")}},
+ Message: providertypes.Message{Parts: []providertypes.ContentPart{providertypes.NewTextPart("{\"task_completion\":{\"completed\":true}}\ndone after many cycles")}},
FinishReason: "stop",
})
return &scriptedProvider{responses: responses}
@@ -2489,14 +2575,7 @@ func TestServiceRunErrorPaths(t *testing.T) {
assert: func(t *testing.T, store *memoryStore, scripted *scriptedProvider, tool *stubTool) {
t.Helper()
if scripted.callCount != 10 {
- t.Fatalf("expected 10 provider calls without loop cap, got %d", scripted.callCount)
- }
- session := onlySession(t, store)
- if got := len(session.Messages); got != 20 {
- t.Fatalf("expected 20 persisted messages after 9 tool cycles and final answer, got %d", got)
- }
- if renderPartsForTest(session.Messages[len(session.Messages)-1].Parts) != "done after many cycles" {
- t.Fatalf("expected final assistant reply to be persisted, got %+v", session.Messages[len(session.Messages)-1])
+ t.Fatalf("expected 10 provider calls (9 tool cycles + 1 completion), got %d", scripted.callCount)
}
},
},
@@ -3935,10 +4014,10 @@ func TestServiceRunPlanModePersistsDraftPlan(t *testing.T) {
if builder.lastInput.CurrentPlan != nil {
t.Fatalf("expected initial plan-mode build input to have nil CurrentPlan")
}
- if len(saved.Messages) != 2 {
- t.Fatalf("message count = %d, want 2", len(saved.Messages))
+ if len(saved.Messages) != 3 {
+ t.Fatalf("message count = %d, want 3", len(saved.Messages))
}
- if got := renderPartsForTest(saved.Messages[1].Parts); !strings.Contains(got, "目标") {
+ if got := renderPartsForTest(saved.Messages[2].Parts); !strings.Contains(got, "目标") {
t.Fatalf("expected rendered plan content, got %q", got)
}
}
@@ -3985,10 +4064,10 @@ func TestServiceRunPlanModeShowsExplanationTextOutsidePlanningJSON(t *testing.T)
if saved.CurrentPlan == nil || saved.CurrentPlan.Spec.Goal != "Preserve prose around planning JSON" {
t.Fatalf("expected current plan to be updated, got %+v", saved.CurrentPlan)
}
- if len(saved.Messages) != 2 {
- t.Fatalf("message count = %d, want 2", len(saved.Messages))
+ if len(saved.Messages) != 3 {
+ t.Fatalf("message count = %d, want 3", len(saved.Messages))
}
- got := renderPartsForTest(saved.Messages[1].Parts)
+ got := renderPartsForTest(saved.Messages[2].Parts)
if strings.Contains(got, "\"plan_spec\"") {
t.Fatalf("expected persisted assistant text to strip planning JSON, got %q", got)
}
@@ -4012,12 +4091,12 @@ func TestServiceRunPlanModeKeepsExistingPlanWhenPlanSpecIsInvalid(t *testing.T)
Spec: agentsession.PlanSpec{
Goal: "Keep previous plan",
Steps: []string{"existing step"},
- Verify: []string{"existing verify"},
+ Verify: acceptText("existing verify"),
},
Summary: agentsession.SummaryView{
Goal: "Keep previous plan",
KeySteps: []string{"existing step"},
- Verify: []string{"existing verify"},
+ Verify: acceptText("existing verify"),
},
}
seed.LastFullPlanRevision = 2
@@ -4030,7 +4109,8 @@ func TestServiceRunPlanModeKeepsExistingPlanWhenPlanSpecIsInvalid(t *testing.T)
{
Message: providertypes.Message{
Role: providertypes.RoleAssistant,
- Parts: []providertypes.ContentPart{providertypes.NewTextPart(`{
+ Parts: []providertypes.ContentPart{providertypes.NewTextPart(`{"task_completion":{"completed":true}}
+{
"plan_spec": {
"goal": "",
"steps": ["bad update"],
@@ -4121,8 +4201,8 @@ func TestServiceRunBuildModeDoesNotRequireCurrentPlan(t *testing.T) {
if saved.CurrentPlan != nil {
t.Fatalf("expected build mode to complete without CurrentPlan, got %+v", saved.CurrentPlan)
}
- if builder.callCount != 1 {
- t.Fatalf("builder call count = %d, want 1", builder.callCount)
+ if builder.callCount != 2 {
+ t.Fatalf("builder call count = %d, want 2", builder.callCount)
}
if builder.builds[0].PlanStage != planStageBuildExecute {
t.Fatalf("PlanStage = %q, want %q", builder.builds[0].PlanStage, planStageBuildExecute)
@@ -4146,7 +4226,8 @@ func TestServiceRunPlanModeInjectsFullPlanOnNextTurnAfterDraftCreation(t *testin
{
Message: providertypes.Message{
Role: providertypes.RoleAssistant,
- Parts: []providertypes.ContentPart{providertypes.NewTextPart(`{
+ Parts: []providertypes.ContentPart{providertypes.NewTextPart(`{"task_completion":{"completed":true}}
+{
"plan_spec": {
"goal": "Introduce plan mode",
"steps": ["persist plan state"],
@@ -4237,7 +4318,7 @@ func TestServiceRunPlanModeUsesSummaryViewForAlignedPlanTurn(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "Keep planning aligned",
Steps: []string{"summarize current plan"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
Todos: []agentsession.TodoItem{
{ID: "todo-aligned", Content: "summarize current plan", Status: agentsession.TodoStatusPending},
},
@@ -4245,7 +4326,7 @@ func TestServiceRunPlanModeUsesSummaryViewForAlignedPlanTurn(t *testing.T) {
Summary: agentsession.SummaryView{
Goal: "Keep planning aligned",
KeySteps: []string{"summarize current plan"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
ActiveTodoIDs: []string{"todo-aligned"},
},
}
@@ -4313,7 +4394,7 @@ func TestServiceRunBuildModeInjectsFullPlanForUnalignedExistingPlan(t *testing.T
Spec: agentsession.PlanSpec{
Goal: "Resume build execution",
Steps: []string{"resume implementation"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
Todos: []agentsession.TodoItem{
{ID: "todo-restored", Content: "resume implementation", Status: agentsession.TodoStatusPending},
},
@@ -4321,7 +4402,7 @@ func TestServiceRunBuildModeInjectsFullPlanForUnalignedExistingPlan(t *testing.T
Summary: agentsession.SummaryView{
Goal: "Resume build execution",
KeySteps: []string{"resume implementation"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
ActiveTodoIDs: []string{"todo-restored"},
},
}
@@ -4379,7 +4460,7 @@ func TestServiceRunBuildModeUsesSummaryViewForAlignedExecuteTurn(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "Execute aligned build",
Steps: []string{"continue implementation"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
Todos: []agentsession.TodoItem{
{ID: "todo-build-aligned", Content: "continue implementation", Status: agentsession.TodoStatusPending},
},
@@ -4387,7 +4468,7 @@ func TestServiceRunBuildModeUsesSummaryViewForAlignedExecuteTurn(t *testing.T) {
Summary: agentsession.SummaryView{
Goal: "Execute aligned build",
KeySteps: []string{"continue implementation"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
ActiveTodoIDs: []string{"todo-build-aligned"},
},
}
@@ -4441,7 +4522,7 @@ func TestServiceRunBuildModeInjectsFullPlanWhenSummaryIsUnusable(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "Follow full plan when summary is missing",
Steps: []string{"review whole plan"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
Todos: []agentsession.TodoItem{
{ID: "todo-full-fallback", Content: "review whole plan", Status: agentsession.TodoStatusPending},
},
@@ -4498,12 +4579,12 @@ func TestServiceApproveCurrentPlanTriggersOneFullPlanAlignment(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "批准并执行当前计划",
Steps: []string{"继续实现"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
},
Summary: agentsession.SummaryView{
Goal: "批准并执行当前计划",
KeySteps: []string{"继续实现"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
},
}
seed.LastFullPlanRevision = 4
@@ -4607,7 +4688,7 @@ func TestServiceApproveCurrentPlanTrimsSessionID(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "trim session id before load",
Steps: []string{"step one"},
- Verify: []string{"verify one"},
+ Verify: acceptText("verify one"),
},
}
if _, err := store.CreateSession(context.Background(), createSessionInputFromSession(seed)); err != nil {
@@ -4647,12 +4728,12 @@ func TestServiceRunBuildModeIgnoresPlanningJSON(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "保持旧计划不被覆盖",
Steps: []string{"旧步骤"},
- Verify: []string{"旧验证"},
+ Verify: acceptText("旧验证"),
},
Summary: agentsession.SummaryView{
Goal: "保持旧计划不被覆盖",
KeySteps: []string{"旧步骤"},
- Verify: []string{"旧验证"},
+ Verify: acceptText("旧验证"),
},
}
seed.LastFullPlanRevision = 1
@@ -4665,7 +4746,8 @@ func TestServiceRunBuildModeIgnoresPlanningJSON(t *testing.T) {
{
Message: providertypes.Message{
Role: providertypes.RoleAssistant,
- Parts: []providertypes.ContentPart{providertypes.NewTextPart(`{
+ Parts: []providertypes.ContentPart{providertypes.NewTextPart(`{"task_completion":{"completed":true}}
+{
"plan_spec": {
"goal": "不应在 build 中落库",
"steps": ["错误改写计划"],
@@ -4717,12 +4799,12 @@ func TestServiceRunCompletedPlanRequestsOneFinalFullReview(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "完成计划后仍需一次全文确认",
Steps: []string{"收尾"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
},
Summary: agentsession.SummaryView{
Goal: "完成计划后仍需一次全文确认",
KeySteps: []string{"收尾"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
},
}
seed.LastFullPlanRevision = 2
@@ -4805,12 +4887,12 @@ func TestServiceCompactMarksPlanContextDirty(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "compact 后重对齐计划",
Steps: []string{"压缩历史"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
},
Summary: agentsession.SummaryView{
Goal: "compact 后重对齐计划",
KeySteps: []string{"压缩历史"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
},
}
if _, err := store.CreateSession(context.Background(), createSessionInputFromSession(session)); err != nil {
@@ -4856,12 +4938,12 @@ func TestServiceRunCompactedSessionRequestsRestoreAlignment(t *testing.T) {
Spec: agentsession.PlanSpec{
Goal: "compact 恢复后重新对齐计划",
Steps: []string{"继续执行"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
},
Summary: agentsession.SummaryView{
Goal: "compact 恢复后重新对齐计划",
KeySteps: []string{"继续执行"},
- Verify: []string{"go test ./internal/runtime"},
+ Verify: acceptText("go test ./internal/runtime"),
},
}
seed.LastFullPlanRevision = 1
diff --git a/internal/runtime/state.go b/internal/runtime/state.go
index 9e65e5a1..2b6f6609 100644
--- a/internal/runtime/state.go
+++ b/internal/runtime/state.go
@@ -6,7 +6,6 @@ import (
providertypes "neo-code/internal/provider/types"
"neo-code/internal/runtime/controlplane"
- "neo-code/internal/runtime/decider"
runtimefacts "neo-code/internal/runtime/facts"
"neo-code/internal/security"
agentsession "neo-code/internal/session"
@@ -14,56 +13,52 @@ import (
// runState 汇总单次 Run 生命周期内会变化的会话与计量状态。
type runState struct {
- mu sync.Mutex
- runID string
- runToken uint64
- session agentsession.Session
- effectiveWorkdir string
- compactCount int
- reactiveCompactAttempts int
- memoRunMessages []providertypes.Message
- rememberedThisRun bool
- planningEnabled bool
- taskID string
- agentID string
- capabilityToken *security.CapabilityToken
- nextAttemptSeq int
- turn int
- baseLifecycle controlplane.RunState
- lifecycle controlplane.RunState
- waitingPermissionCount int
- compactingCount int
- stopEmitted bool
- budgetExceeded bool
- maxTurnsReached bool
- maxTurnsLimit int
- finalInterceptStreak int
- pendingFinalProgress bool
- mustUseToolAfterFinalContinue bool
- noToolAfterFinalContinueStreak int
- lastAcceptanceBlockedReason string
- taskKind decider.TaskKind
- userGoal string
- factsCollector *runtimefacts.Collector
- lastDeciderDecision decider.Decision
- terminalStatus controlplane.TerminalStatus
- terminalStopReason controlplane.StopReason
- terminalStopDetail string
- terminalSet bool
- hasUnknownUsage bool
- completion controlplane.CompletionState
- progress controlplane.ProgressState
- lastEndOfTurnCheckpointID string
- runCheckpointID string
- hasRunWorkspaceWrite bool
- hookAnnotations []string
- hookNotifications []queuedHookNotification
- hookNotificationSeen map[string]time.Time
- hookNotificationOmitted int
- reportedMissingSkills map[string]struct{}
- thinkingOverride *ThinkingOverride
- pendingUserQuestion *UserQuestionRequestedPayload
- disableTools bool
+ mu sync.Mutex
+ runID string
+ runToken uint64
+ session agentsession.Session
+ effectiveWorkdir string
+ compactCount int
+ reactiveCompactAttempts int
+ rememberedThisRun bool
+ planningEnabled bool
+ taskID string
+ agentID string
+ capabilityToken *security.CapabilityToken
+ memoRunMessages []providertypes.Message
+ nextAttemptSeq int
+ turn int
+ baseLifecycle controlplane.RunState
+ lifecycle controlplane.RunState
+ waitingPermissionCount int
+ compactingCount int
+ stopEmitted bool
+ budgetExceeded bool
+ maxTurnsReached bool
+ maxTurnsLimit int
+ userGoal string
+ missingCompletionSignalStreak int
+ pendingSystemReminder string
+ toolTimeoutBackoff map[string]int
+ factsCollector *runtimefacts.Collector
+ terminalStatus controlplane.TerminalStatus
+ terminalStopReason controlplane.StopReason
+ terminalStopDetail string
+ terminalSet bool
+ hasUnknownUsage bool
+ completion controlplane.CompletionState
+ progress controlplane.ProgressState
+ lastEndOfTurnCheckpointID string
+ runCheckpointID string
+ hasRunWorkspaceWrite bool
+ hookAnnotations []string
+ hookNotifications []queuedHookNotification
+ hookNotificationSeen map[string]time.Time
+ hookNotificationOmitted int
+ reportedMissingSkills map[string]struct{}
+ thinkingOverride *ThinkingOverride
+ pendingUserQuestion *UserQuestionRequestedPayload
+ disableTools bool
}
// newRunState 基于持久化会话创建一次运行的内存状态镜像。
@@ -74,9 +69,9 @@ func newRunState(runID string, session agentsession.Session) runState {
nextAttemptSeq: 1,
completion: controlplane.CompletionState{TodoOnlyTaskCandidate: true},
reportedMissingSkills: make(map[string]struct{}),
- taskKind: "",
factsCollector: runtimefacts.NewCollector(),
hookNotificationSeen: make(map[string]time.Time),
+ toolTimeoutBackoff: make(map[string]int),
}
}
diff --git a/internal/runtime/task_kind.go b/internal/runtime/task_kind.go
deleted file mode 100644
index fa28d24b..00000000
--- a/internal/runtime/task_kind.go
+++ /dev/null
@@ -1,27 +0,0 @@
-package runtime
-
-import (
- "strings"
-
- providertypes "neo-code/internal/provider/types"
- "neo-code/internal/runtime/decider"
-)
-
-// inferTaskKindFromInput 基于用户输入文本推断任务类型，避免将简单状态任务误判为通用写入验证任务。
-func inferTaskKindFromInput(parts []providertypes.ContentPart) decider.TaskKind {
- var builder strings.Builder
- for _, part := range parts {
- if part.Kind != providertypes.ContentPartText {
- continue
- }
- text := strings.TrimSpace(part.Text)
- if text == "" {
- continue
- }
- if builder.Len() > 0 {
- builder.WriteString("\n")
- }
- builder.WriteString(text)
- }
- return decider.InferTaskKind(builder.String())
-}
diff --git a/internal/runtime/thinking_callprovider_test.go b/internal/runtime/thinking_callprovider_test.go
index 0a1f2ddf..8f451853 100644
--- a/internal/runtime/thinking_callprovider_test.go
+++ b/internal/runtime/thinking_callprovider_test.go
@@ -13,7 +13,7 @@ import (
func TestCallProviderRetriesWithThinkingDowngrade(t *testing.T) {
t.Parallel()
- scripted := &scriptedProvider{}
+ scripted := &scriptedProvider{requireExplicitCompletion: true}
scripted.chatFn = func(ctx context.Context, req providertypes.GenerateRequest, events chan<- providertypes.StreamEvent) error {
if len(scripted.requests) == 1 {
return errors.Join(provider.ErrThinkingNotSupported, errors.New("upstream rejected thinking"))
@@ -57,6 +57,7 @@ func TestCallProviderEmitsThinkingDeltaEvent(t *testing.T) {
t.Parallel()
scripted := &scriptedProvider{
+ requireExplicitCompletion: true,
chatFn: func(ctx context.Context, req providertypes.GenerateRequest, events chan<- providertypes.StreamEvent) error {
events <- providertypes.NewThinkingDeltaStreamEvent("plan")
events <- providertypes.NewTextDeltaStreamEvent("answer")
diff --git a/internal/runtime/todo_bootstrap.go b/internal/runtime/todo_bootstrap.go
new file mode 100644
index 00000000..9e1d0cd3
--- /dev/null
+++ b/internal/runtime/todo_bootstrap.go
@@ -0,0 +1,92 @@
+package runtime
+
+import (
+ "context"
+
+ agentsession "neo-code/internal/session"
+)
+
+const todoBootstrapRequiredReason = "todo_bootstrap_required"
+
+const todoBootstrapRequiredReminder = `[Runtime Control]
+
+todo_bootstrap_required: This build run has no current plan and no active todos.
+
+Before project analysis, documentation writing, code changes, multi-step debugging, or verification work, call todo_write with action=plan or action=add to create required todos for this run.
+
+Do not update or complete old todo IDs that are not present in the current Todo State.`
+
+// maybeAppendTodoBootstrapReminder 在 direct build 缺少 plan/todo 时注入一次结构化提醒。
+func (s *Service) maybeAppendTodoBootstrapReminder(ctx context.Context, state *runState) error {
+ if !shouldInjectTodoBootstrapReminder(state) {
+ return nil
+ }
+ return s.appendSystemMessageAndSave(ctx, state, todoBootstrapRequiredReminder)
+}
+
+// shouldInjectTodoBootstrapReminder 判断本轮 build 是否需要先创建当前 run 的 todo。
+func shouldInjectTodoBootstrapReminder(state *runState) bool {
+ if state == nil || state.disableTools || !state.planningEnabled {
+ return false
+ }
+ state.mu.Lock()
+ session := state.session
+ state.mu.Unlock()
+
+ if agentsession.NormalizeAgentMode(session.AgentMode) != agentsession.AgentModeBuild {
+ return false
+ }
+ if hasActivePlanForTodoBootstrap(session.CurrentPlan) || len(session.Todos) > 0 {
+ return false
+ }
+ return true
+}
+
+// hasActivePlanForTodoBootstrap 判断当前 plan 是否仍可为 build 继承 todo 所有权。
+func hasActivePlanForTodoBootstrap(plan *agentsession.PlanArtifact) bool {
+ if plan == nil {
+ return false
+ }
+ switch agentsession.NormalizePlanStatus(plan.Status) {
+ case agentsession.PlanStatusDraft, agentsession.PlanStatusApproved:
+ return true
+ default:
+ return false
+ }
+}
+
+const planBootstrapRequiredReason = "plan_bootstrap_required"
+
+const planBootstrapRequiredReminder = `[Runtime Control]
+
+plan_bootstrap_required: You are in plan mode but no current plan exists.
+
+Before research, analysis, or conversational response, you MUST complete the following:
+
+1. Research the codebase as needed using read-only tools.
+2. Output a JSON object containing "plan_spec" and "summary_candidate" that defines the current plan.
+3. plan_spec.todos must be non-empty — include major actionable items with unique IDs and status "pending".
+
+Do not end this turn without producing a plan.`
+
+// maybeAppendPlanBootstrapReminder 在 plan 模式缺少 CurrentPlan 时注入一次结构化提醒。
+func (s *Service) maybeAppendPlanBootstrapReminder(ctx context.Context, state *runState) error {
+ if !shouldInjectPlanBootstrapReminder(state) {
+ return nil
+ }
+ return s.appendSystemMessageAndSave(ctx, state, planBootstrapRequiredReminder)
+}
+
+// shouldInjectPlanBootstrapReminder 判断本轮 plan 模式是否需要先创建 plan。
+func shouldInjectPlanBootstrapReminder(state *runState) bool {
+ if state == nil || state.disableTools || !state.planningEnabled {
+ return false
+ }
+ if resolvePlanningStageForState(state) != planStagePlan {
+ return false
+ }
+ state.mu.Lock()
+ plan := state.session.CurrentPlan
+ state.mu.Unlock()
+ return plan == nil
+}
diff --git a/internal/runtime/todo_bootstrap_test.go b/internal/runtime/todo_bootstrap_test.go
new file mode 100644
index 00000000..61582649
--- /dev/null
+++ b/internal/runtime/todo_bootstrap_test.go
@@ -0,0 +1,164 @@
+package runtime
+
+import (
+ "context"
+ "strings"
+ "testing"
+
+ agentcontext "neo-code/internal/context"
+ providertypes "neo-code/internal/provider/types"
+ agentsession "neo-code/internal/session"
+ "neo-code/internal/tools"
+)
+
+func TestShouldInjectTodoBootstrapReminder(t *testing.T) {
+ t.Parallel()
+
+ required := true
+ cases := []struct {
+ name string
+ state runState
+ want bool
+ }{
+ {
+ name: "direct build without plan or todos injects",
+ state: runState{
+ session: agentsession.Session{AgentMode: agentsession.AgentModeBuild},
+ planningEnabled: true,
+ },
+ want: true,
+ },
+ {
+ name: "direct build does not inspect user text",
+ state: runState{
+ session: agentsession.Session{AgentMode: agentsession.AgentModeBuild},
+ userGoal: "你好",
+ planningEnabled: true,
+ },
+ want: true,
+ },
+ {
+ name: "active plan skips",
+ state: runState{
+ session: agentsession.Session{
+ AgentMode: agentsession.AgentModeBuild,
+ CurrentPlan: &agentsession.PlanArtifact{
+ Status: agentsession.PlanStatusApproved,
+ },
+ },
+ userGoal: "请分析项目并写文档",
+ planningEnabled: true,
+ },
+ want: false,
+ },
+ {
+ name: "existing todo skips",
+ state: runState{
+ session: agentsession.Session{
+ AgentMode: agentsession.AgentModeBuild,
+ Todos: []agentsession.TodoItem{{
+ ID: "todo-1",
+ Content: "existing",
+ Status: agentsession.TodoStatusPending,
+ Required: &required,
+ }},
+ },
+ userGoal: "请分析项目并写文档",
+ planningEnabled: true,
+ },
+ want: false,
+ },
+ {
+ name: "plan mode skips",
+ state: runState{
+ session: agentsession.Session{AgentMode: agentsession.AgentModePlan},
+ userGoal: "请分析项目并写文档",
+ planningEnabled: true,
+ },
+ want: false,
+ },
+ {
+ name: "legacy non planning run skips",
+ state: runState{
+ session: agentsession.Session{AgentMode: agentsession.AgentModeBuild},
+ userGoal: "edit file",
+ },
+ want: false,
+ },
+ }
+
+ for _, tc := range cases {
+ tc := tc
+ t.Run(tc.name, func(t *testing.T) {
+ t.Parallel()
+ got := shouldInjectTodoBootstrapReminder(&tc.state)
+ if got != tc.want {
+ t.Fatalf("shouldInjectTodoBootstrapReminder() = %v, want %v", got, tc.want)
+ }
+ })
+ }
+}
+
+func TestServiceRunDirectBuildInjectsTodoBootstrapReminder(t *testing.T) {
+ t.Parallel()
+
+ manager := newRuntimeConfigManager(t)
+ store := newMemoryStore()
+ builder := &stubContextBuilder{
+ buildFn: func(ctx context.Context, input agentcontext.BuildInput) (agentcontext.BuildResult, error) {
+ return agentcontext.BuildResult{
+ SystemPrompt: "stub system prompt",
+ Messages: append([]providertypes.Message(nil), input.Messages...),
+ }, nil
+ },
+ }
+ scripted := &scriptedProvider{
+ responses: []scriptedResponse{
+ {
+ Message: providertypes.Message{
+ Role: providertypes.RoleAssistant,
+ Parts: []providertypes.ContentPart{
+ providertypes.NewTextPart(`{"task_completion":{"completed":true}}` + "\n完成。"),
+ },
+ },
+ FinishReason: "stop",
+ },
+ },
+ }
+
+ service := NewWithFactory(manager, tools.NewRegistry(), store, &scriptedProviderFactory{provider: scripted}, builder)
+ if err := service.Run(context.Background(), UserInput{
+ RunID: "run-direct-build-todo-bootstrap",
+ Mode: string(agentsession.AgentModeBuild),
+ Parts: []providertypes.ContentPart{providertypes.NewTextPart("请分析项目并写文档")},
+ }); err != nil {
+ t.Fatalf("Run() error = %v", err)
+ }
+
+ saved := onlySession(t, store)
+ foundPersistedReminder := false
+ for _, message := range saved.Messages {
+ if message.Role == providertypes.RoleSystem &&
+ strings.Contains(renderPartsForTest(message.Parts), todoBootstrapRequiredReason) {
+ foundPersistedReminder = true
+ break
+ }
+ }
+ if !foundPersistedReminder {
+ t.Fatalf("expected persisted todo bootstrap reminder, messages=%+v", saved.Messages)
+ }
+ if len(scripted.requests) == 0 {
+ t.Fatalf("expected provider request")
+ }
+ foundRequestReminder := false
+ for _, message := range scripted.requests[0].Messages {
+ if message.Role == providertypes.RoleSystem &&
+ strings.Contains(renderPartsForTest(message.Parts), todoBootstrapRequiredReason) {
+ foundRequestReminder = true
+ break
+ }
+ }
+ if !foundRequestReminder {
+ t.Fatalf("expected provider request to include todo bootstrap reminder, messages=%+v", scripted.requests[0].Messages)
+ }
+}
diff --git a/internal/runtime/todo_run_boundary.go b/internal/runtime/todo_run_boundary.go
index 38dc735a..5df46f46 100644
--- a/internal/runtime/todo_run_boundary.go
+++ b/internal/runtime/todo_run_boundary.go
@@ -2,10 +2,11 @@ package runtime
import (
"context"
- "strings"
+ "reflect"
"time"
runtimefacts "neo-code/internal/runtime/facts"
+ agentsession "neo-code/internal/session"
)
// resetTodosForUserRun 清空新用户 Run 的当前 Todo 状态，避免上一任务遗留的 open todo 阻塞本轮验收。
@@ -13,19 +14,18 @@ func (s *Service) resetTodosForUserRun(ctx context.Context, state *runState) err
if s == nil || state == nil {
return nil
}
- if !shouldResetTodosForUserRun(state.userGoal) {
- return nil
- }
state.mu.Lock()
- if len(state.session.Todos) == 0 {
+ currentTodos := cloneTodosForPersistence(state.session.Todos)
+ nextTodos, reason := todosForUserRunBoundary(state.session, currentTodos)
+ if reflect.DeepEqual(currentTodos, nextTodos) {
state.mu.Unlock()
return nil
}
- state.session.Todos = nil
+ state.session.Todos = nextTodos
state.session.UpdatedAt = time.Now()
if state.factsCollector != nil {
- state.factsCollector.ApplyTodoSnapshot(runtimefacts.TodoSummaryLike{})
+ state.factsCollector.ApplyTodoSnapshot(todoSummaryLikeForItems(nextTodos))
}
sessionSnapshot := cloneSessionForPersistence(state.session)
state.mu.Unlock()
@@ -34,48 +34,51 @@ func (s *Service) resetTodosForUserRun(ctx context.Context, state *runState) err
return err
}
- payload := buildTodoEventPayload(state, "reset", "new_user_run")
+ payload := buildTodoEventPayload(state, "reset", reason)
s.emitRunScoped(ctx, EventTodoSnapshotUpdated, state, payload)
s.emitRuntimeSnapshotUpdated(ctx, state, "todo_reset")
return nil
}
-// shouldResetTodosForUserRun 判断本轮用户输入是否应开启新的 Todo 边界。
-// 策略：默认保留旧 Todo，由 prompt 层 stale_todo_reminder 引导模型自行清理；
-// 仅当用户输入极少且明确的"全新任务"表达时，才主动清空，避免硬编码过度覆盖。
-func shouldResetTodosForUserRun(userGoal string) bool {
- goal := strings.ToLower(strings.TrimSpace(userGoal))
- if goal == "" {
- return false
+// todosForUserRunBoundary 返回新 Run 应继承的 todo 集合；active plan 只继承 plan-owned todo。
+func todosForUserRunBoundary(session agentsession.Session, todos []agentsession.TodoItem) ([]agentsession.TodoItem, string) {
+ if shouldResetTodosForUserRun(session) {
+ return nil, "new_user_run"
+ }
+ return selectPlanOwnedTodos(session.CurrentPlan, todos), "plan_owned_prune"
+}
+
+// shouldResetTodosForUserRun 根据 PlanArtifact 生命周期判断本轮是否开启新的 Todo 边界。
+func shouldResetTodosForUserRun(session agentsession.Session) bool {
+ if session.CurrentPlan == nil {
+ return true
}
- goal = strings.TrimRight(goal, " 。.!！?？,，;；~～")
- if goal == "" {
+ switch agentsession.NormalizePlanStatus(session.CurrentPlan.Status) {
+ case agentsession.PlanStatusDraft, agentsession.PlanStatusApproved:
return false
+ case agentsession.PlanStatusCompleted:
+ return true
+ default:
+ return true
}
- return isExplicitNewTaskIntent(goal)
}
-// newTaskChineseKeywords 中文明确新任务关键词，仅含完全无歧义的表达。
-var newTaskChineseKeywords = []string{"新任务", "换个任务", "换任务", "新需求"}
-
-// newTaskEnglishKeywords 英文明确新任务关键词，仅含完全无歧义的表达。
-var newTaskEnglishKeywords = []string{"new task", "different task", "switch task"}
-
-// isExplicitNewTaskIntent 判断标准化后的 goal 是否含明确的新任务意图。
-// 默认返回 false，仅匹配极少且高度精确的关键词。
-func isExplicitNewTaskIntent(goal string) bool {
- for _, kw := range newTaskChineseKeywords {
- if strings.Contains(goal, kw) {
- return true
- }
- }
- for _, kw := range newTaskEnglishKeywords {
- if goal == kw {
- return true
+// todoSummaryLikeForItems 将保留后的 todo 列表压缩成事实层需要的计数。
+func todoSummaryLikeForItems(items []agentsession.TodoItem) runtimefacts.TodoSummaryLike {
+ var summary runtimefacts.TodoSummaryLike
+ for _, item := range items {
+ if !item.RequiredValue() {
+ continue
}
- if strings.HasPrefix(goal, kw+" ") || strings.HasPrefix(goal, kw+"\t") {
- return true
+ if item.Status.IsTerminal() {
+ if item.Status == agentsession.TodoStatusFailed {
+ summary.RequiredFailed++
+ } else {
+ summary.RequiredCompleted++
+ }
+ continue
}
+ summary.RequiredOpen++
}
- return false
+ return summary
}
diff --git a/internal/runtime/todo_run_boundary_test.go b/internal/runtime/todo_run_boundary_test.go
index aaf9e49e..7dd65895 100644
--- a/internal/runtime/todo_run_boundary_test.go
+++ b/internal/runtime/todo_run_boundary_test.go
@@ -26,7 +26,6 @@ func TestResetTodosForUserRunClearsSessionAndEmitsEmptySnapshot(t *testing.T) {
service := &Service{sessionStore: store, events: make(chan RuntimeEvent, 8)}
state := newRunState("run-boundary", created)
- state.userGoal = "新任务"
if err := service.resetTodosForUserRun(context.Background(), &state); err != nil {
t.Fatalf("resetTodosForUserRun() error = %v", err)
}
@@ -61,15 +60,22 @@ func TestResetTodosForUserRunClearsSessionAndEmitsEmptySnapshot(t *testing.T) {
}
}
-func TestResetTodosForUserRunKeepsTodosForContinuePrompt(t *testing.T) {
+func TestResetTodosForUserRunKeepsTodosForActivePlan(t *testing.T) {
t.Parallel()
store := newMemoryStore()
required := true
- session := agentsession.New("todo-boundary-continue")
+ session := agentsession.New("todo-boundary-plan")
+ session.CurrentPlan = &agentsession.PlanArtifact{
+ ID: "plan-1",
+ Status: agentsession.PlanStatusApproved,
+ Spec: agentsession.PlanSpec{
+ Todos: []agentsession.TodoItem{{ID: "plan-todo", Content: "plan task"}},
+ },
+ }
session.Todos = []agentsession.TodoItem{{
- ID: "old-todo",
- Content: "old task",
+ ID: "plan-todo",
+ Content: "plan task",
Status: agentsession.TodoStatusPending,
Required: &required,
}}
@@ -79,8 +85,7 @@ func TestResetTodosForUserRunKeepsTodosForContinuePrompt(t *testing.T) {
}
service := &Service{sessionStore: store, events: make(chan RuntimeEvent, 8)}
- state := newRunState("run-boundary-continue", created)
- state.userGoal = "继续"
+ state := newRunState("run-boundary-plan", created)
if err := service.resetTodosForUserRun(context.Background(), &state); err != nil {
t.Fatalf("resetTodosForUserRun() error = %v", err)
}
@@ -88,7 +93,39 @@ func TestResetTodosForUserRunKeepsTodosForContinuePrompt(t *testing.T) {
t.Fatalf("state todos = %+v, want preserved", state.session.Todos)
}
if events := collectRuntimeEvents(service.Events()); len(events) != 0 {
- t.Fatalf("continue prompt should not emit reset events, got %+v", events)
+ t.Fatalf("active plan should not emit reset events, got %+v", events)
+ }
+}
+
+func TestResetTodosForUserRunPrunesTodosOutsideActivePlan(t *testing.T) {
+ t.Parallel()
+
+ store := newMemoryStore()
+ required := true
+ session := agentsession.New("todo-boundary-prune")
+ session.CurrentPlan = &agentsession.PlanArtifact{
+ ID: "plan-1",
+ Status: agentsession.PlanStatusApproved,
+ Spec: agentsession.PlanSpec{
+ Todos: []agentsession.TodoItem{{ID: "plan-todo", Content: "plan task"}},
+ },
+ }
+ session.Todos = []agentsession.TodoItem{
+ {ID: "plan-todo", Content: "plan task", Status: agentsession.TodoStatusPending, Required: &required},
+ {ID: "old-todo", Content: "old task", Status: agentsession.TodoStatusPending, Required: &required},
+ }
+ created, err := store.CreateSession(context.Background(), createSessionInputFromSession(session))
+ if err != nil {
+ t.Fatalf("CreateSession() error = %v", err)
+ }
+
+ service := &Service{sessionStore: store, events: make(chan RuntimeEvent, 8)}
+ state := newRunState("run-boundary-prune", created)
+ if err := service.resetTodosForUserRun(context.Background(), &state); err != nil {
+ t.Fatalf("resetTodosForUserRun() error = %v", err)
+ }
+ if len(state.session.Todos) != 1 || state.session.Todos[0].ID != "plan-todo" {
+ t.Fatalf("state todos = %+v, want only plan-owned todo", state.session.Todos)
}
}
@@ -97,49 +134,32 @@ func TestShouldResetTodosForUserRunBoundaryVariants(t *testing.T) {
cases := []struct {
name string
- goal string
+ session agentsession.Session
wantReset bool
}{
- // 空输入 → 保留
- {"empty", "", false},
-
- // 明确新任务 → 清空
- {"chinese exact 新任务", "新任务", true},
- {"chinese 帮我做个新任务", "帮我做个新任务", true},
- {"chinese 换个任务", "换个任务", true},
- {"chinese 新需求", "新需求", true},
- {"english exact new task", "new task", true},
- {"english 新任务", "new task please", true},
- {"english switch task", "switch task", true},
- {"english different task", "different task", true},
-
- // 默认保留：绝大多数输入不再被硬编码清空，交给 prompt 引导模型自行处理
- {"chinese 继续", "继续", false},
- {"chinese 继续修这个", "继续修这个", false},
- {"chinese 接着做", "接着做", false},
- {"chinese 刚才的代码还有问题", "刚才的代码还有问题", false},
- {"chinese 再优化一下", "再优化一下", false},
- {"chinese 补充测试用例", "补充测试用例", false},
- {"chinese 修复登录bug", "修复登录 bug", false},
- {"chinese 开始下一个任务", "开始下一个任务", false},
- {"chinese 重新实现", "重新实现", false},
- {"english continue", "continue", false},
- {"english continue with the failing test", "continue with the failing test", false},
- {"english implement search api", "implement search api", false},
- {"english keep going", "keep going", false},
- {"english keep it simple", "keep it simple please", false},
- {"english resume", "resume task", false},
- {"english go on", "go on please", false},
+ {name: "no plan resets", session: agentsession.New("no plan"), wantReset: true},
+ {name: "draft plan keeps", session: sessionWithPlanStatus(agentsession.PlanStatusDraft), wantReset: false},
+ {name: "approved plan keeps", session: sessionWithPlanStatus(agentsession.PlanStatusApproved), wantReset: false},
+ {name: "completed plan resets", session: sessionWithPlanStatus(agentsession.PlanStatusCompleted), wantReset: true},
}
for _, tc := range cases {
tc := tc
t.Run(tc.name, func(t *testing.T) {
t.Parallel()
- got := shouldResetTodosForUserRun(tc.goal)
+ got := shouldResetTodosForUserRun(tc.session)
if got != tc.wantReset {
- t.Fatalf("shouldResetTodosForUserRun(%q) = %v, want %v", tc.goal, got, tc.wantReset)
+ t.Fatalf("shouldResetTodosForUserRun() = %v, want %v", got, tc.wantReset)
}
})
}
}
+
+func sessionWithPlanStatus(status agentsession.PlanStatus) agentsession.Session {
+ session := agentsession.New("plan-boundary")
+ session.CurrentPlan = &agentsession.PlanArtifact{
+ ID: "plan-1",
+ Status: status,
+ }
+ return session
+}
diff --git a/internal/runtime/tool_diff_helpers_test.go b/internal/runtime/tool_diff_helpers_test.go
index cec450e6..e4071a40 100644
--- a/internal/runtime/tool_diff_helpers_test.go
+++ b/internal/runtime/tool_diff_helpers_test.go
@@ -4,8 +4,8 @@ import (
"context"
"testing"
- "neo-code/internal/repository"
providertypes "neo-code/internal/provider/types"
+ "neo-code/internal/repository"
agentsession "neo-code/internal/session"
"neo-code/internal/tools"
)
diff --git a/internal/runtime/toolexec.go b/internal/runtime/toolexec.go
index 3b28a76d..c6cc6ee0 100644
--- a/internal/runtime/toolexec.go
+++ b/internal/runtime/toolexec.go
@@ -4,6 +4,7 @@ import (
"context"
"encoding/json"
"errors"
+ "fmt"
"os"
"path/filepath"
"strings"
@@ -471,10 +472,27 @@ func hasSuccessfulWorkspaceWriteFact(result tools.ToolResult, execErr error) boo
if execErr != nil || result.IsError {
return false
}
+ return hasConfirmedWorkspaceWriteResult(result)
+}
+
+// hasConfirmedWorkspaceWriteResult 判断工具结果是否带有 runtime 可确认的真实写入。
+func hasConfirmedWorkspaceWriteResult(result tools.ToolResult) bool {
if toolResultNoopWrite(result.Metadata) {
return false
}
- return result.Facts.WorkspaceWrite
+ if !result.Facts.WorkspaceWrite {
+ return false
+ }
+ name := strings.TrimSpace(result.Name)
+ switch {
+ case isFileWriteTool(name):
+ _, ok := buildToolDiffPayload(result)
+ return ok
+ case strings.EqualFold(name, tools.ToolNameBash):
+ return len(toolResultWorkspaceWritePaths(result.Metadata)) > 0
+ default:
+ return false
+ }
}
// toolResultNoopWrite 判断工具结果是否声明了 no-op 写入（内容未变化）。
@@ -505,6 +523,43 @@ func toolResultFilePath(metadata map[string]any) string {
return strings.TrimSpace(p)
}
+// toolResultWorkspaceWritePaths 从工具结果中提取 runtime 确认的写入路径。
+func toolResultWorkspaceWritePaths(metadata map[string]any) []string {
+ if metadata == nil {
+ return nil
+ }
+ raw, ok := metadata["workspace_write_paths"]
+ if !ok || raw == nil {
+ return nil
+ }
+ seen := map[string]struct{}{}
+ var out []string
+ add := func(value any) {
+ text := strings.TrimSpace(fmt.Sprint(value))
+ if text == "" {
+ return
+ }
+ if _, exists := seen[text]; exists {
+ return
+ }
+ seen[text] = struct{}{}
+ out = append(out, text)
+ }
+ switch typed := raw.(type) {
+ case []string:
+ for _, value := range typed {
+ add(value)
+ }
+ case []any:
+ for _, value := range typed {
+ add(value)
+ }
+ case string:
+ add(typed)
+ }
+ return out
+}
+
// isFileWriteTool 判断工具调用是否为文件写入类工具，需在执行前后做 diff。
func isFileWriteTool(name string) bool {
switch strings.TrimSpace(name) {
diff --git a/internal/runtime/turn_control.go b/internal/runtime/turn_control.go
index 6047b447..a2f0b844 100644
--- a/internal/runtime/turn_control.go
+++ b/internal/runtime/turn_control.go
@@ -75,7 +75,7 @@ func applyToolExecutionCompletion(current controlplane.CompletionState, summary
}
}
}
- if result.Facts.WorkspaceWrite && !toolResultNoopWrite(result.Metadata) {
+ if hasConfirmedWorkspaceWriteResult(result) {
current.HasUnverifiedWrites = true
}
if result.Facts.VerificationPerformed && result.Facts.VerificationPassed {
@@ -87,65 +87,19 @@ func applyToolExecutionCompletion(current controlplane.CompletionState, summary
// collectProgressInput 基于执行前后事实组装 progress 评估输入。
func collectProgressInput(
- runState controlplane.RunState,
- beforeTask agentsession.TaskState,
afterTask agentsession.TaskState,
- beforeTodos []agentsession.TodoItem,
afterTodos []agentsession.TodoItem,
summary toolExecutionSummary,
- noProgressLimit int,
repeatLimit int,
) controlplane.ProgressInput {
- evidence := deriveProgressEvidence(beforeTask, afterTask, beforeTodos, afterTodos, summary)
return controlplane.ProgressInput{
- RunState: runState,
- Evidence: evidence,
CurrentToolSignature: computeToolSignature(summary.Calls),
ResultFingerprint: computeToolResultFingerprint(summary.Results),
SubgoalFingerprint: computeSubgoalFingerprint(afterTask, afterTodos, summary.Calls),
- NoProgressLimit: noProgressLimit,
RepeatCycleLimit: repeatLimit,
}
}
-// deriveProgressEvidence 从本轮前后快照和工具摘要中提取结构化 evidence。
-func deriveProgressEvidence(
- beforeTask agentsession.TaskState,
- afterTask agentsession.TaskState,
- beforeTodos []agentsession.TodoItem,
- afterTodos []agentsession.TodoItem,
- summary toolExecutionSummary,
-) []controlplane.ProgressEvidenceRecord {
- var evidence []controlplane.ProgressEvidenceRecord
-
- if computeTaskStateSignature(beforeTask) != computeTaskStateSignature(afterTask) {
- evidence = append(evidence, controlplane.ProgressEvidenceRecord{Kind: controlplane.EvidenceTaskStateChanged})
- }
- if computeTodoStateSignature(beforeTodos) != computeTodoStateSignature(afterTodos) {
- evidence = append(evidence, controlplane.ProgressEvidenceRecord{Kind: controlplane.EvidenceTodoStateChanged})
- }
- if summary.HasSuccessfulWorkspaceWrite {
- evidence = append(evidence, controlplane.ProgressEvidenceRecord{Kind: controlplane.EvidenceWriteApplied})
- }
- if summary.HasSuccessfulVerification {
- evidence = append(evidence, controlplane.ProgressEvidenceRecord{Kind: controlplane.EvidenceVerifyPassed})
- }
- if hasSuccessfulInformationalResult(summary.Results) {
- evidence = append(evidence, controlplane.ProgressEvidenceRecord{Kind: controlplane.EvidenceNewInfoNonDup})
- }
- return evidence
-}
-
-// computeTaskStateSignature 计算 task_state 的结构化签名。
-func computeTaskStateSignature(task agentsession.TaskState) string {
- encoded, err := json.Marshal(task.Clone())
- if err != nil {
- return ""
- }
- hash := sha256.Sum256(encoded)
- return hex.EncodeToString(hash[:])
-}
-
// computeToolResultFingerprint 计算本轮工具结果的聚合指纹。
func computeToolResultFingerprint(results []tools.ToolResult) string {
if len(results) == 0 {
@@ -245,64 +199,6 @@ func hasCompletedRequiredTodos(items []agentsession.TodoItem) bool {
return hasRequired
}
-// hasSuccessfulInformationalResult 判断本轮是否至少获得一个成功的非写入工具结果。
-func hasSuccessfulInformationalResult(results []tools.ToolResult) bool {
- for _, result := range results {
- if result.IsError {
- continue
- }
- switch strings.TrimSpace(result.Name) {
- case tools.ToolNameFilesystemWriteFile, tools.ToolNameFilesystemEdit:
- continue
- default:
- return true
- }
- }
- return false
-}
-
-// shouldPromotePendingFinalProgress 判断本轮执行结果是否可以作为 final 拦截后的有效推进信号。
-func shouldPromotePendingFinalProgress(
- score controlplane.ProgressScore,
- summary toolExecutionSummary,
- completion controlplane.CompletionState,
- lastBlockedReason string,
-) bool {
- if score.HasBusinessProgress {
- return true
- }
- if !score.HasExplorationProgress {
- return false
- }
-
- // 只读 read/glob 首次探索可算推进；同签名/同结果/同子目标且阻塞原因未变化时，不再重置 final 拦截计数。
- if hasSuccessfulReadOrGlobResult(summary.Results) &&
- score.SameToolSignature &&
- score.SameResultFingerprint &&
- score.SameSubgoal == controlplane.SubgoalRelationSame &&
- strings.EqualFold(
- strings.TrimSpace(lastBlockedReason),
- strings.TrimSpace(string(completion.CompletionBlockedReason)),
- ) {
- return false
- }
- return true
-}
-
-// hasSuccessfulReadOrGlobResult 判断本轮是否存在成功的 filesystem_read_file / filesystem_glob 结果。
-func hasSuccessfulReadOrGlobResult(results []tools.ToolResult) bool {
- for _, result := range results {
- if result.IsError {
- continue
- }
- switch strings.TrimSpace(result.Name) {
- case tools.ToolNameFilesystemReadFile, tools.ToolNameFilesystemGlob:
- return true
- }
- }
- return false
-}
-
// hasSuccessfulVerificationResult 判断本轮是否存在显式验证成功的结构化事实。
func hasSuccessfulVerificationResult(results []tools.ToolResult) bool {
if len(results) == 0 {
diff --git a/internal/runtime/turn_control_test.go b/internal/runtime/turn_control_test.go
index 5e7172a3..39fa9130 100644
--- a/internal/runtime/turn_control_test.go
+++ b/internal/runtime/turn_control_test.go
@@ -29,7 +29,7 @@ func TestApplyToolExecutionCompletionTracksWriteAndVerification(t *testing.T) {
written := applyToolExecutionCompletion(controlplane.CompletionState{}, toolExecutionSummary{
Results: []tools.ToolResult{
- {Facts: tools.ToolExecutionFacts{WorkspaceWrite: true}},
+ confirmedFilesystemWriteResult("a.txt"),
},
})
if !written.HasUnverifiedWrites {
@@ -52,7 +52,7 @@ func TestApplyToolExecutionCompletionKeepsUnverifiedWhenVerifyBeforeWrite(t *tes
got := applyToolExecutionCompletion(controlplane.CompletionState{}, toolExecutionSummary{
Results: []tools.ToolResult{
{Facts: tools.ToolExecutionFacts{VerificationPerformed: true, VerificationPassed: true}},
- {Facts: tools.ToolExecutionFacts{WorkspaceWrite: true}},
+ confirmedFilesystemWriteResult("a.txt"),
},
})
if !got.HasUnverifiedWrites {
@@ -65,7 +65,7 @@ func TestApplyToolExecutionCompletionClearsWhenVerifyAfterWrite(t *testing.T) {
got := applyToolExecutionCompletion(controlplane.CompletionState{}, toolExecutionSummary{
Results: []tools.ToolResult{
- {Facts: tools.ToolExecutionFacts{WorkspaceWrite: true}},
+ confirmedFilesystemWriteResult("a.txt"),
{Facts: tools.ToolExecutionFacts{VerificationPerformed: true, VerificationPassed: true}},
},
})
@@ -109,6 +109,100 @@ func TestToolResultNoopWrite(t *testing.T) {
}
}
+func TestHasConfirmedWorkspaceWriteResultRequiresToolDiffEvidence(t *testing.T) {
+ t.Parallel()
+
+ cases := []struct {
+ name string
+ result tools.ToolResult
+ want bool
+ }{
+ {
+ name: "filesystem write with tool diff payload",
+ result: confirmedFilesystemWriteResult("a.txt"),
+ want: true,
+ },
+ {
+ name: "filesystem write without tool diff payload",
+ result: tools.ToolResult{
+ Name: tools.ToolNameFilesystemEdit,
+ Facts: tools.ToolExecutionFacts{WorkspaceWrite: true},
+ },
+ want: false,
+ },
+ {
+ name: "noop write",
+ result: tools.ToolResult{
+ Name: tools.ToolNameFilesystemWriteFile,
+ Facts: tools.ToolExecutionFacts{WorkspaceWrite: true},
+ Metadata: map[string]any{
+ "path": "a.txt",
+ "noop_write": true,
+ },
+ },
+ want: false,
+ },
+ {
+ name: "tool error",
+ result: tools.ToolResult{
+ Name: tools.ToolNameFilesystemEdit,
+ IsError: true,
+ Facts: tools.ToolExecutionFacts{WorkspaceWrite: true},
+ Metadata: map[string]any{
+ "path": "a.txt",
+ },
+ },
+ want: false,
+ },
+ {
+ name: "bash write paths",
+ result: tools.ToolResult{
+ Name: tools.ToolNameBash,
+ Facts: tools.ToolExecutionFacts{WorkspaceWrite: true},
+ Metadata: map[string]any{
+ "workspace_write_paths": []string{"a.txt"},
+ },
+ },
+ want: true,
+ },
+ {
+ name: "bash without write paths",
+ result: tools.ToolResult{
+ Name: tools.ToolNameBash,
+ Facts: tools.ToolExecutionFacts{WorkspaceWrite: true},
+ },
+ want: false,
+ },
+ }
+
+ for _, tc := range cases {
+ tc := tc
+ t.Run(tc.name, func(t *testing.T) {
+ t.Parallel()
+ if got := hasSuccessfulWorkspaceWriteFact(tc.result, nil); got != tc.want {
+ t.Fatalf("hasSuccessfulWorkspaceWriteFact() = %v, want %v", got, tc.want)
+ }
+ })
+ }
+}
+
+func confirmedFilesystemWriteResult(path string) tools.ToolResult {
+ return tools.ToolResult{
+ Name: tools.ToolNameFilesystemEdit,
+ Facts: tools.ToolExecutionFacts{WorkspaceWrite: true},
+ Metadata: map[string]any{
+ "path": path,
+ "tool_diffs": []map[string]any{
+ {
+ "path": path,
+ "diff": "--- a\n+++ b\n@@ -1 +1 @@\n-a\n+b",
+ "kind": FileChangeKindModified,
+ },
+ },
+ },
+ }
+}
+
func TestHasPendingAgentTodosBlocksOnAnyNonTerminalTodo(t *testing.T) {
t.Parallel()
@@ -186,70 +280,6 @@ func TestClassifyToolErrorPrefersExplicitErrorClass(t *testing.T) {
}
}
-func TestShouldPromotePendingFinalProgress(t *testing.T) {
- t.Parallel()
-
- t.Run("business progress always promotes", func(t *testing.T) {
- t.Parallel()
- score := controlplane.ProgressScore{HasBusinessProgress: true}
- if !shouldPromotePendingFinalProgress(score, toolExecutionSummary{}, controlplane.CompletionState{}, "") {
- t.Fatal("expected business progress to promote pending final progress")
- }
- })
-
- t.Run("duplicate read result with same blocked reason does not promote", func(t *testing.T) {
- t.Parallel()
- score := controlplane.ProgressScore{
- HasExplorationProgress: true,
- SameToolSignature: true,
- SameResultFingerprint: true,
- SameSubgoal: controlplane.SubgoalRelationSame,
- }
- summary := toolExecutionSummary{
- Results: []tools.ToolResult{
- {Name: tools.ToolNameFilesystemReadFile, Content: "same result"},
- },
- }
- completion := controlplane.CompletionState{
- CompletionBlockedReason: controlplane.CompletionBlockedReasonPendingTodo,
- }
- if shouldPromotePendingFinalProgress(
- score,
- summary,
- completion,
- string(controlplane.CompletionBlockedReasonPendingTodo),
- ) {
- t.Fatal("expected duplicate informational read to not promote progress")
- }
- })
-
- t.Run("read result still promotes when blocked reason changed", func(t *testing.T) {
- t.Parallel()
- score := controlplane.ProgressScore{
- HasExplorationProgress: true,
- SameToolSignature: true,
- SameResultFingerprint: true,
- SameSubgoal: controlplane.SubgoalRelationSame,
- }
- summary := toolExecutionSummary{
- Results: []tools.ToolResult{
- {Name: tools.ToolNameFilesystemGlob, Content: "same result"},
- },
- }
- completion := controlplane.CompletionState{
- CompletionBlockedReason: controlplane.CompletionBlockedReasonPendingTodo,
- }
- if !shouldPromotePendingFinalProgress(
- score,
- summary,
- completion,
- string(controlplane.CompletionBlockedReasonUnverifiedWrite),
- ) {
- t.Fatal("expected changed blocked reason to allow one more exploration promotion")
- }
- })
-}
-
func TestApplyToolExecutionCompletionTracksTodoStateFacts(t *testing.T) {
t.Parallel()
diff --git a/internal/runtime/verify/command_success.go b/internal/runtime/verify/command_success.go
index 70db7012..cb848f36 100644
--- a/internal/runtime/verify/command_success.go
+++ b/internal/runtime/verify/command_success.go
@@ -90,7 +90,7 @@ func (v CommandSuccessVerifier) VerifyFinal(ctx context.Context, input FinalVeri
}
return VerificationResult{
Name: name,
- Status: VerificationSoftBlock,
+ Status: VerificationFail,
Summary: fmt.Sprintf("verification command failed with exit code %d", result.ExitCode),
Reason: "command exit code is non-zero",
ErrorClass: classifyCommandFailure(name, result),
diff --git a/internal/runtime/verify/command_success_test.go b/internal/runtime/verify/command_success_test.go
index 646aa84a..e6bec2bd 100644
--- a/internal/runtime/verify/command_success_test.go
+++ b/internal/runtime/verify/command_success_test.go
@@ -69,7 +69,7 @@ func TestCommandSuccessVerifier(t *testing.T) {
if err != nil {
t.Fatalf("VerifyFinal() error = %v", err)
}
- if result.Status != VerificationSoftBlock || result.ErrorClass != ErrorClassCompileError {
+ if result.Status != VerificationFail || result.ErrorClass != ErrorClassCompileError {
t.Fatalf("unexpected result: %+v", result)
}
})
diff --git a/internal/runtime/verify/content_match.go b/internal/runtime/verify/content_match.go
index f686d2db..54d30aab 100644
--- a/internal/runtime/verify/content_match.go
+++ b/internal/runtime/verify/content_match.go
@@ -73,7 +73,7 @@ func (ContentMatchVerifier) VerifyFinal(_ context.Context, input FinalVerifyInpu
}
return VerificationResult{
Name: contentMatchVerifierName,
- Status: VerificationSoftBlock,
+ Status: VerificationFail,
Summary: "content rule mismatch detected",
Reason: "content match check did not pass",
Evidence: evidence,
diff --git a/internal/runtime/verify/content_match_test.go b/internal/runtime/verify/content_match_test.go
index 81ca623a..574c546e 100644
--- a/internal/runtime/verify/content_match_test.go
+++ b/internal/runtime/verify/content_match_test.go
@@ -55,7 +55,7 @@ func TestContentMatchVerifier(t *testing.T) {
if err != nil {
t.Fatalf("VerifyFinal() error = %v", err)
}
- if result.Status != VerificationSoftBlock {
+ if result.Status != VerificationFail {
t.Fatalf("status = %q, want soft_block", result.Status)
}
})
diff --git a/internal/runtime/verify/file_exists.go b/internal/runtime/verify/file_exists.go
index 321e7a96..1fb1dd5e 100644
--- a/internal/runtime/verify/file_exists.go
+++ b/internal/runtime/verify/file_exists.go
@@ -21,7 +21,7 @@ func (FileExistsVerifier) VerifyFinal(_ context.Context, input FinalVerifyInput)
if len(paths) == 0 {
return VerificationResult{
Name: fileExistsVerifierName,
- Status: VerificationSoftBlock,
+ Status: VerificationFail,
Summary: "no artifact targets declared",
Reason: "file existence targets are missing",
}, nil
@@ -71,7 +71,7 @@ func (FileExistsVerifier) VerifyFinal(_ context.Context, input FinalVerifyInput)
}
return VerificationResult{
Name: fileExistsVerifierName,
- Status: VerificationSoftBlock,
+ Status: VerificationFail,
Summary: "artifact targets are missing or invalid",
Reason: "file existence check did not pass",
Evidence: evidence,
diff --git a/internal/runtime/verify/file_exists_test.go b/internal/runtime/verify/file_exists_test.go
index 7c8a884e..6aab55db 100644
--- a/internal/runtime/verify/file_exists_test.go
+++ b/internal/runtime/verify/file_exists_test.go
@@ -38,7 +38,7 @@ func TestFileExistsVerifier(t *testing.T) {
if err != nil {
t.Fatalf("VerifyFinal() error = %v", err)
}
- if result.Status != VerificationSoftBlock {
+ if result.Status != VerificationFail {
t.Fatalf("status = %q, want soft_block", result.Status)
}
})
diff --git a/internal/runtime/verify/git_diff.go b/internal/runtime/verify/git_diff.go
deleted file mode 100644
index ef12a61c..00000000
--- a/internal/runtime/verify/git_diff.go
+++ /dev/null
@@ -1,94 +0,0 @@
-package verify
-
-import (
- "context"
- "strings"
-)
-
-const gitDiffVerifierName = "git_diff"
-
-// GitDiffVerifier 校验工作区是否存在真实交付证据。
-type GitDiffVerifier struct {
- Executor CommandExecutor
-}
-
-// Name 返回 verifier 名称。
-func (v GitDiffVerifier) Name() string {
- return gitDiffVerifierName
-}
-
-// VerifyFinal 执行 git status 检查，确保 edit/fix/refactor 任务有真实改动。
-func (v GitDiffVerifier) VerifyFinal(ctx context.Context, input FinalVerifyInput) (VerificationResult, error) {
- executor := v.Executor
- if executor == nil {
- executor = PolicyCommandExecutor{}
- }
- cfg := input.VerificationConfig.Verifiers[gitDiffVerifierName]
- argv := compactStrings(cfg.Command)
- if len(argv) == 0 {
- argv = []string{"git", "status", "--porcelain", "--untracked-files=normal"}
- }
-
- result, err := executor.Execute(ctx, CommandExecutionRequest{
- Argv: argv,
- Workdir: input.Workdir,
- TimeoutSec: cfg.TimeoutSec,
- OutputCapByte: cfg.OutputCapBytes,
- Policy: input.VerificationConfig.ExecutionPolicy,
- })
- if err != nil {
- return VerificationResult{
- Name: gitDiffVerifierName,
- Status: VerificationFail,
- Summary: err.Error(),
- Reason: "git status command execution failed",
- ErrorClass: classifyCommandExecutionError(err),
- Evidence: commandEvidence(argv, result),
- }, nil
- }
- if result.ExitCode != 0 {
- return VerificationResult{
- Name: gitDiffVerifierName,
- Status: VerificationFail,
- Summary: "git status command returned non-zero",
- Reason: "git status command failed",
- ErrorClass: ErrorClassUnknown,
- Evidence: commandEvidence(argv, result),
- }, nil
- }
-
- lines := nonEmptyLines(result.Stdout)
- if len(lines) == 0 {
- return VerificationResult{
- Name: gitDiffVerifierName,
- Status: VerificationSoftBlock,
- Summary: "git status is empty",
- Reason: "no changed files detected",
- Evidence: commandEvidence(argv, result),
- }, nil
- }
- evidence := commandEvidence(argv, result)
- evidence["changed_files"] = lines
- evidence["changed_files_count"] = len(lines)
- return VerificationResult{
- Name: gitDiffVerifierName,
- Status: VerificationPass,
- Summary: "git status contains changed files",
- Reason: "workspace change detected",
- Evidence: evidence,
- }, nil
-}
-
-// nonEmptyLines 返回文本中的非空行列表。
-func nonEmptyLines(text string) []string {
- raw := strings.Split(text, "\n")
- lines := make([]string, 0, len(raw))
- for _, item := range raw {
- item = strings.TrimSpace(item)
- if item == "" {
- continue
- }
- lines = append(lines, item)
- }
- return lines
-}
diff --git a/internal/runtime/verify/git_diff_test.go b/internal/runtime/verify/git_diff_test.go
deleted file mode 100644
index 165d9d78..00000000
--- a/internal/runtime/verify/git_diff_test.go
+++ /dev/null
@@ -1,120 +0,0 @@
-package verify
-
-import (
- "context"
- "errors"
- "testing"
-
- "neo-code/internal/config"
-)
-
-func verifyConfigForGitDiffTests() config.VerificationConfig {
- cfg := config.StaticDefaults().Runtime.Verification
- cfg.ExecutionPolicy = config.VerificationExecutionPolicyConfig{
- Mode: "non_interactive",
- DefaultTimeout: 1,
- DefaultOutputCap: 1,
- }
- return cfg
-}
-
-func TestGitDiffVerifier(t *testing.T) {
- t.Parallel()
-
- t.Run("empty output soft blocks", func(t *testing.T) {
- t.Parallel()
- cfg := verifyConfigForGitDiffTests()
- executor := &stubCommandExecutor{result: CommandExecutionResult{ExitCode: 0, Stdout: ""}}
- result, err := (GitDiffVerifier{Executor: executor}).VerifyFinal(context.Background(), FinalVerifyInput{VerificationConfig: cfg})
- if err != nil {
- t.Fatalf("VerifyFinal() error = %v", err)
- }
- if result.Status != VerificationSoftBlock {
- t.Fatalf("status = %q, want soft_block", result.Status)
- }
- })
-
- t.Run("changed files pass", func(t *testing.T) {
- t.Parallel()
- cfg := verifyConfigForGitDiffTests()
- executor := &stubCommandExecutor{result: CommandExecutionResult{ExitCode: 0, Stdout: "M main.go\n?? new.txt\n"}}
- result, err := (GitDiffVerifier{Executor: executor}).VerifyFinal(context.Background(), FinalVerifyInput{
- Workdir: "/workspace",
- VerificationConfig: cfg,
- })
- if err != nil {
- t.Fatalf("VerifyFinal() error = %v", err)
- }
- if result.Status != VerificationPass {
- t.Fatalf("status = %q, want pass", result.Status)
- }
- if len(executor.requests) != 1 || executor.requests[0].Argv[1] != "status" {
- t.Fatalf("unexpected argv: %+v", executor.requests)
- }
- })
-
- t.Run("staged only pass", func(t *testing.T) {
- t.Parallel()
- cfg := verifyConfigForGitDiffTests()
- executor := &stubCommandExecutor{result: CommandExecutionResult{ExitCode: 0, Stdout: "A staged.txt\n"}}
- result, err := (GitDiffVerifier{Executor: executor}).VerifyFinal(context.Background(), FinalVerifyInput{VerificationConfig: cfg})
- if err != nil {
- t.Fatalf("VerifyFinal() error = %v", err)
- }
- if result.Status != VerificationPass {
- t.Fatalf("status = %q, want pass", result.Status)
- }
- })
-
- t.Run("unstaged only pass", func(t *testing.T) {
- t.Parallel()
- cfg := verifyConfigForGitDiffTests()
- executor := &stubCommandExecutor{result: CommandExecutionResult{ExitCode: 0, Stdout: " M unstaged.go\n"}}
- result, err := (GitDiffVerifier{Executor: executor}).VerifyFinal(context.Background(), FinalVerifyInput{VerificationConfig: cfg})
- if err != nil {
- t.Fatalf("VerifyFinal() error = %v", err)
- }
- if result.Status != VerificationPass {
- t.Fatalf("status = %q, want pass", result.Status)
- }
- })
-
- t.Run("untracked only pass", func(t *testing.T) {
- t.Parallel()
- cfg := verifyConfigForGitDiffTests()
- executor := &stubCommandExecutor{result: CommandExecutionResult{ExitCode: 0, Stdout: "?? untracked.txt\n"}}
- result, err := (GitDiffVerifier{Executor: executor}).VerifyFinal(context.Background(), FinalVerifyInput{VerificationConfig: cfg})
- if err != nil {
- t.Fatalf("VerifyFinal() error = %v", err)
- }
- if result.Status != VerificationPass {
- t.Fatalf("status = %q, want pass", result.Status)
- }
- })
-
- t.Run("ignored only pass", func(t *testing.T) {
- t.Parallel()
- cfg := verifyConfigForGitDiffTests()
- executor := &stubCommandExecutor{result: CommandExecutionResult{ExitCode: 0, Stdout: "!! ignored.log\n"}}
- result, err := (GitDiffVerifier{Executor: executor}).VerifyFinal(context.Background(), FinalVerifyInput{VerificationConfig: cfg})
- if err != nil {
- t.Fatalf("VerifyFinal() error = %v", err)
- }
- if result.Status != VerificationPass {
- t.Fatalf("status = %q, want pass", result.Status)
- }
- })
-
- t.Run("execution error fails", func(t *testing.T) {
- t.Parallel()
- cfg := verifyConfigForGitDiffTests()
- executor := &stubCommandExecutor{err: errors.New("timeout")}
- result, err := (GitDiffVerifier{Executor: executor}).VerifyFinal(context.Background(), FinalVerifyInput{VerificationConfig: cfg})
- if err != nil {
- t.Fatalf("VerifyFinal() error = %v", err)
- }
- if result.Status != VerificationFail || result.ErrorClass != ErrorClassTimeout {
- t.Fatalf("unexpected result: %+v", result)
- }
- })
-}
diff --git a/internal/runtime/verify/orchestrator.go b/internal/runtime/verify/orchestrator.go
index 887e57fa..a903bfff 100644
--- a/internal/runtime/verify/orchestrator.go
+++ b/internal/runtime/verify/orchestrator.go
@@ -8,12 +8,12 @@ import (
"neo-code/internal/runtime/controlplane"
)
-// Orchestrator 按固定顺序执行 verifier 并在首个非 pass 结果处短路。
+// Orchestrator 执行所有 verifier 并聚合结果。
type Orchestrator struct {
Verifiers []FinalVerifier
}
-// RunFinalVerification 执行 verifier 列表并生成统一 gate 决议。
+// RunFinalVerification 执行所有 verifier 并生成统一 gate 决议。任一 Fail → 整体 Failed。
func (o Orchestrator) RunFinalVerification(ctx context.Context, input FinalVerifyInput) (VerificationGateDecision, error) {
results := make([]VerificationResult, 0, len(o.Verifiers))
decision := VerificationGateDecision{
@@ -43,22 +43,8 @@ func (o Orchestrator) RunFinalVerification(ctx context.Context, input FinalVerif
if result.Status == VerificationPass {
continue
}
-
decision.Passed = false
- switch result.Status {
- case VerificationSoftBlock:
- decision.Reason = controlplane.StopReasonTodoNotConverged
- case VerificationHardBlock:
- if result.WaitingExternal {
- decision.Reason = controlplane.StopReasonTodoWaitingExternal
- } else {
- decision.Reason = controlplane.StopReasonTodoNotConverged
- }
- default:
- decision.Reason = stopReasonForVerificationFailure(result)
- }
- decision.Results = results
- return decision, nil
+ decision.Reason = stopReasonForVerificationFailure(result)
}
decision.Results = results
return decision, nil
diff --git a/internal/runtime/verify/orchestrator_test.go b/internal/runtime/verify/orchestrator_test.go
index b48d2d8b..1bc11547 100644
--- a/internal/runtime/verify/orchestrator_test.go
+++ b/internal/runtime/verify/orchestrator_test.go
@@ -27,20 +27,20 @@ func (s stubFinalVerifier) VerifyFinal(ctx context.Context, input FinalVerifyInp
func TestOrchestratorRunFinalVerification(t *testing.T) {
t.Parallel()
- t.Run("short-circuits on first non-pass", func(t *testing.T) {
+ t.Run("runs all verifiers, aggregates all results", func(t *testing.T) {
t.Parallel()
decision, err := (Orchestrator{Verifiers: []FinalVerifier{
- stubFinalVerifier{name: "todo", result: VerificationResult{Name: "todo", Status: VerificationSoftBlock}},
+ stubFinalVerifier{name: "todo", result: VerificationResult{Name: "todo", Status: VerificationFail}},
stubFinalVerifier{name: "build", result: VerificationResult{Name: "build", Status: VerificationFail}},
}}).RunFinalVerification(context.Background(), FinalVerifyInput{})
if err != nil {
t.Fatalf("RunFinalVerification() error = %v", err)
}
- if decision.Passed || decision.Reason != controlplane.StopReasonTodoNotConverged {
+ if decision.Passed {
t.Fatalf("unexpected decision: %+v", decision)
}
- if len(decision.Results) != 1 {
- t.Fatalf("results len = %d, want 1", len(decision.Results))
+ if len(decision.Results) != 2 {
+ t.Fatalf("results len = %d, want 2 (all verifiers run)", len(decision.Results))
}
})
@@ -57,15 +57,15 @@ func TestOrchestratorRunFinalVerification(t *testing.T) {
}
})
- t.Run("hard block waiting external maps correctly", func(t *testing.T) {
+ t.Run("fail with waiting external maps to verification_failed", func(t *testing.T) {
t.Parallel()
decision, err := (Orchestrator{Verifiers: []FinalVerifier{
- stubFinalVerifier{name: "todo", result: VerificationResult{Name: "todo", Status: VerificationHardBlock, WaitingExternal: true}},
+ stubFinalVerifier{name: "todo", result: VerificationResult{Name: "todo", Status: VerificationFail, WaitingExternal: true}},
}}).RunFinalVerification(context.Background(), FinalVerifyInput{})
if err != nil {
t.Fatalf("RunFinalVerification() error = %v", err)
}
- if decision.Reason != controlplane.StopReasonTodoWaitingExternal {
+ if decision.Reason != controlplane.StopReasonVerificationFailed {
t.Fatalf("reason = %q, want %q", decision.Reason, controlplane.StopReasonTodoWaitingExternal)
}
})
diff --git a/internal/runtime/verify/todo_convergence.go b/internal/runtime/verify/todo_convergence.go
index 51fde050..68a710b9 100644
--- a/internal/runtime/verify/todo_convergence.go
+++ b/internal/runtime/verify/todo_convergence.go
@@ -100,7 +100,7 @@ func (TodoConvergenceVerifier) VerifyFinal(_ context.Context, input FinalVerifyI
if len(waitingExternalIDs) > 0 {
return VerificationResult{
Name: todoConvergenceVerifierName,
- Status: VerificationHardBlock,
+ Status: VerificationFail,
Summary: fmt.Sprintf("%d required todo(s) wait for external input", len(waitingExternalIDs)),
Reason: "required todos are blocked by external dependency",
WaitingExternal: true,
@@ -110,7 +110,7 @@ func (TodoConvergenceVerifier) VerifyFinal(_ context.Context, input FinalVerifyI
if len(pendingIDs) > 0 || len(inProgressIDs) > 0 || len(blockedIDs) > 0 {
return VerificationResult{
Name: todoConvergenceVerifierName,
- Status: VerificationSoftBlock,
+ Status: VerificationFail,
Summary: "required todos are not converged",
Reason: "required todos are still pending, in progress, or internally blocked",
Evidence: evidence,
diff --git a/internal/runtime/verify/todo_convergence_test.go b/internal/runtime/verify/todo_convergence_test.go
index 0b6bccf4..9913b3c9 100644
--- a/internal/runtime/verify/todo_convergence_test.go
+++ b/internal/runtime/verify/todo_convergence_test.go
@@ -52,8 +52,8 @@ func TestTodoConvergenceVerifierStates(t *testing.T) {
if err != nil {
t.Fatalf("VerifyFinal() error = %v", err)
}
- if result.Status != VerificationSoftBlock {
- t.Fatalf("status = %q, want %q", result.Status, VerificationSoftBlock)
+ if result.Status != VerificationFail {
+ t.Fatalf("status = %q, want %q", result.Status, VerificationFail)
}
})
@@ -68,8 +68,8 @@ func TestTodoConvergenceVerifierStates(t *testing.T) {
if err != nil {
t.Fatalf("VerifyFinal() error = %v", err)
}
- if result.Status != VerificationSoftBlock {
- t.Fatalf("status = %q, want %q", result.Status, VerificationSoftBlock)
+ if result.Status != VerificationFail {
+ t.Fatalf("status = %q, want %q", result.Status, VerificationFail)
}
})
@@ -83,8 +83,8 @@ func TestTodoConvergenceVerifierStates(t *testing.T) {
if err != nil {
t.Fatalf("VerifyFinal() error = %v", err)
}
- if result.Status != VerificationHardBlock {
- t.Fatalf("status = %q, want %q", result.Status, VerificationHardBlock)
+ if result.Status != VerificationFail {
+ t.Fatalf("status = %q, want %q", result.Status, VerificationFail)
}
if !result.WaitingExternal {
t.Fatalf("expected WaitingExternal=true")
diff --git a/internal/runtime/verify/types.go b/internal/runtime/verify/types.go
index 41cf53a8..594c3026 100644
--- a/internal/runtime/verify/types.go
+++ b/internal/runtime/verify/types.go
@@ -11,11 +11,7 @@ type VerificationStatus string
const (
// VerificationPass 表示验证通过。
VerificationPass VerificationStatus = "pass"
- // VerificationSoftBlock 表示当前不能收尾，但仍可继续推进。
- VerificationSoftBlock VerificationStatus = "soft_block"
- // VerificationHardBlock 表示当前不能收尾且需要外部条件才能继续。
- VerificationHardBlock VerificationStatus = "hard_block"
- // VerificationFail 表示验证明确失败。
+ // VerificationFail 表示验证失败。
VerificationFail VerificationStatus = "fail"
)
@@ -88,14 +84,6 @@ type TaskStateSnapshot struct {
KeyArtifacts []string `json:"key_artifacts,omitempty"`
}
-// RuntimeStateSnapshot 表示 verifier 所需的 runtime 控制面快照。
-type RuntimeStateSnapshot struct {
- Turn int `json:"turn,omitempty"`
- MaxTurns int `json:"max_turns,omitempty"`
- MaxTurnsReached bool `json:"max_turns_reached,omitempty"`
- FinalInterceptStreak int `json:"final_intercept_streak,omitempty"`
-}
-
// FinalVerifyInput 表示一次 final 验证请求的完整输入。
type FinalVerifyInput struct {
SessionID string `json:"session_id,omitempty"`
@@ -107,7 +95,6 @@ type FinalVerifyInput struct {
LastAssistantFinal string `json:"last_assistant_final,omitempty"`
ToolResults []ToolResultLike `json:"tool_results,omitempty"`
TaskState TaskStateSnapshot `json:"task_state,omitempty"`
- RuntimeState RuntimeStateSnapshot `json:"runtime_state,omitempty"`
VerificationConfig config.VerificationConfig `json:"verification_config,omitempty"`
}
diff --git a/internal/session/plan.go b/internal/session/plan.go
index 4bfead1b..1c1f6ab8 100644
--- a/internal/session/plan.go
+++ b/internal/session/plan.go
@@ -1,7 +1,10 @@
package session
import (
+ "encoding/json"
"fmt"
+ "sort"
+ "strconv"
"strings"
"time"
)
@@ -30,6 +33,34 @@ const (
maxSummaryTodoIDs = 20
)
+const (
+ // AcceptCheckOutputOnly 表示仅需要最终回复文本作为交付物。
+ AcceptCheckOutputOnly = "output_only"
+ // AcceptCheckWorkspaceChange 表示需要运行期观测到 agent 产生工作区变更。
+ AcceptCheckWorkspaceChange = "workspace_change"
+ // AcceptCheckCommandSuccess 表示需要运行期命令成功事实。
+ AcceptCheckCommandSuccess = "command_success"
+ // AcceptCheckFileExists 表示需要运行期文件存在或写入事实。
+ AcceptCheckFileExists = "file_exists"
+ // AcceptCheckContentContains 表示需要运行期内容匹配事实。
+ AcceptCheckContentContains = "content_contains"
+ // AcceptCheckToolFact 表示需要运行期工具验证事实。
+ AcceptCheckToolFact = "tool_fact"
+)
+
+// AcceptCheck 声明 plan 阶段模型提出的机器可检查验收项。
+type AcceptCheck struct {
+ ID string `json:"id,omitempty"`
+ Kind string `json:"kind"`
+ Target string `json:"target,omitempty"`
+ Match string `json:"match,omitempty"`
+ Required *bool `json:"required,omitempty"`
+ Params map[string]string `json:"params,omitempty"`
+}
+
+// AcceptChecks 保存 plan 级验收项，并兼容读取旧的 []string 格式。
+type AcceptChecks []AcceptCheck
+
// PlanArtifact stores the current plan persisted in the session.
type PlanArtifact struct {
ID string `json:"id"`
@@ -43,21 +74,21 @@ type PlanArtifact struct {
// PlanSpec is the source of truth for the current plan.
type PlanSpec struct {
- Goal string `json:"goal"`
- Steps []string `json:"steps,omitempty"`
- Constraints []string `json:"constraints,omitempty"`
- Verify []string `json:"verify,omitempty"`
- Todos []TodoItem `json:"todos,omitempty"`
- OpenQuestions []string `json:"open_questions,omitempty"`
+ Goal string `json:"goal"`
+ Steps []string `json:"steps,omitempty"`
+ Constraints []string `json:"constraints,omitempty"`
+ Verify AcceptChecks `json:"verify,omitempty"`
+ Todos []TodoItem `json:"todos,omitempty"`
+ OpenQuestions []string `json:"open_questions,omitempty"`
}
// SummaryView is the compact projection derived from PlanSpec.
type SummaryView struct {
- Goal string `json:"goal"`
- KeySteps []string `json:"key_steps,omitempty"`
- Constraints []string `json:"constraints,omitempty"`
- Verify []string `json:"verify,omitempty"`
- ActiveTodoIDs []string `json:"active_todo_ids,omitempty"`
+ Goal string `json:"goal"`
+ KeySteps []string `json:"key_steps,omitempty"`
+ Constraints []string `json:"constraints,omitempty"`
+ Verify AcceptChecks `json:"verify,omitempty"`
+ ActiveTodoIDs []string `json:"active_todo_ids,omitempty"`
}
// Clone returns a deep copy of the plan artifact.
@@ -76,7 +107,7 @@ func (p PlanSpec) Clone() PlanSpec {
p.Goal = strings.TrimSpace(p.Goal)
p.Steps = append([]string(nil), p.Steps...)
p.Constraints = append([]string(nil), p.Constraints...)
- p.Verify = append([]string(nil), p.Verify...)
+ p.Verify = p.Verify.Clone()
p.OpenQuestions = append([]string(nil), p.OpenQuestions...)
p.Todos = cloneTodoItems(p.Todos)
return p
@@ -87,7 +118,7 @@ func (s SummaryView) Clone() SummaryView {
s.Goal = strings.TrimSpace(s.Goal)
s.KeySteps = append([]string(nil), s.KeySteps...)
s.Constraints = append([]string(nil), s.Constraints...)
- s.Verify = append([]string(nil), s.Verify...)
+ s.Verify = s.Verify.Clone()
s.ActiveTodoIDs = append([]string(nil), s.ActiveTodoIDs...)
return s
}
@@ -159,7 +190,7 @@ func NormalizePlanSpec(spec PlanSpec) (PlanSpec, error) {
spec.Goal = strings.TrimSpace(spec.Goal)
spec.Steps = normalizeTodoTextList(spec.Steps)
spec.Constraints = normalizeTodoTextList(spec.Constraints)
- spec.Verify = normalizeTodoTextList(spec.Verify)
+ spec.Verify = spec.Verify.Normalize()
spec.OpenQuestions = normalizeTodoTextList(spec.OpenQuestions)
todos, err := normalizeAndValidateTodos(spec.Todos)
@@ -180,7 +211,7 @@ func NormalizeSummaryView(summary SummaryView, spec PlanSpec) SummaryView {
normalized.Goal = strings.TrimSpace(normalized.Goal)
normalized.KeySteps = normalizeTodoTextList(normalized.KeySteps)
normalized.Constraints = normalizeTodoTextList(normalized.Constraints)
- normalized.Verify = normalizeTodoTextList(normalized.Verify)
+ normalized.Verify = normalized.Verify.Normalize()
normalized.ActiveTodoIDs = normalizeTodoTextList(normalized.ActiveTodoIDs)
if !summaryViewStructurallyValid(normalized, spec) {
return BuildSummaryView(spec)
@@ -198,7 +229,7 @@ func BuildSummaryView(spec PlanSpec) SummaryView {
Goal: spec.Goal,
KeySteps: clampStringList(spec.Steps, maxSummaryKeySteps),
Constraints: clampStringList(spec.Constraints, maxSummaryConstraints),
- Verify: clampStringList(spec.Verify, maxSummaryVerify),
+ Verify: clampAcceptChecks(spec.Verify, maxSummaryVerify),
ActiveTodoIDs: collectActiveTodoIDs(spec.Todos, maxSummaryTodoIDs),
}
}
@@ -219,7 +250,7 @@ func RenderPlanContent(spec PlanSpec) string {
sections = append(sections, "约束\n"+renderBulletList(spec.Constraints))
}
if len(spec.Verify) > 0 {
- sections = append(sections, "验证\n"+renderBulletList(spec.Verify))
+ sections = append(sections, "验证\n"+renderBulletList(spec.Verify.RenderLines()))
}
activeTodos := collectActiveTodoLines(spec.Todos)
if len(activeTodos) > 0 {
@@ -261,6 +292,185 @@ func clampStringList(items []string, maxItems int) []string {
return append([]string(nil), normalized[:maxItems]...)
}
+// UnmarshalJSON 兼容读取新 AcceptCheck 对象数组与旧字符串数组。
+func (checks *AcceptChecks) UnmarshalJSON(data []byte) error {
+ var structured []AcceptCheck
+ if err := json.Unmarshal(data, &structured); err == nil {
+ *checks = AcceptChecks(structured).Normalize()
+ return nil
+ }
+ var legacy []string
+ if err := json.Unmarshal(data, &legacy); err != nil {
+ return err
+ }
+ migrated := make(AcceptChecks, 0, len(legacy))
+ for _, item := range normalizeTodoTextList(legacy) {
+ migrated = append(migrated, migrateLegacyAcceptCheck(item))
+ }
+ *checks = migrated.Normalize()
+ return nil
+}
+
+// RequiredValue 返回验收项是否为必需项；nil 表示 JSON 省略字段，默认视为必需。
+func (check AcceptCheck) RequiredValue() bool {
+ if check.Required == nil {
+ return true
+ }
+ return *check.Required
+}
+
+// Clone 返回验收项深拷贝，避免调用方共享 Params map。
+func (checks AcceptChecks) Clone() AcceptChecks {
+ if len(checks) == 0 {
+ return nil
+ }
+ out := make(AcceptChecks, 0, len(checks))
+ for _, check := range checks {
+ cloned := check
+ cloned.ID = strings.TrimSpace(cloned.ID)
+ cloned.Kind = strings.TrimSpace(cloned.Kind)
+ cloned.Target = strings.TrimSpace(cloned.Target)
+ cloned.Match = strings.TrimSpace(cloned.Match)
+ if check.Required != nil {
+ required := *check.Required
+ cloned.Required = &required
+ }
+ if len(check.Params) > 0 {
+ cloned.Params = make(map[string]string, len(check.Params))
+ for key, value := range check.Params {
+ key = strings.TrimSpace(key)
+ value = strings.TrimSpace(value)
+ if key == "" && value == "" {
+ continue
+ }
+ cloned.Params[key] = value
+ }
+ }
+ out = append(out, cloned)
+ }
+ return out
+}
+
+// Normalize 规范化验收项文本字段并迁移旧 kind 名称。
+func (checks AcceptChecks) Normalize() AcceptChecks {
+ if len(checks) == 0 {
+ return nil
+ }
+ out := make(AcceptChecks, 0, len(checks))
+ seen := make(map[string]struct{}, len(checks))
+ for _, check := range checks.Clone() {
+ check.ID = strings.TrimSpace(check.ID)
+ check.Kind = normalizeAcceptCheckKind(check.Kind)
+ check.Target = strings.TrimSpace(check.Target)
+ check.Match = strings.TrimSpace(check.Match)
+ if check.Kind == "" && check.Target == "" && check.Match == "" {
+ continue
+ }
+ key := check.Kind + "\x00" + check.Target + "\x00" + check.Match +
+ "\x00" + paramsKey(check.Params) +
+ "\x00" + strconv.FormatBool(check.RequiredValue())
+ if _, exists := seen[key]; exists {
+ continue
+ }
+ seen[key] = struct{}{}
+ out = append(out, check)
+ }
+ if len(out) == 0 {
+ return nil
+ }
+ return out
+}
+
+// paramsKey 将验收参数稳定序列化，用于区分同目标下的不同机器检查。
+func paramsKey(params map[string]string) string {
+ if len(params) == 0 {
+ return ""
+ }
+ keys := make([]string, 0, len(params))
+ for key := range params {
+ keys = append(keys, key)
+ }
+ sort.Strings(keys)
+ parts := make([]string, 0, len(keys))
+ for _, key := range keys {
+ parts = append(parts, strconv.Quote(key)+"="+strconv.Quote(params[key]))
+ }
+ return strings.Join(parts, ";")
+}
+
+// RenderLines 返回面向计划正文的稳定验收项文本。
+func (checks AcceptChecks) RenderLines() []string {
+ normalized := checks.Normalize()
+ if len(normalized) == 0 {
+ return nil
+ }
+ lines := make([]string, 0, len(normalized))
+ for _, check := range normalized {
+ label := check.Kind
+ if check.Target != "" {
+ label += ": " + check.Target
+ }
+ lines = append(lines, label)
+ }
+ return lines
+}
+
+func clampAcceptChecks(items AcceptChecks, maxItems int) AcceptChecks {
+ normalized := items.Normalize()
+ if len(normalized) <= maxItems || maxItems <= 0 {
+ return normalized
+ }
+ return normalized[:maxItems].Clone()
+}
+
+func migrateLegacyAcceptCheck(value string) AcceptCheck {
+ kind := AcceptCheckOutputOnly
+ switch {
+ case looksLikeCommand(value):
+ kind = AcceptCheckCommandSuccess
+ case looksLikePath(value):
+ kind = AcceptCheckFileExists
+ }
+ return AcceptCheck{Kind: kind, Target: strings.TrimSpace(value)}
+}
+
+func normalizeAcceptCheckKind(kind string) string {
+ normalized := strings.ToLower(strings.TrimSpace(kind))
+ switch normalized {
+ case "command":
+ return AcceptCheckCommandSuccess
+ default:
+ return normalized
+ }
+}
+
+func looksLikeCommand(value string) bool {
+ trimmed := strings.ToLower(strings.TrimSpace(value))
+ if trimmed == "" {
+ return false
+ }
+ prefixes := []string{
+ "go ", "go\t", "npm ", "pnpm ", "yarn ", "make", "cargo ", "python ", "pytest", "ruff ",
+ "eslint", "tsc", "golangci-lint", "git ", "powershell ", "pwsh ",
+ }
+ for _, prefix := range prefixes {
+ if strings.HasPrefix(trimmed, prefix) {
+ return true
+ }
+ }
+ return strings.Contains(trimmed, " test ") || strings.Contains(trimmed, " build ")
+}
+
+func looksLikePath(value string) bool {
+ trimmed := strings.TrimSpace(value)
+ if trimmed == "" || strings.Contains(trimmed, " ") {
+ return false
+ }
+ return strings.Contains(trimmed, "/") ||
+ strings.Contains(trimmed, "\\") ||
+ strings.Contains(strings.TrimPrefix(trimmed, "."), ".")
+}
+
func collectActiveTodoIDs(items []TodoItem, limit int) []string {
if len(items) == 0 || limit <= 0 {
return nil
diff --git a/internal/session/plan_test.go b/internal/session/plan_test.go
index 06dff310..f52d5f95 100644
--- a/internal/session/plan_test.go
+++ b/internal/session/plan_test.go
@@ -1,12 +1,73 @@
package session
import (
+ "encoding/json"
"fmt"
"strings"
"testing"
"time"
)
+func acceptText(target string) AcceptChecks {
+ return AcceptChecks{{Kind: AcceptCheckOutputOnly, Target: target}}
+}
+
+func TestAcceptChecksUnmarshalRequiredDefaultAndExplicitFalse(t *testing.T) {
+ t.Parallel()
+
+ var checks AcceptChecks
+ if err := json.Unmarshal([]byte(`[{"kind":"output_only"},{"kind":"tool_fact","required":false}]`), &checks); err != nil {
+ t.Fatalf("Unmarshal() error = %v", err)
+ }
+ if len(checks) != 2 {
+ t.Fatalf("len = %d, want 2", len(checks))
+ }
+ if !checks[0].RequiredValue() {
+ t.Fatalf("omitted required should default to true: %+v", checks[0])
+ }
+ if checks[1].RequiredValue() {
+ t.Fatalf("explicit required=false should stay optional: %+v", checks[1])
+ }
+}
+
+func TestAcceptChecksNormalizePreservesDistinctParams(t *testing.T) {
+ t.Parallel()
+
+ checks := AcceptChecks{
+ {Kind: AcceptCheckContentContains, Target: "README.md", Params: map[string]string{"contains": "NeoCode"}},
+ {Kind: AcceptCheckContentContains, Target: "README.md", Params: map[string]string{"contains": "Todo"}},
+ {Kind: AcceptCheckContentContains, Target: "README.md", Params: map[string]string{"contains": "NeoCode"}},
+ }
+
+ normalized := checks.Normalize()
+ if len(normalized) != 2 {
+ t.Fatalf("Normalize() length = %d, want 2: %+v", len(normalized), normalized)
+ }
+ if normalized[0].Params["contains"] != "NeoCode" || normalized[1].Params["contains"] != "Todo" {
+ t.Fatalf("Normalize() = %+v, want distinct contains params kept", normalized)
+ }
+}
+
+func TestAcceptChecksNormalizePreservesDistinctRequired(t *testing.T) {
+ t.Parallel()
+
+ required := true
+ optional := false
+ checks := AcceptChecks{
+ {Kind: AcceptCheckCommandSuccess, Target: "go test ./...", Required: &required},
+ {Kind: AcceptCheckCommandSuccess, Target: "go test ./...", Required: &optional},
+ }
+
+ normalized := checks.Normalize()
+ if len(normalized) != 2 {
+ t.Fatalf("Normalize() length = %d, want 2: %+v", len(normalized), normalized)
+ }
+ if !normalized[0].RequiredValue() || normalized[1].RequiredValue() {
+ t.Fatalf("Normalize() required flags = [%v %v], want [true false]",
+ normalized[0].RequiredValue(), normalized[1].RequiredValue())
+ }
+}
+
func TestNormalizeSummaryViewFallsBackToBuiltSummaryWhenStructurallyInvalid(t *testing.T) {
t.Parallel()
@@ -14,7 +75,7 @@ func TestNormalizeSummaryViewFallsBackToBuiltSummaryWhenStructurallyInvalid(t *t
Goal: "为 runtime 引入 plan/build 模式",
Steps: []string{"扩展 session", "过滤工具", "调整 runtime"},
Constraints: []string{"plan 模式禁止写工具"},
- Verify: []string{"build 结束后进入 verify"},
+ Verify: acceptText("build 结束后进入 verify"),
Todos: []TodoItem{
{ID: "todo-1", Content: "扩展 session", Status: TodoStatusPending},
{ID: "todo-2", Content: "过滤工具", Status: TodoStatusCompleted},
@@ -27,7 +88,7 @@ func TestNormalizeSummaryViewFallsBackToBuiltSummaryWhenStructurallyInvalid(t *t
got := NormalizeSummaryView(SummaryView{
Goal: " ",
KeySteps: []string{"仅一步"},
- Verify: []string{"验收"},
+ Verify: acceptText("验收"),
ActiveTodoIDs: []string{"missing"},
}, spec)
want := BuildSummaryView(spec)
@@ -49,7 +110,7 @@ func TestBuildSummaryViewUsesActiveNonTerminalTodosOnly(t *testing.T) {
spec, err := NormalizePlanSpec(PlanSpec{
Goal: "整理当前执行摘要",
Steps: []string{"步骤一", "步骤二"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
Todos: []TodoItem{
{ID: "todo-1", Content: "待执行", Status: TodoStatusPending},
{ID: "todo-2", Content: "执行中", Status: TodoStatusInProgress},
@@ -82,7 +143,7 @@ func TestNormalizePlanArtifactDefaultsAndStatusNormalization(t *testing.T) {
Spec: PlanSpec{
Goal: "规范化计划对象",
Steps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
},
})
if err != nil {
@@ -116,7 +177,7 @@ func TestNormalizePlanArtifactPreservesCreatedAtAndNormalizesUpdatedAt(t *testin
Spec: PlanSpec{
Goal: "保留时间字段",
Steps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
},
})
if err != nil {
@@ -136,7 +197,7 @@ func TestNormalizeSummaryViewAllowsEmptyTodoRefsWhenPlanHasNoTodos(t *testing.T)
spec, err := NormalizePlanSpec(PlanSpec{
Goal: "无 todo 计划",
Steps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
})
if err != nil {
t.Fatalf("NormalizePlanSpec() error = %v", err)
@@ -145,7 +206,7 @@ func TestNormalizeSummaryViewAllowsEmptyTodoRefsWhenPlanHasNoTodos(t *testing.T)
summary := NormalizeSummaryView(SummaryView{
Goal: "无 todo 计划",
KeySteps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
}, spec)
if summary.Goal != "无 todo 计划" {
t.Fatalf("Goal = %q", summary.Goal)
@@ -162,7 +223,7 @@ func TestRenderPlanContentIncludesAllSections(t *testing.T) {
Goal: "输出完整计划正文",
Steps: []string{"步骤一", "步骤二"},
Constraints: []string{"约束一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
OpenQuestions: []string{"问题一"},
Todos: []TodoItem{
{ID: "todo-1", Content: "待执行", Status: TodoStatusPending},
@@ -254,7 +315,7 @@ func TestNormalizePlanArtifactEmptyID(t *testing.T) {
Spec: PlanSpec{
Goal: "测试",
Steps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
},
})
if err == nil {
@@ -326,20 +387,20 @@ func TestClampStringListMaxItems(t *testing.T) {
func TestSummaryViewStructurallyValidDetectsInvalid(t *testing.T) {
t.Parallel()
- spec := PlanSpec{Goal: "目标", Steps: []string{"步骤一"}, Verify: []string{"验证一"}}
+ spec := PlanSpec{Goal: "目标", Steps: []string{"步骤一"}, Verify: acceptText("验证一")}
// Empty goal
if summaryViewStructurallyValid(SummaryView{}, spec) {
t.Fatal("expected false for empty summary")
}
// Missing key steps
- if summaryViewStructurallyValid(SummaryView{Goal: "目标", Verify: []string{"v"}}, spec) {
+ if summaryViewStructurallyValid(SummaryView{Goal: "目标", Verify: acceptText("v")}, spec) {
t.Fatal("expected false for missing key steps")
}
// Unknown active todo IDs
if summaryViewStructurallyValid(SummaryView{
Goal: "目标",
KeySteps: []string{"步骤一"},
- Verify: []string{"验证一"},
+ Verify: acceptText("验证一"),
ActiveTodoIDs: []string{"unknown"},
}, spec) {
t.Fatal("expected false for unknown todo IDs")
diff --git a/internal/session/store_test.go b/internal/session/store_test.go
index 90a34af9..b9ec1e4b 100644
--- a/internal/session/store_test.go
+++ b/internal/session/store_test.go
@@ -620,7 +620,7 @@ func TestSQLiteStorePersistsPlanStateRoundTrip(t *testing.T) {
Goal: "落地 plan/build 模式",
Steps: []string{"扩展 session", "扩展 runtime"},
Constraints: []string{"保持 tools 边界"},
- Verify: []string{"go test ./internal/..."},
+ Verify: AcceptChecks{{Kind: AcceptCheckCommandSuccess, Target: "go test ./internal/..."}},
Todos: []TodoItem{
{ID: "todo-plan-1", Content: "补 plan 模型"},
},
@@ -629,7 +629,7 @@ func TestSQLiteStorePersistsPlanStateRoundTrip(t *testing.T) {
Goal: "落地 plan/build 模式",
KeySteps: []string{"扩展 session", "扩展 runtime"},
Constraints: []string{"保持 tools 边界"},
- Verify: []string{"go test ./internal/..."},
+ Verify: AcceptChecks{{Kind: AcceptCheckCommandSuccess, Target: "go test ./internal/..."}},
ActiveTodoIDs: []string{"todo-plan-1"},
},
},
diff --git a/internal/tools/codebase/searchsymbol.go b/internal/tools/codebase/searchsymbol.go
index ccf22101..e5d9286e 100644
--- a/internal/tools/codebase/searchsymbol.go
+++ b/internal/tools/codebase/searchsymbol.go
@@ -25,7 +25,7 @@ func (t *SearchSymbolTool) Name() string {
}
func (t *SearchSymbolTool) Description() string {
- return "Search for symbol definitions across the workspace. Returns file paths, line hints, kind (function/type/method/etc.), and signature. Does NOT return the function body; use codebase_read to view implementation."
+ return "Search for symbol definitions across the workspace. Prefer scope_dir during exploration/plan mode to avoid expensive full-workspace scans. Returns file paths, line hints, kind (function/type/method/etc.), and signature. Does NOT return the function body; use codebase_read to view implementation."
}
func (t *SearchSymbolTool) Schema() map[string]any {
diff --git a/internal/tools/codebase/searchtext.go b/internal/tools/codebase/searchtext.go
index dda70073..f5f670c7 100644
--- a/internal/tools/codebase/searchtext.go
+++ b/internal/tools/codebase/searchtext.go
@@ -25,7 +25,7 @@ func (t *SearchTextTool) Name() string {
}
func (t *SearchTextTool) Description() string {
- return "Search for text occurrences across the workspace. Returns file paths, line hints, and match counts. Does NOT return code snippets; use codebase_read to view content."
+ return "Search for text occurrences across the workspace. Prefer scope_dir during exploration/plan mode to avoid expensive full-workspace scans. Returns file paths, line hints, and match counts. Does NOT return code snippets; use codebase_read to view content."
}
func (t *SearchTextTool) Schema() map[string]any {
diff --git a/internal/tools/filesystem/glob_test.go b/internal/tools/filesystem/glob_test.go
index 439699ed..aae0be3e 100644
--- a/internal/tools/filesystem/glob_test.go
+++ b/internal/tools/filesystem/glob_test.go
@@ -242,6 +242,36 @@ func TestGlobToolFiltersSensitiveAndSymlinkEscapes(t *testing.T) {
}
}
+func TestGlobToolSkipsNoisyDirectories(t *testing.T) {
+ t.Parallel()
+
+ workspace := t.TempDir()
+ mustWriteFile(t, filepath.Join(workspace, "src", "main.go"), "package main\n")
+ for _, dir := range []string{".cache", ".tmp", "tmp", "build", "dist", "out", "target", "coverage", ".next", ".nuxt", ".turbo", ".parcel-cache", ".vite", "vendor", "bin", "obj"} {
+ mustWriteFile(t, filepath.Join(workspace, dir, "skip.go"), "package skip\n")
+ }
+
+ tool := NewGlob(workspace)
+ result, err := tool.Execute(context.Background(), tools.ToolCallInput{
+ Name: tool.Name(),
+ Arguments: mustMarshalFSArgs(t, map[string]string{"pattern": "**/*.go"}),
+ Workdir: workspace,
+ })
+ if err != nil {
+ t.Fatalf("Execute() error = %v", err)
+ }
+
+ content := normalizeSlashPath(result.Content)
+ if !strings.Contains(content, normalizeSlashPath(filepath.Join("src", "main.go"))) {
+ t.Fatalf("expected src match, got %q", result.Content)
+ }
+ for _, dir := range []string{".cache", ".tmp", "tmp", "build", "dist", "out", "target", "coverage", ".next", ".nuxt", ".turbo", ".parcel-cache", ".vite", "vendor", "bin", "obj"} {
+ if strings.Contains(content, dir) {
+ t.Fatalf("expected noisy dir %q to be skipped, got %q", dir, result.Content)
+ }
+ }
+}
+
func TestGlobToolVerificationFacts(t *testing.T) {
t.Parallel()
diff --git a/internal/tools/filesystem/grep_test.go b/internal/tools/filesystem/grep_test.go
index f89cdab8..80d5f277 100644
--- a/internal/tools/filesystem/grep_test.go
+++ b/internal/tools/filesystem/grep_test.go
@@ -239,3 +239,33 @@ func TestGrepToolFiltersSensitiveAndSymlinkEscapes(t *testing.T) {
t.Fatalf("expected symlink_escape reason count, got %#v", reasons)
}
}
+
+func TestGrepToolSkipsNoisyDirectories(t *testing.T) {
+ t.Parallel()
+
+ workspace := t.TempDir()
+ mustWriteFile(t, filepath.Join(workspace, "src", "main.go"), "needle\n")
+ for _, dir := range []string{".cache", ".tmp", "tmp", "build", "dist", "out", "target", "coverage", ".next", ".nuxt", ".turbo", ".parcel-cache", ".vite", "vendor", "bin", "obj"} {
+ mustWriteFile(t, filepath.Join(workspace, dir, "skip.txt"), "needle from "+dir+"\n")
+ }
+
+ tool := NewGrep(workspace)
+ result, err := tool.Execute(context.Background(), tools.ToolCallInput{
+ Name: tool.Name(),
+ Arguments: mustMarshalFSArgs(t, map[string]any{"pattern": "needle"}),
+ Workdir: workspace,
+ })
+ if err != nil {
+ t.Fatalf("Execute() error = %v", err)
+ }
+
+ content := normalizeSlashPath(result.Content)
+ if !strings.Contains(content, normalizeSlashPath(filepath.Join("src", "main.go"))+":1: needle") {
+ t.Fatalf("expected src match, got %q", result.Content)
+ }
+ for _, dir := range []string{".cache", ".tmp", "tmp", "build", "dist", "out", "target", "coverage", ".next", ".nuxt", ".turbo", ".parcel-cache", ".vite", "vendor", "bin", "obj"} {
+ if strings.Contains(content, dir) {
+ t.Fatalf("expected noisy dir %q to be skipped, got %q", dir, result.Content)
+ }
+ }
+}
diff --git a/internal/tools/filesystem/helpers.go b/internal/tools/filesystem/helpers.go
index 5296914f..c7b6366b 100644
--- a/internal/tools/filesystem/helpers.go
+++ b/internal/tools/filesystem/helpers.go
@@ -47,7 +47,9 @@ func skipDirEntry(path string, entry os.DirEntry) bool {
name := strings.ToLower(strings.TrimSpace(entry.Name()))
switch name {
- case ".git", ".idea", ".vscode", "node_modules":
+ case ".git", ".idea", ".vscode", "node_modules",
+ ".cache", ".tmp", "tmp", "build", "dist", "out", "target", "coverage",
+ ".next", ".nuxt", ".turbo", ".parcel-cache", ".vite", "vendor", "bin", "obj":
return true
}
diff --git a/internal/tools/filesystem/helpers_test.go b/internal/tools/filesystem/helpers_test.go
index 3fb75c34..e3c0fdc2 100644
--- a/internal/tools/filesystem/helpers_test.go
+++ b/internal/tools/filesystem/helpers_test.go
@@ -26,6 +26,10 @@ func TestSkipDirEntry(t *testing.T) {
root := t.TempDir()
mustCreateDir(t, filepath.Join(root, ".git"))
mustCreateDir(t, filepath.Join(root, "node_modules"))
+ mustCreateDir(t, filepath.Join(root, ".cache"))
+ mustCreateDir(t, filepath.Join(root, "build"))
+ mustCreateDir(t, filepath.Join(root, "dist"))
+ mustCreateDir(t, filepath.Join(root, "vendor"))
mustCreateDir(t, filepath.Join(root, "keep"))
mustWriteTestFile(t, filepath.Join(root, ".vscode"), "not-a-dir")
@@ -45,6 +49,11 @@ func TestSkipDirEntry(t *testing.T) {
if !got["node_modules"] {
t.Fatalf("node_modules skip = false, want true")
}
+ for _, name := range []string{".cache", "build", "dist", "vendor"} {
+ if !got[name] {
+ t.Fatalf("%s skip = false, want true", name)
+ }
+ }
if got["keep"] {
t.Fatalf("keep skip = true, want false")
}
diff --git a/internal/tools/mode_filter.go b/internal/tools/mode_filter.go
index 4c953a32..697391bd 100644
--- a/internal/tools/mode_filter.go
+++ b/internal/tools/mode_filter.go
@@ -12,6 +12,9 @@ func isReadOnlyVisibleTool(name string) bool {
case ToolNameFilesystemReadFile,
ToolNameFilesystemGrep,
ToolNameFilesystemGlob,
+ ToolNameCodebaseRead,
+ ToolNameCodebaseSearchText,
+ ToolNameCodebaseSearchSymbol,
ToolNameWebFetch,
ToolNameMemoRecall,
ToolNameMemoList,
diff --git a/internal/tools/permission_mapper.go b/internal/tools/permission_mapper.go
index f3dd5f43..0dd71f92 100644
--- a/internal/tools/permission_mapper.go
+++ b/internal/tools/permission_mapper.go
@@ -59,6 +59,27 @@ func buildPermissionAction(input ToolCallInput) (security.Action, error) {
action.Payload.Target = extractStringArgument(input.Arguments, "path")
action.Payload.SandboxTargetType = security.TargetTypePath
action.Payload.SandboxTarget = action.Payload.Target
+ case ToolNameCodebaseRead:
+ action.Type = security.ActionTypeRead
+ action.Payload.Operation = "codebase_read"
+ action.Payload.TargetType = security.TargetTypePath
+ action.Payload.Target = extractStringArgument(input.Arguments, "path")
+ action.Payload.SandboxTargetType = security.TargetTypePath
+ action.Payload.SandboxTarget = action.Payload.Target
+ case ToolNameCodebaseSearchText:
+ action.Type = security.ActionTypeRead
+ action.Payload.Operation = "codebase_search_text"
+ action.Payload.TargetType = security.TargetTypeDirectory
+ action.Payload.Target = extractStringArgument(input.Arguments, "dir")
+ action.Payload.SandboxTargetType = security.TargetTypeDirectory
+ action.Payload.SandboxTarget = action.Payload.Target
+ case ToolNameCodebaseSearchSymbol:
+ action.Type = security.ActionTypeRead
+ action.Payload.Operation = "codebase_search_symbol"
+ action.Payload.TargetType = security.TargetTypeDirectory
+ action.Payload.Target = extractStringArgument(input.Arguments, "dir")
+ action.Payload.SandboxTargetType = security.TargetTypeDirectory
+ action.Payload.SandboxTarget = action.Payload.Target
case ToolNameFilesystemGrep:
action.Type = security.ActionTypeRead
action.Payload.Operation = "grep"
diff --git a/internal/tools/todo/write.go b/internal/tools/todo/write.go
index 2de010cf..716d6478 100644
--- a/internal/tools/todo/write.go
+++ b/internal/tools/todo/write.go
@@ -287,13 +287,22 @@ func (t *Tool) Execute(ctx context.Context, call tools.ToolCallInput) (tools.Too
if resultErr != nil {
reason := mapReason(resultErr)
extra := map[string]any{"action": input.Action}
+ details := resultErr.Error()
if reason == reasonRevisionConflict && input.ID != "" {
if current, ok := call.SessionMutator.FindTodo(input.ID); ok {
extra["current_revision"] = current.Revision
extra["current_status"] = string(current.Status)
}
}
- return errorResult(reason, resultErr.Error(), extra), resultErr
+ if reason == reasonTodoNotFound {
+ details = todoNotFoundRecoveryDetails(call.SessionMutator, input.ID, resultErr)
+ todos := call.SessionMutator.ListTodos()
+ extra["todo_count"] = len(todos)
+ if ids := activeTodoIDsForRecovery(todos); len(ids) > 0 {
+ extra["active_todo_ids"] = ids
+ }
+ }
+ return errorResult(reason, details, extra), resultErr
}
return successResultWithMetadata(input.Action, call.SessionMutator.ListTodos(), dispatchMeta), nil
@@ -394,6 +403,41 @@ func (t *Tool) dispatch(call tools.ToolCallInput, input writeInput) (map[string]
}
}
+// todoNotFoundRecoveryDetails 为缺失 todo 的错误结果补充下一步恢复建议。
+func todoNotFoundRecoveryDetails(mutator tools.SessionMutator, id string, err error) string {
+ base := strings.TrimSpace(err.Error())
+ if base == "" {
+ base = fmt.Sprintf("%s: todo %q", agentsession.ErrTodoNotFound, strings.TrimSpace(id))
+ }
+ if mutator == nil || len(mutator.ListTodos()) == 0 {
+ return base + "; current session has no active todos. Create current run todos first with todo_write action=\"plan\" or action=\"add\", then update or complete those ids."
+ }
+ ids := activeTodoIDsForRecovery(mutator.ListTodos())
+ if len(ids) == 0 {
+ return base + "; current todos are all terminal. Create new current plan todos with todo_write action=\"plan\" or action=\"add\" before updating status."
+ }
+ return base + "; use one of the current active todo ids or recreate the current plan todos. active_todo_ids=" + strings.Join(ids, ",")
+}
+
+// activeTodoIDsForRecovery 收集非终态 todo ID，帮助模型从 todo_not_found 中恢复。
+func activeTodoIDsForRecovery(items []agentsession.TodoItem) []string {
+ if len(items) == 0 {
+ return nil
+ }
+ ids := make([]string, 0, len(items))
+ for _, item := range items {
+ id := strings.TrimSpace(item.ID)
+ if id == "" || item.Status.IsTerminal() {
+ continue
+ }
+ ids = append(ids, id)
+ }
+ if len(ids) == 0 {
+ return nil
+ }
+ return ids
+}
+
// completeTodoWithErgonomics 为 complete 动作提供 pending->in_progress->completed 便捷迁移。
func completeTodoWithErgonomics(call tools.ToolCallInput, input writeInput) (map[string]any, error) {
current, ok := call.SessionMutator.FindTodo(input.ID)
diff --git a/internal/tools/todo/write_test.go b/internal/tools/todo/write_test.go
index db60f3ca..0c11e6df 100644
--- a/internal/tools/todo/write_test.go
+++ b/internal/tools/todo/write_test.go
@@ -435,6 +435,57 @@ func TestToolExecuteReasonMapping(t *testing.T) {
}
}
+func TestToolExecuteTodoNotFoundRecoveryDetails(t *testing.T) {
+ t.Parallel()
+
+ tool := New()
+ emptySession := agentsession.New("todo-not-found-empty")
+ emptyMutator := &stubMutator{session: &emptySession}
+ result, err := tool.Execute(context.Background(), tools.ToolCallInput{
+ Name: tools.ToolNameTodoWrite,
+ SessionMutator: emptyMutator,
+ Arguments: []byte(`{"action":"complete","id":"missing"}`),
+ })
+ if err == nil {
+ t.Fatalf("expected missing todo error")
+ }
+ if result.Metadata["reason_code"] != reasonTodoNotFound {
+ t.Fatalf("reason_code = %v, want %q", result.Metadata["reason_code"], reasonTodoNotFound)
+ }
+ if !strings.Contains(result.Content, "current session has no active todos") ||
+ !strings.Contains(result.Content, `action="plan"`) {
+ t.Fatalf("expected empty-session recovery details, got %q", result.Content)
+ }
+ if result.Metadata["todo_count"] != 0 {
+ t.Fatalf("todo_count = %v, want 0", result.Metadata["todo_count"])
+ }
+
+ existingSession := agentsession.New("todo-not-found-existing")
+ if err := existingSession.AddTodo(agentsession.TodoItem{
+ ID: "todo-1",
+ Content: "current",
+ Status: agentsession.TodoStatusPending,
+ }); err != nil {
+ t.Fatalf("AddTodo() error = %v", err)
+ }
+ existingMutator := &stubMutator{session: &existingSession}
+ result, err = tool.Execute(context.Background(), tools.ToolCallInput{
+ Name: tools.ToolNameTodoWrite,
+ SessionMutator: existingMutator,
+ Arguments: []byte(`{"action":"complete","id":"missing"}`),
+ })
+ if err == nil {
+ t.Fatalf("expected missing todo error")
+ }
+ if !strings.Contains(result.Content, "active_todo_ids=todo-1") {
+ t.Fatalf("expected active todo recovery details, got %q", result.Content)
+ }
+ ids, ok := result.Metadata["active_todo_ids"].([]string)
+ if !ok || len(ids) != 1 || ids[0] != "todo-1" {
+ t.Fatalf("active_todo_ids metadata = %#v", result.Metadata["active_todo_ids"])
+ }
+}
+
func TestParseInput(t *testing.T) {
t.Parallel()
diff --git a/internal/tui/core/app/todo_test.go b/internal/tui/core/app/todo_test.go
index 159cf7b5..997d429e 100644
--- a/internal/tui/core/app/todo_test.go
+++ b/internal/tui/core/app/todo_test.go
@@ -200,7 +200,7 @@ func TestRebuildTodoSanitizesMarkdownTableLikeTitle(t *testing.T) {
ID: "todo-md",
Status: "pending",
Priority: 2,
- Title: "| col1 | col2 |\n| --- | --- |\n| value-a | value-b |",
+ Title: "| col1 | col2 |\n| --- | --- |\n| value-a | value-b |",
},
}
@@ -600,6 +600,35 @@ func TestRuntimeEventTodoHandlers(t *testing.T) {
}
}
+func TestRuntimeEventTodoConflictTodoNotFoundDoesNotRefreshEmptyPanel(t *testing.T) {
+ app, _ := newTestApp(t)
+ app.state.ActiveSessionID = "session-1"
+ app.todoItems = []todoViewItem{{ID: "stale", Title: "stale", Status: "pending"}}
+ app.todoPanelVisible = true
+
+ handled := runtimeEventTodoConflictHandler(&app, agentruntime.RuntimeEvent{
+ SessionID: "session-1",
+ Payload: agentruntime.TodoEventPayload{
+ Action: "complete",
+ Reason: "todo_not_found",
+ },
+ })
+ if handled {
+ t.Fatalf("expected todo conflict handler to return false")
+ }
+ if len(app.todoItems) != 0 || app.todoPanelVisible {
+ t.Fatalf("expected empty todo_not_found conflict to hide stale panel, items=%+v visible=%v", app.todoItems, app.todoPanelVisible)
+ }
+ if len(app.activities) == 0 || app.activities[len(app.activities)-1].Title != "Todo conflict" {
+ t.Fatalf("expected todo conflict activity, got %+v", app.activities)
+ }
+ for _, activity := range app.activities {
+ if activity.Title == "Failed to refresh todo panel" {
+ t.Fatalf("did not expect refresh failure activity for empty todo_not_found conflict: %+v", app.activities)
+ }
+ }
+}
+
func TestParseTodoEventPayload(t *testing.T) {
got, ok := parseTodoEventPayload(agentruntime.TodoEventPayload{Action: "a", Reason: "b"})
if !ok || got.Action != "a" || got.Reason != "b" {
diff --git a/internal/tui/core/app/update.go b/internal/tui/core/app/update.go
index 4af401d2..a83b932b 100644
--- a/internal/tui/core/app/update.go
+++ b/internal/tui/core/app/update.go
@@ -3609,7 +3609,13 @@ func runtimeEventAcceptanceDecidedHandler(a *App, event tuiservices.RuntimeEvent
if status == "" {
status = "unknown"
}
- detail := strings.TrimSpace(payload.UserVisibleSummary)
+ detail := strings.TrimSpace(payload.Summary)
+ if detail == "" {
+ detail = formatAcceptanceResults(payload.Results)
+ }
+ if detail == "" {
+ detail = strings.TrimSpace(payload.UserVisibleSummary)
+ }
if detail == "" {
detail = strings.TrimSpace(payload.InternalSummary)
}
@@ -3627,6 +3633,33 @@ func runtimeEventAcceptanceDecidedHandler(a *App, event tuiservices.RuntimeEvent
return false
}
+// formatAcceptanceResults 将逐项验收结果压缩成活动日志可读的一行摘要。
+func formatAcceptanceResults(results []tuiservices.AcceptanceCheckResult) string {
+ if len(results) == 0 {
+ return ""
+ }
+ parts := make([]string, 0, len(results))
+ for _, result := range results {
+ name := strings.TrimSpace(result.Name)
+ if name == "" {
+ name = strings.TrimSpace(result.Kind)
+ }
+ if name == "" {
+ name = "accept_check"
+ }
+ if result.Passed {
+ parts = append(parts, name+": pass")
+ continue
+ }
+ reason := strings.TrimSpace(result.Reason)
+ if reason == "" {
+ reason = "failed"
+ }
+ parts = append(parts, name+": "+reason)
+ }
+ return strings.Join(parts, "; ")
+}
+
// runtimeEventStopReasonDecidedHandler 处理运行终止原因事件，统一收尾状态与活动日志。
func runtimeEventStopReasonDecidedHandler(a *App, event tuiservices.RuntimeEvent) bool {
payload, ok := event.Payload.(tuiservices.StopReasonDecidedPayload)
@@ -3662,7 +3695,8 @@ func runtimeEventStopReasonDecidedHandler(a *App, event tuiservices.RuntimeEvent
}
case strings.ToLower(string(tuiservices.StopReasonTodoNotConverged)),
strings.ToLower(string(tuiservices.StopReasonTodoWaitingExternal)),
- strings.ToLower(string(tuiservices.StopReasonNoProgressAfterFinalIntercept)),
+ strings.ToLower(string(tuiservices.StopReasonMissingCompletionSignal)),
+ strings.ToLower(string(tuiservices.StopReasonRepeatCycle)),
strings.ToLower(string(tuiservices.StopReasonMaxTurnExceededWithUnconvergedTodos)),
strings.ToLower(string(tuiservices.StopReasonMaxTurnExceededWithFailedVerification)):
detail := strings.TrimSpace(payload.Detail)
@@ -3677,6 +3711,7 @@ func runtimeEventStopReasonDecidedHandler(a *App, event tuiservices.RuntimeEvent
a.state.StatusText = statusCanceled
a.appendActivity("run", "Canceled current run", "", false)
case strings.ToLower(string(tuiservices.StopReasonVerificationFailed)),
+ strings.ToLower(string(tuiservices.StopReasonAcceptCheckFailed)),
strings.ToLower(string(tuiservices.StopReasonRequiredTodoFailed)),
strings.ToLower(string(tuiservices.StopReasonVerificationExecutionDenied)),
strings.ToLower(string(tuiservices.StopReasonVerificationExecutionError)):
@@ -3730,8 +3765,16 @@ func runtimeEventTodoUpdatedHandler(a *App, event tuiservices.RuntimeEvent) bool
}
payload, _ := parseTodoEventPayload(event.Payload)
+ rawReason := strings.TrimSpace(payload.Reason)
+ if rawReason == "" {
+ rawReason = todoConflictReasonFromPayload(event.Payload)
+ }
if len(payload.Items) > 0 {
a.syncTodosFromEventItems(payload.Items)
+ } else if isTodoNotFoundConflict(rawReason) {
+ a.clearTodos()
+ a.applyComponentLayout(false)
+ a.todoPanelVisible = false
} else if err := a.refreshTodosFromSession(sessionID); err != nil {
a.appendActivity("todo", "Failed to refresh todo panel", err.Error(), true)
return false
@@ -3766,14 +3809,22 @@ func runtimeEventTodoConflictHandler(a *App, event tuiservices.RuntimeEvent) boo
}
payload, _ := parseTodoEventPayload(event.Payload)
+ rawReason := strings.TrimSpace(payload.Reason)
+ if rawReason == "" {
+ rawReason = todoConflictReasonFromPayload(event.Payload)
+ }
if len(payload.Items) > 0 {
a.syncTodosFromEventItems(payload.Items)
+ } else if isTodoNotFoundConflict(rawReason) {
+ a.clearTodos()
+ a.applyComponentLayout(false)
+ a.todoPanelVisible = false
} else if err := a.refreshTodosFromSession(sessionID); err != nil {
a.appendActivity("todo", "Failed to refresh todo panel", err.Error(), true)
return false
}
a.state.StatusText = formatTodoSummaryStatus(payload.Summary)
- reason := strings.TrimSpace(payload.Reason)
+ reason := rawReason
if reason == "" {
reason = "todo conflict"
}
@@ -3791,6 +3842,21 @@ func runtimeEventTodoConflictHandler(a *App, event tuiservices.RuntimeEvent) boo
return false
}
+// isTodoNotFoundConflict 判断 todo 冲突是否只是模型操作了不存在的 todo id。
+func isTodoNotFoundConflict(reason string) bool {
+ return strings.EqualFold(strings.TrimSpace(reason), "todo_not_found") ||
+ strings.Contains(strings.ToLower(strings.TrimSpace(reason)), "todo_not_found") ||
+ strings.Contains(strings.ToLower(strings.TrimSpace(reason)), "todo not found")
+}
+
+// todoConflictReasonFromPayload 从未解析的 payload 中兜底提取冲突原因文本。
+func todoConflictReasonFromPayload(payload any) string {
+ if payload == nil {
+ return ""
+ }
+ return strings.TrimSpace(fmt.Sprintf("%v", payload))
+}
+
// runtimeEventTodoSnapshotUpdatedHandler 处理 todo_snapshot_updated 事件并实时同步 Todo 面板。
func runtimeEventTodoSnapshotUpdatedHandler(a *App, event tuiservices.RuntimeEvent) bool {
return runtimeEventTodoUpdatedHandler(a, event)
@@ -4747,8 +4813,6 @@ func humanizeDecisionReason(reason string, missingFacts []map[string]any) string
}
}
return "任务仍缺少关键事实。"
- case "no_progress_after_final_intercept":
- return "连续多轮未产生新事实。"
case "required_todo_failed":
return "存在 required todo 失败。"
default:
diff --git a/internal/tui/core/app/update_runtime_events_test.go b/internal/tui/core/app/update_runtime_events_test.go
index 0015ba03..72231187 100644
--- a/internal/tui/core/app/update_runtime_events_test.go
+++ b/internal/tui/core/app/update_runtime_events_test.go
@@ -799,15 +799,24 @@ func TestRuntimeEventVerificationAndAcceptanceHandlers(t *testing.T) {
}
runtimeEventAcceptanceDecidedHandler(&app, agentruntime.RuntimeEvent{
Payload: agentruntime.AcceptanceDecidedPayload{
- Status: "failed",
- UserVisibleSummary: "",
- InternalSummary: "",
- ContinueHint: "provide missing files",
- CompletionBlockedReason: "unverified_write",
+ Status: "failed",
+ Summary: "command_success: missing successful command evidence",
+ StopReason: agentruntime.StopReasonAcceptCheckFailed,
+ Results: []agentruntime.AcceptanceCheckResult{
+ {
+ Passed: false,
+ Name: "command_success",
+ Kind: "command_success",
+ Target: "go test ./...",
+ Reason: "missing successful command evidence",
+ },
+ },
},
})
acceptance := app.activities[len(app.activities)-1]
- if acceptance.Title != "Acceptance decided (failed)" || !strings.Contains(acceptance.Detail, "reason=unverified_write") || !acceptance.IsError {
+ if acceptance.Title != "Acceptance decided (failed)" ||
+ !strings.Contains(acceptance.Detail, "command_success") ||
+ !acceptance.IsError {
t.Fatalf("unexpected acceptance activity: %+v", acceptance)
}
}
diff --git a/internal/tui/services/runtime_contract.go b/internal/tui/services/runtime_contract.go
index b40e7951..6947c0b8 100644
--- a/internal/tui/services/runtime_contract.go
+++ b/internal/tui/services/runtime_contract.go
@@ -306,12 +306,16 @@ const (
StopReasonVerificationFailed StopReason = "verification_failed"
// StopReasonAccepted 表示双门控通过并被 acceptance 接受。
StopReasonAccepted StopReason = "accepted"
+ // StopReasonMissingCompletionSignal 表示 assistant 停止调用工具但未输出完成信号。
+ StopReasonMissingCompletionSignal StopReason = "missing_completion_signal"
+ // StopReasonAcceptCheckFailed 表示最终 Accept Gate 的验收项失败。
+ StopReasonAcceptCheckFailed StopReason = "accept_check_failed"
// StopReasonTodoNotConverged 表示 required todo 未收敛。
StopReasonTodoNotConverged StopReason = "todo_not_converged"
// StopReasonTodoWaitingExternal 表示 todo 等待外部输入。
StopReasonTodoWaitingExternal StopReason = "todo_waiting_external"
- // StopReasonNoProgressAfterFinalIntercept 表示 final 被拦截后长期无进展。
- StopReasonNoProgressAfterFinalIntercept StopReason = "no_progress_after_final_intercept"
+ // StopReasonRepeatCycle 表示运行重复相同动作或结果。
+ StopReasonRepeatCycle StopReason = "repeat_cycle"
// StopReasonMaxTurnExceededWithUnconvergedTodos 表示 max turn + todo 未收敛。
StopReasonMaxTurnExceededWithUnconvergedTodos StopReason = "max_turn_exceeded_with_unconverged_todos"
// StopReasonMaxTurnExceededWithFailedVerification 表示 max turn + verification 失败。
@@ -367,13 +371,24 @@ type VerificationFailedPayload struct {
// AcceptanceDecidedPayload 描述 acceptance 引擎输出。
type AcceptanceDecidedPayload struct {
- Status string `json:"status"`
- StopReason StopReason `json:"stop_reason,omitempty"`
- ErrorClass string `json:"error_class,omitempty"`
- CompletionBlockedReason string `json:"completion_blocked_reason,omitempty"`
- UserVisibleSummary string `json:"user_visible_summary,omitempty"`
- InternalSummary string `json:"internal_summary,omitempty"`
- ContinueHint string `json:"continue_hint,omitempty"`
+ Status string `json:"status"`
+ StopReason StopReason `json:"stop_reason,omitempty"`
+ ErrorClass string `json:"error_class,omitempty"`
+ CompletionBlockedReason string `json:"completion_blocked_reason,omitempty"`
+ UserVisibleSummary string `json:"user_visible_summary,omitempty"`
+ InternalSummary string `json:"internal_summary,omitempty"`
+ ContinueHint string `json:"continue_hint,omitempty"`
+ Summary string `json:"summary,omitempty"`
+ Results []AcceptanceCheckResult `json:"results,omitempty"`
+}
+
+// AcceptanceCheckResult 描述 Accept Gate 中单个检查项的结果。
+type AcceptanceCheckResult struct {
+ Passed bool `json:"passed"`
+ Name string `json:"name"`
+ Kind string `json:"kind,omitempty"`
+ Target string `json:"target,omitempty"`
+ Reason string `json:"reason,omitempty"`
}
// TokenUsagePayload 描述 runtime 当前 token_usage 事件载荷。