fix(session): preserve logical thread continuity

docs/audits/2026-03-28-session-continuity-post-ship-audit.md

@@ -0,0 +1,188 @@
+---
+title: "Session Continuity Post-Ship Audit"
+date: 2026-03-28
+scope: v0.32.0+ (PR #376)
+runs_sampled: 4 Fro Bot schedule runs (Mar 24-27, 2026)
+---
+
+## Session Continuity Post-Ship Audit
+
+## Executive Summary
+
+The session continuity feature (PR #376, v0.32.0) is **partially functioning**. Prompt restructuring (Non-Negotiable Rules, Constraint Reminder) works correctly. Session titles are set correctly at creation. But **session continuation never succeeds** — every run creates a fresh session because OpenCode auto-renames the title after the first prompt, and the title re-set after prompt was never implemented.
+
+**Impact**: Zero session continuity across runs. The agent cannot recall its prior work on the same entity. The "Thread Identity" section is never populated in prompts.
+
+## Runs Sampled
+
+| Run ID      | Date       | Event    | Session ID          | Title at Creation            |
+| ----------- | ---------- | -------- | ------------------- | ---------------------------- |
+| 23500894440 | 2026-03-24 | schedule | ses_2df4c0cafffe... | `fro-bot: schedule-c757a308` |
+| 23552690701 | 2026-03-25 | schedule | ses_2da20617dffe... | `fro-bot: schedule-c757a308` |
+| 23606276260 | 2026-03-26 | schedule | ses_2d4fc0340ffe... | `fro-bot: schedule-c757a308` |
+| 23655941186 | 2026-03-27 | schedule | ses_2cfed642fffe... | `fro-bot: schedule-c757a308` |
+
+All 4 runs: different session IDs despite identical logical key `schedule-c757a308`.
+
+## Findings
+
+### 1. CRITICAL: Session Title Auto-Overwrite (Root Cause)
+
+**Severity**: Critical — blocks all session continuity
+
+OpenCode auto-renames session titles based on the first message content. The title `fro-bot: schedule-c757a308` set at creation is overwritten (likely to something like "Daily maintenance report") before the cache is saved. The next run's resolver calls `findSessionByTitle()` looking for `fro-bot: schedule-c757a308` and finds no match.
+
+**Evidence**:
+
+- All 4 runs log: `Session continuity: no existing session found`
+- All 4 runs log: `continueSessionId: null`
+- All 4 runs: `Created new OpenCode session` with the correct title
+- Cache IS restored from previous run (confirmed: `Cache hit for restore-key: opencode-storage-github-fro-bot-agent-main-Linux-23606276260`)
+- 10 sessions exist at resolver time, none match the title
+
+**Root cause**: The plan specified "After successful prompt, **re-set title** via `session.update()` to guard against OpenCode's auto-title behavior" — but this was **never implemented** in `execution.ts`. There is no `session.update()` call anywhere in the codebase.
+
+**Fix**: Add `session.update()` call after each prompt to restore the logical key title. The OpenCode PATCH endpoint (`/session/{id}`) accepts `{ title }` — confirmed in `routes/session.ts:265`.
+
+### 2. WORKING: Prompt Restructuring (Instruction Sandwich)
+
+**Severity**: N/A — functioning correctly
+
+The prompt shows:
+
+- **Position 1**: `## Critical Rules (NON-NEGOTIABLE)` — 5-line hard constraints ✅
+- **Last position**: `## Reminder: Critical Rules` — 1-line recency reminder ✅
+- Task and Trigger Comment positioned near the top ✅
+- Agent Context demoted below task content ✅
+
+### 3. WORKING: Logical Key Computation
+
+**Severity**: N/A — functioning correctly
+
+The logical key `schedule-c757a308` is:
+
+- Deterministic (same hash across all 4 runs) ✅
+- Correctly derived from the schedule event ✅
+- Used for session title and search fallback ✅
+
+### 4. WORKING: Observability (Artifact Upload)
+
+**Severity**: N/A — functioning correctly
+
+All 4 Fro Bot schedule runs produced downloadable `opencode-logs` artifacts containing:
+
+- OpenCode server log (4,000-8,700 lines each)
+- Prompt artifact file (197 lines each)
+- Unique artifact names (`opencode-logs-{runId}-1`) ✅
+
+### 5. ISSUE: Thread Identity Section Missing from Prompts
+
+**Severity**: High — dependent on session continuity
+
+The prompt shows NO "Thread Identity" section. This is expected given continuity never succeeds — the section is only populated when `logicalKey` and `isContinuation` are provided to the prompt builder. When `continueSessionId` is null, no thread identity is injected.
+
+**Fix**: Will resolve automatically once session continuity works. However, consider showing thread identity even on fresh sessions (e.g., "Fresh conversation — no prior thread found for schedule-c757a308") to aid debugging.
+
+### 6. ISSUE: Task Content Duplicated in Prompt
+
+**Severity**: Low — token waste
+
+The schedule task description appears verbatim in both:
+
+- `## Task` section (lines 8-37)
+- `## Trigger Comment` section (lines 38-70)
+
+This is ~30 lines / ~400 tokens of pure duplication. For schedule events, the Task and Trigger Comment are identical because the `prompt` input IS the task.
+
+**Fix**: In `buildAgentPrompt()`, skip the Trigger Comment section when its content is identical to the Task section.
+
+### 7. NOISE: `tool.registry ... invalid` Entries
+
+**Severity**: Low — cosmetic noise
+
+Every run logs 14-24 instances of:
+
+```
+INFO service=tool.registry status=started invalid
+INFO service=tool.registry status=completed duration=N invalid
+```
+
+The `invalid` here is NOT an error — it's the tool name. OpenCode's tool registry initializes a tool called `invalid` (likely a placeholder/sentinel). This is benign startup chatter from OpenCode internals.
+
+**Recommendation**: No action needed. Document as expected noise.
+
+### 8. NOISE: `Blocked 3 postinstalls`
+
+**Severity**: Low — cosmetic noise
+
+Every run logs exactly once:
+
+```
+Blocked 3 postinstalls. Run `bun pm untrusted` for details.
+```
+
+This is Bun's security policy for oMo's dependencies. Expected behavior.
+
+**Recommendation**: No action needed. Document as expected noise.
+
+### 9. OBSERVATION: Prior Session Context Shows Stale Sessions
+
+**Severity**: Medium — reduced context quality
+
+The "Prior Session Context" table shows sessions from March 3-22 (pre-v0.32.0), none of which are schedule-related maintenance sessions. The recent schedule sessions (Mar 24-26) don't appear because they likely got renamed by OpenCode and are no longer matching any useful search query.
+
+**Fix**: Will improve automatically once session continuity works. The "Current Thread Context" section will show the actual prior work from the same schedule thread.
+
+## Prioritized Action Items
+
+### P0: Fix session title persistence (blocks all continuity)
+
+Add `session.update()` call in `execution.ts` after successful prompt to re-set the logical key title. Without this, no session continuity can ever work.
+
+```typescript
+// After successful prompt, re-set title to guard against auto-rename
+if (sessionTitle != null) {
+  try {
+    await client.session.update({
+      path: {id: sessionId},
+      body: {title: sessionTitle} as Record<string, unknown>,
+    })
+  } catch {
+    logger.warning("Failed to re-set session title", {sessionId, sessionTitle})
+  }
+}
+```
+
+### P1: Show Thread Identity on fresh sessions too
+
+Currently Thread Identity only shows when `isContinuation` is true. Show it on fresh starts too so the logical key is visible in the prompt artifact for debugging:
+
+```
+## Thread Identity
+**Logical Thread**: `schedule-c757a308` (schedule)
+**Status**: Fresh conversation — no prior thread found for this entity.
+```
+
+### P2: Deduplicate Task / Trigger Comment for schedule events
+
+Skip the Trigger Comment section when it's identical to the Task section. Saves ~400 tokens per schedule run.
+
+### P3: Add debug logging for session title state
+
+Log the titles of sessions returned by `listSessionsForProject()` during resolution so we can see what titles the sessions actually have (vs. what we're searching for).
+
+## Metrics
+
+| Metric                                  | Value                                                        |
+| --------------------------------------- | ------------------------------------------------------------ |
+| Runs sampled                            | 4                                                            |
+| Session continuity success rate         | **0%** (0/4 runs)                                            |
+| Prompt restructuring working            | **Yes** (Non-Negotiable Rules + Constraint Reminder present) |
+| Artifact upload working                 | **Yes** (all 4 runs have artifacts)                          |
+| Logical key computation working         | **Yes** (deterministic `schedule-c757a308` across all runs)  |
+| OpenCode errors in logs                 | **0**                                                        |
+| OpenCode warnings in logs               | **0**                                                        |
+| tool.registry "invalid" entries per run | 14-24 (benign noise)                                         |
+| Avg run duration                        | ~4-10 minutes                                                |
+| Avg OpenCode log size                   | 4,000-8,700 lines                                            |
+| Prompt size                             | 197 lines (~2,500 words)                                     |

src/features/agent/execution.ts

@@ -7,6 +7,7 @@ import * as crypto from 'node:crypto'
 import * as fs from 'node:fs/promises'
 import * as path from 'node:path'
 import {createOpencode} from '@opencode-ai/sdk'
+import {reassertSessionTitle} from '../../services/session/title-reassert.js'
 import {sleep} from '../../shared/async.js'
 import {DEFAULT_AGENT, DEFAULT_TIMEOUT_MS} from '../../shared/constants.js'
 import {getGitHubWorkspace, getOpenCodeLogPath, isOpenCodePromptArtifactEnabled} from '../../shared/env.js'
@@ -116,23 +117,17 @@ export async function executeOpenCode(
 
       const prompt = attempt === 1 ? initialPrompt : CONTINUATION_PROMPT
       const files = attempt === 1 ? promptOptions.fileParts : undefined
-      const result = await sendPromptToSession(client, sessionId, prompt, files, directory, config, logger)
+      const result = await (async () => {
+        try {
+          return await sendPromptToSession(client, sessionId, prompt, files, directory, config, logger)
+        } finally {
+          await reassertSessionTitle(client, sessionId, config?.sessionTitle, logger)
+        }
+      })()
+
       if (result.success) {
         final = result.eventStreamResult
 
-        // Best-effort title re-assertion: OpenCode may auto-overwrite session titles
-        // based on first message content. Re-set to preserve deterministic lookup.
-        if (config?.sessionTitle != null) {
-          try {
-            await (client.session as unknown as {update: (args: Record<string, unknown>) => Promise<unknown>}).update({
-              sessionID: sessionId,
-              title: config.sessionTitle,
-            })
-          } catch {
-            logger.debug('Best-effort session title re-assertion failed', {sessionId})
-          }
-        }
-
         return {
           success: true,
           exitCode: 0,

src/features/agent/opencode.test.ts

@@ -110,6 +110,7 @@ function createMockClient(options: {
       create: options.throwOnCreate
         ? vi.fn().mockRejectedValue(new Error('Session creation failed'))
         : vi.fn().mockResolvedValue({data: {id: 'ses_123', title: 'Test', version: '1'}}),
+      update: vi.fn().mockResolvedValue({data: {id: 'ses_123', title: 'Test', version: '1'}}),
       promptAsync: options.throwOnPrompt
         ? vi.fn().mockRejectedValue(new Error('Prompt failed'))
         : vi.fn().mockResolvedValue({data: options.promptResponse}),
@@ -444,6 +445,70 @@ describe('executeOpenCode', () => {
     expect(result.duration).toBeGreaterThanOrEqual(0)
   })
 
+  it('re-asserts session title with SDK update payload after prompt attempts', async () => {
+    // #given
+    const mockClient = createMockClient({
+      promptResponse: {parts: [{type: 'text', text: 'Agent response'}]},
+    })
+    const mockOpencode = createMockOpencode({client: mockClient})
+    vi.mocked(createOpencode).mockResolvedValue(mockOpencode as unknown as Awaited<ReturnType<typeof createOpencode>>)
+    const config: ExecutionConfig = {
+      agent: 'sisyphus',
+      model: null,
+      timeoutMs: 1800000,
+      omoProviders: {
+        claude: 'no',
+        copilot: 'no',
+        gemini: 'no',
+        openai: 'no',
+        opencodeZen: 'no',
+        zaiCodingPlan: 'no',
+        kimiForCoding: 'no',
+      },
+      sessionTitle: 'fro-bot: schedule-c757a308',
+    }
+
+    // #when
+    await executeOpenCode(createMockPromptOptions(), mockLogger, config)
+
+    // #then
+    expect(mockClient.session.update).toHaveBeenCalledWith({
+      path: {id: 'ses_123'},
+      body: {title: 'fro-bot: schedule-c757a308'},
+    })
+  })
+
+  it('re-asserts session title even when prompt attempt fails', async () => {
+    // #given
+    const mockClient = createMockClient({throwOnPrompt: true})
+    const mockOpencode = createMockOpencode({client: mockClient})
+    vi.mocked(createOpencode).mockResolvedValue(mockOpencode as unknown as Awaited<ReturnType<typeof createOpencode>>)
+    const config: ExecutionConfig = {
+      agent: 'sisyphus',
+      model: null,
+      timeoutMs: 1800000,
+      omoProviders: {
+        claude: 'no',
+        copilot: 'no',
+        gemini: 'no',
+        openai: 'no',
+        opencodeZen: 'no',
+        zaiCodingPlan: 'no',
+        kimiForCoding: 'no',
+      },
+      sessionTitle: 'fro-bot: schedule-c757a308',
+    }
+
+    // #when
+    await executeOpenCode(createMockPromptOptions(), mockLogger, config)
+
+    // #then
+    expect(mockClient.session.update).toHaveBeenCalledWith({
+      path: {id: 'ses_123'},
+      body: {title: 'fro-bot: schedule-c757a308'},
+    })
+  })
+
   it('returns failure result when prompt fails', async () => {
     // #given
     const mockClient = createMockClient({throwOnPrompt: true})

src/features/agent/prompt.test.ts

@@ -136,6 +136,25 @@ describe('buildAgentPrompt', () => {
     expect(prompt).toContain('**Status**: Continuing previous conversation thread.')
   })
 
+  it('shows fresh thread identity status when logical key exists without continuation', () => {
+    // #given
+    const options: PromptOptions = {
+      context: createMockContext(),
+      customPrompt: null,
+      cacheStatus: 'hit',
+      logicalKey: createMockLogicalKey(),
+      isContinuation: false,
+    }
+
+    // #when
+    const prompt = buildAgentPrompt(options, mockLogger)
+
+    // #then
+    expect(prompt).toContain('## Thread Identity')
+    expect(prompt).toContain('**Logical Thread**: `pr-42` (pr #42)')
+    expect(prompt).toContain('**Status**: Fresh conversation — no prior thread found for this entity.')
+  })
+
   it('places current thread context above environment and historical context for continuation runs', () => {
     // #given
     const sessionContext: SessionContext = {
@@ -570,6 +589,37 @@ describe('buildAgentPrompt', () => {
     expect(prompt).toContain('## GitHub Operations (Use gh CLI)')
   })
 
+  it('skips Trigger Comment section when schedule prompt text matches trigger comment', () => {
+    // #given
+    const duplicatedTask = 'Run daily maintenance tasks and update the report issue.'
+    const context = createMockContext({
+      eventName: 'schedule',
+      issueNumber: null,
+      issueTitle: null,
+      issueType: null,
+      commentBody: duplicatedTask,
+    })
+    const triggerContext = createMockTriggerContext({
+      eventType: 'schedule',
+      commentBody: duplicatedTask,
+      target: undefined,
+    })
+    const options: PromptOptions = {
+      context,
+      customPrompt: `  ${duplicatedTask}  `,
+      cacheStatus: 'hit',
+      triggerContext,
+    }
+
+    // #when
+    const prompt = buildAgentPrompt(options, mockLogger)
+
+    // #then
+    expect(prompt).toContain('## Task')
+    expect(prompt).toContain(duplicatedTask)
+    expect(prompt).not.toContain('## Trigger Comment')
+  })
+
   describe('session context', () => {
     it('includes session context section when sessionContext is provided', () => {
       // #given

src/features/agent/prompt.ts

@@ -156,7 +156,16 @@ Respond to the trigger comment above. Follow all instructions and requirements l
 `)
   }
 
-  if (context.commentBody != null) {
+  const trimmedCustomPrompt = customPrompt?.trim() ?? null
+  const trimmedCommentBody = context.commentBody?.trim() ?? null
+  const triggerCommentDuplicatesTask =
+    trimmedCustomPrompt != null &&
+    trimmedCustomPrompt.length > 0 &&
+    trimmedCommentBody != null &&
+    trimmedCommentBody.length > 0 &&
+    trimmedCustomPrompt === trimmedCommentBody
+
+  if (context.commentBody != null && !triggerCommentDuplicatesTask) {
     parts.push(`## Trigger Comment
 **Author:** ${context.commentAuthor ?? 'unknown'}
 

src/harness/phases/execute.ts

@@ -43,6 +43,9 @@ export async function runExecute(
       recentSessions: sessionPrep.recentSessions,
       priorWorkContext: sessionPrep.priorWorkContext,
     },
+    logicalKey: sessionPrep.logicalKey ?? null,
+    isContinuation: sessionPrep.isContinuation,
+    currentThreadSessionId: sessionPrep.continueSessionId ?? null,
     triggerContext: routing.triggerResult.context,
     fileParts: sessionPrep.attachmentResult?.fileParts,
   }