Feat/self verifying turns#262
Merged
Merged
Conversation
One resolvePermissionDecision pipeline (hard-deny -> deny rules -> task lease -> allow sources -> boundary asks -> mode policy) now backs the native/API preflight, the XML and eager tool paths, the companion buildOnApproval gate, the MCP side-channel watcher, and the delegated agent callback. The three divergent per-gate permissionMode reads are retired; notably the smart-mode blanket auto-approve for delegated agents is replaced by an auto-edit floor clamp (file edits run free, Bash mutations still prompt), and an Always rule now suppresses the lease prompt instead of double-gating. New agonPermissionMode config key (ask|auto-edit|auto) migrates from the legacy permissionMode (auto->auto, smart->auto-edit, ask/deny-all->ask); deny-all stays a hard-deny override checked first. Leaseless delegated pushes and other external side effects still ask even in auto mode. The approval prompt fallback without a dispatch surface now denies instead of approving. 34-case golden decision table in tests/unit/permission-resolver.test.ts. ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
Shift+Tab now cycles ask -> auto-edit -> auto (Claude Code parity); Ctrl+A and /auto stay as the direct AUTO toggle, mapped onto the same mode enum (auto <-> auto-edit) through one applyPermissionMode writer so cesarAutoMode, agonPermissionMode, the lease, and the footer never disagree. The footer's AUTO slot becomes an always-visible permission segment: 'AUTO' when autonomous, otherwise the mode label with a shift+tab hint — restoring the approval-posture visibility the status bar rewrite dropped. New /mode command (bare cycles, /mode <mode> sets, /mode status explains). ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
The permission prompt grows to Yes / Yes-for-this-session / Always / No / Never. Always persists a validated scoped rule (two-token Bash scope like Bash(git push:*), exact resolved path for file tools) into permissions.allow — never a bare base token; when no safe rule can be synthesized (bare verbs, compounds, substitution, rendered previews) the Always/Never choices simply don't appear. Never persists a deny rule. Yes-for-session stores the rule in a session-scoped store so one answer drains coalesced sibling prompts across a multi-engine run; the queue drain now also honors persisted and session rules, not just allowedCommands. /permissions lists persisted + session rules, the legacy allowedCommands compat list, and the active mode, and gains add allow|deny <rule> / remove <rule> subcommands. ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
Full-roster review (4/6 engines) confirmed two blockers, both fixed: legacy bare allow sources (allowedCommands base tokens, tool-level toolPermissions allow, session base tokens) could cover a dangerous lease boundary — they now rank BELOW the boundary ask, and only scoped permission rules (the deliberate Always artifact) remain strong enough to auto-approve a push. And fileTargetInsideWorkspace treated an empty cwd as universal containment, so delegated agent runs floor- clamped to auto-edit could silently approve edits anywhere on disk; empty-cwd now passes only relative non-escaping paths, the delegated callback passes the real workspace root (team worktrees live under .agon/agent-worktrees inside it), and auto mode without a lease fences file mutations to the workspace itself. Also from the review: subcommand synthesis skips key=value option tokens (git -c user.name=x push -> Bash(git push:*)), persisting a rule evicts it from the opposite bucket so Always-after-Never actually wins, the remaining inline approval-string checks use isApprovedPermissionResponse, ask-mode prompts carry the resolver reason, and dead rule-engine imports are dropped. Refuted (no change): 'auto_off' in boundaryReasons is reachable via the lease reason carry-through; parsePermissionRule strips ':*' so synthesized rules validate (covered by tests). ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
When Cesar claims done after write-work (legacy done-claim or the agentic 'verifying' state), the harness now executes the discovered gate directly whenever the unified permission resolver already allows the command (mode auto, or an Always/session rule; resolver 'ask' falls back to the prompt nudge). Green runs verify the claim deterministically; red runs feed the real exit code and output tail back as a [SYSTEM] continuation so the engine fixes actual failures. Results route through recordToolUse so all existing verify bookkeeping flips uniformly. Runs are capped per turn (cesarGateAutoRunLimit, default 3) and each re-run requires fresh mutations. New tunables: cesarGateAutoRun, cesarGateAutoRunLimit, cesarGateTimeoutSec, cesarGateOutputTailChars. The verify nudge round-trip is extracted into a shared _injectSystemContinuation helper reused by the gate-failure feedback path. ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
… turns Full-roster review (6/6 engines) findings, each verified against source: (1) codex — agentic fail → fix → re-run never re-triggered because verificationFailed routes evaluateAgenticTaskState to 'running', never back to 'verifying'; gateAutoRunTriggered (new pure fn, unit-tested) now also fires on the settled verification_failed_fix_required state. (2) kimi/claude — a done-claim after a red harness run ended the turn looking done; the turn now ends with an explicit RED-gate warning and a [HARNESS] note in the transcript. (3) codex/minimax — the read-only classifier strips redirections, so 'npm test > file' resolved allow under an ask posture; redirecting gates are now refused outside mode auto (> /dev/null and fd-dups stay allowed). (4) minimax — engine-error state is reset at the top of _injectSystemContinuation so a stale error cannot short-circuit a healthy round-trip. (5) zai — the harness success note is [HARNESS]-tagged so verification is not misattributed to the engine, and the README documents the read-only-gate posture semantics. Refuted (no change): lifecycle-hook/exec-tree expansion — the gate command comes from the repo's own package.json/fitness line and the engine could already run the identical command under the identical posture; silent tool-loop catch — pre-existing pattern kept for parity. ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
# Conflicts: # packages/cli/src/generated/cesar/brain.ts # packages/cli/src/generated/signals/intent.ts # packages/cli/src/kern/signals/intent.kern # packages/core/src/generated/models/types.ts
# Conflicts: # packages/cli/src/generated/cesar/brain.ts # packages/core/src/generated/models/types.ts # packages/core/src/kern/models/types.kern
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.