fix(autonomous): persist PolicyEngine state across pause/resume (AUDIT-C3)

[konard]

Summary

Closes #256 (AUDIT-C3). Before this change, AutonomousTaskManager.runLoop() constructed a brand-new PolicyEngine on every start and resume, so every pauseTask() + resumeTask() cycle silently reset:

• toolCallTimestamps → bypassing rateLimit.toolCallsPerHour (default 100)
• apiCallTimestamps → bypassing rateLimit.apiCallsPerMinute (default 30)
• recentActions → disarming the 5-identical-actions loop detector
• consecutiveUncertainCount → disarming the uncertainty escalator

The loop now hydrates a single PolicyEngine from a new policy_state table keyed by task_id, and registers a write-through callback that flushes a snapshot on every mutation. Snapshots are cleared only on terminal task states (completed / failed / cancelled); paused tasks keep their state for the next resume.

Changes

Persistence

• New table policy_state (task_id PK, state JSON, updated_at, FK → autonomous_tasks ON DELETE CASCADE)
• Schema migration 1.23.0 (src/memory/schema.ts + src/memory/migrations/1.23.0.sql)
• AutonomousTaskStore.savePolicyState() / getPolicyState() / clearPolicyState()

PolicyEngine

• New PolicyEngineState interface (toolCallTimestamps, apiCallTimestamps, consecutiveUncertainCount, recentActions)
• New serialize(), hydrate(), setOnStateChange(), recordAction(), getRecentActions() methods
• recordToolCall / recordApiCall / recordUncertain / resetUncertainCount / recordAction now trigger the onStateChange callback. resetUncertainCount no-ops (and does not write to DB) when the counter is already zero, which is the common non-stuck path.

Loop

• AutonomousLoop.run() hydrates the engine from policy_state before the first policy check and wires the write-through callback.
• recentActions moved from an in-memory field on the loop into the engine so it is covered by the same snapshot contract.
• Terminal transitions (completed, failed, cancelled, max_iterations, rate_limit, planning failure, crash) clear policy_state. The abort path (e.g. graceful stop during pause) preserves state.

Housekeeping

• CURRENT_SCHEMA_VERSION bumped to 1.23.0
• package.json version bumped to 0.8.11 (to trigger the release workflow)
• CHANGELOG.md entry under Unreleased / Fixed

Acceptance criteria coverage (from the issue)

• Migration / schema for policy_state created (migration 1.23.0 + schema entry).
• PolicyEngine is hydrated from storage on resume; state is persisted on every record* call.
• Unit test: 10 pause/resume cycles do not reset toolCallsPerHour.
• Unit test: identical-action detector persists through pause/resume.
• Unit test: consecutiveUncertainCount is not cleared by pause/resume.
• Regression integration test added under src/autonomous/__tests__/.

Reproduction

Before the fix, this would let an agent make unlimited tool calls:

for (let i = 0; i < 1000; i++) {
  // burn through the 100/hour budget
  manager.pauseTask(id);
  manager.resumeTask(id); // fresh PolicyEngine — counter reset to 0
}

After the fix, the rate-limit check trips on attempt #101 regardless of how many pause/resume cycles are injected. The new test tool-call rate limit still fires after 10 pause/resume cycles encodes exactly that scenario.

Test plan

• npm run typecheck — clean
• npm run lint — clean (max-warnings 0)
• npm run format:check — clean
• npm test — 2942 passed / 140 files
• New suite src/autonomous/__tests__/policy-persistence.test.ts — 11/11 passed
• Manager regression test for issue audit-c3-pause-resume-policy-bypass #256 — passing

[konard]

Solution summary

PR #257 is now ready at #257.

Summary

Fixed AUDIT-C3 (issue #256): pause/resume no longer bypasses PolicyEngine sliding-window limits.

Core change — AutonomousLoop.run() now hydrates its PolicyEngine from a new per-task policy_state table and writes a snapshot on every state mutation. A fresh PolicyEngine on resume used to wipe toolCallTimestamps, apiCallTimestamps, recentActions, and consecutiveUncertainCount; those now survive any number of pause/resume cycles. Terminal states clear the snapshot.

Files touched — policy-engine.ts (serialize/hydrate + callback), loop.ts (hydrate on start, wire callback, clear on terminal), memory/schema.ts + migration 1.23.0.sql (new policy_state table, version bump to 1.23.0), memory/agent/autonomous-tasks.ts (save/get/clear methods), package bump to 0.8.11, CHANGELOG entry, 11 new regression tests in src/autonomous/__tests__/policy-persistence.test.ts plus a manager-level integration test.

Verification — npm run typecheck, npm run lint, npm run format:check all clean; full suite 2942 / 2942 passing.

---

This summary was automatically extracted from the AI working session output.

[konard]

🤖 Solution Draft Log

This log file contains the complete execution trace of the AI solution draft process.

💰 Cost: $7.929922

📊 Context and tokens usage:

• 167.4K / 1M (17%) input tokens, 43.8K / 128K (34%) output tokens

Total: (169.4K + 11.6M cached) input tokens, 43.8K output tokens, $7.929922 cost

🤖 Models used:

• Tool: Anthropic Claude Code
• Requested: opus
• Model: Claude Opus 4.7 (claude-opus-4-7)

📎 Log file uploaded as Gist (3010KB)

• View complete solution draft log

---

Now working session is ended, feel free to review and add any feedback on the solution draft.

[konard]

🔄 Auto-restart triggered (iteration 1)

Reason: Merge conflicts detected

Starting new session to address the issues.

---

Auto-restart-until-mergeable mode is active. Will continue until PR becomes mergeable.

[konard]

🔄 Auto-restart-until-mergeable Log (iteration 1)

This log file contains the complete execution trace of the AI solution draft process.

💰 Cost: $3.157243

📊 Context and tokens usage:

• 87.3K / 1M (9%) input tokens, 19.6K / 128K (15%) output tokens

Total: (84.7K + 4.3M cached) input tokens, 19.6K output tokens, $3.157243 cost

🤖 Models used:

• Tool: Anthropic Claude Code
• Requested: opus
• Model: Claude Opus 4.7 (claude-opus-4-7)

📎 Log file uploaded as Gist (4774KB)

• View complete solution draft log

---

Now working session is ended, feel free to review and add any feedback on the solution draft.