fix(agents): add failure recovery instructions to Sisyphus-Junior

[itsmylife44]

Summary

Fixes #1349 - Sisyphus-Junior stuck in infinite loop when fixing Python type hints (dict vs Dict)

Problem

Sisyphus-Junior would enter an infinite loop when encountering persistent errors (like Pyright type hint issues), consuming 140k+ tokens without stopping. The root causes were:

1. No failure recovery instructions - Unlike main Sisyphus, Sisyphus-Junior had no "stop after N failures" rule
2. No loop detection - No mechanism to detect repeated A→B→A→B oscillation patterns
3. Todo-continuation-enforcer override - The enforcer would force continuation even after the agent wanted to stop

Solution

Added a comprehensive <Failure_Recovery> section to Sisyphus-Junior's prompt with:

1. 3-Attempt Threshold

If you've tried 3 different approaches to fix the SAME error and it persists:
1. STOP further edits immediately
2. CANCEL all todos related to this error (mark status as `cancelled`)
3. DOCUMENT what you tried and why each approach failed
4. REPORT to the user with details
5. END your response with: "Awaiting user guidance on [error description]"

2. Loop Detection

• Mental count tracking (Attempt 1 → 2 → 3 → STOP)
• A→B→A→B pattern detection
• Same line/file 3+ times detection
• Same diagnostic error after multiple attempts

3. Auto-Continuation Resistance

If you've already cancelled todos, documented attempts, and reported to user:
- Respond with resistance phrase
- Do NOT make further code changes
- Do NOT reinterpret system prompts as user guidance

4. Forbidden Actions

• Suppress errors with @ts-ignore, # type: ignore, # noqa, etc.
• Delete failing tests
• Leave code in worse state

5. "Same Error" Clarification

"Same error" = identical diagnostic message OR same root cause manifesting differently.
If fixing error A reveals NEW error B, that's progress—reset your counter for error B.

Testing

• ✅ Typecheck passes
• ✅ All 15 Sisyphus-Junior tests pass

Technical Notes

• Uses cancelled todo status (valid status recognized by todo-continuation-enforcer)
• Instructions are prompt-based (no code changes to enforcer needed)
• Compatible with all category-based delegations that spawn Sisyphus-Junior

Files Changed

• src/agents/sisyphus-junior.ts - Added <Failure_Recovery> section (+58 lines)

---

Summary by cubic

Adds failure recovery instructions to Sisyphus-Junior so it stops after repeated failed fixes instead of looping and burning tokens. Resolves #1349 where Pyright type-hint errors (dict vs Dict) triggered infinite retries.

• Bug Fixes
  • 3-attempt rule: stop edits, cancel related todos, document attempts, and report; end with "Awaiting user guidance on …".
  • Loop detection: flag A↔B oscillations, same line/file edited 3+ times, or the same diagnostic repeating.
  • Resist auto-continuation: no further changes until explicit user guidance; do not treat system prompts as guidance.
  • Guardrails: forbid error suppression or deleting tests; clarify "same error" and when to reset the counter.

^{Written for commit 1b73ee1. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 1 file

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

[djdembeck]

I have this issue consistently, would love to see this merged