Review this session. Reflection was pushing to write code. But it is wrong. The reflection has to classify that is the research task and code shouldn't be writter.n't

[dzianisv]

Review this session. Reflection was pushing to write code. But it is wrong. The reflection has to classify that is the research task and code shouldn't be writter.n't

Please create an evaluation test with similar research task, and make sure that reflection doesn't push to do things that are not expected.

## Reflection-3: STOP: Planning Loop Detected
You have been reading files, checking git status, and creating todo lists without writing any code.
DO NOT:
- Run git status or git log again
- Create another todo list
- Read more files "for context"
- Say "let me get right to work" without actually working
DO NOW:
Pick the FIRST item from your existing todo list and implement it. Open a file with Edit or Write and make changes. If you don't know where to start, create the simplest possible file first.
Start coding NOW. No more planning.

Full session

session-ses_392b.md

[dzianisv]

Update: I’ve implemented the fix and added coverage so the planning-loop toast only triggers for coding tasks. New unit test validates research tasks don’t get the “Start coding NOW” feedback.

Changes:

• reflection-3.ts: gate planning-loop feedback behind task type (only coding).
• reflection-3.test-helpers.ts: added shouldApplyPlanningLoop helper.
• test/reflection-3.unit.test.ts: new test for research case.
• test/tts.test.ts: added jest.setTimeout(15000) for Whisper integration + deps checks to stabilize CI (existing tests were timing out at 5s).

Tests:

• npm test ✅
• npm run eval:judge ✅ (23/23)
• npm run eval:compression ✅ (12/12)
• npm run eval:stuck ❌ still 2 failing (unchanged from baseline).

I can dig into the stuck eval failures next if you want.

[dzianisv]

Follow-up: I dug into eval:stuck and fixed the two failing cases by tightening the prompt’s priority and planning-loop rule.

Changes:

• evals/prompts/stuck-detection.txt: enforce WORKING when any tool is running/pending, and explicitly forbid “complete” when the planning-loop read-only rule applies.

Tests:

• npm run eval:stuck ✅ (13/13)
• npm test ✅
• npm run eval:judge ✅ (23/23)
• npm run eval:compression ✅ (12/12)

Let me know if you want a PR.