Add plan for visualizing experiment results in the debugger

[smirnovlad]

No description provided.

[smirnovlad]

(Reposting with proper formatting — ignore the previous review)

Nice implementation — follows the plan well and the app.js changes for offline mode / sample navigation are solid. One critical bug and a couple of minor things.

Critical: import before sys.path (convert_results_to_debugger.py:23)

from service_app.core.debugger_events import ... runs before sys.path.insert on line 26, so Python can't find the service_app package → ModuleNotFoundError on every run.

Fix: move PROJECT_ROOT / sys.path.insert above the import:

PROJECT_ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(PROJECT_ROOT))

from service_app.core.debugger_events import convert_strategy_result_to_debugger_run

Minor: import yaml dependency

pyyaml is not in the base deps (only in service extras). Consider adding a note or using omegaconf (already a dep via hydra).

Nit: speculative keys in run_tts_eval.py

all_traces and all_trajectory_steps — I don't see any strategy returning these keys. Not harmful (the if key in result guard skips them), but consider removing to avoid confusion.

Serialization: looks good

Checked: save_results_json uses safe_serialize() which handles StepCandidate objects. And for beam search, step_candidates entries are already plain dicts. No issues here.

app.js offline mode: looks good

Graceful fallback to cached examples when API is down, auto-enable cached mode toggle, sample navigation with prev/next and incorrect filter — all solid.

[smirnovlad]

One more thing: scripts/README_debugger.md feels misplaced — we keep docs in docs/. Please move it to docs/service/debugger.md (next to the existing docs/service/SERVICE_API_GUIDE.md).

[smirnovlad]

I've pushed two follow-up commits on this branch:

1. 1069ce5 — Fixed the critical import order bug in convert_results_to_debugger.py (moved sys.path.insert before the service_app import)
2. 7d3e57e — Restored all_traces and all_trajectory_steps keys that I incorrectly removed (they are used by debugger_events.py for self-consistency and offline BoN visualization)

Remaining suggestion: move scripts/README_debugger.md to docs/service/debugger.md (see comment above).

[smirnovlad]

Final review after fixes

All previously identified issues have been resolved:

• Import order — fixed (sys.path.insert now comes before service_app import)
• Linting — flake8 (noqa: E402), black, isort all pass
• Tree data keys — all 7 keys restored; verified each is used by at least one strategy + debugger_events.py
• Docs location — moved to docs/service/debugger.md

What this PR does

1. run_tts_eval.py — saves tree visualization data that strategies already compute but was previously discarded
2. convert_results_to_debugger.py — new converter script that transforms experiment results into debugger-compatible JSON
3. app.js + index.html — file upload, sample navigation (prev/next), incorrect only filter, offline mode fallback
4. docs/service/debugger.md — usage docs
5. Test config — quick beam search config for MATH-500 with OpenRouter

Non-blocking note

pyyaml dependency: the converter uses import yaml which requires pyyaml. It is available via hydra (already a dep for eval), but worth mentioning in docs if someone runs the converter in a minimal env.

Should be tested end-to-end by the author: run an experiment with updated run_tts_eval.py, then run the converter and verify the debugger renders the tree.

LGTM from my side.

[smirnovlad-test]

1. Not clear names of entries in selector of cached examples
2. "Выберите файл" -> add styles
3. Hardcoded "Online bon" for each cached example? It is like a constant
4. More tests (with bigger beam size, number of candidates and so on)