Skip to content

docs: root CLAUDE.md + benchmark guide overhaul + code-health report - #37

Open
duanyiqun wants to merge 1 commit into
mainfrom
claude/pr-review-arv8t7
Open

docs: root CLAUDE.md + benchmark guide overhaul + code-health report#37
duanyiqun wants to merge 1 commit into
mainfrom
claude/pr-review-arv8t7

Conversation

@duanyiqun

Copy link
Copy Markdown
Contributor

Summary

Docs-only change establishing the repo's development guide and cleaning up benchmark documentation. No src/ code touched.

  • CLAUDE.md (new) — the root development guide that docs/BENCHMARK.md and docs/PHILOSOPHY.md already declared themselves extensions of (it never existed until now). Covers: commands (with the pinned ruff/pyright versions and lint scope), architecture map + dependency-layering rules, code conventions (typing, logging, Pydantic-vs-dataclass split, node ID prefixes), testing reality (pytest is smoke-only; what real verification looks like), benchmark quick-reference with the judge/branch locks, branch & deployment model, and gotchas.
  • docs/BENCHMARK.md (overhauled)
    • fixed the --event-stream guidance: the flag exists only in the locomo runner (README's description was correct; the old "ALWAYS pass it to every benchmark" instruction was an argparse error on 8 of 9 runners)
    • removed 5 dead links to the nonexistent docs/benchmark/ directory
    • aligned the File Map / status tables with the 9 real benchmark dirs (7 retired ones removed; longmemeval's different entry point documented)
    • corrected the false "datasets auto-download" claim and the MODEL default vs paper-stack mismatch
    • added an environment-variable table and a step-by-step "Contributing a new benchmark" guide (mutual = template; lists all four hardcoded lists a new benchmark must be wired into)
  • docs/CODE_HEALTH.md (new) — evidence-based structural findings with file:line citations, three tiers, fix directions, and a suggested sequencing. Tier-1 items are also filed as issues.
  • .env.example — documented the benchmark env vars (previously undocumented) and fixed the duplicate COGNIFOLD_SESSION_BACKEND definition.

Verification

  • Every command referenced in CLAUDE.md checked against Makefile / scripts/ (all exist)
  • grep for the removed dead links returns zero matches
  • CODE_HEALTH.md line references spot-checked against source (dispatch dict query/agent.py:789-808, config.py:161 service import, session.py:346/365 key persistence, serve.py uvicorn workers, register_domain:1494, deprecated embeddings call sites in executor/runner.py)
  • Docs-only + .env.example comments — outside the ruff/pyright/CI lint scope

🤖 Generated with Claude Code

https://claude.ai/code/session_01VnAaidzJkganncouErUp3m


Generated by Claude Code

- CLAUDE.md (new): commands, architecture map + layering rules, conventions,
  testing reality, benchmark quick-ref, branch/PR rules, gotchas. Two docs
  already declared themselves "extensions of CLAUDE.md" — it now exists.
- docs/BENCHMARK.md: remove 5 dead links to nonexistent docs/benchmark/,
  align File Map/status tables with the 9 real benchmark dirs (7 retired
  ones removed), fix the --event-stream guidance (flag exists only in
  locomo; other runners consolidate unconditionally in base_runner),
  document longmemeval entry points, correct the auto-download claim, add
  an environment-variable table and a "Contributing a new benchmark" guide.
- docs/CODE_HEALTH.md (new): evidence-based structural findings (file:line)
  in three tiers with fix directions and sequencing.
- .env.example: document benchmark env vars; fix duplicate
  COGNIFOLD_SESSION_BACKEND definition.

Docs-only change; no src/ code touched.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VnAaidzJkganncouErUp3m
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants