chess-review-bot 0.1.0

Classify git diffs using chess.com-style move categories (Brilliant, Blunder, and everything between).

chess-review-bot

chess.com's Game Review vocabulary — Brilliant!!, Great!, Best, Excellent, Good, Book, Inaccuracy?!, Mistake?, Blunder??, Miss — pointed at a git diff instead of a chess move.

chess-review-bot  ──────────────────────────────────────────────
  Diff:        HEAD~1..HEAD  ·  3 files  ·  +142 / -38 lines
  Accuracy:     71/100
──────────────────────────────────────────────────────────────

  src/auth/session.py                                  Blunder??
    → touches critical path with no matching test changes
    → "chess-review-bot says: hanging your queen in the auth module
       without a single test is how production incidents are born."

  src/utils/formatting.py                              Book
    → dependency/formatting-only change, nothing to see here

  tests/test_session.py                                Great!
    → tests added alongside a meaningful source change

──────────────────────────────────────────────────────────────
Result: FAIL (1 Blunder)  ·  Exit code: 1

The one design decision that matters

Category is never LLM-decided. Deterministic signals (size, test-file overlap, critical-path membership, force-push, revert, commit message quality, credential leakage) run through a fixed rule table. Delete commentary.py; classification doesn't change.

LLM's only job (optional): phrase an already-decided category in one dry sentence. Same lesson as position-evaluator: LLMs are bad at inventing a severity scale, good at picking from a fixed one.

Install

pip install chess-review-bot

Zero required runtime deps. Commentary is an opt-in extra:

pip install "chess-review-bot[commentary]"

No extra / no Gemini key: still classifies and reports correctly, just prints the deterministic reason instead of a generated sentence.

CLI usage

chessreview                       # HEAD~1..HEAD in current repo
chessreview HEAD~5..HEAD          # wider range
chessreview my-change.diff        # saved diff file
git diff | chessreview -          # stdin

Option	Description
--critical PATTERN	Critical-path glob. Repeatable, adds to defaults.
--only-critical PATTERN	Replace default critical-path globs. Repeatable.
--large-threshold N	Lines changed = "large". Default 400.
--moderate-threshold N	Lines changed = "moderate". Default 100.
--force-pushed	Flag as force-pushed (local testing only).
--revert	Flag as a revert.
--enable-commentary	Gemini one-line commentary. Requires a key.
--gemini-key TEXT	Or set GEMINI_API_KEY.
--format [text|json|markdown]	Default: text.
--summary	One-line output.
--debug	Show per-file signals.

Exit codes: 0 clean, 1 Blunder found, 2 tool error.

GitHub Action

- uses: actions/checkout@v4
  with:
    fetch-depth: 0   # needed for force-push ancestry check

- uses: SemTiOne/chess-review-bot@v1
  with:
    github-token: ${{ secrets.GITHUB_TOKEN }}
    gemini-key: ${{ secrets.GEMINI_API_KEY }}   # optional
    fail-on-blunder: 'true'

Input	Description	Default
github-token	pull-requests: write + contents: read.	required
gemini-key	Optional, enables commentary.	''
critical-paths	Newline-separated globs.	built-in defaults
large-threshold	Lines changed = "large".	400
moderate-threshold	Lines changed = "moderate".	100
fail-on-blunder	Fail check on any Blunder??.	true
post-comment	Post/update the PR review card.	true

One comment per PR, updated on every push (idempotency marker <!-- chess-review-bot-managed-comment -->), never a second comment.

### ♟️ chess-review-bot — PR Game Review

**Accuracy: 71/100** · 3 files · +142/-38 · 1 Blunder??

| File | Category | Why |
|---|---|---|
| `src/auth/session.py` | Blunder?? | critical path, no tests, no description |
| `src/utils/formatting.py` | Book | dependency/formatting-only change |
| `tests/test_session.py` | Great! | tests added alongside a real change |

What this deliberately does NOT do

• Never scores or names a person. Diffs and commits only, never authors. An earlier concept scored people ("most toxic collaborator"); rejected on purpose: a screenshotted report naming someone is an HR incident, and no manager installs a tool that can start a fight on their team.
• Doesn't find bugs. Severity communication, not a correctness/security scanner. Pair with your existing linters/CI.
• Commit-message quality is a blunt denylist heuristic, not sentiment analysis.
• Force-push detection is Action-mode only. Compares synchronize event's before/after SHAs via git merge-base --is-ancestor (needs fetch-depth: 0). CLI can't infer this from reflog, pass --force-pushed to test it locally. (Earlier draft assumed a forced payload field; that field only exists on push events, not pull_request. See CHANGELOG.md.)

Exit codes

Code	Meaning
0	No file classified Blunder??
1	At least one Blunder??
2	Tool error (bad input, invalid config, git failure)

Companion tools

• position-evaluator — the original: chess terminology for personal decisions, not code.
• envreap — dead environment variables.
• stalebot-local — stale TODO/FIXME/HACK via git-blame age.

License

MIT