[Dripar]

What this is

A Git pre-commit hook that reads the staged diff, reasons about it with Claude, and refuses the commit when it finds a runtime hazard a linter structurally cannot see — with a ready-to-paste patch in the terminal.

It targets three defect classes, deliberately narrow:

Hazard	Shape
TRANSACTION_FAULT	Durable write that can fail partway through with no rollback or compensating action
IDEMPOTENCY_RISK	At-least-once handler applying side effects with no de-duplication key
RACE_CONDITION	Unsynchronised shared state, or a check-then-act (TOCTOU) window

Style, naming and formatting are explicitly not blocking — the linter already owns those.

Design decisions worth reviewing

Two-tier gate. Tier 1 is deterministic and free: empty diffs, docs, config, images and lockfiles never reach the API. Only code-bearing diffs cost a call.

Fail open, always. Missing key, dead network, timeout, malformed response, crashed interpreter — every failure path exits 0 with a warning. A guardrail that blocks a commit because its own dependency is down trains the team to run --no-verify reflexively, which disables it permanently. Blocking is reserved for the one case where the model actively asserts a hazard.

Structured outputs, not a forced tool call. output_config.format with a JSON Schema guarantees a schema-valid verdict, and composes cleanly with adaptive thinking — which is where the quality on this task actually comes from. No regex or string matching anywhere in the decision path.

Dashboard. A blocked commit leaves no trace in git history, so the tool's most valuable moments are invisible to git log. Verdicts are appended to .git/sentinel-audit.jsonl and dashboard.py joins them against the DAG. Stdlib only — no Flask, no npm, no build step.

Bugs caught while testing this

Both were found by running it, not by reading it:

1. The hook blocked every commit on Windows. Interpreter selection used command -v python3, which resolves to the Microsoft Store stub — it exists, is executable, prints "Python was not found" and exits non-zero, so Git treated it as a hook failure. Now each candidate is probed, and any exit status other than 0/1 fails open.
2. Docs-only commits were billed for an API call. os.path.splitext("LICENSE") and splitext(".gitignore") both return an empty extension, so extensionless files and dotfiles fell straight through the zero-token gate. Fixed with a filename allowlist; both cases are now regression-tested.

A third, caught by a cross-check over the examples: event_id is a parameter in the race-condition fixture, so it matched the idempotency rule first and reported the wrong hazard. Rules are now ordered by specificity.

Testing

• 24 unit tests covering the deterministic gate, all three hazard classifications, and audit-log failure modes
• Both demo scenarios run end-to-end through a real git commit (reject, then pass)
• Fail-open verified with no credentials present
• CI runs the suite on Ubuntu + Windows against Python 3.10 and 3.12

Not tested: the live API path. Every run used SENTINEL_MOCK=1 — there was no API key on the build machine. The request shape is built to the current SDK, but it has never made a real call. That is the first thing a reviewer should exercise.

Try it

pip install -r requirements.txt
export ANTHROPIC_API_KEY="sk-ant-..."
./install.sh

cp examples/transaction_fault_unsafe.py payment_service.py
git add payment_service.py && git commit -m "Add payment deduction"   # refused

git reset
cp examples/transaction_fault_safe.py payment_service.py
git add payment_service.py && git commit -m "Add payment deduction"   # passes

python dashboard.py