Review notes protocol

Date: 2026-10-02 Status: approved

Purpose

Record code-review results as git notes under refs/notes/review so that the next reviewer — human or AI — can read the current verdict, see open findings with stable IDs, and continue from where the last review stopped. This replaces a CI reviewer: reviews happen locally, results travel with the commits they judged.

Storage model

One note per reviewed commit, in the review notes ref.
The tip of a reviewed branch always carries the latest verdict.
Re-reviewing after fixes writes a new note on the new tip commit. Old notes stay attached to old commits as history.
Each note links to the previously reviewed commit via previous, forming a walkable chain of review rounds per branch.
Merging to main needs no special handling: the approve note on the final tip is the record.

Note format

Line-based header, then a blank line, then free markdown body:

branch: <branch name>
round: <integer, starting at 1>
verdict: approve | request-changes
previous: none | <sha of previously reviewed tip>
findings:
  - id: F<N>
    severity: blocker | major | minor
    file: <path>:<line>
    title: <one line>
    status: open | resolved | wontfix
## Details

<markdown prose>

Rules:

Finding IDs are unique within a note and carry forward across rounds: a finding resolved since the last note keeps its ID and moves to status: resolved; an unfixed finding keeps its ID and status: open. New findings get fresh IDs that do not collide with carried-forward ones.
The body is free markdown; the ## Details heading in the example is illustrative, not required.
Headers are plain key: value lines so the helper script can parse them with sed/awk.

Helper script: `scripts/review-note`

Bash, git plumbing only. Subcommands:

review-note show [commit] — print the note for a commit (default: current tip). Exit 1 with a message on stderr if there is no note.
review-note write <file> — attach the note in <file> to the current tip commit. Refuses to replace an existing note unless --force is given; a proper re-review targets a new commit, so overwriting is always deliberate. Validates that branch, round, and verdict are present and that verdict uses the vocabulary above. No other validation.
review-note latest [branch] — print the newest note for a branch: start at the branch tip (default: current branch), and walk back through previous until a note is found. Exit 1 if the chain is exhausted without a note.
review-note push — git push origin refs/notes/review. Plain git push ignores notes, so this is the only way results travel.

push pushes without force; if the remote rejects, report the failure and let the user decide.

Lifecycle

Reviewer reads review-note latest <branch> (nothing printed means first review).
Reviewer reviews the branch tip per REVIEW.md.
Reviewer writes the note with review-note write and runs review-note push.
If the verdict is request-changes, the author pushes fixes as new commits. The next review round starts at the new tip with round+1, previous set to the last reviewed tip, and carried-forward findings.
When the verdict is approve and the branch merges, the notes remain on the branch commits as the review record.

Documentation integration

REVIEW.md gains a "Recording reviews" section describing the protocol and the four script commands. This is the single source of truth for the protocol.
AGENTS.md "Reviewing" section gains one line pointing at the section, so any AI session picks up the flow automatically.

Seeding the experiment

The failing review of home_assistant_controller_cleanup (commit ba6e0e8) becomes note round 1, with these findings:

F1 (blocker, src/AFRP.hs:273): hlint's \_ -> x → const x rewrites break the rank-2 Mealy type; library does not compile.
F2 (blocker, test/AFRPLawsSpec.hs:98): seven law tests reduced to tautologies by hlint rewrites; see :103, :212, :221, :228, :235, :256.

Verification

The script is verified by round-trip checks in a throwaway temp clone: write, show, latest (including chain walk across two rounds), force-overwrite refusal and acceptance, and push. No Haskell or cabal changes, so no cabal2nix regeneration.

4.3 KiB Raw Blame History