hsablonniere/paper-review

Paper Review

Conversational code review as a living markdown file. The agent writes findings as quoted lines; the reviewer annotates in plain text between or after those lines; the agent re-reads, acts on clear decisions, asks follow-ups, and rescans for new issues. The file is the conversation.

Protocol (non-negotiable)

• > quoted lines = agent voice (findings, questions, suggestions, status notes).
• Plain, unquoted text = the human reviewer's annotations.
• Never edit or delete the human's plain-text lines. Only add or update > lines.
• One sentence per line inside agent blockquotes.
• Item status lives in the item title, one of: [OPEN] (default), [DONE], [WONTFIX], [DISCUSS]. The agent owns it.

File location

• .reviews/<branch>.md — slashes in the branch name become - (e.g. feat/foo → feat-foo.md).
• .reviews/ is expected to be ignored via the user's global gitignore (git config --global core.excludesfile). Do not modify the project's .gitignore. If a review file shows up in git status, tell the user to add .reviews/ to their global excludes once — never commit review files.

Engine — thin orchestrator, one fat sub-agent

The whole job runs inside one sub-agent (Task tool). The main thread is a dispatcher, nothing more.

The orchestrator (main thread) does ONLY this:

1. Read the current branch (git branch --show-current) to name the file.
2. Parse any modifier from the invocation: working tree, staging/staged, rescan, or an explicit ref/range/PR (default: branch vs base).
3. Spawn one sub-agent, passing: the branch name, the modifier, and the instructions below (point it at references/directives.md and references/format.md — do not read them yourself).
4. Relay the sub-agent's short summary to the reviewer.

The orchestrator must NOT — these all belong to the sub-agent:

• grep or read source files, resolve symbols, or map call-sites
• compute or look at the diff (never let diff content enter the main thread)
• read the review file or the reference files
• re-verify the sub-agent's work afterwards — trust its summary; if confidence matters, tell the sub-agent to self-verify and report it

Anything beyond the four orchestrator steps above is a leak. Resist the urge to "just check".

Sub-agent brief

Give the sub-agent this whole brief. It does everything: resolve, explore, decide, write, self-verify.

Setup

• File path: .reviews/<branch>.md — slashes in the branch become -.
• Get the date itself: date +%F (use in the title).
• Detect the mode with a single check: test -f .reviews/<branch>.md → exists = apply, else = write. Treat a later read failure as "fall back to write".
• Resolve the target:
  • default → base = whichever of main/master exists (git rev-parse --verify), diff <base>...HEAD
  • working tree → git diff HEAD plus contents of untracked files (git ls-files --others --exclude-standard)
  • staging/staged → git diff --cached
  • explicit arg → that ref/range/PR
• Shell may be fish: avoid bash-isms and unescaped globs (grep --include=*.js fails); prefer git grep / git ls-files.

Write mode

• Read the diff and only the surrounding context needed to judge it.
• Produce items across four axes: bugs/correctness, simplification/readability, architecture/design, open questions.
• Apply the false-positive filter from references/directives.md — keep signal, drop nitpicks.
• Write .reviews/<branch>.md per references/format.md.
• Return counts (items per axis) + a one-line summary.

Apply mode — do all localization yourself (grep/read source as needed; do not expect the orchestrator to have done it). For each item:

• Human gave a clear decision/instruction → make the change in the working tree, set [DONE], add a > done: <what changed> line citing file:line.
• Human decided not to act (with reason) → set [WONTFIX], add > wontfix: <reason restated>.
• Human asked a question or the answer is ambiguous → set [DISCUSS], add a > follow-up question or answer.
• Untouched item → leave [OPEN].

Then rescan the diff against the same base as the original review (working-tree edits just made are included) and append genuinely new items as [OPEN], numbered from max(existing number) + 1. Never renumber existing items, even when non-contiguous after a [WONTFIX].

Self-verify the edits (e.g. a final git grep that the changes landed and nothing stale remains) and state the result in the summary. Return: what changed in code + what now awaits the reviewer ([DISCUSS] / [OPEN] counts).

Boundaries (sub-agent)

• Edit only the working tree. Never commit — the reviewer commits.
• Never touch the human's plain-text lines; only add/update > lines.
• Preserve item numbers across runs.
• Do not run build/typecheck/lint as part of review (CI handles those).

References

• references/directives.md — review axes, context to pull (CLAUDE.md, git history, prior PR comments, code comments), the anti-false-positive list, and the confidence filter.
• references/format.md — exact file template plus a worked, annotated round-trip example.