pr-review-local

pr-review-local

Pre-PR local code review. Output to terminal only — no GitHub interaction. Optionally apply fixes once with --fix, or loop review→fix→re-review with --goal until the review passes cleanly.

Usage

/pr-review-local                       # review current branch vs default base
/pr-review-local <BASE_BRANCH>         # review against an explicit base branch
/pr-review-local --fix                 # review and apply fixes once (unstaged; refuses on dirty tree)
/pr-review-local --goal                # loop review->fix->re-review, commit each iteration, until clean
/pr-review-local --goal --max-iters 8  # raise the loop ceiling (default 5)
/pr-review-local --fast                # skip the documentation agent (cheapest meaningful cut)
/pr-review-local <BASE_BRANCH> --fix   # flags combine freely

<BASE_BRANCH> is positional and must NOT begin with --. Flag order is otherwise free.

--fast excludes the documentation agent via the engine's <EXCLUDE_AGENTS> input — the most expensive lens per launch and the most likely to return clean on code-focused diffs. Use the default (full panel) when the diff touches Markdown, the §10 inventory, or public-API docs.

--goal is the autonomous loop: it reviews, fixes critical/high/medium findings, re-gates (format → lint → typecheck → test, scoped to the changed surface — see Goal mode), commits, and re-reviews until no actionable findings remain. It commits each iteration and therefore refuses on a dirty tree. --goal supersedes --fix (loop-and-commit beats single-shot-unstaged); if both are passed, --goal wins. There is no post-convergence runtime check / --no-runtime here — this is a headless SDK with no route-reachable UI, so no dev server is ever booted.

Validating end-to-end

A maintainer changing this skill should verify each outcome shape:

Scenario	Expected last line
Clean branch, no findings	Sentinel: REVIEW_CLEAN — no issues found in <HEAD_BRANCH> vs <BASE_BRANCH>.
Findings present	Sentinel: REVIEW_DONE_LOCAL — <N> findings (X critical, Y high, Z medium, W low) on <HEAD_BRANCH> vs <BASE_BRANCH>.
Findings + agent crash	Sentinel: REVIEW_DONE_LOCAL — <N> findings (X critical, Y high, Z medium, W low) on <HEAD_BRANCH> vs <BASE_BRANCH>. (with a WARNING: <FAILED_AGENTS> of <TOTAL_AGENTS_LAUNCHED> agents failed (<names>) — review may be incomplete. line prepended to the findings output)
Zero findings + agent crash	Sentinel: REVIEW_INCOMPLETE — <FAILED_AGENTS> of <TOTAL_AGENTS_LAUNCHED> agents failed (<names>); no findings does NOT mean clean.
--fix happy path	Sentinel: FIX_DONE_LOCAL — <X> applied, <Y> skipped (Local-only, unstaged). plus git diff shows the unstaged edits.
--fix aborted on dirty tree	Sentinel: FIX_ABORTED — working tree is not clean. Commit or stash before --fix.
--goal converges	Sentinel: GOAL_CLEAN — review passes cleanly after <i> iteration(s) on <HEAD_BRANCH> vs <BASE_BRANCH>; <K> low finding(s) triaged (not auto-fixed). plus one fix(review): iteration N commit per fixing pass.
--goal on already-clean branch	Sentinel: GOAL_CLEAN — ... after 1 iteration(s) ... (idempotent: no commits made).
--goal aborted on dirty tree	Sentinel: GOAL_ABORTED — working tree is not clean; commit or stash before --goal.
--goal aborted on detached HEAD	Sentinel: GOAL_ABORTED — detached HEAD; check out a branch before --goal.
--goal aborted on red base gate	Sentinel: GOAL_ABORTED — base gate is red (<TEST_CMD> fails before any fix); fix it or run without --goal.
--goal same findings twice	Sentinel: GOAL_STUCK — identical findings on iteration <i> and <i-1>; stopping for user input.
--goal hits the ceiling	Sentinel: GOAL_MAXED — <N> actionable finding(s) remain after <MAX_ITERS> iteration(s); extend, accept, or stop?
Re-run, input unchanged (cache hit)	cached findings reprinted under a (cached — input unchanged since <head_sha>) header + a reuse/re-review prompt; on reuse, the matching REVIEW_* sentinel from the cached counts (Step 2c — single-shot only)