pr-review

name	pr-review
description	Review a GitHub pull request chunk-by-chunk with inference-driven findings (security, correctness, performance, etc.), persisted per review and browsable as a two-pane diff viewer on the mosaic dashboard. Use when asked to review a PR, generate a PR review, or check past PR review findings.
version	3.0.0
kind	hybrid
triggers	["review this PR","review pull request","review <github-url>","pr review for <url>","next chunk","set verdict for <review>"]
intent	code-review
config_dir	~/.config/skill-config/pr-review
created_at	2026-05-30T00:00:00.000Z
updated_at	2026-07-30T00:00:00.000Z
guardrails	["Do not post, comment, or push the review anywhere — findings are persisted locally only, browsable on the mosaic dashboard","Do not reproduce secrets/tokens/credentials found in a diff inside a finding's comment — reference that one is present, don't repeat it","Every finding must carry a concrete side+line the diff renderer can place — never a free-text line range or description","Read-only against the reviewed repo/PR — never push, comment, or merge on GitHub"]
resources	["<SKILL_PATH>/scripts/run.sh","<SKILL_PATH>/webapp/app.json","<SKILL_PATH>/references/schema.md","<SKILL_PATH>/references/review-criteria.md"]
tools	["bash","jq","gh","awk"]

PR review

Hybrid skill: scripts/run.sh handles every deterministic step — fetching a PR via gh, chunking its diff, validating and persisting findings, rolling up counts into the dashboard manifest. The only judgment call — reading a diff chunk and writing findings — is the agent's. Diff parsing (unified diff → line-numbered two-pane rows) is likewise deterministic and lives entirely in webapp/static/js/app.js, never in the agent's output.

Workflow

1. Fetch.

   <SKILL_PATH>/scripts/run.sh fetch <PR_URL> [-s <story_file>]
   

   Prints review_id, scratch_dir (holds chunk_N.diff files to read), chunks (count), and the meta path. First run onboards the skill into mosaic automatically (idempotent after).

2. Review each chunk (agent judgment). For chunk N:

   • Read <scratch_dir>/chunk_N.diff with the host agent's file-read capability.
   • Apply references/review-criteria.md — 9 categories, must-fix / should-fix / suggestion severities.
   • Anchor every finding to a concrete line per references/schema.md § Line anchoring (side: RIGHT + new-file line for an added/context line, side: LEFT + old-file line for a removed line) — read the chunk's @@ -a,b +c,d @@ headers to count accurately, don't guess.
   • Write the findings JSON (shape: references/schema.md § "Agent-authored findings file") to a scratch path, then persist:

     <SKILL_PATH>/scripts/run.sh persist-chunk <review_id> <N> <findings_file>
     

     This validates the shape, merges in the raw diff, and rolls counts into manifest.json.

3. Set the overall verdict once every chunk is persisted:

   <SKILL_PATH>/scripts/run.sh set-verdict <review_id> <APPROVE|REQUEST_CHANGES|NEEDS_DISCUSSION>
   

   Refuses if any chunk is still unreviewed.

4. Deliver. Tell the user to start mosaic if it isn't already running ($TOOLS_PATH/mosaic/quick-start.sh start) then open http://localhost:<mosaic-port>/mosaic/apps/pr-review/ (default port 47500). There is no standalone file to hand over — the dashboard is the only deliverable. A review's detail page is paginated one chunk per page — each page lazily fetches only its own , never the whole review at once, so a large PR never means loading its entire diff into one view. writes every chunk's diff immediately (), so all pages are browsable right away; a page whose chunk hasn't been reviewed yet still shows the diff, just with no findings and a "not yet reviewed" banner.