| name | pr-report |
| description | This skill should be used when the user asks for a PR report, PR audit, PR delta analysis, /pr-report, or wants per-PR summary with delta files/lines. Generates a structured report for one or more open PRs covering purpose, files changed, +/− line counts, /code-standards + /thermo review findings, /ponytail simplification opportunities, file-overlap combination candidates, and /green status per pr-green-definition. Output: machine-readable JSONL + human-readable markdown. Read-only by default; can apply simplifications if user opts in. |
/pr-report — Per-PR Delta Audit & Quality Review
Use when the user invokes /pr-report [<PR#s>] or asks for a structured audit of one or more open pull requests. The skill defines the methodology, lane protocol, model routing, and output format. The Claude command at ~/.claude/commands/pr-report.md dispatches the work.
Inputs
- PR#s (space-separated) — defaults to all open prompt PRs in the repo if omitted. Resolved via
gh pr list --state open --json number,title,headRefName,files --limit 100.
- Repo — defaults to current
git remote get-url origin | sed ... derivation; can be overridden.
--apply (optional) — when set, the audit may apply non-destructive simplifications (extract shared rules, add unit tests). Default = read-only.--push (optional) — when set, write the human-readable report to ~/roadmap/<repo-name>-pr-report-YYYY-MM-DD.md and push to the roadmap repo. Default = local-only.--baseline (optional, default origin/main) — the merge-base to diff against for delta stats.
Methodology (4 lanes, parallel)
The Claude command spawns one Sonnet sidekick teammate per the /sidekick skill. The sidekick fans out 4 anonymous lanes per the /swarm rules. All lanes set model: explicitly per CLAUDE.md "Subagent model routing":
| Lane | Model | Job |
|---|
| A — per-PR delta stats | haiku (mechanical) | title, purpose 1-paragraph, base/head SHA, files-changed count + list, +lines, -lines, net, top 3 files by lines added |
| B — /code-standards + /thermo review | sonnet (review) | for PRs >50 lines added OR >3 files changed: prompt consistency with $PROJECT_ROOT/prompts/AGENTS.md (setting-agnostic, no banned entities, generic placeholders, no meta-commentary, no version history), repetition/mirroring (shared/ candidate), redundant phrasing, anti-patterns |
| C — /ponytail simplification analysis | sonnet (review) | per-prompt-PR ponytail ladder: (1) does the rule need to exist at all, (2) duplicate elsewhere, (3) one-liner possible, (4) move to shared/, (5) minimum text |
| D — combination analysis | haiku (mechanical) | file-level overlap matrix across all PRs in scope, cluster PRs that touch same files, propose merge candidates |
Output
- Machine-readable:
/tmp/pr_report.jsonl — one JSONL line per PR with {pr, title, purpose, files, additions, deletions, net, top_files, review_findings, simplifications, combination_candidates}
- Human-readable:
/tmp/pr_report.md — per-PR table + review findings + simplification recommendations + combination candidates + suggested merge order
- Optional push: when
--push is set, copy pr_report.md to ~/roadmap/<repo>-pr-report-YYYY-MM-DD.md and push to the roadmap repo (commit message includes the date + the repo name).
Lane protocol (per lane)
- Lane A (haiku): mechanical
gh api calls only. No file modifications. Append to /tmp/pr_report.jsonl.
- Lane B (sonnet): read-only diff inspection via
gh pr diff <N>. Output structured findings to /tmp/pr_report.jsonl. Cite file:line for every claim.
- Lane C (sonnet): read-only prompt-rule inspection. Cite
$PROJECT_ROOT/prompts/AGENTS.md rule number for every anti-pattern. Output to /tmp/pr_report.jsonl.
- Lane D (haiku): file-overlap matrix via
gh pr view <N> --json files. Output to /tmp/pr_report.jsonl.
After all 4 lanes complete, the sidekick synthesizes /tmp/pr_report.md and surfaces to the main session.
/green integration
For each PR, also compute the canonical 2-gate /green status (per pr-green-definition skill):
gate1=$(gh pr view N --json statusCheckRollup --jq '[.statusCheckRollup[] | select((.__typename == "StatusContext" and (((.state // "") | ascii_upcase) != "SUCCESS")) or (.__typename == "CheckRun" and .name != "Green Gate" and .name != "Cursor Bugbot" and ((((.status // "") | ascii_upcase) != "COMPLETED") or (((.conclusion // "") | ascii_upcase) != "SUCCESS"))))] | length')
gate2=$(gh pr view N --json mergeable --jq '.mergeable')
gate1 == 0 AND gate2 == "MERGEABLE" → MERGE_READYgate1 == 0 AND gate2 == "UNKNOWN" → PENDING (re-poll)gate1 == 0 AND gate2 == "CONFLICTING" → BLOCKED_CONFLICTgate1 > 0 → BLOCKED_CI (real test failure)- Otherwise →
BLOCKED
Green Gate and Cursor Bugbot checks are advisory and excluded from Gate 1; CodeRabbit/Bugbot feedback is advisory across all phases (per draft-first-pr + pr-green-definition skills).
Conventions
- Read-only by default.
--apply is the explicit opt-in.
- No
gh pr merge — user issues MERGE APPROVED separately.
- Cost-route — haiku for mechanical stats, sonnet for review/narrative. No top-tier (opus/fable) unless user explicitly requests.
- Max 4 parallel lanes (per
/swarm rule 4 — bursts kill, ramps don't).
- Sidekick durability: STATE.md at
$CLAUDE_STATE_DIR/<repo-slug>/sidekick/pr-report-<date>/STATE.md; resumption bead P1.
Example
/pr-report # audit all open prompt PRs
/pr-report 8564 8628 8443 # audit specific PRs
/pr-report --apply --push # audit + apply simplifications + push to roadmap repo
Related skills
/code-standards — code/PR review standards/thermo (thermo-nuclear-code-quality-review) — strict structural audit/ponytail — lazy senior dev ladder/pr-green-definition — canonical 2-gate /green status check/draft-first-pr — PR lifecycle (draft → /es → /er → /advice → /green)/sidekick — durable Sonnet teammate wrapper/swarm — multi-agent orchestration playbook/e — cost-aware model routing
