gh-review-triage

PR Review Triage

Triage your PR review queue into a single prioritized table. Combines a mechanical decision tree (draft status, CI, reviewer assignment, staleness) with LLM judgment (type classification, cognitive load estimation, priority narration) to produce an opinionated ordering.

Includes team-assigned reviews by default. Use --no-team to show only PRs where you are individually named as a reviewer.

Phase 1: Discovery

Fetch all open PRs where review is requested from you (directly or via team membership).

Identify current user

gh api user --jq '.login'

Store the result as $ME for use in all subsequent phases.

Fetch review-requested PRs

gh search prs --review-requested=@me --state=open \
  --json repository,title,number,url,author,createdAt,updatedAt,labels,commentsCount,isDraft \
  --limit 100

Sort results by updatedAt descending (most recently active first).

If --no-team is specified, filtering happens in Phase 2 after enrichment — the search API does not distinguish individual vs team assignment.

Phase 2: Enrich

For each PR from Phase 1, fetch detailed data in parallel (batch 5–10 at a time):

gh pr view <number> --repo <owner/repo> \
  --json additions,deletions,reviewRequests,reviews,statusCheckRollup,body

Derived fields

Compute the following from raw PR data:

• is_bot: author.login ends with [bot] or matches a known bot (renovate, dependabot, github-actions, etc.)
• has_your_prior_review: $ME appears in reviews[].author.login
• author_responded_after_you: find your latest review timestamp, check if updatedAt is after it AND latest activity is from the author
• days_since_update: today minus updatedAt in days
• sole_reviewer: reviewRequests has exactly one User entry and it is $ME
• other_individual_reviewers: User entries in reviewRequests excluding $ME
• you_individually_assigned: $ME appears as a User (not Team) in reviewRequests
• ci_status: aggregate statusCheckRollup — "pass" if all pass, "fail" if any fail, "pending" if any pending and none fail
• is_fresh: days_since_update ≤ 3
• is_stale: days_since_update ≥ 7

Team filtering

If --no-team is specified: drop PRs where you_individually_assigned is false. Only PRs where you are explicitly named as a User reviewer survive.

Phase 3: Classify

Mechanical decision tree. Evaluate conditions top-to-bottom; first match assigns the tier.

Priority cascade

See references/cascade.md for full field definitions and threshold defaults.

Quick type detection

Classify PR type mechanically before falling back to LLM:

1. Labels containing "bug", "hotfix", "fix", "Type: Bug" → bug
2. Title starts with fix:, fix(, hotfix:, hotfix( → bug
3. Labels containing "security" → security (handled by step 5)
4. Otherwise → not a bug (skip step 4, continue cascade)

Phase 4: Evaluate (LLM judgment)

Only runs on 🔴 and 🟡 candidates. Skip ⚪ and 🔵 — they are already classified and don't benefit from LLM refinement.

Input per PR

• Non-bot PRs: read PR body (capped at 500 chars from data fetched in Phase 2)
• Bot PRs: title + labels only (body is auto-generated boilerplate)

LLM evaluates per PR

• Type classification (fallback chain: labels → title prefix → LLM → "unknown"): bug | hotfix | feat | chore | docs | test
• Cognitive load estimation:
  • quick: < 100 lines, single file or test-only, docs, config
  • moderate: 100–500 lines, 2–5 files, straightforward logic
  • heavy: > 500 lines, core logic, multiple components, concurrency
  • Factor in file types — 200 lines of tests are easier than 50 lines of concurrency
• Why narration: one sentence combining cascade reason with title/body context
• Tier override (rare): promote 🟡 → 🔴 if body says "blocking release", "urgent", "production issue"; demote if exploratory/WIP despite not being marked draft

Stale PR judgment (step 9)

For PRs that matched the stale rule, the LLM provides nuanced judgment:

• Abandoned: no comments, author not active → "Stale — likely abandoned"
• Slow-moving: periodic updates, design doc → "Slow-moving, not urgent"
• Needs attention: was active, then went silent → "Stale — may need a ping"

This does NOT change the tier (stays ⚪) but improves the "why" column.

Phase 5: Output

Merge tree tier with LLM refinements. Sort by tier (🔴 → 🟡 → 🔵 → ⚪), then within each tier by cognitive load (quick → moderate → heavy).

Table format

## PR Review Triage ({date}) — {total_count} PRs

| P | PR | Author | Type | Size | Updated | Load | Why |
|---|---|---|---|---|---|---|---|
| 🔴 | [owner/repo#123](https://github.com/owner/repo/pull/123) | alice | fix | +5/-2 | 1h ago | quick | Author responded, your turn |
| 🔴 | [owner/repo#456](https://github.com/owner/repo/pull/456) | renovate | security | +12/-4 | today | quick | Security patch, review soon |
| 🟡 | [owner/repo#789](https://github.com/owner/repo/pull/789) | bob | feat | +200/-30 | 1d ago | moderate | Assigned to you, others can help |
| 🔵 | [owner/repo#101](https://github.com/owner/repo/pull/101) | carol | chore | +8/-0 | today | quick | Others are covering this |
| ⚪ | [owner/repo#201](https://github.com/owner/repo/pull/201) | dave | feat | +300/-40 | 30d ago | heavy | Stale — likely abandoned |
| ⚪ | [owner/repo#301](https://github.com/owner/repo/pull/301) | eve | feat | +1200/-0 | 5d ago | — | Draft |

Footer

🔴 {n}  🟡 {n}  🔵 {n}  ⚪ {n} — Review 🔴 first, ~{estimate} estimated

Time estimate: quick = 5 min, moderate = 15 min, heavy = 30 min. Sum 🔴 PRs only, round to nearest 5 min.

Limit

If --limit N is specified: show top N rows but still show full tier counts in the footer so the user knows what was omitted.

Key Principles

• Freshness over age — sort by updatedAt, not createdAt
• Re-reviews are always top priority — the author is actively blocked on you
• Quick wins first within tiers — 5-line fix takes 2 min, do it before 500-line feature
• Don't waste tokens on obvious cases — LLM only refines where judgment adds value
• Deprioritize when others can cover — diminishing value when 3 others assigned
• Read-only — this skill never modifies PRs, merges, or comments