agentsclimarketplace

Triage pr

Skill UMEP-dev/SUEWS/.claude/skills/triage-pr

Use whenever auditing, prioritising, or batch-triaging SUEWS draft and stalled pull requests so each gets a clear disposition. Trigger for "triage PRs", "which draft PRs need attention", "what's the disposition of these PRs", "groom the PR backlog", "defer this PR with a reason", "which PRs are ready to advance to merge", or a scheduled PR-backlog sweep. Routes live PRs to prep-pr/split-pr/queue-pr; defers or escalates the rest; never merges or closes.From its SKILL.md

Install
npx -y skills add UMEP-dev/SUEWS --skill triage-pr

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

12.2 KB, as published. Nobody here has run it

Triage PR

The PR-axis sibling of triage-issue. Where triage-issue grooms issue descriptions into maintainer-ready engineering work items, triage-pr grooms the draft and stalled PR backlog so every open PR has a clear, reasoned disposition.

This skill is a router, not a reviewer. It decides what happens next for each PR; it does not re-review code, order the merge queue, or carve diffs. Those responsibilities stay with their owners:

  • Code review -> audit-pr
  • Merge queue ordering -> queue-pr
  • Oversized diff carving -> split-pr

Verdict routes:

  • advance -> route to prep-pr, split-pr, or queue-pr
  • continue -> route to fix-issue or author
  • defer -> parked with a named revisit trigger
  • close -> human-gated: propose, never act
  • escalate -> human-gated: surface conflicting signals, never guess

triage-pr never merges and never closes a PR.


Triggers

Use this skill when asked to:

  • Audit, prioritise, or groom the SUEWS draft/stalled PR backlog.
  • Assign a disposition to one or many draft or stale PRs.
  • Find which PRs are ready to advance, which should be deferred, and which need a maintainer decision.
  • Run a scheduled PR-backlog sweep (autonomous batch mode).

Capabilities

  • Scan the open PR backlog using the bundled read-only scanner.
  • Compute one of five verdicts (advance, continue, defer, close, escalate) per PR using the signal-to-verdict rules in references/rubric.md.
  • Apply a 0-pr:* disposition label (if it exists in the repo) and an idempotency label (0-auto:pr-audited) to each processed PR.
  • Post (or upsert) a machine-readable marker comment containing the verdict block.
  • Record defer verdicts with a named revisit trigger (revisit-by: or blocked-on:).
  • Route advance PRs to the correct downstream skill (prep-pr, split-pr, or queue-pr) based on CI, merge state, and size signals.
  • Escalate close and genuinely ambiguous escalate verdicts to a maintainer without guessing or acting.

Quick Start

Run the read-only scanner to collect candidates:

bash .claude/skills/triage-pr/scripts/scan-prs.sh --stale-days 14

The scanner prints a Scope: line followed by one TSV row per candidate PR:

PR  STATE  CI  MERGE  REVIEW  AGE_D  CHURN  FILES  AUTHOR  TITLE

Then compute a verdict for each row using references/rubric.md and apply the procedure in references/methodology.md.

For a different staleness threshold or a non-default repository:

bash .claude/skills/triage-pr/scripts/scan-prs.sh --repo UMEP-dev/SUEWS --stale-days 7

For JSON output (for scripting):

bash .claude/skills/triage-pr/scripts/scan-prs.sh --json 2>/dev/null | jq .

Safety Rules

  • Drafting or scanning is not permission to post, label, or flip PR state. In interactive mode, present the verdict block for each PR and wait for explicit approval before any write.
  • Reading a PR's signals is not permission to act on the PR.
  • Privacy is a hard invariant on every outward write. Apply the Privacy scrub gate (below) before any comment or label rationale reaches GitHub.
  • triage-pr NEVER closes a PR, NEVER merges, and NEVER enqueues. These are always human-gated.
  • Do not create labels. Apply labels only if they already exist in the repository.

Label Model

Disposition labels (mutually exclusive, one per PR per run)

VerdictLabel
advance0-pr:advance
continue0-pr:continue
defer0-pr:deferred
close0-pr:close-proposed
escalate0-pr:escalate

These labels are mutually exclusive. Before applying any new 0-pr:* label, remove every other 0-pr:* label already on the PR (there should be at most one, but remove all if more than one is present). Never leave two 0-pr:* labels on a PR simultaneously.

Bookkeeping label

0-auto:pr-audited -- marks that an autonomous pass has processed this PR. It is additive: apply it once and leave it. Do not remove it on a re-run. It enables idempotency: batch re-runs skip PRs carrying this label whose last substantive activity predates the previous audit timestamp.

Classification label: 0-physics:change

Independent of the disposition label, apply 0-physics:change (if it exists in the repo) when the PR diff touches physics source (suews_phys_*.f95, the Rust physics backend), moves a reference fixture (test/fixtures/data_test/sample_output.csv.gz, .../stebbs_test/sample_output_stebbs.csv, test/fixtures/benchmark1/*.pkl), or changes a physics-affecting data-model default. It is additive (not part of the mutually-exclusive 0-pr:* family) and activates the evidence requirements in .claude/rules/physics-change-evidence.md: a ## Scientific evidence PR-body section, domain-owner sign-off, and the full -m physics CI tier. When you apply it, note in the marker comment that the PR needs the evidence section and owner sign-off (STEBBS -> @yiqing1021) before it can advance.

Priority

Priority (P0/P1/P2) is not a label. It lives in the verdict block only.

Maintainer one-off setup

Create these six labels once in the repository:

0-auto:pr-audited
0-pr:advance
0-pr:continue
0-pr:deferred
0-pr:close-proposed
0-pr:escalate

triage-pr applies labels that already exist and NEVER creates them. If a label is absent, the verdict is recorded in the marker comment and handoff block only, with a note that the label is missing.


Privacy Scrub Gate

Before any outward write (marker comment, label rationale, or any text posted to GitHub), scan the model-authored distilled text for:

  • Internal tracker IDs (PER-\d+ and similar private references).
  • Absolute local or home-directory paths (/Users/..., worktree paths).
  • Private host or infrastructure names (internal compute servers and the like).

On a hit in distilled text, redact the offending token before posting.

If a private token appears inside a verbatim quote (for example from the PR description or a commit message) that must be preserved to make the finding actionable, replace it with neutral phrasing or escalate to a maintainer rather than post as-is.

This operationalises "Privacy is a hard invariant on every outward write" from .claude/rules/autonomous-workflow.md.


Gating Model -- Three Tiers

Auto-applicable (additive, reversible, or self-scoped)

May run without human approval:

  • Run the read-only scanner (scan-prs.sh).
  • Compute verdicts per references/rubric.md.
  • Post or upsert the marker comment on each PR (idempotent: replace an existing <!-- triage-pr:verdict ts=... --> comment rather than duplicating it).
  • Apply the matching 0-pr:* label using the mutual-exclusion swap described above, if that label exists in the repo.
  • Apply the 0-auto:pr-audited label if it exists in the repo.

Confidence-gated

Applied without human approval only when all stated gates hold; escalated otherwise:

  • Flipping a draft PR to "ready for review" -- only when ALL advance gates in references/rubric.md (Part A) hold AND the PR is one the agent created this run (not a third-party PR).

Human-gated (always escalate; never auto-apply)

  • Closing or superseding a PR.
  • Any merge or queue-entry (queue-pr --enqueue or gh pr merge).
  • Posting to, marking ready, or force-pushing a PR the agent does not own.
  • Editing the body or metadata of a PR the agent does not own.

The verdict-vs-action split

A verdict is a disposition; enacting it is a separate tiered action.

An advance verdict on a third-party PR still gets the 0-pr:advance label (auto) and the recommendation in the marker comment, but the ready-flip and downstream handoff are human-gated. The Next: field in the verdict block says "recommend marking ready / hand to <route>". The verdict is NOT downgraded to escalate because the enacting action is human-gated -- only the action waits; the verdict is recorded immediately.


Workflow

Read references/methodology.md for the full step-by-step procedure. The summary is:

  1. Run scan-prs.sh to collect the candidate set and the Scope: line.
  2. For each candidate PR, check idempotency (already audited with no new activity and no fired revisit trigger -> record as noop, stop).
  3. Compute the verdict using references/rubric.md.
  4. Apply the auto-tier: upsert the marker comment; swap the 0-pr:* label; apply 0-auto:pr-audited.
  5. State the route and next step in the verdict block:
    • advance -> route to prep-pr, split-pr, or queue-pr; record priority.
    • continue -> note live WIP; route is fix-issue or author notification.
    • defer -> record Revisit-by: or Blocked-on: in the verdict block.
    • close -> present 0-pr:close-proposed for maintainer decision; no further action.
    • escalate -> state conflicting signals explicitly; present for maintainer resolution.

Autonomous Mode

By default triage-pr is interactive: it drafts and waits for explicit approval before any write. Autonomous mode is an opt-in contract for unattended orchestration (a scheduled backlog sweep or an upstream orchestrator). It never asks a human mid-run.

Opt-in standing approval covers ONLY the auto-applicable tier. The human-gated tier (close, merge/enqueue, third-party PR edits) always escalates in autonomous mode; there is no standing approval for those actions.

In autonomous batch mode:

  • Bound the input set explicitly with --limit. Emit the scanner's Scope: line verbatim so a silent cap is never read as full coverage.
  • Apply the idempotency check per PR before computing a verdict. A noop PR (already audited, no new activity) is counted as noop in coverage, never as skipped.
  • Accumulate verdict counts across the batch and emit a Summary: line at the end.
  • A PR that cannot be fetched or processed is recorded as skipped with a reason in the Next: field. Never silently drop a PR.
  • Coverage equation (must hold exactly): advance + continue + defer + close + escalate + noop + skipped = K where K is the fetched set size. Full coverage of the open backlog is assertable only when truncated: no.

Any review -> revise loop inside a batch run follows .claude/rules/review-convergence.md (round cap + oscillation guard). A single PR must not consume the batch; escalate and move on if it does not converge within the cap.


Output Format

Use the machine-readable handoff verdict block defined in references/methodology.md. The block is reproduced here for convenience; keep it identical to the canonical form in that file:

[triage-pr] verdict
PR: #<n>
Verdict: advance | continue | defer | close | escalate
Priority: P0 | P1 | P2 | -        (advance only)
Route: prep-pr | split-pr | queue-pr | fix-issue | -
Reason: <one line>
Revisit-by: YYYY-MM-DD | -        (defer only)
Blocked-on: #<n> | -             (defer only)
Applied: comment | label:<0-pr:...> | ready | none
Next: <downstream action / the human question>

Downstream skills (prep-pr, split-pr, queue-pr, fix-issue) branch on Verdict and Route. They do NOT need to parse prose.


References

  • references/rubric.md -- Signal-to-verdict decision rule (five verdicts, priority, near-green CI definition).
  • references/methodology.md -- Step-by-step procedure, marker and idempotency mechanics, machine-readable verdict block, batch coverage accounting.
  • scripts/scan-prs.sh -- Read-only PR signal scanner; emits TSV candidates for the router.
  • .claude/rules/autonomous-workflow.md -- Pipeline contract: gating tiers, the always-human-gated list, and the merge gate.
  • .claude/rules/work-sizing.md -- Right-sized-PR criteria; referenced by the advance route decision (size gate -> split-pr vs queue-pr vs prep-pr).
  • .claude/rules/review-convergence.md -- Bounded loop discipline for any review -> revise cycle inside a batch run.

Keep looking

Skills are one crate of 614,194. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.