AGENT SKILL

triage

Classify work into Quick, Standard, or Full tiers from deterministic signals. Use when routing new work or re-scoring scope mid-flight. Does not run the doc pipeline or implementation phases.

License 许可
MIT
最近更新
2026-11-03
Stars
★ 2
Forks
0

01 前言 Front Matter

name
triage
metadata
{ "shipwright-capability": { "version": 1, "metadata": { "skill": "triage", "selectionFamily": "workflow" }, "triggers": [ { "type": "phase_default", "command": "sw-triage", "selectionFamily": "workflow" } ] } }
description
Classify work into Quick, Standard, or Full tiers from deterministic signals. Use when routing new work or re-scoring scope mid-flight. Does not run the doc pipeline or implementation phases.

03 SKILL.md 正文

Triage rubric (/sw-triage)

Deterministic tier classifier. Auditable — same inputs → same tier. Not model judgment.

Model tier: cheap — resolve via python3 scripts/sw_bootstrap.py resolve-model-tier.py -- --skill triage. When using the Task tool for subagent dispatch, resolve concrete model IDs from models.tiers in config (never semantic tier names in subagent model: frontmatter).

Inputs

Collect before scoring:

  1. File count — number of files likely touched (user estimate or git diff --stat scope).
  2. Risk keywords — scan description + file paths for triggers (case-insensitive).
  3. Ambiguity markers — vague scope language in the request.
  4. Override — optional --tier <quick|standard|full> forces tier (record override).
  5. Misroute re-entry — --re-score when a Quick item's scope grew mid-flight.

Risk triggers (hard floor → ≥ Standard)

Each keyword is tagged. Any tag match forces at least Standard regardless of file count. security-tagged entries flow into doc-review selection via signal_context.derived_tags and manifest text_token triggers on sw-security-reviewer (see core/sw-reference/capability-manifest.md — not duplicated here).

Keyword Category
auth security
authn security
authz security
authentication security
authorization security
login security
session security
oauth security
jwt security
payment security
payments security
billing security
PII security
credentials security
token security
encryption security
public api security
public endpoint security
external api security
webhook security
stripe billing-routing
paddle billing-routing
subscription billing-routing
migration data-migration
data migration data-migration
schema migration data-migration
backfill data-migration

Ambiguity markers (bias upward)

Any match adds +1 to the score:

  • maybe, possibly, not sure, unclear, TBD, figure out
  • explore, investigate, spike, prototype
  • missing acceptance criteria on a multi-file change

File-count score

Files Base tier
0–1 Quick
2–5 Standard
6+ Full

Resolution algorithm

1. If override flag set → use override tier; record "override: <tier>".
2. If any risk trigger matches (any category) → floor = Standard.
3. Compute base tier from file count.
4. If ambiguity markers present → bump one tier (Quick→Standard, Standard→Full).
5. If mixed/insufficient signals (no file count, empty description) → Standard (conservative).
6. final = max(base tier, floor) using order Quick < Standard < Full.
7. When advisory rigor is supplied, merge via `scripts/triage_lib.merge_tier_monotonic` (union/max-rigor).
   Tier reductions require an authorized R7 reduction path — never file-count alone for Full.
8. On --re-score with prior tier Quick and new score > Quick → promote; report misroute recovery.

Mechanical scoring lives in `scripts/triage_lib.py` (`classify_triage`, `merge_triage_monotonic`).
**Monotonic merge:** union/max-rigor of mechanical set vs advisory rigor; reductions below the
mechanical floor require R7 authorization (detector no-fire or human waiver) — file-count alone
(including rename-only churn) is not a sole Full trigger.

Output contract

Report:

Tier: <Quick|Standard|Full>
Signals:
  - file_count: <n> → <base tier>
  - risk_triggers: [<matches>] → floor Standard (if any)
  - ambiguity: [<matches>] → bumped (if any)
  - override: <tier> (if set)
  - misroute_reentry: promoted from Quick (if applicable)
Next: <route>

Routes:

  • Quick → implementation workstream (/sw-execute when available); no doc artifacts.
  • Standard → /sw-prd (skip brainstorm) or /sw-doc.
  • Full → /sw-brainstorm or /sw-doc.

Misroute re-entry

When implementation reveals scope growth on a Quick-classified item:

  1. Re-run /sw-triage --re-score with updated file list and description.
  2. If new tier > Quick, route into Standard/Full pipeline.
  3. Record prior Quick classification in the output for audit.

Test matrix (structural)

Case Input Expected
Trivial 1 file, no risk keyword Quick
Bounded feature 4 files, no risk Standard
Risk floor 1 file + "auth" Standard (not Quick)
Billing floor only 1 file + "stripe" Standard (not Quick); does not fire security persona
Migration floor only 1 file + "migration" Standard (not Quick); does not fire security persona
Ambiguous 2 files + "maybe refactor" Standard or Full
Conservative default empty file count Standard
Override --tier full on 1-file change Full + override recorded
Misroute --re-score, was Quick, now 6 files Standard or Full

Quick tier and graph-native /sw-ship (PRD 271 R7/R15)

Quick-classified work has no frozen task list — /sw-deliver does not apply. Entry remains /sw-ship: the same operator command compiles Quick work to a fixed WorkflowGraph via scripts/graph/quick_ship_compile.py (canonicalPhaseChains.sw-ship in kernel-classification.json). Topology is configuration-declared only — no adaptive capability selection.

Graph-native Quick does not add /sw-graph-* slash commands; live node progress and explain stay on /sw-status. On misroute (--re-score above Quick), route to Standard or Full — do not stretch Quick onto multi-phase deliver.

04 同仓库更多 Skills