triage

name	triage
description	Classify readiness, blockers, severity, AFK/HITL, lifecycle-state need, closeout, or whether incomplete information should be blocked or handed off. Not for shaping, slicing, code edits, planning, verification reports, dispatch packages, or wiki.

triage

Trigger Contract

Use for task state, readiness, blockers, AFK/HITL, existing issue ready-for-agent/ready-for-human classification, lifecycle-state, closeout classification, or deciding whether incomplete information should stop as blocked, continue as needs-info, or be preserved for handoff.

Use triage when the prompt says verify/tests/done evidence already passed and asks whether an issue/task can be closed. That is a closeout state transition decision; do not route it to verify unless the user is asking to first verify whether the evidence passes.

Route away:

• New PRD/spec -> to-prd.
• Accepted requirements to tasks -> to-issues.
• Implementation plan after readiness -> write-plan.
• Code edits -> implement.
• Release/UAT/evidence proof -> verify.
• Runtime/subagent/worktree routing -> dispatch.

Required Evidence

Inspect task source, PRD/spec, conversation, prior state, blockers, source refs, lifecycle state, and verification expectations. Unknown source -> needs-info.

For Groundwork repo maintenance, apply repo-local AGENTS.md before reporting complete.

Every verdict includes Severity and State Transition Reason. Use SEVERITY.md; severity is current blocker/gap impact, not product priority. AFK can proceed from available source/acceptance/scope/first inspection/verification without new human decision/access/approval. HITL needs human decision, authority/access, manual validation, risky approval, or product/design choice.

Closeout requires verify evidence, already-done evidence, or explicit wontfix owner decision. Otherwise keep open as verification, needs-info, or ready-for-human.

Use skills/_shared/LIFECYCLE-STATE.md only to decide if workstream-scoped state is justified. Do not duplicate external issue/PR ownership unless Groundwork recovery needs remain.

Executable ready-for-agent + AFK must include a Goal Contract using skills/_shared/GOAL-CONTRACT.md field names, preferred runtime recommendation, and expected result package. Do not emit child goals for needs-info, ready-for-human, or HITL-only tasks.

Severity Derivation

Identify current blocker/gap, strongest affected boundary, blast radius/dependencies/workaround, then assign the highest matching severity. If source, boundary, or workaround is unknown, classify needs-info and name missing evidence.

Workflow

Gather source/outcome; classify state; assign severity and transition reason; separate evidence added/missing; apply readiness contracts; for ready-for-agent + AFK, produce agent brief with Goal Contract/runtime recommendation; for HITL, output decision/options/risks/next action; decide lifecycle-state need; recommend next route.

Hard Stops

• Stop before ready-for-agent unless acceptance criteria, source/first inspection step, expected output, stop condition, AFK/HITL decision points, blockers, and boundaries are explicit.
• Stop before state change unless previous/current state, severity, transition reason, evidence added, and evidence missing tie to source.
• Stop before lifecycle state, closeout, or external ownership changes when source truth is unclear, mixed, or duplicated.
• Do not treat severity as product priority.
• Do not generate Goal Contracts for missing-info/HITL tasks.
• Do not claim runtime/model/reasoning selector enforcement; recommendations are advisory for dispatch.
• Reject project-global .planning, .gsd, GSD clones, or task databases; redirect only when workstream-scoped artifacts/<workstream-slug>/STATE.md is justified.

Output Shape

Use Triage Verdict with state, severity, AFK/HITL, previous state, transition reason, evidence added/missing, blockers, readiness check, lifecycle recommendation, agent brief or human decision, next action, and artifact recommendation.

For needs-info, blocked, HITL, risky approval, or any other gate-bearing triage, include these exact machine-readable gate labels before the next action detail:

Proposed Action:
Target:
Risk:
Rollback/Undo:
Approval Needed:

Keep these gate labels in English even when the report body follows a non-English session locale.

Stop Condition

Stop when state, severity, transition reason, blockers, AFK/HITL, readiness reason, evidence added/missing, lifecycle-state recommendation, and next action are explicit.

Artifact Rule

Do not create OUT-OF-SCOPE.md in MVP. Write an agent brief only when reused for execution, review, or handoff. Redact secrets, credentials, PII, sensitive logs, screenshots, private payloads, and database rows.