issue-triage

name	issue-triage
description	Rules for triaging new issues, classifying severity, and routing to the correct agent.
metadata	{"version":"1.0.0","type":"procedure"}

Issue Triage

Covers the triage_new_issue (github:issue_opened) and evaluate_and_delegate (github:issue_labeled) tasks. Load this when an issue event arrives.

Task: triage_new_issue (github:issue_opened)

A new issue was created. Your ONLY job is to apply labels. Do NOT delegate work. Do NOT publish architect:build_directive.

Step 0: Check Repo Correctness (Before Labeling)

1. Call repo:list to get all registered repos.
2. Read each repo's registry entry (description, tech_stack, deployment type).
3. Compare the issue title/body against repo topology to determine where the work belongs.

If the issue is on the WRONG repo:

1. Post comment: "This work belongs in {correct-repo}. Moving."
2. Create the issue on the correct repo via github:create_issue (copy title, body, context).
3. Apply label needs-human to the original issue with a redirect note.
4. Do NOT label the original as ao-eligible or bug — do NOT delegate from the wrong repo.
5. Stop. Return.

If the issue is on the correct repo: proceed to Step 1.

Step 1: Apply Labels

1. Read issue title and body from the event payload.
2. Decide labels based on content:

Label	When to Apply
bug	Describes a bug, defect, or incorrect behavior
QA	Describes a test gap, missing test, or quality issue
ao-eligible	Task AO can handle but isn't a bug or QA issue
needs-human	Requires human judgment, ambiguous, or touches security/credentials
coordination	Requires cross-agent or cross-repo coordination
UI	Requires frontend/design work
security-sensitive	Touches auth, credentials, permissions, or safety gates
P1	High priority (production impact, blocking other work)
P2	Normal priority

3. Apply labels with github:add_labels.
4. Do NOT assign the issue.
5. Do NOT publish any events.

Label Conflict Rules

• If you apply needs-human, do NOT also apply bug, QA, or ao-eligible. needs-human takes absolute priority.
• Bot-created issues: If created by yclaw-agent-orchestrator[bot] or contains "follow-up from #", it's a bot-created follow-up. These are almost always bug or QA — label accordingly and let the delegation path handle them.

---

Task: evaluate_and_delegate (github:issue_labeled)

A label was just added to an issue. Your job: check if eligible for AO delegation, and if so, publish an architect:build_directive.

Eligibility Contract (STRICT — do not deviate)

An issue is eligible if ALL of these are true:

• Has at least one eligible label: bug, QA, or ao-eligible (match emoji-prefixed variants like 🐛 bug, 🧪 QA, 🤖 ao-eligible)
• Does NOT have any exclusion label: needs-human, coordination, UI, security-sensitive (emoji: 🙅 needs-human, 🔗 coordination, 🎨 UI, 🔒 security-sensitive)
• Does NOT have in-progress (emoji: 🚧 in-progress) — already being worked
• The label just added (from label_added field) is an eligible label — don't re-evaluate old label additions

If not eligible, stop. Return immediately.

If eligible:

1. Call ao:status to check AO health. If degraded or unavailable, stop.
2. Call github:get_issue to fetch full issue details.
3. Create a structured directive:
   • investigation_summary: What the issue is, root cause analysis
   • key_files: Which files likely need changes (use github:get_contents to verify paths)
   • constraints: What NOT to change, safety boundaries
   • acceptance_criteria: How to verify the fix is correct
4. Publish event:publish with event architect:build_directive containing:
   • All structured fields above
   • repo (MUST be full slug owner/repo, e.g., YClawAI/yclaw)
   • issueNumber (integer)

What NOT to do

• Do NOT check assignees. Assignee is irrelevant.
• Do NOT check comments or branches. You don't have tools for that and don't need them.
• Do NOT delegate more than 1 issue per invocation. You're handling a single label event.
• Do NOT apply labels — the delegation path in the runtime handles in-progress.

See Also

• delegation-policy/SKILL.md — which agent to route eligible work to
• stale-management/SKILL.md — safety-net sweep that catches missed issues