Github Triage

GitHub Triage

• Use explicit issues, pull requests, filters, or repositories when supplied.
• For direct invocation without explicit scope, use the current GitHub repository.
• Require an explicit request before broadening across repositories.
• Treat labels, assignment, milestones, and project status as signals, never readiness proof.
• Remain read-only: do not comment, label, assign, close, merge, patch, change project state, or spawn agents.
• Stop after routing. Do not implement work, perform the final code review, or decide landing.

Triage loop

Scope

• Resolve scope in this order: explicit items or filters → explicit repositories → current repository.
• Never broaden silently.
• Inventory the requested scope before selecting items for deeper inspection.
• Inspect every item when feasible; otherwise inspect the slice most likely to change routing.
• Report exact coverage and anything skipped.

Understand

• Read repository rules and the product or architecture sources they identify.
• Read each surfaced item's body, material comments, and current GitHub state.
• Inspect code, diff, CI, tests, history, or dependency evidence only as far as needed for a reliable route.
• State the observable desired outcome, user or repository impact, owning surface, and important dependencies.
• Separate issue or PR claims from evidence you verified.
• Name material risk from blast radius, reversibility, contracts, data, security, concurrency, lifecycle, or compatibility.
• Classify risk:
• LOW — localized, reversible, no contract, data, security, or concurrency boundary.
• MEDIUM — multiple callers or meaningful failure-path or operability risk.
• HIGH — contract, data, migration, concurrency, lifecycle, compatibility, or broad blast radius.
• CRITICAL — security boundary, irreversible data or infrastructure, or production-wide blast radius.
• Stop before autoreview-level analysis unless correctness determines whether the item is actionable at all.

Test readiness

Route an item as ready only when all hold:

• Outcome: the next action and observable completion condition are clear.
• Decision completeness: a worker need not choose product or contract behavior.
• Boundaries: owning surfaces and dependencies are identifiable and bounded.
• Authority: the action fits repository delegation rules.
• Proof: a credible verification path exists.
• Risk: material failure surfaces are known well enough to name review lenses.

Without repository delegation rules, treat only reversible, behavior-preserving, verifiable work as ready. Route user-visible, contract, data, security, infrastructure, or irreversible changes to the owner.

Route

• READY — passes the readiness test. Name the next action: IMPLEMENT, REVIEW, FIX, or PROVE.
• REFINE — missing technical evidence or boundaries that an agent can obtain without an owner decision.
• OWNER — needs product or contract choice, owner-only access, irreversible authority, or an explicit waiver.
• DEFER — stale, duplicate, superseded, invalid, out of scope, or not worth acting on.

Do not use REFINE for evidence available through a cheap read-only check in the current pass. Use it when the investigation exceeds triage scope or needs dedicated work.

Credential rule:

• Access required to understand or implement safely → OWNER.
• Access required only for final proof, with an exact owner-run proof path → READY with a proof dependency.
• No reliable proof path → OWNER.

Output contract

Use URL-first items in this order. Omit empty sections.

SCOPE
Repository/filter: 
Coverage: 
Skipped: 

READY
 — 
Action: IMPLEMENT | REVIEW | FIX | PROVE
Outcome: 
Impact: 
Basis: 
Boundaries: 
Risk: 
Proof: 

REFINE
 — 
Missing: 
Investigate: 

OWNER
 — 
Decision/access: 
Why owner: 

DEFER
 — 
Reason: 

• Route code execution to the orchestrator; do not prescribe worker count or reviewer topology.
• Do not include type, size, priority, author trust, or lifecycle fields unless the caller explicitly needs them.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: raul-cgk
• Source: raul-cgk/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.