task-triage

name	task-triage
description	Task triage and complexity classification. Use when selecting the OpenCode workflow pipeline for a user request.

Task Triage

Classify each request before expensive work.

Levels

• level-0: answer only; no code change.
• level-1: genuinely tiny, isolated, reversible, low-risk change; no plan.
• level-2: small isolated implementation; compact plan.
• level-3: medium feature or multi-file change; plan plus critique.
• level-4: high-risk, architecture, auth, data, payment, security, migration, or large refactor; deep plan and approval gate.

Pipeline choice

• Level 0 is answer-only and needs no review artifact.
• Level 1 may omit a reviewer/artifact only when it is genuinely tiny, isolated, reversible, and low-risk; record that exemption in the assignment/wrap-up.
• Level 2+ implementation requires at least a final .agents/REVIEW_code-<n>.md; Level 3+ also requires a .agents/REVIEW_plan-<n>.md artifact.
• If any reviewer is delegated at any level, its assigned durable artifact is mandatory.
• Skip planning only for Levels 0 and 1.
• Use templates/PLAN.md as the source for .agents/PLAN.md starting at Level 2.
• Use compact .agents/PLAN.md for Level 2.
• Use planner and an assigned .agents/REVIEW_plan-<n>.md plan critique for Level 3+.
• Use specialized reviewers for security, performance, migration, or large diffs.
• Ask the user before risky or destructive phases.

Execution mode selection

• continuous: tiny changes, small fixes, medium low-risk tasks, and well-understood implementation work. No user pause unless blocked, failures are ambiguous, scope expands, or risk increases.
• phase-gated: large features, UX-heavy changes, full-stack work with manual testing value, ambiguous product behavior, or features where the user should inspect intermediate results. Pause after each main phase checkpoint.
• approval-gated: security-sensitive changes, database migrations, auth/permission changes, payment/billing logic, destructive operations, and large refactors. Ask for explicit approval before implementation or risky phases.

For Level 2+ plans, record the selected mode in .agents/PLAN.md and mirror the current pause state in .agents/PROGRESS.md when progress tracking is used.

Escalate when

• Scope is unclear.
• Multiple systems are touched.
• Data shape changes.
• Authorization or secrets are involved.
• Tests are missing or failures are ambiguous.

User input

When the next step requires user input, use the canonical question standard: use question for scope ambiguity, product decisions, approval requests, and unclear ownership. Provide 2–4 concise options. Put the recommended option first and mark it (Recommended) when a safe default exists. Never end a run with a plain-text question when the question tool is available.