# Task workflow SOP

Use for multi-step work, blockers or reporting. No registry/manager/automatic queue
is required. Miegyver owns platform analysis; Thinker owns QA. No automatic dispatch.

1. State objective, scope, acceptance criteria and risk. Ask only for decisions that
   materially affect the task. Do not require PRD/design artifacts for simple
   diagnostics; request the missing design when the task genuinely depends on it.
2. Check tool permissions and safe evidence sources. If execution is denied, deliver
   a plan/review rather than attempting a workaround. User approval does not grant
   capabilities absent from runtime policy.
3. Read minimum relevant evidence. Separate observations, hypotheses and unknowns.
   Do not invent fresh telemetry or trust stale model/service success.
4. Before a mutation, require explicit task authorization, least privilege and
   verification/rollback criteria. The operator executes system/repository/file
   changes. Exception: task-relevant platform-kb create/update is authorized;
   follow knowledge-memory.md, preserve history and verify saved content.
5. Verify the exact outcome where tools permit. Otherwise label NOT VERIFIED and
   provide the operator a safe verification step. Never translate missing evidence
   into PASS or a fabricated successful operation.
6. Bound retries to useful new evidence. Stop repeated failures at the same approach
   and report evidence/new options; checkpoint lengthy work and respect user deadlines.
7. Report: outcome; evidence/reference; changes executed versus proposed; validation;
   blockers/risks; next owner/action. Do not include raw secrets or personal data.

Failure template:
- Objective and failed step:
- Observed vs expected (sanitized):
- Attempts and evidence:
- Scope/permission limitation:
- Recommended next action and operator decision: