Agency Runtime v3

Use this skill for substantial specialist-agent work where the output must be auditable and release-safe.

Operating contract

1. Build a TaskProfile with task id, artifact type, risk, mutation mode, acceptance criteria, evidence requirements, and technologies.
2. Route deterministically and explainably. Select one primary specialist; add support or an independent reviewer only when it can change the result.
3. Keep conflicting mutations serialized. Read-only evidence collection may be parallelized by the host implementation.
4. Require structured deliverables, evidence, verification, quality dimensions, and hard-gate results from the worker.
5. Record provenance for material claims and evidence. Never put secrets or credentials into the ledger.
6. Evaluate the result through the deterministic quality runtime. A specialist cannot self-certify release readiness.
7. Material/blocker findings require repair and re-verification. Repair cycles must have a hard upper bound.
8. If evidence, verification, acceptance, or permissions are insufficient, stop with a failed/blocked receipt rather than reporting success.

Failure boundaries

• Do not expand scope without an explicit task change.
• Do not treat model text as evidence of execution.
• Do not claim tests, builds, deployments, reviews, or external actions that were not recorded.
• Do not run arbitrary commands merely to satisfy a quality field; verification must be performed by the host execution layer.

Output

Return the final deliverables, evidence references, verification summary, findings, deterministic quality receipt, and provenance ledger. The caller should expose uncertainty and unresolved risks rather than hiding them.