{"id":17198,"plugin_id":"plugins_6a746bbb64d48191942c65a16a1eb19f","kind":"skill","collection_source":null,"comparison_source":null,"observed_at":"2026-09-30T23:14:01.361Z","digest":"caa0e4dd34545aec9c4c74ce4b9d577ee6ab3bc6f751f83b599bbeab8328533e","against":null,"payload":{"description":"Architect a new AI agent harness or redesign an unreliable loop, including trust roles, durable state, transition shape, worker isolation, verification, governance, ports, budgets, runtime health, threats, and evidence.","included_files":[{"relative_path":"agents/openai.yaml","size_in_bytes":214}],"name":"design-agent-harness","skill_md_contents":"---\nname: design-agent-harness\ndescription: Architect a new AI agent harness or redesign an unreliable loop, including trust roles, durable state, transition shape, worker isolation, verification, governance, ports, budgets, runtime health, threats, and evidence.\n---\n\n# Design an agent harness\n\nDesign the user's target application. The coding assistant is its architect\nand implementer, not a runtime worker or governed subject.\n\nRead these references before designing:\n\n- `../../references/philosophy.md`\n- `../../references/system-architecture.md`\n- `../../references/pattern-catalog.md`\n- `../../references/threat-model.md`\n\nUse `../../assets/harness-design-dossier.md` as the output structure.\n\n## Workflow\n\n1. Define the user outcome, durable mutations, explicit non-goals, tolerated latency/cost, and failure history.\n2. Classify the shape. Choose `machine` only when the driver can be model-free. Choose `orchestrator` when a model selects from a closed transition set and the choice is validated and write-ahead persisted.\n3. Separate contract proposer, contract ratifier, worker, result verifier, operator, signer, and receipt sink. Record any unavoidable shared principal or failure domain.\n4. Model durable states, allowed edges, terminal states, and exact persist-before-effect ordering.\n5. Define narrow ports for queue, worker, verifier, state, governance, receipts, alerts, and ground-truth views.\n6. Specify the contract grammar, context reconstruction, idempotency identity, budgets, quarantine, escalation, and override behavior.\n7. Design governance with `design-agent-governance`; treat reversibility as verified data.\n8. Add continuous component health, a watched watcher, anomaly fixtures, and a quarterly obligation retirement test.\n9. Map each important claim to a self-test, chaos fixture, or external check. Keep the intended conformance claim `UNSCORED` until evidence exists.\n\n## Output\n\nReturn the completed dossier, a short architecture decision summary, the first thin vertical slice, its acceptance/chaos tests, and the controls deliberately deferred. Flag assumptions that materially change trust or authority.\n\nMake the target runtime's trust roles, mutation paths, state machine, and\nverification source concrete enough to test, then implement the requested\nslice. This design prerequisite does not ratify or govern the coding session.\n"},"changes":[],"summary":"First saved snapshot. No earlier version is available for comparison.","summary_kind":"deterministic","summary_metadata":{}}