{"id":17194,"plugin_id":"plugins_6a746bbb64d48191942c65a16a1eb19f","kind":"skill","collection_source":null,"comparison_source":null,"observed_at":"2026-09-30T23:14:01.275Z","digest":"b2911331ec3f381a1c94d4a031fd9559f0c186064d7b21ca38b8a39ba99e01f5","against":null,"payload":{"description":"Build an application or agent harness with the published Frontier SDK packages. Use when implementing a governed agent system, wiring runtime health, agent-readable state, or signed receipts into a product, or when the user asks to use @frontier-infra packages, the Frontier SDK, or \"build this the Frontier Infra way\".","included_files":[{"relative_path":"agents/openai.yaml","size_in_bytes":275}],"name":"build-with-frontier-sdk","skill_md_contents":"---\nname: build-with-frontier-sdk\ndescription: Build an application or agent harness with the published Frontier SDK packages. Use when implementing a governed agent system, wiring runtime health, agent-readable state, or signed receipts into a product, or when the user asks to use @frontier-infra packages, the Frontier SDK, or \"build this the Frontier Infra way\".\n---\n\n# Build with the Frontier SDK\n\nThe SDK is installable plumbing — use the published packages instead of\nhand-rolling equivalents. If you find yourself writing a health reducer, an\nagent-view generator, a receipt signer, or a conformance scorer from scratch,\nstop: it ships.\n\nYou are the software developer using the SDK. You are not a worker, verifier,\noperator, or subject inside the target application. Apply the requirements\nbelow to the runtime being built, not to the interactive coding session.\n\n| Package | Job | Get it |\n|---|---|---|\n| `@frontier-infra/protocol` | Four-layer runtime health record + deterministic fail-closed reducer (`evaluateRuntimeHealth`, `runtimeHealthExitCode`) | `npm install @frontier-infra/protocol` |\n| `@frontier-infra/avl` | Agent-readable ground truth for your app; validator CLI (`validate <url>`, `validate-file <path>`, `--level`) | `npm install @frontier-infra/avl` |\n| `@frontier-infra/audit` | Conformance scoring, evidence packets, AAR sign/verify (bundles the canonical `aar.mjs`) | `npx -y @frontier-infra/audit` |\n\n## Application workflow\n\n1. **Map the product** — inspect the target application's existing runtime,\n   state, provider calls, effects, and failure modes. Choose only the Frontier\n   capabilities it needs.\n2. **Install the plumbing** — add the selected published packages as ordinary\n   application dependencies. Preserve the application's package manager,\n   framework, provider API, storage, and deployment conventions.\n3. **Build the harness shape** — implement durable state, deterministic driver, fresh workers,\n   single mutation gate (`machine-deployment` for the obligations). Do not\n   design the loop from a blank page: clone the reference driver\n   (`git clone https://github.com/frontier-infra/machine-driver`) and adapt\n   its `driver.py` + `goal.json` contract to the deployment; `conductor-public`\n   is the orchestrator-shaped starting point for ops/triage pipelines.\n4. **Implement runtime contracts** — add contract proposal, ratification, and\n   enforcement inside the target runtime when its workload needs bounded\n   delegated work. Do not create a contract for the coding session by default.\n5. **Ground truth** — emit an AVL view of real application state so verifiers\n   read declared state, not the worker's narrative (`avl-adoption`).\n6. **Health** — emit the four-layer health record; evaluate with the published\n   reducer; aggregate must fail closed (`design-agent-governance`).\n7. **Receipts** — sign target-runtime outcomes as AARs when the product needs\n   portable proof and the operator\n   provides local key paths (`aar-attestation`).\n8. **Test the product** — run its ordinary unit, integration, failure, and\n   runtime tests. If the user explicitly requests a conformance claim or the\n   target's release criteria require one, audit the target deployment with\n   `machine-conformance`; otherwise do not add an audit ceremony.\n\n## Boundaries\n\n- The builder is outside the target runtime. Host install prompts, tool\n  permissions, commits, and the builder's own patch are not Frontier\n  governance events unless the user deliberately installs that integration.\n- The model is a replaceable worker; everything trustworthy lives in the\n  target harness around its runtime model calls. If the target later invokes\n  Codex or Claude through a worker adapter, those bounded invocations are\n  workers; the coding assistant building the adapter is not.\n- The object of governance is the application's agents — never the SDK's own\n  supply chain. Do not build machinery that attests to package distribution:\n  no vendored copies with hash locks, no bootstrap approval flows, no evals of\n  your own paperwork. Host permission systems own install trust.\n- Static audits never establish Machine-L3; live chaos evidence does.\n- `same_principal` receipts are self-attestation — say so.\n"},"changes":[],"summary":"First saved snapshot. No earlier version is available for comparison.","summary_kind":"deterministic","summary_metadata":{}}