Build Workflow Skill

Overview

Build the approved callable skill deliverable. A skill is a substantive SKILL.md plus matching agents/openai.yaml; code, commands, helpers, apps, or references cannot substitute for it. Leave tracker writes, phase transitions, and publishing with the invoking lifecycle task.

Use GPT-5.6 Sol for every model-run evaluation. If it is unavailable, stop with eval-runner-unavailable; never select a fallback model.

Required Context

Read root and project policy, project state, every non-done task, the selected task, Matt Pocock skill guidance, and the exact product repository policy when the target is standalone. Then read:

• the complete bundle contract for identity, semantics, ownership variants, optional resources, and completion;
• the behavior evaluation contract for RED/GREEN isolation, evidence fields, stable failures, and release gates.

Input Contract

Require one selected task at matt_phase: "implement" with approved version-1 deliverable migration, an exact create/update operation, ownership, target surface, visibility, runtime targets, required artifacts, guidance, evals, completion conditions, artifact bindings, and deliverable/role/artifact-bound phase approvals. Require a complete dependency write plan for every dependency step.

For update, resolve exactly one existing skill identity. Stop on zero or multiple matches. Never invent a new slug as an update shortcut.

Workflow

1. Ask $continue-task for the selected task and projected blockers. Do not repair tracker state here.
2. Match the contract to exactly one ownership branch: canonical Foundry, project-packaged, standalone product, or workflow-pack member. In a mixed task, select only contracts whose kind is a skill. A UI, package, validator, document, or tracker contract remains with its own builder even when its paths are separately approved.
3. Load the shared scenario contract. Confirm it separates lifecycle-intent routing from implementation-ready construction cases.
4. Capture a fresh first-attempt RED run without this builder. Keep the exact prompt, session, output, diff, and validation result.
5. Implement the smallest complete bundle permitted by the approved paths.
6. Run every required scenario variant independently with GPT-5.6 Sol in a clean disposable repository. Give the runner only the raw prompt and implementation-ready repository context—not the expected route or rubric.
7. Grade emitted contracts, file tree, contents, diff, ownership, visibility, catalog state, phase result, and dependency outcome. A narrative claim is not evidence.
8. Run deterministic bundle, routing, boundary, catalog, dependency, and command/helper regressions. Re-run the affected family after each fix.
9. Return evidence and readiness to the lifecycle task. Never mark it done.

Decision Points

Approved branch	Target	Visibility	Foundry catalog
Canonical	.agents/skills/<slug>/	canonical-active	Include when complete
Project packaged	projects/<project>/skills/<slug>/	project-local-inactive	Exclude
Standalone	Exact product-owned surface	standalone-only	Exclude
Unresolved	None	unresolved	Stop with zero writes

Each workflow-pack member needs its own skill contract, identity, complete bundle, and invocation. Shared files remain declared support. Promotion from a packaged or standalone surface requires a separate contract.

Never place SKILL.md or agents/openai.yaml under a non-skill target to make that contract look complete. In particular, do not write a skill bundle into a UI application path; return that independent deliverable as pending.

Add optional references, scripts, or assets only when the contract declares the exact support artifact, SKILL.md directly consumes it, policy permits it, and verification covers it. Keep operator invocation as $skill-name. Commands are secondary and allowed only under Developer Verification, Package Smoke Test, Deterministic Validation, Query Helper, or Internal Support.

Failure Handling

Stop with zero new writes for pending migration; missing or contradictory contracts; unresolved ownership; unbound approvals; ambiguous update identity; undeclared resources; unsafe dependency or external writes; incomplete, helper-only, thin-wrapper, command-first, or code-only bundles; unavailable GPT-5.6 Sol; or failed deterministic, fresh-agent, security, or review gates.

Return the stable code, observed evidence, and recovery owner. Do not weaken a contract or rewrite a failed first attempt into a pass.

Human Boundaries

The operator approves contracts, product ownership, phase paths, external writes, promotion, publishing, and consequential dependency actions. This skill may write only exact task-approved paths inside the declared boundary.

Output Contract

Return deliverable identity and ownership; primary and justified support files; catalog result; per-scenario GPT-5.6 Sol session evidence; deterministic results; dependency provenance; validation and review; blockers; and the lifecycle task that owns the handoff.

Completion Gate

Apply every check in the linked bundle and behavior contracts. Completion requires substantive semantics, exact identities, boundary-safe artifacts, complete per-variant attempt chains, deterministic verification, dependency reconciliation when present, and a Standards-plus-Spec review with no critical or high findings.

Developer Verification

node scripts/validate-workflow-state.mjs
node scripts/query-workflow-state.mjs --skill-catalog