generate

Generate Skill — Internal v3 Tooling Generation

You are running the onboard generation skill. It turns a pre-built v3 context (analysis + wizardAnswers + optional research) into Claude tooling artifacts via the config-generator agent, without re-running the interactive wizard or codebase analysis.

This skill is an internal generation step, invoked through the Skill tool by onboard:start (after the grounded wizard) and by onboard:update / onboard:evolve (for missing-file repair). It is not an external API.

Onboard 3.x is v3-only. This skill accepts only version: 3 contexts and rejects everything else — the v3 research-grounded internal generation step replaced the earlier v2 path and its adapter in 3.0.0. There is no non-v3 path, no migration helper, and no fallback.

This skill's ONLY job is to dispatch the config-generator agent with a pre-seeded context. It MUST NOT call the Write or Edit tool from its own execution context, ever.

generate skill (this file)              config-generator agent
─────────────────────                   ──────────────────────
1. Read context input                   1. (dispatched by generate)
2. Validate                             2. Run full generation pipeline
3. Map to onboard format                3. Emit ALL artifacts
4. Build agent prompt                   4. Self-audit telemetry
5. DISPATCH AGENT  ───────────────────► 5. Return structured JSON response
6. Parse JSON response, return summary

FORBIDDEN patterns (every one observed in the 2026-04-16 release-gate run):

• FORBIDDEN: Writing CLAUDE.md inline via Write tool from this skill's execution context.
• FORBIDDEN: Calling Write or Edit tools from this skill at all (any file).
• FORBIDDEN: Skipping the agent dispatch and running generation pipeline steps directly.
• FORBIDDEN: Treating this skill as a Write-tool wrapper.

REQUIRED pattern:

• REQUIRED: A single Agent dispatch with subagent_type: "config-generator" and the pre-seeded context object (Step 3 below).

If you find yourself reaching for the Write tool while executing this skill, STOP — that is the bug this contract is designed to prevent. The artifacts must be written by the dispatched agent, not by this skill.

Hard-fail safety net: The config-generator agent itself checks for dispatchedAsAgent: true in its context. If a caller bypasses dispatch (somehow invoking the agent's logic from the main session inline), the agent refuses to write anything and reports the violation. This is defense in depth — but it does NOT excuse violations of the contract above.

Plugin detection fallback: If callerExtras.installedPlugins is absent in the provided context, the generation skill probes the filesystem using the shared procedure in ../generation/references/plugins/plugin-drift-detection.md § Probe Procedure (generate runs probe-only — no baseline diff). This means internal callers that don't compile plugin data will still get Plugin Integration output if plugins are installed.

---