Extension Foundry
Evolve capability through promotion, testing, and rollback rather than uncontrolled self-modification.
- Run
evopilot_analyze_sequences; use evopilot_analyze_habits only to inspect individual actions. Ignore patterns supported by fewer than three observations.
- At three observations, describe a candidate workflow. At five with at least three successful outcomes, it becomes
draft_ready. At eight with at least three successes, consider it stable.
- Prefer a Skill when existing tools are sufficient. Add a deterministic script for repeatable mechanics.
- Draft an MCP server only when live data, authentication, controlled external actions, or a missing tool genuinely requires one.
- Use
evopilot_compile_skill only for a draft_ready or stable fingerprint. Run both evopilot_validate_skill and evopilot_assess_skill_quality, then refresh QUALITY_REPORT.md with evopilot_annotate_skill_quality after edits. Review SKILL.md, evopilot.json, and every quality annotation; keep the bundle uninstalled, version it, and test it in an isolated workspace. Structural validation is not proof that the workflow executes correctly.
- Compare success rate, retries, corrections, time, and user acceptance against the previous workflow.
- Do not promote pure generic tool sequences or adjacent duplicate actions. When a meaningful workflow reaches
draft_ready or stable, proactively explain its evidence before generating it. Compile, validate, assess, and annotate locally, then present the exact installation approval only if both gates pass. After the user explicitly confirms, consume the one-time approval and install the Skill. Never install silently or treat an earlier general preference as installation approval.
- Roll back when the new workflow performs worse or routes unrelated requests.
Never modify EvoPilot's approval policy through learned behavior. Explicit user instructions outrank learned memories.
When the PreToolUse gate blocks an exact action, show its approval ID and ask the user. Call evopilot_authorize_once only after the user explicitly confirms that action, then retry it once.
1---2name: extension-foundry3description: Analyze repeated EvoPilot observations and compile proven workflows into portable, versioned Agent Skills or other reviewed extensions. Use when improving or extending the agent itself.4---56# Extension Foundry78Evolve capability through promotion, testing, and rollback rather than uncontrolled self-modification.9101. Run `evopilot_analyze_sequences`; use `evopilot_analyze_habits` only to inspect individual actions. Ignore patterns supported by fewer than three observations.112. At three observations, describe a candidate workflow. At five with at least three successful outcomes, it becomes `draft_ready`. At eight with at least three successes, consider it stable.123. Prefer a Skill when existing tools are sufficient. Add a deterministic script for repeatable mechanics.134. Draft an MCP server only when live data, authentication, controlled external actions, or a missing tool genuinely requires one.145. Use `evopilot_compile_skill` only for a `draft_ready` or `stable` fingerprint. Run both `evopilot_validate_skill` and `evopilot_assess_skill_quality`, then refresh `QUALITY_REPORT.md` with `evopilot_annotate_skill_quality` after edits. Review `SKILL.md`, `evopilot.json`, and every quality annotation; keep the bundle uninstalled, version it, and test it in an isolated workspace. Structural validation is not proof that the workflow executes correctly.156. Compare success rate, retries, corrections, time, and user acceptance against the previous workflow.167. Do not promote pure generic tool sequences or adjacent duplicate actions. When a meaningful workflow reaches `draft_ready` or `stable`, proactively explain its evidence before generating it. Compile, validate, assess, and annotate locally, then present the exact installation approval only if both gates pass. After the user explicitly confirms, consume the one-time approval and install the Skill. Never install silently or treat an earlier general preference as installation approval.178. Roll back when the new workflow performs worse or routes unrelated requests.1819Never modify EvoPilot's approval policy through learned behavior. Explicit user instructions outrank learned memories.2021When the PreToolUse gate blocks an exact action, show its approval ID and ask the user. Call `evopilot_authorize_once` only after the user explicitly confirms that action, then retry it once.
sdfdu