Agent-Skills.md

Agent Skills: Moonshot Orchestrator

Use for bounded implementation work that already has enough context and does not need the phase harness.

UncategorizedID: munlucky/claude-settings/moonshot-orchestrator

Repository

munlucky/claude-settings
munlucky

Install this agent skill to your local

pnpm dlx add-skill https://github.com/munlucky/moonshot-relay/tree/HEAD/skills/moonshot-orchestrator

Skill Files

Browse the full folder contents for moonshot-orchestrator.

Download Skill

Loading file tree…

skills/moonshot-orchestrator/SKILL.md

Skill Metadata

Name
moonshot-orchestrator
Description
Use for bounded implementation work that already has enough context and does not need the phase harness.

Moonshot Orchestrator

Use When

Use for a bounded implementation objective with enough accepted context to execute and verify now.

Role

Own and prove one implementation slice. Architecture-derived work consumes only the bounded selected ADR and traceability slice.

Route Away

  • Use moonshot-phase-runner for multi-phase plans, long-running harness work, or staged adoption packages.
  • Use product-orchestrator when the user is still defining product scope.
  • Use moonshot-architecture before implementation when the request lacks an accepted architecture package for non-trivial architecture decisions.
  • Stop for clarification only when a wrong assumption would change scope, security, data shape, or user-visible behavior.

Hard Stops

  • Do not broaden scope beyond the user request.
  • Do not skip code review for non-trivial code changes.
  • Do not claim completion with stale, missing, or smoke-only evidence.
  • Do not execute a blocked ARCHITECTURE_HANDOFF, and do not bypass a ready handoff by copying raw KG, ontology, MemoryGraph, log, transcript, or browser scrape payloads into the attempt prompt.
  • When runtime-state completion authority is available, do not claim clean finish from chat output, markdown reports, phase status, or verifier JSON alone. Require scripts/runtime-state.mjs assess-completion to produce an accepted DB decision.
  • Before approval-required operations or writes near protected runtime paths, classify the operation with tools/sandbox/policy.mjs check --json; unauthorized blocking events must stop clean completion.
  • Do not mutate unrelated files or revert user changes.

Procedure

  1. Confirm bounded scope and ready handoff guards; never replace architecture evidence with chat summaries.
  2. Apply docs/public/guidelines/minimal-correct-implementation.md before choosing the implementation shape, and apply docs/public/guidelines/untrusted-content-boundary.md.
  3. Implement, review, rerun invalidated checks, and report evidence or a typed blocker.
  4. For moonshot-architecture work, consume selected ADR ADR/*.md, TRACEABILITY_MATRIX.md, ARCHITECTURE_REVIEW.md, and the traceability slice; pass only ARCHITECTURE_HANDOFF.promptBlock. On violation use scripts/architecture-feedback-render.mjs, never raw KG.

Output Contract

  • Affected file list and rationale.
  • Minimality decision: reused existing surface, added new surface, or skipped lower-rung options, with reason.
  • Fresh test/build/lint or targeted verification output.
  • Review evidence when behavior, shared contracts, or harness logic changes.
  • Agent operating policy evidence when applicable: retrieval, assumptions/blockers, untrusted content disposition, artifact routing, skill readiness, and cumulative risk. This evidence does not replace runtime-state completion authority.
  • Explicit blocker classification if a required check cannot run.

References

  • references/bounded-flow.md: stage order, scope control, and output contract.
  • references/review-and-verification.md: review gate, verifier expectations, and failure taxonomy.

Project Knowledge Context Contract

Pass only the staged compact projectKnowledgeContext.promptBlock and status metadata. Raw knowledge records, logs, transcripts, and secrets are forbidden.