AGENT SKILL

ai-catapult-init

Bootstrap AI-ready governance, traceability, cascade, catalog audits, and validation. Use when setting up AI SDLC; aliases: init-ai-repo and ai-sdlc-init.

License 许可
Apache-2.0
所属仓库
r3dlex/skills
最近更新
2026-10-29
Stars
★ 0
Forks
0

01 前言 Front Matter

name
ai-catapult-init
description
Bootstrap AI-ready governance, traceability, cascade, catalog audits, and validation. Use when setting up AI SDLC; aliases: init-ai-repo and ai-sdlc-init.

02 资源文件 86

03 SKILL.md 正文

Init AI Repo

Quick Start

Run this skill in the target repo. ai-catapult-init is the canonical skill name; init-ai-repo and ai-sdlc-init remain deprecated compatibility aliases. Use dry-run first when scope or host choices are unclear. Keep hosted settings as checklist output unless the user explicitly requests an admin/credentialed action.

Workflow

Phase 1 — Discover & Decide

Inspect repository state, host tooling, trackers, CI, and runtime conventions. Choose greenfield bootstrap, brownfield adoption, hosted-tracker-first, or local fallback. Emit .ai/matrix.json, .ai/init/repo-profile.json, .ai/init/sdlc-path.md, and .ai/phases/01-discover-decide/. OMX surfaces: $deep-interview, $plan, $ralplan; OMC surfaces must produce the same artifact contract.

Phase 2 — Govern & Plan

Generate or refresh AGENTS.md, RULES.md, PLANS.md, CONTRIBUTING.md, active/archived specs, ADRs, .ai/work-intake/, .ai/plans/, and .ai/phases/02-govern-plan/. Ensure a hosted issue/ticket when configured and authorized. Local fallback is allowed before coding, but it must be reconciled before final PR merge. Require active spec/PRD, plan, and acceptance criteria before implementation.

Phase 3 — Configure & Generate

Generate command/runtime surfaces and policy automation under .ai/bin/, .ai/policies/, .ai/commands/omx/, .ai/commands/omc/, .ai/language-packs/, optional Makefile/justfile, and .ai/phases/03-configure-generate/. OMX surfaces: $ralph, $team, $ultragoal, $ultrawork; OMC aliases/commands delegate to the same generated structures rather than duplicate semantics.

Phase 4 — Validate & Handoff

Run local validation, drift checks, generated smoke tests, and hosted/local ticket reconciliation. Emit .ai/validation/report.md, .ai/drift/migration-manifest.json, .ai/handoff/init-ai-repo-handoff.md, and .ai/phases/04-validate-handoff/. OMX surfaces: $doctor, $code-review, $team, $ralph. The handoff records done, verified, remaining, and reconciliation status.

Internal checkpoints

The public workflow is four phases, but the generator preserves the original eight internal checkpoints for compatibility and traceability:

  1. Detect repo state
  2. Choose SDLC path
  3. Scaffold foundation and v3 layout
  4. Scaffold work intake
  5. Configure host adapters
  6. Configure CI and policy
  7. Select language packs
  8. Validate and emit handoff

PR Merge Gate

Every implementation initialized by this skill must assume protected main and PR-only delivery. Emit provider-specific branch-policy checklist/config artifacts unless the user explicitly authorizes hosted mutation with credentials. Admin users may self-approve only when host policy permits it and the host/runtime explicitly supports an admin approve/admin bypass lane for the same actor; all required checks must still pass. If the host rejects same-actor review approval (for example, GitHub hosted PR review), use a distinct admin reviewer or explicit host admin bypass/admin merge. Record the actor, authority, reason, checks, and approval mode, and fail closed when same-actor admin approval support cannot be proven.

When this skill creates or updates PR workflow guidance, require merge only after:

  1. The architect confirms the implementation still matches ADRs, module boundaries, branch policy, and acceptance criteria.
  2. The reviewer confirms code quality, safety, documentation, and drift checks have no blocking findings.
  3. The executor confirms the requested change is implemented, cleanup is complete, and all required checks are green.
  4. All actionable PR comments are resolved and local CI plus host SCM CI (GitHub Actions, Azure Pipelines, or GitLab CI as applicable) are green.
  5. The loop reaches explicit agreement across architect, reviewer, and executor; if any role disagrees, comments remain actionable, checks are not green, or branch policy forbids merge, do not merge or auto-merge.

modules/ci-policy.md adds a mechanical eval-coverage check to this PR merge gate: a changed shippable skill must declare a structurally valid evalset. See modules/evals.md.

Module Map

  • modules/README.md — read when choosing which Layer 3 module applies.
  • modules/phases/README.md — read when mapping the four public phases to internal checkpoints.
  • modules/topology.md — read when generating .ai/matrix.json or validating umbrella depth.
  • modules/ci-adapters.md — read when an umbrella renders selected GitHub, ADO, or GitLab CI from its execution profile.
  • modules/documentation-blueprint.md — read when generating the v3 .ai/, .memory/, and docs/ trees.
  • modules/memory.md — read when defining .memory/human-override/ and .memory/self-learned schemas.
  • modules/sync.md — read when implementing physical-copy sync and drift detection.
  • modules/host-policy-automation.md — read when applying hosted branch/PR/policy mutations safely.
  • modules/validation.md — read when validating generated artifacts, depth rules, and golden fixtures.
  • REFERENCE.md — read only for legacy full template bodies that have not yet moved into focused modules.
  • modules/readme-documentation.md — read when initializing, augmenting, or rewriting README.md.
  • modules/release-versioning.md — read when initializing release tagging, versioning, or CI/CD release workflows.
  • modules/workflow.md — read when generating repo workflow docs, workflow/status manifests, entry-surface links, and handoff indexes.
  • modules/traceability.md — read when generating stable traceability IDs, graph schema, backlink validation, graph fixtures, or cross-skill requirement/work-artifact links.
  • modules/cascade.md — read when generating multi-repo cascade plans, first-run confirmation gates, idempotent linked updates, host adapter contracts, audits, and reconciliation reports.
  • modules/skill-modernization.md — read when auditing compact descriptions, progressive disclosure, trigger boundaries, cross-skill links, and AI-SDLC compatibility.
  • modules/evals.md — read when generating the .ai/evals/ scaffold or the offline eval-coverage gate that modules/ci-policy.md adds to the PR merge gate.
  • modules/mcp-a2a.md — read when generating the .ai/mcp/ registry stub and the A2A cross-agent handoff convention.
  • modules/graph-automation.md — read when scaffolding automatic background knowledge-graph refresh via git hooks and harness hooks; covers the lock+coalesce wrapper, engine knob (default graphify, alternative graphwiki), hook-body contract, and ai-catapult graph-hooks install.

Safety Rules

  • Keep SKILL.md under 100 body lines; move variants to modules.
  • Treat GitHub/ADO/GitLab/Jira branch and policy settings as confirmation-gated mutations; do not apply them without explicit opt-in even with admin credentials.
  • Reference setup-skills and publish-semver host docs instead of copying their full semantics.
  • Preserve existing GitLab/local support unless a later module explicitly migrates it.
  • Physical-copy is the canonical sync strategy; symlinks and git submodules are not.

04 同仓库更多 Skills