ac

Commit the current working-tree changes as atomic commits using the `ac:git-master` skill.

Execute an approved plan from `.ac/plans/<slug>.md`. Wave-by-wave parallel `ac:plan-worker` spawn with tier-aware model routing (quick→haiku, junior→sonnet, senior→opus), per-step done-when verification, tier escalation retry on UNMET, wisdom accumulation across waves, complexity-gated post-execution review (`ac:plan-code-review` for Standard, `ac:plan-code-deep-review` for Complex), 3-strike halt rule, and `/ac:commit` handoff. Use after `/ac:plan` approves a plan. Accepts `--loop` for auto-mode.

Deep project investigation followed by an optimized CLAUDE.md, optional CLAUDE.local.md, and up to five path-scoped `.claude/rules/*.md` files. Fuses the `init-deep` 4-phase shape with the gap-fill interview from CC native `/init`, and drives the `ac:claude-md-rules-creator` skill as the writing playbook.

Interview-driven planner with pre-investigation, double-check via re-verify pass, and plan generation through Metis → Prometheus → Momus chain. Stage 0 investigates feasibility (parallel `ac:explore` and `ac:librarian` fan-out plus codebase fit plus initial ambiguity scoring) before asking any questions. Stage 1 interactive co-decision interview (4-dim weighted Socratic with 5 rotating perspectives plus gray-area multiSelect discussion). Stage 2 re-verify cross-check against research findings and other decisions, surface conflicts via AskUserQuestion, apply 6 thinking models. Stage 3 plan generation through `ac:plan-metis` (or deep), then `ac:plan-prometheus`, then `ac:plan-momus` (or deep) with revision loop max 3 and stall detection. Stage 4 deliver via AskUserQuestion for Execute / Deep Review / Adjust, handoff to `/ac:execute`. Accepts `.ac/tasks/*.yaml` task file. Flags `--loop` (auto-execute after approval) and `--deep` (force Complex tier review).

Full E2E autonomous entry point. Runs plan, execute, and commit end-to-end with state-file checkpoint for resumability and an audit log. Interviews the user to lock requirements, generates a plan via the Metis/Prometheus/Momus chain, executes wave-by-wave with tier-aware worker routing, runs complexity-gated review, and commits. Pass `resume <slug>` to continue an interrupted run from its last saved checkpoint.

Deep, parallel-friendly codebase research specialist. Use proactively when exploration needs more than three queries, covers multiple naming conventions, requires deep file traversal, or benefits from LSP/AST-grep precision. Triggers on questions like "where is X defined", "who calls Y", "find all usages of Z", "how does W work", "find the regression in...", "search the codebase for X". Caller may pass a thoroughness hint "quick", "medium", or "thorough". Read-only. Returns `file_path:line_number` citations with a short synthesis. Use aggressively; undertriggering is the failure mode.

External documentation and OSS research specialist. Use proactively when the question involves an unfamiliar library, framework, or external API; needs official documentation; references third-party code patterns; or asks about open-source implementations. Triggers on questions like "how do I use [library]?", "what's the best practice for [framework feature]?", "show me [library] source for X", "find [library] usage examples", "how does [framework] implement Y?", "why does [package] behave this way?". Caller may pass a thoroughness hint "quick", "medium", or "thorough". Read-only. Returns URL/permalink citations with code-snippet evidence and a short synthesis. Use aggressively; undertriggering is the failure mode.

Strategic technical advisor for architecture decisions, deep debugging stalls, second-opinion review, and security or performance concerns. Use when the question requires elevated reasoning beyond the orchestrator's bandwidth: multi-system tradeoffs, unfamiliar patterns, 2+ failed fix attempts on the same bug, post-implementation self-review on significant work, security or performance hot paths. Trigger phrases - "should we refactor X to Y?", "review this implementation before I ship", "I've tried 3 fixes and the bug persists", "is this architecture sound?", "what's the right pattern for Z?". Read-only consultant; advises, never executes. Returns a 2-3 sentence bottom line, a numbered action plan, effort estimate (Quick/Short/Medium/Large), and confidence (high/medium/low). Use proactively before implementing load-bearing decisions.

4-stage deep post-execution reviewer for Complex-complexity plans. Reads plan file path plus modified files list, runs Stages 1-3 identical to `ac:plan-code-review` (compliance L1/L2/L3 + spec + quality), then Stage 4 cross-layer integration (integration trace across module boundaries + caller impact via LSP `findReferences` + architectural compliance against CLAUDE.md conventions). Single Opus agent consolidating verifier + code-review + deep-code-review chain. Returns APPROVED or BLOCKED. Single-shot stateless. Use after `ac:execute` completes all waves of a Complex plan; mandatory, no skip.

3-stage post-execution reviewer for Standard-complexity plans. Reads the plan file path plus modified files list passed in the prompt, verifies implementation against plan claims. Stage 1 compliance (L1 Exists / L2 Substantive / L3 Wired + Must NOT Have + Scope Fidelity, gates everything), Stage 2 spec compliance against acceptance criteria, Stage 3 code quality (logic errors, null/undefined, anti-patterns, SOLID violations, missing error handling) with severity + confidence tagging. Returns APPROVED or BLOCKED. Single-shot stateless. Use after `ac:execute` completes all waves of a Standard plan, before commit.

Authors Claude Code custom subagents (`.claude/agents/<name>.md`, `~/.claude/agents/<name>.md`, or `<plugin>/agents/<name>.md`). Use whenever a new subagent is being designed, an existing agent is being edited, a tool allowlist or denylist is being chosen, a model and effort budget are being set, `permissionMode` is being picked, persistent `memory` (user/project/local) is being added, `skills:` are being preloaded into a subagent, `isolation: worktree` is being configured, the agent's system-prompt body is being structured (Identity/Execution/Output/Failures/Constraints), or an agent that fails to fire or returns the wrong shape is being debugged. Triggers on "create a custom subagent", "write a new agent", "add an `.claude/agents/` file", "make a `code-reviewer`", "tool restrictions for an agent", "agent system prompt", "preload skills into an agent", "agent memory", "agent worktree", "fix this agent". Use even when the user does not say the word "agent" but is asking to delegate a recurring task to a fresh isolated context. Pair with `ac:skill-creator` for the surrounding file-shape and scope decisions, and with `ac:prompt-writer` for the system-prompt body content. Target is Opus 4.7; Sonnet 4.6 follows the same shape at lower effort. Undertriggering is the failure mode, lean in when the request implies an isolated worker.

Authors CLAUDE.md, CLAUDE.local.md, and `.claude/rules/*.md` (the standing-instruction layer Claude Code prepends as a `<system-reminder>` user message at every session start). Covers the four scopes (managed / user-global / project-team / project-personal), `paths:` for path-scoped rules, `@path` imports (depth 5, external-approval), splitting bloated files, `claudeMdExcludes`, HTML-comment maintainer notes, AGENTS.md interop, monorepo + worktree handling, and debugging "Claude is not following my CLAUDE.md". Use proactively when the user says "add to CLAUDE.md", "write a rule", "init memory", "global instructions", "user-level instructions", "team conventions", "project memory", "claude.md", "rules file", "agent instructions", or asks to refactor an oversized CLAUDE.md or fix instructions the model is ignoring. Triggers even when the user does not say "skill". Pair with `ac:skill-creator` for the surrounding shape, `ac:prompt-writer` for body content, `ac:agent-creator` for subagent CLAUDE.md inheritance.

Authors Claude Code slash commands (`/name [args]` markdown files at `.claude/commands/<name>.md` or `<plugin>/commands/<name>.md`, plus the equivalent `.claude/skills/<name>/SKILL.md` directory format). Use whenever a new slash command is being designed, an existing command is being edited, an argument-driven workflow is being captured, shell injection is being added to gather live context (git diff, PR data, env), multi-phase command bodies are being structured, `argument-hint` and `arguments` frontmatter are being chosen, approval gates with AskUserQuestion are being inserted before destructive actions, or a command that fires the wrong way is being debugged. Triggers on "create a slash command", "build a /command", "add a plugin command", "make a /pr-summary", "write a command body", "shell injection in command", "argument design for command", "fix this command". Use even when the user does not say the word "command" but is asking for `/name` invocation with arguments, a side-effect workflow, or a context-gathering recipe. Pair with `ac:skill-creator` for the surrounding skill shape (frontmatter, scope, invocation control, bundling) and `ac:prompt-writer` for the body content. Target is Opus 4.7; Sonnet 4.6 follows the same shape at lower effort. Undertriggering is the failure mode, lean in when the request implies a slash command.

Git expert - atomic commits with style detection, rebase/squash, and history archaeology (blame, bisect, pickaxe). Use for git operations.

Writes high-signal prompts for Claude Opus 4.7 (system prompts, subagent briefings, skill bodies, command bodies, agent bodies, CLAUDE.md files, .claude/rules/*.md) and audits existing ones. Use whenever instructions are being authored or edited for any Claude to execute, even when the user does not say the word "prompt". Triggers on "write a system prompt", "brief a subagent", "draft an agent body", "skill content", "command body", "CLAUDE.md", "rules file", "audit this prompt", "improve this instruction", "make this prompt better". Sibling creator skills (skill-creator, command-creator, agent-creator, claude-md-rules-creator) call this skill for the prompt body itself. Use aggressively; undertriggering is the failure mode.