plan
Plan ONE feature thoroughly before any code. Use when starting a feature in ha and you want a rigorous, vetted plan you approve before implementation — for a non-trivial change where getting the design and the tests right up front matters more than speed. Accepts a GitHub issue, spec, @file, or free-text task. Invoke explicitly with /ha:plan; hands off to /ha:implement. Requires the superpowers plugin.
clean-worktrees
Reclaim worktrees after features land — remove every merged worktree in the repo (ca/sa/ha and any other, regardless of location) and delete their merged branches. Use once a ca PR is merged. Safe by default — never removes a worktree with uncommitted changes, never deletes an unmerged branch, and never touches the main checkout (it removes the current worktree too, but only if its branch is merged). Invoke explicitly with /ca:clean-worktrees.
code-review
Standards for writing and reviewing code — quality, test rigor, security, and codebase consistency. Use whenever implementing, changing, or reviewing code, or when checking a PR or diff for bugs, missing test coverage, vulnerabilities, missed propagation, or style drift.
dual-review
Run the ca dual-model review (Codex second opinion + blind Claude review + Claude synthesis) standalone on any PR — no implement loop required. Use to get a Claude-and-Codex joint review of a pull request, to re-review a PR after the ca loop finished, or when the user says "dual review this PR", "review with both Claude and Codex", or invokes /ca:dual-review. Emits the same ca_claude_review.v1 verdict as the loop's final review.
implement
Kick off the ca implement loop for a saved plan — create an isolated git worktree for it and print the exact Codex command to run there. Use when the user has a saved plan and wants to build it with the ca loop, or says things like "implement this plan with ca", "start building this plan", "kick off implementation of", "set up the worktree for", or invokes /ca:implement. Hands off to a Codex session; it does not implement the plan itself.
merge-pr
Merge a ca PR after confirming it is genuinely ready — open, not draft (in ca a draft means the Codex-Claude review loop has not approved it), green CI, mergeable, no unresolved blocking feedback. Use once the ca loop marked the PR ready and you want it merged from the CLI. Refuses to merge drafts, red checks, or conflicts. Invoke explicitly with /ca:merge-pr.
plan-loop
Draft an implementation plan for an epic or feature, then spar with Codex to harden it before saving. Use when the user wants to plan a sizable change and have a second model stress-test the plan, or says things like "plan this feature", "draft a plan and spar with Codex", "write an implementation plan for", or invokes /ca:plan-loop. Produces a saved task-by-task plan ready for the ca implement loop.
resolve-conflicts
Resolve merge conflicts between a ca feature branch and its base branch, in an isolated worktree, preserving the intent of both sides. Use when a ca PR or feature branch is behind or conflicting with the base branch and you want the conflicts resolved, verified, and pushed. Invoke explicitly with /ca:resolve-conflicts.
review-pr
Review a pull request against the implementation plan it was built from, then emit a structured ca_claude_review.v1 JSON verdict (approve / request_changes / blocked) with blocking findings. Supports mid-implementation checkpoint reviews (mode=checkpoint) that judge only the milestones built so far. Use when the ca loop or a user asks to review a Codex-built PR before it is merged, or says things like "review this PR against the plan", "review what Codex built", "is this implementation correct", or invokes /ca:review-pr. Reviews the PR diff; does not edit code.
synthesize-review
Synthesize a blind Claude review and an advisory Codex second-opinion review into the single ca_claude_review.v1 verdict for a ca final review round. Use when the ca loop invokes /ca:synthesize-review with blind, second_opinion, plan, pr, round, and out paths. Does not edit code.
ca-implement-plan
Build a saved task-by-task implementation plan in an isolated worktree, milestone by milestone, with Claude checkpoint reviews between milestones and Claude-reviewed final rounds gating the PR before it is marked ready — the Codex half of the ca (Cooperate Agents) loop. Use when the user points at a plan file under docs/ca/plans or any task-by-task plan and says things like "implement this plan", "build the plan at this path", "run the ca loop on this plan", or "have Claude review what you built". Human confirmation is required at key decision points and before cleanup.
adversarial-verification
Multi-pass adversarial verification harness that guarantees the accuracy of a change before it ships. Use whenever you need high confidence that an implementation is correct, safe, and complete — before opening a PR, before approving a review, or whenever a change touches risky surface area such as auth, crypto, data migration, money, or external input.
apply-feedback
Address review feedback on a PR and push the fixes back, rigorously — feedback is evaluated against the code, not obeyed blindly. Use after /ha:review-pr requests changes, or to act on human review comments on a PR. Invoke explicitly with /ha:apply-feedback. Requires the superpowers plugin.
simple-implement
Takes a plan — a file path or plain-text task — and drives ONE simple feature to an open PR fast, with a human approval gate before any code is written. Use for a small-to-medium, well-scoped implementation task where you want to move quickly and review separately afterward. Stops at PR; run /sa:review-pr to review. Invoke explicitly with /sa:simple-implement.