roborev-superpowers

Carry a feature from intent to verified implementation with Superpowers, Kata, and Roborev. Keep the user's outcome and constraints visible throughout the work: Superpowers supplies the spec and linked implementation plan, Kata records the work and evidence, and Roborev critiques intent and implementation.

Usage

/roborev-superpowers <feature or existing spec/plan>

Explicit invocation only

Invocation must be explicit: literal personal /roborev-superpowers, or structured Factory skill selection. Requests such as “help me implement this feature” without one of these explicit mechanisms must use native behavior and must not run roborev.

Honor existing session authorization and project AGENTS.md instructions. This invocation authorizes the review loop; commits, branches, pushes, external messages, and changes to requirements still follow the user's authorized scope. Resume from existing artifacts, open tasks, and evidence rather than restarting completed stages. Use the installed Superpowers skills for their respective stages and preserve their human decision points unless already waived by the user. If Superpowers, Kata, or a goal-capable reviewer is unavailable, report the missing prerequisite and the work that remains unverified.

Select a configured Pi (pi) or supported bare-mode Claude (claude) reviewer and use its name as <reviewer> below. The implementing agent can be Codex or another agent; goal review currently supports only Pi and Claude.

Connect intent to tracked work

1. Use Superpowers brainstorming to establish the intended outcome, constraints, non-goals, and observable success criteria. Keep its spec authoritative.

2. Search Kata first; reuse or create a parent for this outcome. Record the spec path, attention state, and existing branch when applicable. Review the spec with roborev review --from-skill --type goal --agent <reviewer> --wait --spec <spec-path>. Quote paths as shell arguments.

3. Resolve findings against the captured artifacts. Correct incomplete intent or explain a false positive with evidence. A changed requirement returns to brainstorming and the user's decision; it is not an implementation shortcut. Request a new goal review after edits. Reruns review the old frozen snapshot.

4. Use Superpowers writing-plans to produce a plan linked through **Spec:**. Map each implementable task to a Kata child with its spec/plan references, deliverable, verification command, and expected result. Parent links group work; dependency links represent actual prerequisites. Mark deliberate deferrals explicitly. Leave unrelated project work intact.

5. Review the linked plan and the populated open graph together:

   roborev review --from-skill --type goal --agent <reviewer> --wait \
     --spec docs/superpowers/specs/feature-design.md \
     --plan docs/superpowers/plans/feature.md
   

   Address in-scope findings in the spec, plan, or task graph. Record the review snapshot and outcome on the parent before executing the agreed plan.

For proposed Kata creates/edits, an installed Kata integration can call POST /api/goal-review with the candidate before committing it. Respect its blocked result and the user's gate policy. The endpoint only simulates a write; atomic review-and-write and freshness checks belong to that integration. Without such an integration, author the graph normally and run the explicit goal review above. Do not claim that direct Kata writes were gated.

Execute, review, and keep evidence

Use the selected Superpowers execution skill, inline or with authorized agents. For each task, claim its Kata issue and maintain attention state. When delegating, create tracked children and assign one writer to each issue's work metadata.

Follow Superpowers TDD: observe the intended test fail, implement the smallest change that passes, and run the task's verification. Commit coherent slices when authorized. Record commands, results, and commit references on the task. Read Roborev code reviews at the configured cadence, verify each finding, fix valid in-scope defects with regression tests, and record the disposition. Intent review findings require artifact/task edits; they are not code-fix candidates.

When implementation exposes a plan defect or scope change, reconcile the spec, plan, and Kata first and request a fresh goal review. Proceed against the updated intent. Automatic watching is useful feedback, but a queued job is not a pass and checkbox changes alone do not request another review.

Complete the outcome

Use Superpowers verification-before-completion and finishing-a-development-branch within the user's authorized delivery scope. Completion evidence includes the success criteria exercised, verification commands and results, delivered commit or PR, and review dispositions tied to the delivered code or intent snapshot. Await required CI and current-head reviews before claiming them green. Close Kata children with their evidence, then the parent once all required work is complete. Leave both attention state and an explanation current for unfinished handoffs.

Sandbox access

roborev uses a local daemon. If a command fails with permission denied, the sandbox may be blocking its loopback port or Unix socket. Retry the same command with the runtime's supported sandbox escalation mechanism when supported and authorized. Do not start or restart the daemon because a sandboxed status probe cannot reach it. Never replace a user's installed roborev binary to run this workflow.