pi

Work through review comments on the current PR. Triages each comment as Address / Push-back / Escalate and executes immediately — no approval gate. Addressed comments produce code changes, commits, and silent thread resolution via GraphQL (no reply text). Pushed-back comments stay unresolved. Escalations invoke /interview or /architect inline. Watches GitHub Actions after push — the loop is not complete until all checks pass.

Re-derive the simplest, cleanest architecture from first principles. Enforces minimal code, composition, single source of truth, deep modules, clear abstractions, clean interfaces, functional core, state machines for lifecycle, ports/adapters, typed recovery, no silent fallbacks, Effect primitives in Effect-owned code, and game-theoretic incentive compatibility. Output uses ASCII diagrams and code contracts, never mermaid, and fits one to two screens. Hands off to /review.

Fan out six focused reviewer personas against the current PR — correctness, testing, maintainability, project-standards, security, and previous-findings. Each reviewer applies first-principles and game-theoretic pressure: invariants, incentives, information asymmetry, local shortcuts, repeated-game health, and adversarial behavior. Dedupe, validate, verify diff-line commentability, then post one GitHub review with summary plus inline comments where possible. No autofix, no approval, no request-changes. Hands off to /address.

This skill encodes Emil Kowalski's philosophy on UI polish, component design, animation decisions, and the invisible details that make software feel great.

Explore a codebase to find opportunities for architectural improvement, focusing on making the codebase more testable by deepening shallow modules. Use when user wants to improve architecture, find refactoring opportunities, consolidate tightly-coupled modules, or make a codebase more AI-navigable.

Interview the user in five-question rounds until the problem, values, constraints, non-goals, incentives, failure modes, and solution direction are locked. Uses first-principles decomposition, game-theoretic incentive analysis, concrete examples, 5 Whys, laddering, QFT-style question selection, and Socratic pressure. Hands off to /architect.

Open a GitHub issue (or epic) from the locked architecture using a Linear + Shape Up + Larson-synthesized structure. Header (Owner · Appetite · Status · Part of), Problem, Outcome, Solution sketch with ASCII diagrams and code contracts, Files, Acceptance criteria, Rabbit holes, No-gos, References, Depends on. No mermaid. Hands off to /work.

Capture the post-mortem for a compound-engineering cycle. Reads the issue, PR, review comments, git log, and final diff. Extracts what worked, what changed, what incentives caused friction, what mechanisms improved code health, and what durable rule should be considered. Writes docs/learnings/YYYY-MM-DD-slug.md, commits it, pushes the branch, and hands off to /merge when the PR is still open.

Merge the current PR after /learn has committed and pushed the learning file. Verifies a clean tree, open PR, up-to-date pushed head, passing required checks, and reviewability before merging with the repository's safe strategy, deleting the branch, checking out the base branch, and pulling it forward.

Open a pull request with a strict minimal body — Summary (2-4 sentences) plus an ASCII Flow diagram plus `Closes

Pre-/interview grounding for unfamiliar domains. Maps one concrete library, module, external service, API, repo area, or pattern using repo search, tests, node_modules source, official docs, context7, gh_grep, exa, shelf, structural search, first-principles decomposition, and game-theoretic incentive analysis. Produces a one-screen brief with at least one real code snippet and, where structure matters, one ASCII diagram. Hands off to /interview.

Audit and harden tests after /work and before /docs. Reads the linked issue architecture, changed files, commits, diff, and existing tests; derives an edge-case test matrix; adds only missing tests; runs the tightest verification loop; commits test-only changes; hands off to /docs.

Execute the architecture in a GitHub issue on the current branch. Reads the issue, derives tasks from the Modules section, commits referencing the issue number. Refuses to run on main/master/trunk without explicit confirmation. No worktrees, no new branches. Hands off to /test.