dev-kit

0-arg restart-safe PR check, diagnose, fix, and hand-off loop.

0-arg local-mode PR babysitter. Pre-push pytest gate + local LLM judge verdict loop; replaces `gh pr checks --watch` with `bin/review-local.sh`.

0-arg setup for CLAUDE.md, AGENTS.md, hooks, and optional CI.

0-arg. Per-step sub-agent delegation + self-fix loop (MUST-36~38). Uses harness-runner engine. TDD + verify + debug integrated.

4-phase systematic debugging. No fix proposal before Phase 1 (reproduce) completes (MUST-L2). Root-cause-first Iron Law. Standalone invocation hands the root cause to /dev-kit:plan instead of fixing inline.

Explicit version bump of `.claude-plugin/plugin.json` + push of `chore/bump-vX.Y.Z`. Mirrors the auto-bump in `.github/workflows/version-bump.yml` but user-triggered for race recovery and pre-PR explicit bumps. For catching a branch up to origin/main (==, not +1) use `/dev-kit:sync-version` instead.

Read-only CI readiness audit. Prints one PASS/FAIL summary across files, marker, provider file, secrets, and gh auth. Hand-off answer to "would CI succeed on my next PR?"

Install dev-kit's reusable CI workflow templates into a target project. Idempotent via `.dev-kit/ci-config.json` presence, no version gate. Hand-off to /dev-kit:build.

Triage failing GitHub Actions runs across recent commits, dedupe against a persisted case store, judge new failures against a model/context/harness taxonomy with a required repro + regression test, and record them without re-analyzing repeats.

Detect + selectively apply drift between installed CI templates and current dev-kit source. 4-state per-file classification with backup-before-overwrite.

0-arg generic plugin-architecture visualizer. Walks any target repo, emits self-contained HTML with multi-level views (architecture / code / skill / hook / tools-lib / external) + domain pillar map (DB · Cloud · API · MCP · Skill · Hook · Network · Security · Build · Test · Storage · LLM) + per-skill workflows (multi-strategy extraction incl.

Refresh the dev-kit Codex marketplace checkout and synchronize the versioned plugin cache. Use when Codex reports the marketplace is current but the installed cache may be stale, or after a dev-kit merge.

skill + hook + methodology picker (multiSelect).

0-arg cost-gate status. Prints current session spend, threshold distance, and a two-line git-trailer block to include in commits so the PR-level cost flag can aggregate.

Audit repository documentation with the project README as the highest-priority document. The README is always audited AND verified every run, and updated when needed.

0-arg eval extension. Replays transcripts and consumes workflow evidence against registered rubrics, preserving legacy Agent Behavior D1–D7 and reporting five harness-effectiveness components plus the nested measurement-integrity submetric (issue

Idea → cited research → HTML proposal (human confirms) → /dev-kit:plan hand-off. The proposal is rendered and reviewed BEFORE the expensive 5-gate PRD work runs, not after.

Create or delete managed GitHub Actions gate artifacts backed by gates.json.

Unified 3-dimension picker for project / session / AI-judge gates. Reads .dev-kit/gates.json + .dev-kit/ci-config.json + .dev-kit/harness-mode.session.json and dispatches writes through `python -m lib.gates_state ...`.

Session-scoped on/off toggle for the tdd-guard, worktree-guard, and git-guard hard-block hooks.

0-arg harness-effectiveness report. Wraps `lib.harness_effectiveness.build_report` and prints the five-component (prevention / first-pass / recovery / learning / measurement-integrity) scorecard as JSON + a one-line status verdict. The measurement-integrity component also reports a nested stability submetric (issue

Session-scoped local-hook mode picker — fast (all optional local hooks off), full (default, all on), or custom (interactive per-local-hook picker via AskUserQuestion).

Diagnose failed Claude Code or Codex hooks, repair safe cache and registration drift, and report the exact restart step.

0-arg read-only code health audit. 8-dim fan-out (dead, dup, smell, overeng, overarch, cleancode, tokenbudget, slop) + --secrets/--slop aliases to the audit family (lib/analysis_core/dimensions.py).

5-field safety-contract interview that gates plan emission. Drives `lib.interview_engine` through one Ralph loop, enforces `safety_valve=8`, `narrowed_delta`, `dedup_metric` (identical-ambiguity-cycle=2), and `user_interrupt`.

Optional Linear task tracker. Reconcile the current repository task with a canonical project and non-duplicate issue. Auto-syncs on every Claude Code edit when configured. Owner-gated auto-triggers also fire on worktree create, session start, and task change.

Refresh docs/llm-info/<provider>.json from each vendor's official pricing page via WebFetch extraction. Diff-then-commit; manual like set-provider.sh.

Toggle /log setup|on|off|status — install/remove loghooks from ~/dev/loghooks into the current project's Claude/Codex settings.

Read or write the active DEV_KIT_MODE (full | lite | undev) for the current project. Picker by default; --show to display current mode; --scope=local to write to .claude/settings.local.json instead of .claude/settings.json.

0-arg plan stage. Take 1-line idea → PRD.md + phases/<name>/{index.json, step<N>.md} in 5 gates. Quantified value (cost/LTV) + ambiguity loop (0-10) replace the old 5-question grill-me.

Deterministic PR verification — fresh `gh pr view` + check + comment fetches on every call. Catches the "stale CI / LLM-judge still in progress" false positive the babysit flow had.

0-arg YAML-to-HTML renderer for reviewable design proposals.

0-arg orchestrator-first GitHub backlog triage. Gathers open PRs + issues, scores (bottleneck / risk / change containment), orders by orchestrator critical path, and writes a proposal YAML + HTML via the existing `/dev-kit:proposal` renderer.

0-arg slop-removal chain. One slash wraps inspect → 3-pass delete sweep → review. Gated phases for deleting AI slop and dead features (not refactoring).

0-arg skill — usage telemetry dump + per-skill delete proposal. User approves each deletion explicitly.

End-to-end autonomous loop with 4 user gates + unattended build/babysit/ship.

0-arg cleanup chain. One slash wraps inspect -> cleanup -> review. 3 gated phases with quoted exit codes between each.

0-arg research gate. Run Phase 0-3 escalation (cache / direct / multi / human) + verify() + enforce_citations(). /dev-kit:research <claim> [--max-phase N].

Parallel multi-dimension code review with a false-positive filter. Fans out to per-dim experts (correctness, security, architecture) that run in parallel and return evidence-backed findings; a verifier pass confirms/rejects each candidate before rendering per-line inline comments plus a PR-style summary with a verdict.

Local equivalent of the GH-Actions review workflow. Runs /dev-kit:review + /dev-kit:security + /dev-kit:maintenance (via local `claude` CLI) with the same verdict extraction + combined gate + L3-evidence enforcement + optional auto-approve as `.github/workflows/review.yml`. Saves Action minutes when private repos hit the GH-Actions budget cap.

Security fan-out — OWASP Top 10 2025 (A01–A10) plus a separate LLM01 Prompt Injection dimension. Eleven parallel subagents, one per category, return evidence-backed findings; a verification pass confirms or rejects each before a per-category breakdown table + verdict.

Calculate a deterministic 0-100 security scorecard for the current repository and render an evidence-backed Markdown table for OWASP Top 10 categories.

0-arg. Release tag emit. Gate check only (hooks auto). Requires Review verdict=Approve + main-block pass.

Run the skill usage telemetry CLI and inspect turns, invocations, and per-project usage.

Interview-based Single Source of Truth harness document writer (5 rounds × 2-3 evidence-backed recommendations, full traceability, hands off to /dev-kit:plan).

HOTL visualization. Current loop progress + cumulative cycles + hand-off chain + eval score on one screen.

DEPRECATED. The GitHub Merge Queue now owns version sync at merge time; this skill is a no-op wrapper around bin/sync-version.sh that preserves the CLI surface for callers that haven't migrated yet. See docs/proposals/release/plugin-version-bump-via-merge-queue.yaml.

Read or write the team collaboration toggle (DEV_KIT_TEAM on|off). Default OFF, independent of DEV_KIT_MODE. When ON, team roles/dependency-aware planning is enabled and .dev-kit/ stays tracked in git.

0-arg token-efficiency dashboard. Runs tools/token_efficiency_analyzer.py over logs/{claude-code,codex}/*.jsonl to produce an HTML report (+ lazy per-worktree transcript sidecars) -- 4-dim session scoring, 6 anti-pattern warnings, USD savings estimate.

0-arg interactive prune of stale worktrees. Counts registered worktrees, lists them oldest-first by branch-tip age, asks how many to remove, then dispatches `bin/worktree-remove-safe.sh` per row after a y/N gate.