Codex Toolkit
Automatic workflow routing for 20 Codex skills and 6 agents — bug hunting, implementation, completion gates, and release verification.
Say what you want done, not which skill to run.
Codex Toolkit gives Codex specialist workflows for the parts of software work that are easy to get wrong: understanding an unfamiliar repository, finding unknown bugs, proving root causes, controlling migrations and refactors, finishing large tasks completely, coordinating parallel agents, and deciding whether an integrated change is actually ready to ship.
The full installer adds a small global routing policy so Codex can select one primary specialist and compose supporting skills only when their trigger becomes relevant.
For example:
Find and fix important bugs in this repository.
can become:
repository-intelligence?
→ bug-finder
→ debugging-investigator?
→ owning implementation specialist?
→ unlazy?
→ verification-and-release?
? means conditional. A workflow does not activate every skill mechanically.
Install once
npx --yes github:cmdr-chara/codex-toolkit setup
That installs:
- all 20 Codex skills;
- 6 Mission Control agents;
- automatic workflow routing in a managed section of your global Codex
AGENTS.md; - the detailed workflow catalog under
~/.codex/codex-toolkit/workflows.md; - a user-level auto-updater pinned to the latest published GitHub Release.
Future releases can add or update skills and routing without another manual install. The updater never follows unreleased main commits.
The installer never replaces your whole AGENTS.md. It owns only:
<!-- codex-toolkit:start -->
...
<!-- codex-toolkit:end -->
Everything outside that block is preserved. Malformed or duplicate markers fail closed instead of being guessed around.
Check or disable automatic updates:
npx --yes github:cmdr-chara/codex-toolkit auto-update status
npx --yes github:cmdr-chara/codex-toolkit auto-update remove
See Automatic updates for Windows, macOS, Linux, custom CODEX_HOME, and scheduler details.
Use it normally
You usually do not need to name a skill after the full setup.
Find important bugs we have not noticed yet.
Fix this intermittent reconnect regression properly.
Improve this repository and take the best improvement through completion.
Make this slow path faster, but prove the improvement with comparable measurements.
Upgrade this framework without breaking compatibility.
Build this approved feature and do not stop half-finished.
Tell me whether this release candidate is actually safe to ship.
The routing layer chooses the smallest workflow that owns the decision. Repository-local AGENTS.md instructions remain authoritative for project-specific rules.
Common workflows
Unknown bugs
repository-intelligence?
→ bug-finder
→ debugging-investigator? # when causal proof is still missing
→ implementation specialist?
→ unlazy? # substantial accepted remediation
→ verification-and-release? # integrated release candidate
bug-finder discovers previously unknown correctness defects and proves or retires candidates. debugging-investigator starts from a concrete failure and establishes the causal chain.
Known bug
repository-intelligence?
→ debugging-investigator
→ focused fix
→ unlazy?
→ verification-and-release?
Improve a codebase
repository-intelligence
→ codebase-improvement-planner
→ selected specialist
→ unlazy?
→ verification-and-release?
Build a feature
repository-intelligence?
→ owning web / mobile / evolution specialist
→ unlazy?
→ documentation-synchronizer?
→ verification-and-release?
Parallel work
repository-intelligence
→ multi-agent-work-coordinator
→ delegate-with-mission-cards
→ isolated reader/writer missions
→ integration
→ unlazy?
→ verification-and-release?
The router orchestrates. The specialist remains authoritative for its domain decisions, approvals, and stop conditions. unlazy cannot bypass a safety or approval gate, and verification-and-release remains the final owner of ship/no-ship judgment.
Skill catalog
Understand, investigate, and improve
| What you need | Skill |
|---|---|
| Map an unfamiliar codebase or determine change blast radius | repository-intelligence |
| Find important bugs you do not know about yet | bug-finder |
| Find the root cause of a known bug or regression | debugging-investigator |
| Decide what the codebase should improve next | codebase-improvement-planner |
| Review code or refactor it safely | review-and-refactor-code |
| Make a slow path faster using measurements | optimize-codebase-performance |
| Tighten TypeScript types and lint rules without hiding errors | typescript-quality-enforcer |
| Inspect or remove hidden provenance/metadata from files you own | content-provenance-hygiene |
Build and evolve
| What you need | Skill |
|---|---|
| Upgrade a dependency, framework, API, schema, or runtime safely | codebase-evolution-controller |
| Keep documentation synchronized with code changes | documentation-synchronizer |
| Define product UX and visual direction | product-design-director |
| Rebuild a UI from screenshots or visual references | screenshot-to-interface |
| Build or audit a production web feature/app | production-web-builder |
| Choose the right mobile stack | mobile-architecture-director |
| Build or audit a Flutter feature/app | flutter-production-builder |
| Build or audit an Expo/React Native feature/app | expo-react-native-builder |
Finish, verify, and coordinate
| What you need | Skill |
|---|---|
| Finish a substantial already-scoped task without premature “done” claims | unlazy |
| Decide whether an integrated change is safe to ship | verification-and-release |
| Split work across agents without write conflicts | multi-agent-work-coordinator |
| Send approved missions to the toolkit's custom agents | delegate-with-mission-cards |
Every skill remains independently installable. To list the collection:
npx skills add https://github.com/cmdr-chara/codex-toolkit --list
Install one skill globally for Codex:
npx skills add https://github.com/cmdr-chara/codex-toolkit --skill repository-intelligence -g -a codex
Selective npx skills add installs do not install the toolkit's global routing or automatic-update setup.
Mission Control
Mission Control provides six optional agents for work that is already understood well enough to split safely.
| Agent | Best for | Route |
|---|---|---|
| pathfinder-reader | Fast file, symbol, and fact lookup | Luna Max |
| patcher-writer | Small isolated edits | Luna Max |
| investigator-reader | Debugging, tracing, and focused reviews | Luna Max |
| builder-writer | Features, tests, fixes, docs, and configuration | Luna Max |
| sentinel-reader | Security, privacy, migrations, and other high-risk analysis | Sol High |
| architect-writer | Difficult architecture and failure-sensitive changes | Sol Max |
The coordinator owns decomposition, exclusive write scopes, dependency order, and integration. Mission Control chooses an appropriate agent for each approved mission. The parent Codex task still owns the integrated result.
The recommended setup command installs Mission Control automatically. To install or refresh only Mission Control:
npx --yes github:cmdr-chara/codex-toolkit
What setup adds
~/.codex/
├── AGENTS.md # user content + small managed routing block
├── skills/ # 20 installable skills
├── agents/ # 6 Mission Control agent configs
└── codex-toolkit/
├── workflows.md # conditional multi-skill workflows
├── state.json # installed release state
└── auto-update.* # short-lived updater support
Repository-local instructions still take precedence for repository-specific constraints.
Validation
The repository ships structural validation, routing/overlap evaluation cases, helper smoke tests, release metadata verification, installer tests, and auto-update tests.
python scripts/validate_skill_pack.py . --as-of 2026-08-17
python scripts/run_smoke_tests.py . --as-of 2026-08-17
CI additionally verifies:
- all 20 skills and 19 production routes;
- vendored anti-slop integrity and TypeScript regressions;
- package contents;
- isolated full-toolkit installation;
- preservation and idempotence of user-authored
AGENTS.mdcontent; - fail-closed behavior for malformed managed markers;
- release-pinned auto-update planning;
- canonical social-preview/release metadata.
See the evaluation guide for routing and workflow tests.
Repository layout
| Folder | Contents |
|---|---|
agents |
Six Mission Control agents |
docs |
Design decisions, boundaries, updater docs, and research sources |
evaluations |
Routing, overlap, workflow, and smoke-test cases |
orchestration |
Managed routing instructions and multi-skill workflow catalog |
scripts |
Installers, update runner, validation, and smoke tests |
skills |
Twenty installable skills |
Research and credit
Product design and screenshot reconstruction include adaptations from Leonxlnx's MIT-licensed Taste Skill project. THIRD_PARTY_NOTICES.md contains the source mapping and preserved notices.
unlazy adapts Leonxlnx's MIT-licensed completion-gate and Depth Tree method for Codex Toolkit's safety, approval, and specialist-handoff model. Its provenance and modification boundaries are recorded in skills/unlazy/references/upstream-provenance.md.
The TypeScript quality enforcer vendors the deterministic Oxlint runtime from Dillon Mulroy's MIT-licensed anti-slop project at a pinned upstream revision. Attribution and the upstream license are preserved in THIRD_PARTY_NOTICES.md.
Content provenance hygiene was designed after inspecting Guillaume Meyer's MIT-licensed watermarks-remover service and skill. Codex Toolkit does not vendor that runtime; the optional protocol reference is pinned in skills/content-provenance-hygiene/references/service-protocol.md.
Contributing
Bug reports, routing examples, and focused skill improvements are welcome. Read CONTRIBUTING.md before opening a pull request.
License
MIT Copyright 2026 cmdr-chara
No comments yet
Be the first to share your take.