Skip to content

Repository files navigation

Axiom

Release Distribution and publication guards License: MIT

Think before AI thinks.

Workflow guardrails for Codex and Claude Code.

Axiom is a safety-first workflow router for high-impact coding-agent actions. It loads one focused, inspectable workflow when scope, authorization, evidence, or rollback needs to be explicit, while ordinary requests continue through the host normally.

A route never grants mutation authority. Selecting one does not by itself permit an edit, commit, push, deployment, deletion, credential use, or external action. Axiom is a public beta, not a sandbox or a guarantee that an agent cannot make a mistake.

Start Safely

Request Expected boundary
"Perform a read-only audit of this repository's AGENTS.md instruction system. Report findings only; do not modify files." Select agents-architect, report evidence, and stop without changes
"Summarize the purpose of this README. Do not modify files." Select no Axiom route and continue normally

Install one host, inspect its installed Hook, and then try both requests. These are expected contracts, not claims that your host has already reproduced them.

Codex:

codex plugin marketplace add wheakerd/axiom
codex plugin add axiom@axiom

Claude Code:

/plugin marketplace add wheakerd/axiom
/plugin install axiom@axiom
/reload-plugins

Plugin Hooks execute commands in the host session. Before trusting Axiom, compare the installed /hooks entry with the exact declarations and wrapper in the Hook reference. Stop if they differ. Continue with Getting Started for first use, or use Managing an Installation for updates, removal, and non-destructive troubleshooting.

What Axiom Routes

Outcome Route Core boundary
Audit or maintain repository instructions agents-architect Inspect first; limit changes to the authorized instruction system
Design or audit packaged agent-plugin architecture agent-plugin-architect Require explicit package intent; keep ordinary plugin code outside
Reduce Codex usage overhead optimize-codex-usage Preserve required quality and safety; never invent hidden usage data
Review an Axiom-guided task review-axiom-task Review observable evidence independently and label unavailable facts
Confirm a consequential external action confirm-external-action Bind actor, target, payload, disclosure, count, and retry semantics
Make Git publication independently traceable traceable-git-submit Separate checkpoint, push, history, and cleanup authority
Plan or execute a persistent system change reversible-system-change Separate planning, rehearsal, activation, rollback, and retention

Machine-credential lifecycle work composes confirm-external-action and reversible-system-change when provider effects and persistent consumers are both in scope; it does not add another public route.

Ordinary coding, documentation, explanation, status, local commits, and conceptual requests continue through the host normally when no route clearly matches. A no-route result is not a safety certification: host permissions and repository instructions still apply. See Examples for more routed requests and controls.

Current Support Boundary

Codex and Claude Code have separate manifests and Hooks over the same shared Skill source. The checked-in package, route contracts, and validation fixtures are statically testable, but fresh-session behavior still depends on the exact host version, operating system, policy, installation method, and installed snapshot.

Host Checked-in support Current v0.10.0 observation boundary
Codex Manifest, marketplace wrapper, SessionStart Hook, and shared Skills Installed-host observation is NOT-RUN
Claude Code Manifest, marketplace wrapper, SessionStart Hook, and shared Skills Authenticated observation is UNAVAILABLE / NOT-RUN

The current release-status record remains STATIC-ONLY; static checks do not create host evidence. Read Compatibility for the bounded matrix and known limitations, Field Validation to report a result, and the v0.10.0 notes for version-specific detail. Historical observations remain under evidence/ and evals/results/ with their original identities and terminal statuses.

Runtime and repository identity

The installed package version, repository-policy revision, and deterministic runtime digest are separate identities. See Runtime and Repository Identity for the complete input and version policy.

Git Boundary

Ordinary named-remote, non-force staging, commits, and pushes stay host-native when they include neither a tag nor a traceable trigger. A combined commit, tag, and push of an already-prepared plugin release selects traceable-git-submit's hardened phase. The traceable triggers are an explicit $traceable-git-submit invocation, checkpoint, baseline, consolidation, recovery, multi-target, force, and history replacement. Merely mentioning submit, publish, or push does not select the route.

Route selection never authorizes Git mutation. Read the Trust Model and Architecture for the broader authority and execution boundaries.

Package Shape

Both hosts install the same checked-in skills/ source. Supporting references load on demand and are not separate routes.

Shared skills

  • using-axiom, the session-start routing gate.
  • agents-architect, the repository-instruction workflow.
  • agent-plugin-architect, the packaged agent-plugin architecture workflow.
  • optimize-codex-usage, the explicit Codex consumption workflow.
  • review-axiom-task, the read-only Axiom task-review workflow.
  • confirm-external-action, the consequential external-action workflow.
  • traceable-git-submit, the checkpoint and Git submission workflow.
  • reversible-system-change, the persistent-change workflow.

Platform wrappers

  • Codex uses .codex-plugin/plugin.json, .agents/plugins/marketplace.json, and hooks/codex-hooks.json.
  • Claude Code uses .claude-plugin/plugin.json, .claude-plugin/marketplace.json, and hooks/claude-hooks.json.

Both manifests point to ./skills/. Axiom installs no daemon, network service, watcher, background updater, or bundled runtime dependency.

Documentation and Support

Contributors should read CONTRIBUTING.md and the public Documentation Policy. Repository validators and CI checks are contributor tooling; they are not installed runtime dependencies or evidence that a host behavior was observed.

License

MIT

About

Safety-first workflow routing for Codex and Claude Code: explicit scope, authorization, evidence, and rollback for high-impact agent actions.

Topics

Resources

Contributing

Security policy

Stars

3 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages