Communitygithub.com

wheakerd/axiom

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

O que é axiom?

axiom is a Claude Code agent skill that safety-first workflow routing for Codex and Claude Code: explicit scope, authorization, evidence, and rollback for high-impact agent actions.

Funciona comClaude CodeCodex CLI~Cursor
npx skills add wheakerd/axiom

Perguntar na sua IA favorita

Abre um novo chat com esta habilidade de agente já pré-carregada.

Documentação

Using Axiom

Axiom is a routing gate. Select the smallest installed workflow set that applies without turning ordinary work into an Axiom task.

Route Once

  1. Honor higher-priority system, developer, user, and repository instructions.
  2. Select a route only when Axiom is explicitly invoked or the request clearly matches a bundled skill description.
  3. Load only the smallest necessary skill set, then only its active-phase references. Do not inspect candidate skill bodies before selection.
  4. Normalize unambiguous non-English wording to the canonical English route. Ask one concise question only when the route or permitted action would materially differ.
  5. On no match, continue through the host normally without mentioning Axiom. A no-match result is not a denial, does not create authorization, and does not manufacture a repository-state conflict.

Bundled Routes

  • agents-architect: create, audit, split, migrate, or maintain a target repository's AGENTS.md system, routed .agents/ guidance, or supporting repo-local skills; handle the explicit current-task modes effective-instructions, effective-instructions:preview, effective-instructions:refactor, and effective-instructions:force <instruction>; and reconcile existing guidance with current implementation only when the user explicitly invokes effective-instructions:reconcile, effective-instructions:reconcile-preview, or an explicit unambiguous request for that operation. Packaged plugin skills are outside this route.
  • agent-plugin-architect: design or audit packaged Codex or Claude Code plugin architecture across shared Skills, routes, manifests, wrappers, hooks, and compatibility evidence. Repo-local AGENTS systems and ordinary plugin code stay outside.
  • optimize-codex-usage: explicitly reduce or diagnose Codex credits, tokens, context, Skill/AGENTS/MCP loading, tool churn, or output overhead while preserving the required quality and safety bar.
  • review-axiom-task: review the routing, scope, authorization, actions, evidence, stops, and outcome of the current or an explicitly identified Axiom-guided task when the user explicitly requests that retrospective.
  • confirm-external-action: prepare, authorize, execute once, and verify an explicitly requested consequential external action such as send, publish, invite, purchase, trade, delete, or an external app/account change when its actor, target, payload, disclosure, cost, or retry boundary is material. Read-only lookup and draft-only work stay host-native.
  • traceable-git-submit: create traceable checkpoints or baseline metadata, consolidate or recover their history, or perform an explicitly invoked, hardened, multi-target, or otherwise independently traceable Git push. A combined commit, tag, and push of an already-prepared plugin release selects this route. Ordinary named-remote non-force staging, commits, and pushes without a tag, checkpoint, baseline, consolidation, recovery, hardening, multiple targets, or history replacement stay host-native; merely mentioning submit, publish, or push does not select this route.
  • reversible-system-change: plan, rehearse, or execute a persistent install, upgrade, deployment, migration, destructive retention, or promotion with rollback, data, service, or activation risk. Plans remain read-only.

Resolve cross-route ownership from this table before inspecting either candidate body. A deployment, promotion, migration, destructive retention, or similar persistent change that also causes a consequential external app or account effect, including publish, delete, or remote-state mutation, selects both confirm-external-action and reversible-system-change. Keep the exact external action envelope and the persistent write-set and rollback gates independent; authorization under either route never satisfies the other. Publication of an already-prepared artifact alone selects only confirm-external-action; publication alone is not a persistent system change.

When a request delegates a choice among mutually exclusive implementations and the alternatives would select materially different route sets, write surfaces, or authorization or safety boundaries, routing MUST NOT choose an alternative for the user. Select no route yet and ask exactly one concise clarification question. Wording such as "choose one" does not remove the ambiguity. Once the user chooses an unambiguous implementation, resume normal route selection.

An explicit usage-reduction goal selects optimize-codex-usage. Add another route only when the requested implementation also needs that route's distinct authorization or safety contract. Ordinary AGENTS audits select only agents-architect; ordinary performance work does not select usage optimization.

An explicit retrospective selects review-axiom-task; any implementation must be separately authorized and routed.

An external action selects confirm-external-action only when the user asks to cause the effect. Preparation does not authorize execution, and an exact current request need not be reconfirmed unless a material envelope field is missing or changes. Keep independently traceable Git workflows under traceable-git-submit; ordinary named-remote Git remains host-native.

For an ordinary combined commit-and-push request, an expected staged set that exactly matches the current authorized payload is normal host state, not a conflict. If the actor, repository, branch, named remote, command, payload, and non-force policy remain unchanged and inspection confirms that set, continue host-native without asking again. Stop before commit only on concrete current evidence of a material repository, branch, remote, command, payload, force, or instruction conflict that changes the frozen action or genuinely requires user input. Extra or unknown staged paths, a changed or missing target or branch, an in-progress Git operation, force, widened or multiple targets, an instruction conflict, or known divergence qualifies; mere possibility, ordinary staged state, stale tracking information, or no Axiom route does not.

A persistent change with no distinct consequential external effect stays under reversible-system-change; apply the cross-route rule above when both effects are present.

Boundaries

  • Routing selects instructions; it never authorizes edits, commits, pushes, deployments, deletion, credentials, remote writes, or scope expansion.
  • Startup routing is foreground and read-only. It must not write files, contact a network, start a service or background process, collect telemetry, or check for updates.
  • On resume or compaction, reselect every still-active route from current direct evidence before any new mutation. If route or phase cannot be reconstructed, perform zero new mutations; let each selected route's handoff contract resolve prior attempts.
  • Do not load every Axiom skill, route from broad topical similarity, edit protected plugin metadata without explicit scope, or persist one-off task discoveries as durable instructions.
  • Ordinary coding, documentation, explanation, status, local commits, named-remote non-force pushes, and conceptual requests continue normally unless a route description clearly matches.

Explicit Refresh

Only for an explicit Axiom update or refresh request, read references/updating.md. Never check, fetch, install, or announce an update automatically.

Habilidades Relacionadas