using-axiom

Invocation Notes

Optional, per user verbosity preference and host behavior. When explaining, name bundled skills as Axiom:<skill-name> and other skills, plugins, MCP tools or repo-local skills by their own name, all in inline code. Give a brief reason in the user's language; no fixed layout, repetition or notice for no-route or startup.

Route Once

1. Honor higher-priority instructions.
2. Route only explicit Axiom invocations or clear bundled-description matches.
3. Load the smallest matching installed skill set and active-phase references; inspect no candidate bodies before selection.
4. Map unambiguous non-English to its canonical English route. Clarify only material unresolved intent; tentative wording alone is no trigger.
5. No match: continue normally in the host. A no-match result is not a denial, does not create authorization, and does not manufacture a repository-state conflict.

Before web research from local or uncertain execution locations, including
cloud-search fallbacks, reassess and select local-web-search; routing stays
read-only.

Bundled Routes

• clarify-intent: resolve ambiguity with options and a custom answer; never reopen settled choices or ask routine details.
• delegate-simple-task: assess simple-task models only with this gate active and the host's verified top reasoning level. Before loading, require user confirmation or explicit invocation consent; else ask and wait. Missing gate or level evidence also blocks direct requests, even in permission-bypass mode.
• task-planning: create/revise current task/implementation plans and scope. Exclude scheduling, status, corrections and execution; keep specialized planning ownership.
• continuous-progress: explicit invocation only; supervise fresh workers, handoffs and compaction via verified native scheduling; exclude ordinary continuation and skill discussion.
• local-web-search: constrain planned web search and page access from the user's machine or local browser/client; classify an uncertain execution location first. Exclude confirmed cloud-hosted search and unrelated web app actions.
• agents-architect: audit repository AGENTS.md; handle explicit effective-instructions, effective-instructions:preview, effective-instructions:refactor, effective-instructions:force, effective-instructions:reconcile, and effective-instructions:reconcile-preview modes. Exclude packaged plugin skills.
• 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-agent-usage: explicitly reduce/diagnose agent usage of credits, tokens and context; preserve required quality and safety.
• review-axiom-task: review an identified Axiom task's observable routing, authorization and evidence. Prior refusal governs neither review nor narrower requests.
• verify-completion: verify completion or progress claims against current evidence after compaction, resume or long work; advise before complying on conflicting evidence. Exclude status questions and routine continuation.
• confirm-external-action: prepare, authorize, execute once and verify requested consequential external actions when actor, target, payload, disclosure, cost or retry boundaries matter; lookups and drafts stay host-native.
• traceable-git-submit: create traceable checkpoints/baseline metadata, consolidate/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/execute persistent installs, upgrades, deployments, migrations, destructive retention, or promotions with rollback, data, service, or activation risk. Includes backup preparation; plans are read-only.

Resolve cross-route ownership from this table before inspecting either
candidate body. Deployment, promotion, migration, destructive retention, or
similar persistent changes causing consequential external app/account effects,
including publish/delete/remote-state mutation, select both confirm-external-action and reversible-system-change. The exact
external action envelope and persistent write-set/rollback gates stay independent; authorization under either route never satisfies the other.
Publishing an already-prepared artifact alone selects only
confirm-external-action; it is not a persistent system change.

Explicit machine-credential lifecycle work composes the existing owners; read
references/credential-lifecycle.md. Inventory/planning/consumer activation/cleanup select reversible-system-change; provider creation/revocation/disclosure select confirm-external-action; end-to-end uses both.
Human login, conceptual help, secret reveal stay no-route. Keep each owner's
authority, write-set, rollback and verification gates.

When a request delegates a choice among mutually exclusive implementations
with materially different route sets, write surfaces, or authorization or safety boundaries, routing MUST NOT choose an alternative
for the user. Select only clarify-intent and ask exactly one concise
clarification question before selecting an action route. Wording such as
"choose one" does not remove the ambiguity. Once the
user chooses an unambiguous implementation, resume normal route selection.

Assess ambiguity before useful delegation; only this gate selects
clarify-intent then, never delegate-simple-task. A permission-bypass mode resolves no intent and expands no authority.

An explicit usage-reduction goal selects optimize-agent-usage; ordinary
performance work does not. Ordinary AGENTS audits select only agents-architect.

Exact existing authorization needs no reconfirmation unless scope changes. Prior refusal/assistant prose creates no policy. Ordinary named-remote Git remains host-native.

Boundaries

• Routing selects instructions, never edits, commits, pushes, deployments, deletion, credentials, remote writes or scope expansion.
• Startup routing is foreground and read-only: no writes, network, service, background process, telemetry or update check.
• On resume (including a fork) 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; use each selected route's handoff contract for prior attempts. A summary is a claim, not evidence.
• Never load all skills, route by topical similarity, edit protected metadata without scope or persist one-off discoveries as durable instructions.

Refresh

Read references/updating.md only on explicit Axiom update/refresh requests; never check, fetch, install or announce updates.