Agent Plugin Architect

Design or audit the smallest coherent packaged-plugin control plane for Codex
and Claude Code. Keep each host's experience native and preserve a single
shared Skill tree when both hosts can discover it without weakening either host.

Route The Work

1. For an unresolved choice between mutually exclusive workflows with material
   differences in route ownership, write surfaces, persistent metadata or
   authority, ask one clear question and keep the workflow unselected until the
   user chooses. Restate each option's original end goal and explain permission
   limits separately. Preparation, review, planning or preflight may support a
   chosen goal, but cannot replace it without the user's choice. Read-only
   authority, lower risk, lighter work or current feasibility cannot justify
   defaulting to one option, proceeding with it, or requiring the user to object
   to stop.
   Explicit architecture requests continue below; explicit installation stays
   with its owner, explaining unavailable authority without substituting an
   architecture audit. Conceptual comparisons and choices without material
   boundary differences do not require this gate. After selection, do not
   reconfirm unchanged conditions; ask only for necessary missing input or a
   material conflict.

   Confirm that the selected request explicitly concerns packaged agent-plugin
   architecture. Repo-local AGENTS.md, .agents/skills, or .claude/skills work belongs to
   agents-architect; ordinary plugin code or documentation has no Axiom
   route solely because it lives in a plugin repository.

2. Inventory the package before editing. Read
   references/package-inventory.md.

3. For a read-only audit of one packaged release candidate, read
   references/release-readiness.md. It classifies impact, version and runtime
   identity, local and remote evidence, and the one next decision, then stops
   before every mutation phase.

4. Define route ownership and trigger boundaries with
   references/route-and-trigger-contracts.md.

5. For shared Skill structure and direct discovery, read
   references/packaged-skill-architecture.md.

6. For startup hooks, untrusted content, credentials, or runtime effects, read
   references/hooks-and-trust-boundaries.md.

7. For manifests, wrappers, and Codex/Claude Code parity, read
   references/cross-host-packaging.md.

8. For routing cases, context cost, or host observations, read
   references/evaluation-and-evidence.md.

9. Before completion, read references/validation-reporting.md.

Load only references needed for the active phase. A request may compose this
route with at most one other Axiom route when both contracts are independently
necessary. An explicit usage-cost goal may add optimize-agent-usage;
retrospective review may add review-axiom-task. Git submission, installation,
publication, deployment, and consequential external effects remain separate
active phases and select their owning route instead of inheriting authority
from this Skill.

Deliverable

Return the package inventory, route and trust decisions, changed surfaces,
validation evidence, compatibility class, and every host or lifecycle check
that is NOT-RUN or UNAVAILABLE. File presence and static parity are not
host execution evidence.