Agent Catalog

This catalog lists all agents and their roles. Codex is the Tier 1, first-class runtime; normal child-model selection comes from the workspace profile while the provider remains inherited from the parent Codex session. Generated role profiles may override only the model, never the provider. Claude Code .claude/agents/*.md and VS Code Copilot .agent.md files are Tier 2 best-effort exports without a feature-parity guarantee. OpenCode support ended at the frozen OpenCode-first release, v0.26.1. Runtime-neutral source lives in agents/, protocols/, skills/, and tools/. Generated runtime outputs must not be hand-edited. Mode aliases and their orchestrator targets are defined in modes.json. $run-adaptive is intentionally a skill-only Simple/Flow/Pipeline router with route-independent execution presets, not an agent role or manifest-backed compatibility mode. Before managed role or Web session agent Job dispatch, read protocols/REASONING_POLICY.md and follow its profile/tier, shared-resolver, saved-configuration, dispatch-surface evidence, and result-label rules. Exported roles are leaf workers and must not dispatch another agent. Never claim native managed-role identity, workspace routing, a verified model, or effective effort without the evidence that protocol requires. Configured workspace profiles require health = ok, profile_eligibility = eligible, and catalog_state = current before new dispatch or resume. A pinned or retired catalog requires explicit workspace set --model-set openai or clear; it does not silently inherit the parent model. For an explicit request to dispatch a registered managed role outside a $run-* workflow, read the Ad-hoc managed-role dispatch section of protocols/REASONING_POLICY.md before spawning. Keep this path adaptive, pass tier unknown when the profile/runtime cannot prove it, and create no workflow artifacts; unverifiable or unhealthy profile status and resolver or selector conflicts stop dispatch. Use the smallest implementation and verification sufficient for the stated requirement. Rigor means proving the requested behavior, not adding abstractions, checks, or polish. Before resuming a run, editing after a failed check, or starting any repair, reviewer followup, capability recovery, or new Goal continuation round, read protocols/MATERIALITY_GATE.md. Apply its requirement-authority, failure-classification, validation-infrastructure, budget, stop, and resume/strategy-delta rules; only material work tied to an unmet original condition may continue. Before capability recovery or a later recovery-stage dispatch, also read protocols/CAPABILITY_RECOVERY.md. Keep its effort-first default, sole qualified LSA v2 exception, eligible-role and operational-failure limits, task-scoped target binding, and existing counter rules unchanged. Keep validation bounded to the requested delivery. Workflow-generated artifacts cannot expand scope, and validation infrastructure remains forbidden unless the original request or a pre-existing repository contract authorized it before dispatch. Before any authorized commit or push, including orchestrator helper actions, follow protocols/COMMIT_GUARD.md. Invoke its deterministic guard directly; hook presence does not replace the check. A passing check never grants commit/push permission, and an explicit no-commit instruction remains binding. Repository-specific Goal policy: do not create or activate a Goal for work in this repository unless the user's current message explicitly requests Goal mode. A run-* workflow request alone is not Goal consent. For the conceptual UI/UX layer, use ui-ux-designer; see protocols/UI_UX_WORKFLOW.md plus the ui-ux-bundle schema/example bundle at protocols/schemas/ui-ux-bundle.schema.json and protocols/examples/ui-ux-bundle.valid.json. The same surface also covers communication-first redesign and critique work via skills/ui-communication-designer/SKILL.md. For frontend implementation or polish after a conceptual handoff, use skills/frontend-aesthetic-director/SKILL.md; it preserves the upstream wireframe/flow and focuses on visual direction, tokens, responsive behavior, accessibility, and rendered QA.