Agent Lifecycle Manager

Manage agent work as a lifecycle, not as isolated edits to agent.json, claude-code.md, or codex.toml. This skill is a thin router: it identifies which lifecycle stage matters now, runs only those stages, and keeps agent contract work, trigger tuning, install work, and governance concerns separate.

The design bias is:

• Use skill-lifecycle-manager discipline for routing, validation-before-claims, and explicit next steps
• Reuse skill-lifecycle-manager tooling for validation, projection, and eval harness work unless agent-specific behavior truly needs its own script path
• Use repo-local agent conventions for agents/<name>/, agent.json, claude-code.md, codex.toml, CHANGELOG.md, install targets, and generated catalogs
• Use manual-first posture for meta agent-management requests unless a narrower posture is clearly better

Outputs

Depending on the lifecycle stage, produce one or more of:

• A new or updated agent package (agent.json, claude-code.md, codex.toml, CHANGELOG.md) with an explicit invocation posture and agent-contract brief
• A validation report with concrete schema, structure, or cross-surface parity failures
• An evaluation plan or benchmark note with realistic prompts under workspaces/agent-lifecycle-manager/
• A Discover handoff brief capturing delegated research before authoring begins