feat(models): name delegated agents by model tier

[gsdevme]

Summary

Subagents launched without a name all render under their type label. A batch of parallel workers therefore shows up as several identical general-purpose rows with nothing to tell them apart, and none of them can be addressed with SendMessage.

This adds an Agent names section to models.md requiring a name on every Agent call, with the model tier in it (opus-agent-1, sonnet-agent-1, …) and numbering that runs monotonically across the session rather than restarting per batch, so a number is never reused.

The Implementation bullet now cross-references the new section, since that is where parallel agents get launched.

Why the name, and not a custom agent type

A custom type in ~/.claude/agents/ would change the type label itself, but it would need one definition file per distinguishable worker and still provide no numbering. It would also require apply.sh, the Makefile, and tests/apply.bats to grow an install path for agent definitions, which they do not have today. The per-call name gets the same result with a doctrine line and no tooling change.

Verification

• Behaviour confirmed against Claude Code before writing the doctrine: a call passing name: "opus-agent-1" is accepted and the agent registers and runs as opus-agent-1, displayed alongside its type rather than replacing it. Names must start with a letter or digit and may contain only letters, digits, underscores, and hyphens.
• make test — all 13 bats cases pass (apply.sh is untouched by this change).

Note for a follow-up, not addressed here

If CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 is ever enabled, a named subagent spawned from the main conversation launches as a teammate instead of a subagent. Agent teams are off by default and there is no env key in settings.json, so this doctrine is safe as written — but naming agents is the thing that would activate that coupling later.

🤖 Generated with Claude Code