8 of 423 shipped agent definitions fail to start: the model: value is not a model id

Eight of the agent definitions in cli-tool/components/agents/ cannot run. A subagent using one of them dies before it emits a token, with the model-selection error, and the failure surfaces to whoever invoked it.

I measured all 423, and tested every distinct model: spelling against a live Claude Code rather than reasoning about them.

The eight

The first five are display names sitting in a field that takes a model id. The last three are different, and more interesting: claude-sonnet-4 is spelled correctly and used to work. That generation has since been retired, so those three definitions broke without anyone touching them.

The error, verbatim:

There's an issue with the selected model (Claude Sonnet 4.5).
It may not exist or you may not have access to it.

The subagent records model = <synthetic>, stop_reason = stop_sequence, and zero input, output and cache tokens.

What I ran

Claude Code 2.1.226, macOS arm64, main thread on claude-opus-5. For each distinct spelling in the repo I wrote a project-level definition containing only that model: value, invoked it once, and read the model back out of the transcript at ~/.claude/projects/**/subagents/**/agent-*.jsonl.

Two notes for anyone reproducing this:

• Read the version field in the transcript, not claude --version. On my machine those disagree by 74 patch releases, because --version was answering from a stale install on PATH rather than the process that actually ran.
• The subagent registry is snapshotted at session start. A definition file created inside a running session is not visible to it; you get Agent type '...' not found. Restart before testing.

The full census, for context

Suggested fixes, in the order I would do them

The eight. One line each. Claude Sonnet 4.5 → sonnet, and claude-sonnet-4 → sonnet. I can send a PR if that is useful, but the substitution is a judgement about intent and you may prefer to make it yourself.

Worth considering separately: alias over exact id. The three definitions that pin an exact model still work today, and the three that pinned claude-sonnet-4 are the same pattern one generation earlier. A pinned model has a shelf life. An alias does not, because tiers do not get retired. If a definition does not actually need a specific model, sonnet ages better than any id.

The 356 with no model: are not a bug, and I am not reporting them as one. They inherit, which may well be what you want. It is worth knowing that they inherit rather than defaulting to something modest: a user running Opus gets Opus for every one of them, including the ones that read a file and summarise it.

What I did not test

• The 7 definitions with no frontmatter. Whether they register as subagents at all is a different question from which model they route to, and needs a different probe.
• Any version other than 2.1.226. Subagent model behaviour has changed at least once already — the built-in Explore agent stopped being hardcoded to Haiku in v2.1.198 — so a result without a version number is not reproducible.
• claude-sonnet-4-5 still resolves, so please do not read this as "old ids are dead". Only the sonnet-4 generation is gone. The retirement boundary falls between those two, which is exactly why this needs measuring rather than assuming.