Skip to content
MantleMantle docs

Release 0.223

0.223.3: An assistant that answers to its own name (branch feat/agent-name-token)

A copied assistant introduced itself as the one it was copied from. Give a login its own assistant called Tommy and his prompt still opened “You are Mira — a specialist assistant to a Risk-Based Inspection team”, because cloning copies the prompt verbatim and the name lived in the prose. Caught on a live box the day per-login assistants shipped.

The name and the prompt were always two separate columns with nothing keeping them in step, so the same bug was already there without any cloning: renaming an agent in Settings → Agents writes name and never touches system_prompt, so it kept introducing itself by the old one.

An assistant’s name is now a token in its prompt — {{name}} — resolved once per turn from the agent actually running. Rename it and the prompt follows; copy it and the copy is itself.

Three things had to agree for that to be true:

  • Resolution happens at the composition seam, the one function every surface routes through — real turns, delegated specialists, heartbeats, runner workers, the Studio sandbox, and Studio’s composed-prompt preview. Substituting any deeper would have let the model see a name the preview didn’t, which is the hidden prompt that seam exists to prevent. The name is a required argument, so a future call site cannot quietly omit it — which immediately earned itself: it caught a sixth call site (delegated specialists) that a search for the function had missed.
  • The persona bank stops baking names in. New assistants are name-agnostic from the start. The token is declared in two packages because the bank is a browser-safe leaf that the resolver depends on, so importing back would cycle; a tripwire test fails if the two literals ever drift, and caught a missing export the first time it ran.
  • Cloning rewrites the source’s name to the token, not to the new name. Baking “Tommy” in would recreate the bug the moment anyone renamed Tommy. Whole-word and case-sensitive, so an assistant called Max doesn’t turn “maximum” into a template.

A prompt that never mentions its own name is unchanged, byte for byte — the cached prefix every turn depends on is untouched until a brain opts in. The existing {{secret:service/label}} refs are a different mechanism resolved in the HTTP tool dispatcher, and are never matched: their syntax appears verbatim in the toolsmith skill’s own instructions, and a greedy matcher would have eaten the example it teaches from.