The Foundation Models seed in macOS 27 beta 4 (26A5388g) answers a plain
`respond(to:)` as if it were driving a tool-calling harness, so free-form
replies now arrive wrapped: a bare JSON object, a fenced ```json block, a
`tool_call: {…}` literal, `[start_x]`/`[No tools needed]` markers, or a
persona preamble ahead of the answer. Every consumer takes the reply's
first non-empty line, so the wrapper landed verbatim in chat titles and
collapsed summaries — dev even carries a commit named
"Nucleic: {task:Update Website Hero Text}". Guided generation
(@Generable) still binds its schema and is untouched.
Add ModelOutput to NucleicProtocol — the one layer the Mac app, core and
the iOS client all link — and run it at each provider's text choke point,
before the existing line-oriented parsing. Plain replies pass through
byte-identical; an all-scaffolding reply yields nil, which is the "no
result" every caller already handles by falling back to its heuristic, so
no new failure path. DelegatedIntelligence unwraps again on receipt: the
mesh is mixed-version and the agent-CLI backend never unwraps.
Also retune the two prompts the seed broke, measured against the live
on-device model rather than by inspection:
- classifyTurn: the old wording let the agentically-tuned seed reason
about what the agent should do NEXT, so it answered AWAITING for
plainly finished turns — 12/17 on a labeled set (3 samples, majority
vote) with 4 false AWAITING, which silently stops autoship. Reframing
it as "a classifier, NOT an assistant" scores 17/17 with none. Guided
generation was tried here and is worse (81% struct, 68% enum). Applied
to both hand-duplicated twins.
- sessionKeywords: 9 of 9 runs returned a JSON/tool-call envelope, one
inventing a fake Package.swift to "search". Anchoring extraction to
terms already in the input returns 9 of 9 clean comma lists.
Tests pin the captured beta 4 shapes verbatim and the classifier clauses
that earned the accuracy, since they read like boilerplate.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
abkslmandClaude Opus 4.8
ef1fc5af1f