[arshlibruh-code]

What

• Condensing runs on macOS 27's on-device fm (Apple Foundation Models), with none as the fallback. The Ollama provider, its model setting and doctor check are removed.
• Long replies, tables and turns with evidence go raw to the condenser; short ones are spoken as-is minus markdown.
• Dropped the word-budget "lead with the required action" instruction.
• New default condense prompt: at most 25 words, keeps numbers and names, never answers a question it relays, no filler.
• Spoken contract kept (trivial success to earcon, meta speech, false passes) and now also rejects "I am a coding agent". strip_markdown drops list bullets.

Why

• The log showed 52 of 52 Ollama condenses failing: the server was rarely running.
• On a 30-reply side-by-side of real replies, the old pipeline turned questions and finished work into spoken orders ("Delete the job board scanner" when the reply asked whether to). The raw-to-fm path kept them as questions and reports.

Checks

• ./smoke.sh: SMOKE PASS (one earlier run failed on voice inbox briefing with a curl timeout and broken pipe; the rerun passed).
• Unit tests: 28/28 OK.

Known

• prepare_speech used to raise KeyError when called without verification; the rewrite reads it with .get.
• build-packet.html still describes the Ollama setup; left as the dated v7 record.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LmQVErtihWNfyrQK91MaRw