Peter SteinbergerandGitHub 756a491a51 fix(speech): strip markdown before TTS and route code-heavy replies sensibly (#110948) ...

The talk.speak synthesis path passed raw agent text to providers, so talk
mode could read fenced code, links, and tables aloud. Speech normalization
now happens once at the shared synthesis boundary (speech-mode stripMarkdown:
label-only links, decorative bullet/glyph cleanup, punctuation collapsing).
Code-heavy replies (>=50% fenced content, CommonMark-aware scanner incl.
blockquote/list containers and unclosed fences) get routed by surface:
talk.speak speaks a short on-screen fallback line, channel auto-TTS skips
the voice note since the visible text already carries the content, and
explicit conversions (tts.convert, tagged hidden TTS, chat Listen) always
speak stripped best-effort content.

2026-07-18 18:16:43 -07:00

..

node_modules

feat(agents): add managed git worktree lifecycle (create/provision/snapshot/restore/GC) (#100535)

2026-07-06 05:24:58 +01:00

src

fix(speech): strip markdown before TTS and route code-heavy replies sensibly (#110948)

2026-07-18 18:16:43 -07:00

package.json

refactor(speech-core): remove redundant API back-edge

2026-07-13 22:11:48 -07:00

runtime-api.ts

refactor(speech-core): share TTS request pre-resolution across discord and voice-call (#109640)

2026-07-16 21:47:06 -07:00

speaker.ts

docs: document speech core facades

2026-06-04 01:22:31 -04:00

tsconfig.json

…

voice-models.ts

refactor(packages): remove dead internal helpers (#107077)

2026-07-14 11:48:14 +08:00