feat(channels): voice transcription hook in Chat SDK bridge

[mtichikawa]

Summary

Adds an opt-in voice transcription pass to the shared Chat SDK bridge. When WHISPER_BIN is set and an inbound attachment looks like audio, the bridge runs whisper.cpp on the buffer after fetchData() and appends the transcript to the message content as [Voice: <transcript>].

When WHISPER_BIN is unset (the default), the new code path is a no-op — zero behavior change for existing installs.

Why one bridge hook instead of per-adapter

The Chat SDK bridge is shared by Discord, Slack, Teams, Webex, Google Chat, and every future Chat SDK-supported platform. Hooking transcription here means all bridge-based channels gain voice support from a single integration point. The bridge already handles attachment download centrally (messageToInbound → att.fetchData()), so the transcription pass slots in naturally right after the buffer is in hand.

Sibling effort: @ira-at-work's #2317 (/add-voice-transcription-free-whisper) patches Signal/Telegram/WhatsApp adapters directly (they don't go through the Chat SDK bridge). That PR + this one would together cover every channel.

Demand signal: #2426 (just opened by @b1ek) — "LLM cant see the image in discord" — voice transcription is the same kind of gap.

Files

• src/transcription.ts (new, ~95 lines): transcribeAudioBuffer(Buffer): Promise<string | null> plus isAudioAttachment(att) helper. Channel-agnostic. Shells out to ffmpeg for input normalization (any container → 16 kHz mono WAV) and whisper-cli for the transcript. Returns null on any failure or empty output; the bridge only injects [Voice: ...] when the transcript is non-empty.
• src/channels/chat-sdk-bridge.ts (+15 lines): gated transcription pass inside messageToInbound(). Runs only when process.env.WHISPER_BIN is set and the attachment passes isAudioAttachment (mimeType: audio/* or coarse type: 'audio'/'voice').
• src/transcription.test.ts (new, 8 tests): isAudioAttachment truth table; transcribeAudioBuffer env-gate, trim, empty-output, and execFile-failure paths.

Total: 3 files, +236 lines.

Pairs with

A sibling PR against main adds the user-facing skill metadata: .claude/skills/add-discord-voice-transcription/{SKILL.md, REMOVE.md, VERIFY.md} following the @ddaniels Signal template (#1953). It instructs users to fetch the modified chat-sdk-bridge.ts and the new transcription.ts from origin/channels after this PR lands.

Test plan

• pnpm run build — clean (3 pre-existing errors in deltachat/slack/telegram unrelated to this change)
• npx vitest run src/transcription.test.ts — 8/8 passing
• npx vitest run src/channels/chat-sdk-bridge.test.ts — 7/7 passing (unchanged)
• Live test: send a Discord voice memo with WHISPER_BIN=whisper-cli set, confirm the agent receives [Voice: <transcript>] inline. Will run this once the channels-branch checkout is up.
• Live test: same with WHISPER_BIN unset, confirm voice attachments arrive as plain audio placeholders (no regression).

🤖 Generated with Claude Code

[mtichikawa]

Closing — wrong target. After re-reading CONTRIBUTING.md, feature-skill PRs should be a single PR against main containing both the SKILL.md trio and the source code, with maintainers extracting the code to a skill/<name> branch on merge. Also src/channels/chat-sdk-bridge.ts exists on main (not just channels), so the modification belongs there.

Folded both changes into #2459 against main — that's now a single self-contained feature-skill PR following the template you used for @ddaniels' merged Signal PR (#1953).