@bojackduy/opencode-voice 0.11.1 → 0.11.3

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -18,13 +18,13 @@ One package runs on both OpenCode 1.x and 2.x. The plugin detects the host
18
18
  by capability (never by version string) and adapts; the v1 path is unchanged.
19
19
 
20
20
  > [!IMPORTANT]
21
- > **Submit-only mode on v2.** OpenCode v2.0.26 exposes no plugin API for
22
- > appending text to the composer for editing, so the v1 STT loop of
23
- > transcribe → append → edit → submit is unimplementable there. On v2 every
24
- > STT turn **transcribes and submits immediately** via `session.prompt` —
25
- > both `/stt-record` and `/stt-submit` submit — and the turn toast says
26
- > submitted, never "added to prompt". A warning toast on first load says the
27
- > same. There is deliberately no edit step; nothing is faked.
21
+ > **Dictation inserts; it does not send.** On both v1 and v2 the transcript
22
+ > is typed at the cursor of whichever editable field currently has focus —
23
+ > the session composer, a dialog/search input, or another plugin's text
24
+ > field — leaving the text editable. Nothing is submitted unless you asked
25
+ > for it (`/stt-submit`, `sttAutoSubmit`, or conversation mode). If no
26
+ > editable field is focused, v2 keeps the transcript in a recovery dialog
27
+ > instead of sending it as a prompt.
28
28
 
29
29
  ### v1 (OpenCode 1.x)
30
30
 
@@ -84,21 +84,26 @@ capability detection already picks the right path per host.
84
84
  every one reliably; the `Ctrl+P` palette shows them for matching queries
85
85
  but its search occasionally reports no results for queries that should
86
86
  match — when in doubt, use `/slash`.
87
- - **STT (batch):** submits immediately, no edit step (see above). Recording
88
- from home creates a session only when final text is ready, opens it, and
89
- submits once. The selected model is carried into the new session; existing
90
- sessions retain their agent/model. On home, v2's public plugin API exposes
91
- the selected model but no selected-agent getter, so session creation uses
92
- the host's default agent. Create/submit failures show the full transcript in
93
- a recovery dialog. Cancelling before submission drops the captured text.
87
+ - **STT (batch):** the transcript is typed at the cursor of the focused
88
+ editable field and nothing is sent unless submission is requested.
89
+ On home the composer is focused, so dictation works before a session
90
+ exists; an explicit submit creates the session, carries the selected
91
+ model into it, and sends once. Existing sessions retain their
92
+ agent/model. On home, v2's public plugin API exposes the selected model
93
+ but no selected-agent getter, so session creation uses the host's
94
+ default agent. Create/submit failures show the full transcript in a
95
+ recovery dialog. With no editable field focused, the transcript is kept
96
+ in that dialog instead of being sent as a prompt.
94
97
  - **STT (streaming dictation):** enable with `"sttMode": "streaming"` in plugin
95
98
  options. Local whisper streaming displays a bounded live transcript above
96
- the prompt footer on home and session screens. Stable words are plain;
99
+ the prompt footer on home and session screens, with the live input waveform
100
+ beside it. Stable words are plain;
97
101
  the tentative tail is bracketed and may change. Press the record key again
98
- to finalize and submit once, or cancel to clear the preview without sending.
99
- No composer edit step is promised on v2. Batch remains the default and
100
- produces text after recording stops; the hosted `sttEndpoint` option is for
101
- batch transcription, not this local streaming preview.
102
+ to finalize: the full text is inserted once at the focused cursor (and
103
+ submitted when `sttAutoSubmit` is set), or cancel to clear the preview
104
+ without sending. Batch remains the default and produces text after
105
+ recording stops; the hosted `sttEndpoint` option is for batch
106
+ transcription, not this local streaming preview.
102
107
  - **TTS / conversation / live notes / cancel:** same semantics as v1 (auto
103
108
  TTS on idle/permission/question gates, conversation barge-in, live-notes
104
109
  background saves, one-key cancel).
@@ -747,7 +752,7 @@ Idle invocations toast `Nothing to cancel`. Keybind override:
747
752
 
748
753
  Live local dictation for short prompts: the mic stays on, partial
749
754
  transcriptions stream in as you speak, and one press of `/stt-record`
750
- finalizes the text into the prompt box on v1, or submits it once on v2.
755
+ finalizes the text at the focused cursor on both hosts.
751
756
  Set `"sttMode": "streaming"` in the
752
757
  plugin options (default `"batch"` keeps the record-then-transcribe flow
753
758
  exactly as before).
@@ -763,19 +768,20 @@ there is never a silent gap — `Loading speech model…` during cold start,
763
768
  `● Streaming dictation — listening` with an elapsed timer while live
764
769
  (plus a `(listening…)` note when inference runs slow), and `Finalizing…`
765
770
  between the second keypress and finalization. On v2, a bounded transcript
766
- above the prompt footer updates during recording: stable words are plain,
767
- tentative words are bracketed. Finalization submits directly to the active
768
- session (or creates one from home); cancel clears the preview without sending.
769
- On v1, the OpenCode renderer only
771
+ above the prompt footer updates during recording, with the live input
772
+ waveform beside it: stable words are plain, tentative words are bracketed.
773
+ Finalization inserts the full text once at the focused cursor; cancel clears
774
+ the preview without sending.
775
+ The OpenCode renderer only
770
776
  exposes `insertText`/`submit`, with no live range-replacement API, so
771
777
  partials render as preview toasts (`🎙 ...`) alongside (not instead of)
772
778
  the sticky status, and the full text is inserted EXACTLY ONCE at
773
- finalize. If no editable field is focused, the transcript is kept via
774
- toast instead of being redirected into the chat. `/stt-submit` submits
775
- only through the dictated field's own `submit()` — a field that cannot
776
- submit keeps its text un-submitted rather than falling through to the
777
- primary chat. `/stt-stop` cancels and removes only plugin-dictated
778
- text, never your typed prefix/suffix.
779
+ finalize. If no editable field is focused, the transcript is kept visibly
780
+ (v2 recovery dialog, v1 toast) instead of being redirected into the chat.
781
+ `/stt-submit` submits only through the dictated field's own `submit()` — a
782
+ field that cannot submit keeps its text un-submitted rather than falling
783
+ through to the primary chat. `/stt-stop` cancels and removes only
784
+ plugin-dictated text, never your typed prefix/suffix.
779
785
 
780
786
  Warm server: the whisper model stays loaded across dictations (finalize
781
787
  and cancel keep the lease); it reloads only on model/language change,
@@ -796,9 +802,8 @@ Options:
796
802
  - `sttMode` _(optional, default `"batch"`)_ - `"streaming"` selects this mode
797
803
  - `sttStreamWindowMs` _(optional, default `10000`)_ - rolling audio window
798
804
  - `sttStreamStepMs` _(optional, default `1000`)_ - live transcription cadence
799
- - `sttAutoSubmit` applies after streaming finalize on v1 (submit via the
800
- dictated field only, as above). v2 always submits final text because there
801
- is no composer append/edit API.
805
+ - `sttAutoSubmit` applies after streaming finalize on both hosts (submit via
806
+ the dictated field only, as above).
802
807
 
803
808
  Measured latencies (Apple M3 Pro, 36 GB, whisper.cpp 1.9.4 with Metal,
804
809
  `ggml-large-v3-turbo-q5_0.bin`, scripted speech through the real
@@ -839,10 +844,11 @@ real streaming controller in `test/v2-preview.test.js`.
839
844
  3. LLM normalizes the transcription: fixes punctuation, removes filler words,
840
845
  corrects software engineering homophones ("Jason" to "JSON", "bullion" to
841
846
  "boolean", etc.)
842
- 4. Cleaned text is appended to the OpenCode prompt, or submitted immediately
843
- when `/stt-submit` is used. If normalization fails (e.g. LLM endpoint
844
- unreachable), the raw transcription is used as a fallback so you never lose
845
- your input
847
+ 4. Cleaned text is inserted at the cursor of the focused editable field
848
+ (composer, search, or plugin input), or submitted immediately when
849
+ `/stt-submit` (or `sttAutoSubmit`) is used. If normalization fails (e.g.
850
+ LLM endpoint unreachable), the raw transcription is used as a fallback
851
+ so you never lose your input
846
852
 
847
853
  ### TTS pipeline
848
854
 
package/index.js CHANGED
@@ -1,9 +1,9 @@
1
1
  // opencode-voice: Speech-to-text and text-to-speech for OpenCode.
2
2
  //
3
3
  // STT: Record voice via sox, transcribe with whisper-cpp, normalize with
4
- // an OpenAI-compatible LLM, append to the TUI prompt (v1) or submit
5
- // immediately via session.prompt (v2 submit-only mode, see README
6
- // "OpenCode v1 and v2").
4
+ // an OpenAI-compatible LLM, and insert the text at the cursor of the
5
+ // focused input (composer, dialog search, or plugin field) on both
6
+ // hosts. Explicit submits only: /stt-submit, sttAutoSubmit, conversation.
7
7
  //
8
8
  // TTS: Auto-speak assistant responses (or read on demand) via Piper,
9
9
  // with LLM normalization for natural speech.
@@ -26,7 +26,7 @@
26
26
  // /stt-mic - select microphone
27
27
  // /tts-speak - read last response aloud (default: <leader>] = ctrl+x, ])
28
28
  // /tts-mode - toggle auto TTS on/off (palette-only)
29
- // /tts-stop - stop playback (default: <leader>; , also palette)
29
+ // /tts-stop - stop playback (palette-only; global <leader>. already stops TTS)
30
30
  // /tts-voice - select TTS voice
31
31
  // /voice-conversation - toggle hands-free voice conversation (default: <leader>v = ctrl+x, v)
32
32
  // /voice-conversation-stop - exit voice conversation mode (palette-only)
@@ -46,7 +46,8 @@ import { adaptV2Context } from "./lib/v2-compat.js";
46
46
  // Dual-compat export: v1's loader reads `tui`, v2's reads `setup`
47
47
  // (`Plugin.define` is an identity wrapper, so a plain object carries both).
48
48
  // Each host ignores the other's key. The core (`bootVoice`) is shared; v2
49
- // runs it over the v2-compat adapted api in submit-only mode.
49
+ // runs it over the v2-compat adapted api, which forwards the live renderer
50
+ // so dictation lands in whatever input currently has focus.
50
51
  export default {
51
52
  id: "opencode-voice",
52
53
  tui: async (api, options) => {
@@ -55,16 +56,16 @@ export default {
55
56
  setup: async (ctx) => {
56
57
  const adapted = adaptV2Context(ctx);
57
58
  await bootVoice(adapted.api, ctx.options, {
58
- submitOnly: true,
59
+ v2: true,
59
60
  onReady: async ({ logger }) => {
60
- // First-run submit-only notice (required UX disclosure): v2 has no
61
- // append-to-composer API, so every STT turn submits immediately.
61
+ // First-run v2 notice (required UX disclosure): text is typed into
62
+ // the focused field and is never sent automatically.
62
63
  const notice =
63
- "Voice on OpenCode v2: transcripts submit immediately as-is " +
64
- "(submit-only mode — append-for-edit is unavailable on v2). " +
65
- "For live preview, set sttMode to streaming.";
64
+ "Voice on OpenCode v2: transcripts are inserted at the cursor of the focused text field " +
65
+ "(prompt, search, or dialog input) and are not sent automatically. " +
66
+ "Set sttAutoSubmit to submit after recording.";
66
67
  adapted.api.ui.toast({ message: notice, variant: "warning", duration: 8000 });
67
- logger?.log?.("plugin", `v2 submit-only mode: ${notice}`, "warn");
68
+ logger?.log?.("plugin", `v2 focused-input mode: ${notice}`, "warn");
68
69
  // v2 has no client.app.log, so the logger above is a no-op: compat
69
70
  // warnings must surface as a toast or they are lost entirely.
70
71
  if (adapted.warnMessages.length > 0) {
package/lib/boot.js CHANGED
@@ -4,9 +4,9 @@
4
4
  // an `api` object (raw v1 `TuiPluginApi` vs the v2-compat adapted `Context`).
5
5
  // Everything below runs identically on both, so there is exactly one core to
6
6
  // maintain. The v1 `tui` path calls `bootVoice(api, options)` with no extras
7
- // and behaves exactly as before; the v2 `setup` path passes
8
- // `{ submitOnly: true }` for the first-run notice. No STT/TTS/routing/
9
- // keybind behavior branches live here.
7
+ // and behaves exactly as before; the v2 `setup` path passes `{ v2: true }`
8
+ // for the first-run notice. No STT/TTS/routing/keybind behavior branches
9
+ // live here.
10
10
 
11
11
  import fs from "node:fs";
12
12
  import os from "node:os";
@@ -98,7 +98,7 @@ export async function bootVoice(api, options, extras = {}) {
98
98
  ...cancel.commands,
99
99
  ]);
100
100
 
101
- if (extras?.submitOnly && typeof extras?.onReady === "function") {
101
+ if (extras?.v2 && typeof extras?.onReady === "function") {
102
102
  await extras.onReady({ logger, stt, tts, conversation, liveNotes, voiceModel, cancel });
103
103
  }
104
104
 
@@ -373,7 +373,7 @@ export function registerConversation(api, opts, logger, { stt, tts, isLiveNotesA
373
373
  // v2 prompt dispatch may emit the entire reply before its promise returns.
374
374
  // Home creation navigates before prompting, so resolve the ID at event time.
375
375
  let earlyWait = null;
376
- if (api.client?.tui?.submitOnly) {
376
+ if (api.client?.tui?.v2) {
377
377
  phase = "waiting";
378
378
  startReplyStream(currentSessionID, turnStart);
379
379
  earlyWait = waitForReply(currentSessionID);
package/lib/llm-client.js CHANGED
@@ -102,7 +102,9 @@ export function resolveProviderConfig(provider, model) {
102
102
  const exported = envVars.find((name) => name && process.env[name]);
103
103
  return {
104
104
  endpoint,
105
- model: model.id,
105
+ // v2 keeps the upstream provider's model id in `modelID` (`id` is the
106
+ // selection id and can differ for configured aliases), so prefer it.
107
+ model: model.modelID ?? model.id,
106
108
  apiKeyEnv: exported || envVars[0] || null,
107
109
  };
108
110
  }
@@ -436,6 +436,16 @@ export function createSoxRollingCapture({
436
436
  });
437
437
  }
438
438
 
439
+ // Cheap tail read for live level meters: copies only the requested bytes
440
+ // (no WAV wrap, no snapshot sequence bookkeeping). Returns empty when the
441
+ // capture has no audio yet.
442
+ function peekTail(nBytes = 2048) {
443
+ const want = Math.max(0, Math.min(Math.floor(Number(nBytes) || 0), capacityBytes));
444
+ if (want === 0) return null;
445
+ const pcm = ring.snapshotLast(want);
446
+ return pcm.length > 0 ? pcm : null;
447
+ }
448
+
439
449
  function snapshot() {
440
450
  const pcm = ring.snapshotLast(Math.ceil((windowMs / 1000) * STREAMING_DEFAULTS.bytesPerSecond));
441
451
  if (pcm.length === 0) return null;
@@ -551,6 +561,7 @@ export function createSoxRollingCapture({
551
561
  return {
552
562
  start,
553
563
  snapshot,
564
+ peekTail,
554
565
  stopFinal,
555
566
  cancel,
556
567
  dispose,
@@ -1316,6 +1327,15 @@ export function createStreamingController({
1316
1327
  getState,
1317
1328
  getServerPort,
1318
1329
  getTranscript: currentTranscript,
1330
+ // Live level meter source: newest PCM tail (16-bit mono). Null when the
1331
+ // capture is gone or has no audio yet; never throws.
1332
+ peekPcm: (nBytes) => {
1333
+ try {
1334
+ return capture?.peekTail?.(nBytes) || null;
1335
+ } catch {
1336
+ return null;
1337
+ }
1338
+ },
1319
1339
  getFallbackAudio: () => fallbackAudio,
1320
1340
  getFallbackCoverage: () => fallbackCoverage,
1321
1341
  };
package/lib/stt-log.js ADDED
@@ -0,0 +1,39 @@
1
+ // Local file log for STT/streaming diagnosis (user-requested).
2
+ //
3
+ // Why this exists: v2 toasts are one-shot and the slot preview's fate is
4
+ // invisible from the UI. This log records which branch ran (batch vs
5
+ // streaming), what the preview did, and toast emissions, so the next
6
+ // recording answers "what happened" without guessing.
7
+ //
8
+ // Privacy: local /tmp only, never shipped. Free text (partials, messages)
9
+ // is truncated to 120 chars; options/endpoints/keys are never logged.
10
+ // Disable with OPENCODE_VOICE_DEBUG_LOG=0. Rotates at ~512KB.
11
+ import fs from "node:fs";
12
+
13
+ const LOG_PATH = "/tmp/opencode-voice-stt.log";
14
+ const MAX_BYTES = 512 * 1024;
15
+ const MAX_TEXT = 120;
16
+
17
+ function enabled() {
18
+ return process.env.OPENCODE_VOICE_DEBUG_LOG !== "0";
19
+ }
20
+
21
+ export function trunc(value) {
22
+ const s = String(value ?? "");
23
+ return s.length > MAX_TEXT ? `${s.slice(0, MAX_TEXT)}…` : s;
24
+ }
25
+
26
+ export function sttLog(event, details = "") {
27
+ if (!enabled()) return;
28
+ try {
29
+ let header = "";
30
+ try {
31
+ if (fs.statSync(LOG_PATH).size > MAX_BYTES) {
32
+ header = `--- rotated ${new Date().toISOString()} ---\n`;
33
+ fs.writeFileSync(LOG_PATH, header);
34
+ }
35
+ } catch {}
36
+ const line = `${new Date().toISOString()} pid=${process.pid} ${event}${details ? ` ${trunc(details)}` : ""}\n`;
37
+ fs.appendFileSync(LOG_PATH, header + line);
38
+ } catch {}
39
+ }