@mindstudio-ai/agent 0.1.186 → 0.1.187

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/CHANGELOG.md CHANGED
@@ -4,6 +4,19 @@ What changed in `@mindstudio-ai/agent`, newest first, each version under its own
4
4
 
5
5
  Versions before 0.1.163 aren't recorded here.
6
6
 
7
+ ## 0.1.187
8
+
9
+ ### New
10
+
11
+ - `evaluateDecision` can run on Drex Max Finance, Nace.AI's finance tuning of Drex Max, served by DeepInfra: `decisionModelOverride.model` set to `drex-max-finance`. It takes up to 64 questions, where the other Drex models take 16, and reads up to 4 PNG, JPEG, WebP or GIF images, 9 MiB together. The usage notes say so. Check: nothing.
12
+ - `transcribeAudio` can run on ElevenLabs' Scribe v3: `transcriptionModelOverride.model` set to `elevenlabs-scribe-v3`, with the options Scribe v2 takes (speakers, word timings, key terms, redaction). The usage notes list it. Check: nothing.
13
+
14
+ ### Changed behavior
15
+
16
+ - `transcribeAudio`'s default model is `elevenlabs-scribe-v3`, where it was `elevenlabs-scribe-v2`, at the same price and with the same options. Check: a call that depends on Scribe v2's exact wording or timings names `elevenlabs-scribe-v2` in `transcriptionModelOverride.model`.
17
+
18
+ - `renderVideoFromHtmlComposition` renders with HyperFrames 0.9.19, and its usage notes name that version, where they named 0.8.217. Among its fixes: a render matches the preview when a `set` interrupts a running tween, where it came out blank; an image that loads late is in the frame; a video with no source no longer holds capture for 45 seconds; and one loud moment turns down only its own peaks, not the whole mix. HyperFrames 0.9.0's breaking changes are in its own cloud rendering, which the platform doesn't use, and a composition's format didn't change. App code isn't affected.
19
+
7
20
  ## 0.1.186
8
21
 
9
22
  ### Fixed
package/dist/cli.js CHANGED
@@ -306,7 +306,7 @@ var init_metadata = __esm({
306
306
  "evaluateDecision": {
307
307
  stepType: "evaluateDecision",
308
308
  description: "Ask structured questions about text, JSON, or images and get calibrated answers from a decision model. The model returns a probability, a selected option, or a score. It does not write a reply. Each question is a noul (yes or no, as a probability from 0 to 1), a choice (one option from a set you define, with a probability for each), or a score (a rating against ordered levels). Put every question in the same call. The question id is yours: answers come back under the same ids. The model defaults to Jev (jev-latest).",
309
- usageNotes: "- state is the content to judge: a string, or a JSON object or array (a record, a chat log, application state).\n- questions is a map. The key is an id you choose. It is not shown to the model. The answer comes back under the same id.\n- A noul asks a yes/no question and returns noul, a probability from 0 (no) to 1 (yes). Optional criteria.true and criteria.false say what a yes and a no mean.\n- A choice picks one option. criteria is a map of option id to a short description (or null). At most 255 options. The answer includes the winning choice, a probability for every option, and confidence.\n- A score rates the state on an ordered rubric. criteria is 2 to 10 level descriptions, low to high. The answer's score can land between levels, and includes a probability per level plus confidence.\n- An answer of type refusal means the model declined that question (GPT-6 Luna Decisions can).\n- instructions may be a string, or an object that keeps the question in one field and reference data in the others. Refer to those fields by name in backticks.\n- Send every question you need in one call. The model reads the state once and answers them together.\n- Probabilities from different models are not comparable: tune a threshold against the model, and the version, it will run on.\n- decisionModelOverride.model picks the model:\n - jev-latest (the default) is TypeSafe's current stable release, text only, up to 64k tokens. jev-preview tracks TypeSafe's newest release. Pin a version such as jev-1.13.0 when a threshold was tuned against that version.\n - clef and clef-flash run Cloudflare Clef, which also reads up to 4 PNG, JPEG, or WebP images, each at most 4 MiB and 8 MiB together, and takes at most 64 questions, each id using only letters, digits, underscore, dot, and hyphen, up to 100 characters. clef is the larger model, clef-flash the faster one. @cf/cloudflare/clef and @cf/cloudflare/clef-flash name the same two.\n - mercury-decide runs Inception's Mercury Decide, which reads up to 64k tokens and up to 16 PNG, JPEG, WebP, or GIF images, each at most 10 MiB and 14 MiB together.\n - drex-max, drex-v2.0 and drex-v1.5 run Nace.AI's Drex, served by DeepInfra, and take at most 16 questions. drex-max reads up to 262k tokens and up to 4 PNG, JPEG, WebP, or GIF images, 9 MiB together. drex-v2.0 and drex-v1.5 read text only, up to 128k tokens. nace-ai/drex-max, nace-ai/drex-v2.0 and nace-ai/drex-v1.5 name the same three.\n - gpt-6-luna-decisions runs OpenAI's Decisions API (in beta), which reads up to 922k tokens and PNG, JPEG, WebP, or GIF images of up to 20 MB each. It takes up to 200 questions, and a choice needs at least 2 options. gpt-6-luna names it too.\n - gemini-decision-1 runs Google's Gemini Decision 1, which reads up to 262k tokens and PNG, JPEG, WebP, or GIF images, up to 100 MB of image data in one request. It takes 1 to 1,000 questions. A question id uses only letters, digits, underscore, and hyphen, up to 64 characters, and a choice needs 2 to 255 options. A yes/no probability is between 0.01 and 0.99.\n- images is for a model that reads them. Pass URLs or data URLs. A model that reads none refuses the call.\n- The response model field is the version that answered.\n- Billed per input token. Output tokens are free. Each model has its own rate, and GPT-6 Luna Decisions bills a prompt over 272k tokens at twice its rate.",
309
+ usageNotes: "- state is the content to judge: a string, or a JSON object or array (a record, a chat log, application state).\n- questions is a map. The key is an id you choose. It is not shown to the model. The answer comes back under the same id.\n- A noul asks a yes/no question and returns noul, a probability from 0 (no) to 1 (yes). Optional criteria.true and criteria.false say what a yes and a no mean.\n- A choice picks one option. criteria is a map of option id to a short description (or null). At most 255 options. The answer includes the winning choice, a probability for every option, and confidence.\n- A score rates the state on an ordered rubric. criteria is 2 to 10 level descriptions, low to high. The answer's score can land between levels, and includes a probability per level plus confidence.\n- An answer of type refusal means the model declined that question (GPT-6 Luna Decisions can).\n- instructions may be a string, or an object that keeps the question in one field and reference data in the others. Refer to those fields by name in backticks.\n- Send every question you need in one call. The model reads the state once and answers them together.\n- Probabilities from different models are not comparable: tune a threshold against the model, and the version, it will run on.\n- decisionModelOverride.model picks the model:\n - jev-latest (the default) is TypeSafe's current stable release, text only, up to 64k tokens. jev-preview tracks TypeSafe's newest release. Pin a version such as jev-1.13.0 when a threshold was tuned against that version.\n - clef and clef-flash run Cloudflare Clef, which also reads up to 4 PNG, JPEG, or WebP images, each at most 4 MiB and 8 MiB together, and takes at most 64 questions, each id using only letters, digits, underscore, dot, and hyphen, up to 100 characters. clef is the larger model, clef-flash the faster one. @cf/cloudflare/clef and @cf/cloudflare/clef-flash name the same two.\n - mercury-decide runs Inception's Mercury Decide, which reads up to 64k tokens and up to 16 PNG, JPEG, WebP, or GIF images, each at most 10 MiB and 14 MiB together.\n - drex-max, drex-max-finance, drex-v2.0 and drex-v1.5 run Nace.AI's Drex, served by DeepInfra. drex-max-finance takes up to 64 questions, and the others at most 16. drex-max and drex-max-finance, its finance tuning, read up to 262k tokens and up to 4 PNG, JPEG, WebP, or GIF images, 9 MiB together. drex-v2.0 and drex-v1.5 read text only, up to 128k tokens. nace-ai/drex-max, nace-ai/drex-max-finance, nace-ai/drex-v2.0 and nace-ai/drex-v1.5 name the same four.\n - gpt-6-luna-decisions runs OpenAI's Decisions API (in beta), which reads up to 922k tokens and PNG, JPEG, WebP, or GIF images of up to 20 MB each. It takes up to 200 questions, and a choice needs at least 2 options. gpt-6-luna names it too.\n - gemini-decision-1 runs Google's Gemini Decision 1, which reads up to 262k tokens and PNG, JPEG, WebP, or GIF images, up to 100 MB of image data in one request. It takes 1 to 1,000 questions. A question id uses only letters, digits, underscore, and hyphen, up to 64 characters, and a choice needs 2 to 255 options. A yes/no probability is between 0.01 and 0.99.\n- images is for a model that reads them. Pass URLs or data URLs. A model that reads none refuses the call.\n- The response model field is the version that answered.\n- Billed per input token. Output tokens are free. Each model has its own rate, and GPT-6 Luna Decisions bills a prompt over 272k tokens at twice its rate.",
310
310
  inputSchema: { "type": "object", "properties": { "state": { "anyOf": [{ "type": "string" }, { "type": "object", "properties": {}, "required": [] }, { "type": "array", "items": {} }] }, "questions": { "type": "object", "properties": {}, "required": [], "description": "Questions to ask about the state, keyed by an id you choose." }, "images": { "type": "array", "items": { "type": "string" }, "description": "Images for a model that reads them (Clef, Clef Flash, GPT-6 Luna Decisions, Gemini Decision 1). Each entry is a URL or a data URL. A model that reads none refuses the call." }, "decisionModelOverride": { "type": "object", "properties": { "model": { "type": "string", "description": "Decision model: a catalog id, its vendor's name for it (`@cf/cloudflare/clef`, `gpt-6-luna`), or a version it pins (`jev-1.13.0`)" }, "config": { "type": "object", "properties": {}, "required": [], "description": "Additional model-specific configuration" } }, "required": ["model"], "description": "Optional model override. Uses the workspace's default decision model (Jev) if not specified" } }, "required": ["state", "questions"] },
311
311
  outputSchema: { "type": "object", "properties": { "model": { "type": "string", "description": "Versioned model id that answered, such as jev-1.13.0." }, "answers": { "type": "object", "properties": {}, "required": [], "description": "One answer per question, under the same ids." }, "usage": { "type": "object", "properties": { "input_tokens": { "type": "number" }, "output_tokens": { "type": "number" } }, "required": ["input_tokens", "output_tokens"], "description": "Token counts for the call. Input tokens are what the model bills." } }, "required": ["model", "answers", "usage"] }
312
312
  },
@@ -832,7 +832,7 @@ var init_metadata = __esm({
832
832
  "renderVideoFromHtmlComposition": {
833
833
  stepType: "renderVideoFromHtmlComposition",
834
834
  description: "Render an HTML composition (HTML, CSS, media and seekable animation on a timeline) to a video file",
835
- usageNotes: "- Renders an HTML composition in HyperFrames' format (https://hyperframes.heygen.com) with HyperFrames 0.8.217, in Chrome off the app's own sandbox; author for that version\n- Pass the composition as `html`, the page itself (up to 10 MB), or as `project`, the URL of a zip of a project's folder (`index.html` at its root, or the file `entryFile` names) or of one HTML file, up to 2 GiB: an app file or any public URL\n- `variables` overrides the composition's `data-composition-variables` defaults, which it reads with `window.__hyperframes.getVariables()`\n- Whatever the composition loads itself (fonts, scripts, images, video) is fetched as it renders, so each is in the project or at a public URL; a private app file needs a signed URL\n- fps 24, 30 or 60 (default 30); quality draft, standard or high (default standard); format mp4 (default), webm or mov (both keep transparency), or gif\n- An mp4, webm or mov renders in chunks of about 10 seconds side by side, so a long video takes little longer than a short one; a call finishes within about 30 minutes, and reports its progress as it goes. A gif renders whole, within 15 minutes\n- A composition HyperFrames can't render fails with HyperFrames' own reason",
835
+ usageNotes: "- Renders an HTML composition in HyperFrames' format (https://hyperframes.heygen.com) with HyperFrames 0.9.19, in Chrome off the app's own sandbox; author for that version\n- Pass the composition as `html`, the page itself (up to 10 MB), or as `project`, the URL of a zip of a project's folder (`index.html` at its root, or the file `entryFile` names) or of one HTML file, up to 2 GiB: an app file or any public URL\n- `variables` overrides the composition's `data-composition-variables` defaults, which it reads with `window.__hyperframes.getVariables()`\n- Whatever the composition loads itself (fonts, scripts, images, video) is fetched as it renders, so each is in the project or at a public URL; a private app file needs a signed URL\n- fps 24, 30 or 60 (default 30); quality draft, standard or high (default standard); format mp4 (default), webm or mov (both keep transparency), or gif\n- An mp4, webm or mov renders in chunks of about 10 seconds side by side, so a long video takes little longer than a short one; a call finishes within about 30 minutes, and reports its progress as it goes. A gif renders whole, within 15 minutes\n- A composition HyperFrames can't render fails with HyperFrames' own reason",
836
836
  inputSchema: { "type": "object", "properties": { "html": { "type": "string", "description": "The composition's HTML, up to 10 MB; give this or project" }, "project": { "type": "string", "description": "URL of a zip of the project's folder, or of one self-contained HTML file; give this or html" }, "entryFile": { "type": "string", "description": "The entry HTML inside the zip, relative to its root; index.html when omitted" }, "variables": { "type": "object", "properties": {}, "required": [], "description": "Values over the composition's declared variables" }, "fps": { "enum": [24, 30, 60], "type": "number", "description": "Frames per second: 24, 30 (default) or 60" }, "quality": { "enum": ["draft", "standard", "high"], "type": "string", "description": "Encode quality: draft, standard (default) or high" }, "format": { "enum": ["mp4", "webm", "mov", "gif"], "type": "string", "description": "mp4 (default), webm or mov (both keep transparency), or gif" }, "intermediateAsset": { "type": "boolean", "description": "When true, the asset is created but hidden from the user's gallery (tagged as intermediate)" } } },
837
837
  outputSchema: { "type": "object", "properties": { "videoUrl": { "type": "string", "description": "URL of the rendered video" }, "durationSec": { "type": "number", "description": "Its length in seconds" }, "frames": { "type": "number", "description": "How many frames it has" } }, "required": ["videoUrl", "durationSec", "frames"] }
838
838
  },
@@ -1081,7 +1081,7 @@ var init_metadata = __esm({
1081
1081
  "transcribeAudio": {
1082
1082
  stepType: "transcribeAudio",
1083
1083
  description: "Convert an audio file to text using a transcription model.",
1084
- usageNotes: "- The default model is elevenlabs-scribe-v2: speakers, word timings, key terms and redaction, at any length.\n- The prompt field gives context (speaker names, domain) to OpenAI's and DeepInfra's models. Other models ignore it; give them `keyterms` instead.\n- Options go in `transcriptionModelOverride.config`, and support varies by model:\n - `language` \u2014 the spoken language as a code such as `en` or `es`; left out, the model detects it. Every model takes it. deepgram-nova-3 assumes `en` unless told, and also takes `multi` for speech that switches language (billed at its multilingual rate).\n - `detectLanguage: 'yes'` \u2014 detect the language instead of assuming English (deepgram-nova-3 only).\n - `includeSpeakers: 'yes'` \u2014 label each passage with a speaker. Supported by elevenlabs-scribe-v2, elevenlabs-scribe-v2-medical, deepgram-nova-3, grok-stt, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3 and gemini-3.5-transcribe; gpt-4o-transcribe-diarize always does. Scribe and Voxtral also switch `text` to a timestamped, speaker-labelled block format; Deepgram, Gemini and gpt-4o-transcribe-diarize to `Speaker N:` lines.\n - `numSpeakers` \u2014 the most people talking, to help tell them apart (Scribe).\n - `keyterms` \u2014 a list of names, products and terms the audio contains, so they're spelled right. Supported by gpt-transcribe, Scribe, deepgram-nova-3, grok-stt, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3 and gemini-3.5-transcribe. Billed as an add-on on Scribe and Deepgram.\n - `redactPii: 'yes'` \u2014 replace names, phone numbers, addresses and other personal information with a label of what was there (Scribe, deepgram-nova-3). Billed as an add-on.\n - `includeWordTimings: 'yes'` \u2014 add per-word timings and confidence inside each segment. Supported by Scribe, deepgram-nova-3, whisper-1, grok-stt, voxtral-mini-transcribe-2 and gemini-3.5-transcribe. Off by default because an hour of speech is roughly ten thousand words.\n - `subtitles: 'srt' | 'vtt'` \u2014 also return the transcript as a subtitle file, in the `subtitles` output, timed by the segments. Any model that returns segments.\n- gemini-3.5-transcribe can't take `keyterms` together with speakers, word timings or subtitles; asked for both, it's refused.\n- The `segments` output carries timings, confidence and speakers as structured data, so there is no need to slice the audio up to recover them. It is populated by Scribe, deepgram-nova-3, grok-stt, whisper-1, gpt-4o-transcribe-diarize, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3, gemini-3.5-transcribe (with word timings or subtitles) and the DeepInfra-served whisper and qwen models, and absent for models that return nothing but text (gpt-transcribe, gpt-4o-transcribe, gpt-4o-mini-transcribe, the DeepInfra-served voxtral models).\n- In a workflow, `destinationVar` receives the transcript text. Name `segmentsVar` to also keep the `segments` array. A model that returns no segments leaves that variable unset.\n- A recording of any length can be transcribed. One longer than a model takes at once (OpenAI's and DeepInfra's models, Gemini, Voxtral 3) is transcribed in pieces and joined, its timings running from the start of the recording. Speakers can't be followed from one piece to the next, so with speakers on, such a recording is refused; Scribe and Deepgram label speakers at any length.\n- `confidence` is always 0..1 with higher meaning more confident, whichever model produced it.",
1084
+ usageNotes: "- The default model is elevenlabs-scribe-v3: speakers, word timings, key terms and redaction, at any length.\n- The prompt field gives context (speaker names, domain) to OpenAI's and DeepInfra's models. Other models ignore it; give them `keyterms` instead.\n- Options go in `transcriptionModelOverride.config`, and support varies by model:\n - `language` \u2014 the spoken language as a code such as `en` or `es`; left out, the model detects it. Every model takes it. deepgram-nova-3 assumes `en` unless told, and also takes `multi` for speech that switches language (billed at its multilingual rate).\n - `detectLanguage: 'yes'` \u2014 detect the language instead of assuming English (deepgram-nova-3 only).\n - `includeSpeakers: 'yes'` \u2014 label each passage with a speaker. Supported by elevenlabs-scribe-v2, elevenlabs-scribe-v2-medical, elevenlabs-scribe-v3, deepgram-nova-3, grok-stt, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3 and gemini-3.5-transcribe; gpt-4o-transcribe-diarize always does. Scribe and Voxtral also switch `text` to a timestamped, speaker-labelled block format; Deepgram, Gemini and gpt-4o-transcribe-diarize to `Speaker N:` lines.\n - `numSpeakers` \u2014 the most people talking, to help tell them apart (Scribe).\n - `keyterms` \u2014 a list of names, products and terms the audio contains, so they're spelled right. Supported by gpt-transcribe, Scribe, deepgram-nova-3, grok-stt, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3 and gemini-3.5-transcribe. Billed as an add-on on Scribe and Deepgram.\n - `redactPii: 'yes'` \u2014 replace names, phone numbers, addresses and other personal information with a label of what was there (Scribe, deepgram-nova-3). Billed as an add-on.\n - `includeWordTimings: 'yes'` \u2014 add per-word timings and confidence inside each segment. Supported by Scribe, deepgram-nova-3, whisper-1, grok-stt, voxtral-mini-transcribe-2 and gemini-3.5-transcribe. Off by default because an hour of speech is roughly ten thousand words.\n - `subtitles: 'srt' | 'vtt'` \u2014 also return the transcript as a subtitle file, in the `subtitles` output, timed by the segments. Any model that returns segments.\n- gemini-3.5-transcribe can't take `keyterms` together with speakers, word timings or subtitles; asked for both, it's refused.\n- The `segments` output carries timings, confidence and speakers as structured data, so there is no need to slice the audio up to recover them. It is populated by Scribe, deepgram-nova-3, grok-stt, whisper-1, gpt-4o-transcribe-diarize, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3, gemini-3.5-transcribe (with word timings or subtitles) and the DeepInfra-served whisper and qwen models, and absent for models that return nothing but text (gpt-transcribe, gpt-4o-transcribe, gpt-4o-mini-transcribe, the DeepInfra-served voxtral models).\n- In a workflow, `destinationVar` receives the transcript text. Name `segmentsVar` to also keep the `segments` array. A model that returns no segments leaves that variable unset.\n- A recording of any length can be transcribed. One longer than a model takes at once (OpenAI's and DeepInfra's models, Gemini, Voxtral 3) is transcribed in pieces and joined, its timings running from the start of the recording. Speakers can't be followed from one piece to the next, so with speakers on, such a recording is refused; Scribe and Deepgram label speakers at any length.\n- `confidence` is always 0..1 with higher meaning more confident, whichever model produced it.",
1085
1085
  inputSchema: { "type": "object", "properties": { "audioUrl": { "type": "string", "description": "URL of the audio file to transcribe" }, "prompt": { "type": "string", "description": "Optional context to improve transcription accuracy (e.g. language, speaker names, domain terms)" }, "transcriptionModelOverride": { "type": "object", "properties": { "model": { "type": "string", "description": "Audio transcription model identifier" }, "config": { "type": "object", "properties": {}, "required": [], "description": "Additional model-specific configuration" } }, "required": ["model"], "description": "Optional model configuration override. Uses the workflow's default transcription model if not specified" } }, "required": ["audioUrl", "prompt"] },
1086
1086
  outputSchema: { "type": "object", "properties": { "text": { "type": "string", "description": "The transcribed text from the audio file" }, "segments": { "type": "array", "items": { "type": "object", "properties": { "start": { "type": "number", "description": "Seconds from the start of the audio." }, "end": { "type": "number" }, "text": { "type": "string", "description": "The passage, spacing normalized." }, "confidence": { "type": "number", "description": "How sure the model was, always `0..1` with higher meaning more confident \u2014 whatever the provider reported.\n\nNormalizing this is the whole point of the type. Providers disagree on both the scale and the direction: Deepgram reports a probability in `0..1`, ElevenLabs and Whisper report a log probability in `(-inf, 0]`. Handing those through untouched would make `confidence < 0.5` mean opposite things depending on which model the caller happened to pick. Log probabilities are converted with `Math.exp`; probabilities pass through." }, "speaker": { "type": "string", "description": "Present only when diarization was asked for and the provider supports it." }, "language": { "type": "string", "description": "Present only for providers that report language per passage rather than per request. Most report it once for the whole audio \u2014 see `TranscriptionResponse.language`." }, "words": { "type": "array", "items": { "type": "object", "properties": { "text": { "type": "string" }, "start": { "type": "number", "description": "Seconds from the start of the audio." }, "end": { "type": "number" }, "confidence": { "type": "number", "description": "See the confidence convention on `TranscriptSegment`." }, "speaker": { "type": "string", "description": "Provider-assigned speaker id, present only when diarization was asked for." } }, "required": ["text"], "description": "One recognized token.\n\n`text` rather than `word` because an entry is not always a word \u2014 providers emit punctuation, spacing and non-speech audio events (laughter, footsteps) through the same channel." }, "description": "Word-level detail, when the caller asked for it (`includeWordTimings`). Off by default: an hour of speech is roughly ten thousand words, an order of magnitude more payload than the segments that contain them." } }, "required": ["start", "end", "text"], "description": "A contiguous passage of one speaker's speech, bounded in time.\n\nThe unit a review UI works in: short enough to seek to and highlight, coarse enough that an hour of audio is a few hundred entries rather than ten thousand." }, "description": "Timed passages with per-segment confidence, speaker attribution and optional word-level detail \u2014 see the step's usage notes for which models populate it. Absent for models that return nothing but text." }, "language": { "type": "string", "description": "Detected or declared language of the audio, when the model reports it." }, "durationSeconds": { "type": "number", "description": "Length of the transcribed audio, when the model reports it." }, "subtitles": { "type": "string", "description": "The transcript as an SRT or WebVTT file, when `subtitles` was asked for in the model's config and the model returned segments to time it by." } }, "required": ["text"] }
1087
1087
  },
@@ -10271,7 +10271,7 @@ async function startMcpServer(options) {
10271
10271
  capabilities: { tools: {} },
10272
10272
  serverInfo: {
10273
10273
  name: "mindstudio-agent",
10274
- version: "0.1.186"
10274
+ version: "0.1.187"
10275
10275
  },
10276
10276
  instructions: 'Welcome to MindStudio \u2014 a platform with 200+ AI models, 850+ third-party integrations, and pre-built agents.\n\nGetting started:\n1. Call `ask` with any question about the SDK \u2014 it knows every action, model, and connector and returns working code with real model IDs and config options. Examples: ask("generate an image with FLUX"), ask("what models support vision?"), ask("how do I send a Slack message?").\n2. Call `changeName` to set your display name \u2014 use your name or whatever your user calls you. This is how you\'ll appear in MindStudio request logs.\n3. If you have a profile picture, pass its public image URL to `changeProfilePicture`.\n4. For manual browsing, call `listActions` to discover all available actions.\n\nThen use the tools to generate text, images, video, audio, search the web, work with data sources, run agents, and more.\n\nImportant:\n- AI-powered actions (text generation, image generation, video, audio, etc.) cost money. Before running these, call `estimateActionCost` and confirm with the user before proceeding \u2014 unless they\'ve explicitly told you to go ahead.\n- Not all agents from `listAgents` are configured for API use. Do not try to run an agent just because it appears in the list \u2014 it will likely fail. Only run agents the user specifically asks you to run.'
10277
10277
  });
@@ -11311,7 +11311,7 @@ function isNewerVersion(current, latest) {
11311
11311
  return false;
11312
11312
  }
11313
11313
  async function checkForUpdate() {
11314
- const currentVersion = "0.1.186";
11314
+ const currentVersion = "0.1.187";
11315
11315
  if (!currentVersion || process.env.MINDSTUDIO_NO_UPDATE_CHECK) {
11316
11316
  return null;
11317
11317
  }
@@ -11346,7 +11346,7 @@ async function checkForUpdate() {
11346
11346
  }
11347
11347
  }
11348
11348
  function printUpdateNotice(latestVersion) {
11349
- const currentVersion = "0.1.186";
11349
+ const currentVersion = "0.1.187";
11350
11350
  process.stderr.write(
11351
11351
  `
11352
11352
  ${ansi2.cyanBright("Update available")} ${ansi2.gray(currentVersion + " \u2192")} ${ansi2.cyanBold(latestVersion)}
@@ -11359,7 +11359,7 @@ function isStandaloneBinary() {
11359
11359
  return !argv1.includes("node_modules");
11360
11360
  }
11361
11361
  async function cmdUpdate() {
11362
- const currentVersion = "0.1.186";
11362
+ const currentVersion = "0.1.187";
11363
11363
  process.stderr.write(
11364
11364
  ` ${ansi2.gray("Current version:")} ${currentVersion}
11365
11365
  `
@@ -11500,7 +11500,7 @@ async function cmdLogin(options) {
11500
11500
  process.stderr.write("\n");
11501
11501
  printLogo();
11502
11502
  process.stderr.write("\n");
11503
- const ver = "0.1.186";
11503
+ const ver = "0.1.187";
11504
11504
  process.stderr.write(
11505
11505
  ` ${ansi2.bold("MindStudio Agent")} ${ver ? " " + ansi2.gray("v" + ver) : ""}
11506
11506
  `
@@ -11830,7 +11830,7 @@ async function main() {
11830
11830
  const updatePromise = command !== "mcp" && command !== "login" && command !== "update" ? checkForUpdate() : Promise.resolve(null);
11831
11831
  try {
11832
11832
  if (command === "version" || command === "-v") {
11833
- process.stdout.write("0.1.186\n");
11833
+ process.stdout.write("0.1.187\n");
11834
11834
  return;
11835
11835
  }
11836
11836
  if (command === "login") {
@@ -642,7 +642,7 @@ export interface StepMethods {
642
642
  * - jev-latest (the default) is TypeSafe's current stable release, text only, up to 64k tokens. jev-preview tracks TypeSafe's newest release. Pin a version such as jev-1.13.0 when a threshold was tuned against that version.
643
643
  * - clef and clef-flash run Cloudflare Clef, which also reads up to 4 PNG, JPEG, or WebP images, each at most 4 MiB and 8 MiB together, and takes at most 64 questions, each id using only letters, digits, underscore, dot, and hyphen, up to 100 characters. clef is the larger model, clef-flash the faster one. @cf/cloudflare/clef and @cf/cloudflare/clef-flash name the same two.
644
644
  * - mercury-decide runs Inception's Mercury Decide, which reads up to 64k tokens and up to 16 PNG, JPEG, WebP, or GIF images, each at most 10 MiB and 14 MiB together.
645
- * - drex-max, drex-v2.0 and drex-v1.5 run Nace.AI's Drex, served by DeepInfra, and take at most 16 questions. drex-max reads up to 262k tokens and up to 4 PNG, JPEG, WebP, or GIF images, 9 MiB together. drex-v2.0 and drex-v1.5 read text only, up to 128k tokens. nace-ai/drex-max, nace-ai/drex-v2.0 and nace-ai/drex-v1.5 name the same three.
645
+ * - drex-max, drex-max-finance, drex-v2.0 and drex-v1.5 run Nace.AI's Drex, served by DeepInfra. drex-max-finance takes up to 64 questions, and the others at most 16. drex-max and drex-max-finance, its finance tuning, read up to 262k tokens and up to 4 PNG, JPEG, WebP, or GIF images, 9 MiB together. drex-v2.0 and drex-v1.5 read text only, up to 128k tokens. nace-ai/drex-max, nace-ai/drex-max-finance, nace-ai/drex-v2.0 and nace-ai/drex-v1.5 name the same four.
646
646
  * - gpt-6-luna-decisions runs OpenAI's Decisions API (in beta), which reads up to 922k tokens and PNG, JPEG, WebP, or GIF images of up to 20 MB each. It takes up to 200 questions, and a choice needs at least 2 options. gpt-6-luna names it too.
647
647
  * - gemini-decision-1 runs Google's Gemini Decision 1, which reads up to 262k tokens and PNG, JPEG, WebP, or GIF images, up to 100 MB of image data in one request. It takes 1 to 1,000 questions. A question id uses only letters, digits, underscore, and hyphen, up to 64 characters, and a choice needs 2 to 255 options. A yes/no probability is between 0.01 and 0.99.
648
648
  * - images is for a model that reads them. Pass URLs or data URLs. A model that reads none refuses the call.
@@ -1817,7 +1817,7 @@ export interface StepMethods {
1817
1817
  * Render an HTML composition (HTML, CSS, media and seekable animation on a timeline) to a video file
1818
1818
  *
1819
1819
  * @remarks
1820
- * - Renders an HTML composition in HyperFrames' format (https://hyperframes.heygen.com) with HyperFrames 0.8.217, in Chrome off the app's own sandbox; author for that version
1820
+ * - Renders an HTML composition in HyperFrames' format (https://hyperframes.heygen.com) with HyperFrames 0.9.19, in Chrome off the app's own sandbox; author for that version
1821
1821
  * - Pass the composition as `html`, the page itself (up to 10 MB), or as `project`, the URL of a zip of a project's folder (`index.html` at its root, or the file `entryFile` names) or of one HTML file, up to 2 GiB: an app file or any public URL
1822
1822
  * - `variables` overrides the composition's `data-composition-variables` defaults, which it reads with `window.__hyperframes.getVariables()`
1823
1823
  * - Whatever the composition loads itself (fonts, scripts, images, video) is fetched as it renders, so each is in the project or at a public URL; a private app file needs a signed URL
@@ -2412,12 +2412,12 @@ export interface StepMethods {
2412
2412
  * Convert an audio file to text using a transcription model.
2413
2413
  *
2414
2414
  * @remarks
2415
- * - The default model is elevenlabs-scribe-v2: speakers, word timings, key terms and redaction, at any length.
2415
+ * - The default model is elevenlabs-scribe-v3: speakers, word timings, key terms and redaction, at any length.
2416
2416
  * - The prompt field gives context (speaker names, domain) to OpenAI's and DeepInfra's models. Other models ignore it; give them `keyterms` instead.
2417
2417
  * - Options go in `transcriptionModelOverride.config`, and support varies by model:
2418
2418
  * - `language` — the spoken language as a code such as `en` or `es`; left out, the model detects it. Every model takes it. deepgram-nova-3 assumes `en` unless told, and also takes `multi` for speech that switches language (billed at its multilingual rate).
2419
2419
  * - `detectLanguage: 'yes'` — detect the language instead of assuming English (deepgram-nova-3 only).
2420
- * - `includeSpeakers: 'yes'` — label each passage with a speaker. Supported by elevenlabs-scribe-v2, elevenlabs-scribe-v2-medical, deepgram-nova-3, grok-stt, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3 and gemini-3.5-transcribe; gpt-4o-transcribe-diarize always does. Scribe and Voxtral also switch `text` to a timestamped, speaker-labelled block format; Deepgram, Gemini and gpt-4o-transcribe-diarize to `Speaker N:` lines.
2420
+ * - `includeSpeakers: 'yes'` — label each passage with a speaker. Supported by elevenlabs-scribe-v2, elevenlabs-scribe-v2-medical, elevenlabs-scribe-v3, deepgram-nova-3, grok-stt, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3 and gemini-3.5-transcribe; gpt-4o-transcribe-diarize always does. Scribe and Voxtral also switch `text` to a timestamped, speaker-labelled block format; Deepgram, Gemini and gpt-4o-transcribe-diarize to `Speaker N:` lines.
2421
2421
  * - `numSpeakers` — the most people talking, to help tell them apart (Scribe).
2422
2422
  * - `keyterms` — a list of names, products and terms the audio contains, so they're spelled right. Supported by gpt-transcribe, Scribe, deepgram-nova-3, grok-stt, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3 and gemini-3.5-transcribe. Billed as an add-on on Scribe and Deepgram.
2423
2423
  * - `redactPii: 'yes'` — replace names, phone numbers, addresses and other personal information with a label of what was there (Scribe, deepgram-nova-3). Billed as an add-on.
package/dist/index.js CHANGED
@@ -5278,7 +5278,7 @@ var stepMetadata = {
5278
5278
  "evaluateDecision": {
5279
5279
  stepType: "evaluateDecision",
5280
5280
  description: "Ask structured questions about text, JSON, or images and get calibrated answers from a decision model. The model returns a probability, a selected option, or a score. It does not write a reply. Each question is a noul (yes or no, as a probability from 0 to 1), a choice (one option from a set you define, with a probability for each), or a score (a rating against ordered levels). Put every question in the same call. The question id is yours: answers come back under the same ids. The model defaults to Jev (jev-latest).",
5281
- usageNotes: "- state is the content to judge: a string, or a JSON object or array (a record, a chat log, application state).\n- questions is a map. The key is an id you choose. It is not shown to the model. The answer comes back under the same id.\n- A noul asks a yes/no question and returns noul, a probability from 0 (no) to 1 (yes). Optional criteria.true and criteria.false say what a yes and a no mean.\n- A choice picks one option. criteria is a map of option id to a short description (or null). At most 255 options. The answer includes the winning choice, a probability for every option, and confidence.\n- A score rates the state on an ordered rubric. criteria is 2 to 10 level descriptions, low to high. The answer's score can land between levels, and includes a probability per level plus confidence.\n- An answer of type refusal means the model declined that question (GPT-6 Luna Decisions can).\n- instructions may be a string, or an object that keeps the question in one field and reference data in the others. Refer to those fields by name in backticks.\n- Send every question you need in one call. The model reads the state once and answers them together.\n- Probabilities from different models are not comparable: tune a threshold against the model, and the version, it will run on.\n- decisionModelOverride.model picks the model:\n - jev-latest (the default) is TypeSafe's current stable release, text only, up to 64k tokens. jev-preview tracks TypeSafe's newest release. Pin a version such as jev-1.13.0 when a threshold was tuned against that version.\n - clef and clef-flash run Cloudflare Clef, which also reads up to 4 PNG, JPEG, or WebP images, each at most 4 MiB and 8 MiB together, and takes at most 64 questions, each id using only letters, digits, underscore, dot, and hyphen, up to 100 characters. clef is the larger model, clef-flash the faster one. @cf/cloudflare/clef and @cf/cloudflare/clef-flash name the same two.\n - mercury-decide runs Inception's Mercury Decide, which reads up to 64k tokens and up to 16 PNG, JPEG, WebP, or GIF images, each at most 10 MiB and 14 MiB together.\n - drex-max, drex-v2.0 and drex-v1.5 run Nace.AI's Drex, served by DeepInfra, and take at most 16 questions. drex-max reads up to 262k tokens and up to 4 PNG, JPEG, WebP, or GIF images, 9 MiB together. drex-v2.0 and drex-v1.5 read text only, up to 128k tokens. nace-ai/drex-max, nace-ai/drex-v2.0 and nace-ai/drex-v1.5 name the same three.\n - gpt-6-luna-decisions runs OpenAI's Decisions API (in beta), which reads up to 922k tokens and PNG, JPEG, WebP, or GIF images of up to 20 MB each. It takes up to 200 questions, and a choice needs at least 2 options. gpt-6-luna names it too.\n - gemini-decision-1 runs Google's Gemini Decision 1, which reads up to 262k tokens and PNG, JPEG, WebP, or GIF images, up to 100 MB of image data in one request. It takes 1 to 1,000 questions. A question id uses only letters, digits, underscore, and hyphen, up to 64 characters, and a choice needs 2 to 255 options. A yes/no probability is between 0.01 and 0.99.\n- images is for a model that reads them. Pass URLs or data URLs. A model that reads none refuses the call.\n- The response model field is the version that answered.\n- Billed per input token. Output tokens are free. Each model has its own rate, and GPT-6 Luna Decisions bills a prompt over 272k tokens at twice its rate.",
5281
+ usageNotes: "- state is the content to judge: a string, or a JSON object or array (a record, a chat log, application state).\n- questions is a map. The key is an id you choose. It is not shown to the model. The answer comes back under the same id.\n- A noul asks a yes/no question and returns noul, a probability from 0 (no) to 1 (yes). Optional criteria.true and criteria.false say what a yes and a no mean.\n- A choice picks one option. criteria is a map of option id to a short description (or null). At most 255 options. The answer includes the winning choice, a probability for every option, and confidence.\n- A score rates the state on an ordered rubric. criteria is 2 to 10 level descriptions, low to high. The answer's score can land between levels, and includes a probability per level plus confidence.\n- An answer of type refusal means the model declined that question (GPT-6 Luna Decisions can).\n- instructions may be a string, or an object that keeps the question in one field and reference data in the others. Refer to those fields by name in backticks.\n- Send every question you need in one call. The model reads the state once and answers them together.\n- Probabilities from different models are not comparable: tune a threshold against the model, and the version, it will run on.\n- decisionModelOverride.model picks the model:\n - jev-latest (the default) is TypeSafe's current stable release, text only, up to 64k tokens. jev-preview tracks TypeSafe's newest release. Pin a version such as jev-1.13.0 when a threshold was tuned against that version.\n - clef and clef-flash run Cloudflare Clef, which also reads up to 4 PNG, JPEG, or WebP images, each at most 4 MiB and 8 MiB together, and takes at most 64 questions, each id using only letters, digits, underscore, dot, and hyphen, up to 100 characters. clef is the larger model, clef-flash the faster one. @cf/cloudflare/clef and @cf/cloudflare/clef-flash name the same two.\n - mercury-decide runs Inception's Mercury Decide, which reads up to 64k tokens and up to 16 PNG, JPEG, WebP, or GIF images, each at most 10 MiB and 14 MiB together.\n - drex-max, drex-max-finance, drex-v2.0 and drex-v1.5 run Nace.AI's Drex, served by DeepInfra. drex-max-finance takes up to 64 questions, and the others at most 16. drex-max and drex-max-finance, its finance tuning, read up to 262k tokens and up to 4 PNG, JPEG, WebP, or GIF images, 9 MiB together. drex-v2.0 and drex-v1.5 read text only, up to 128k tokens. nace-ai/drex-max, nace-ai/drex-max-finance, nace-ai/drex-v2.0 and nace-ai/drex-v1.5 name the same four.\n - gpt-6-luna-decisions runs OpenAI's Decisions API (in beta), which reads up to 922k tokens and PNG, JPEG, WebP, or GIF images of up to 20 MB each. It takes up to 200 questions, and a choice needs at least 2 options. gpt-6-luna names it too.\n - gemini-decision-1 runs Google's Gemini Decision 1, which reads up to 262k tokens and PNG, JPEG, WebP, or GIF images, up to 100 MB of image data in one request. It takes 1 to 1,000 questions. A question id uses only letters, digits, underscore, and hyphen, up to 64 characters, and a choice needs 2 to 255 options. A yes/no probability is between 0.01 and 0.99.\n- images is for a model that reads them. Pass URLs or data URLs. A model that reads none refuses the call.\n- The response model field is the version that answered.\n- Billed per input token. Output tokens are free. Each model has its own rate, and GPT-6 Luna Decisions bills a prompt over 272k tokens at twice its rate.",
5282
5282
  inputSchema: { "type": "object", "properties": { "state": { "anyOf": [{ "type": "string" }, { "type": "object", "properties": {}, "required": [] }, { "type": "array", "items": {} }] }, "questions": { "type": "object", "properties": {}, "required": [], "description": "Questions to ask about the state, keyed by an id you choose." }, "images": { "type": "array", "items": { "type": "string" }, "description": "Images for a model that reads them (Clef, Clef Flash, GPT-6 Luna Decisions, Gemini Decision 1). Each entry is a URL or a data URL. A model that reads none refuses the call." }, "decisionModelOverride": { "type": "object", "properties": { "model": { "type": "string", "description": "Decision model: a catalog id, its vendor's name for it (`@cf/cloudflare/clef`, `gpt-6-luna`), or a version it pins (`jev-1.13.0`)" }, "config": { "type": "object", "properties": {}, "required": [], "description": "Additional model-specific configuration" } }, "required": ["model"], "description": "Optional model override. Uses the workspace's default decision model (Jev) if not specified" } }, "required": ["state", "questions"] },
5283
5283
  outputSchema: { "type": "object", "properties": { "model": { "type": "string", "description": "Versioned model id that answered, such as jev-1.13.0." }, "answers": { "type": "object", "properties": {}, "required": [], "description": "One answer per question, under the same ids." }, "usage": { "type": "object", "properties": { "input_tokens": { "type": "number" }, "output_tokens": { "type": "number" } }, "required": ["input_tokens", "output_tokens"], "description": "Token counts for the call. Input tokens are what the model bills." } }, "required": ["model", "answers", "usage"] }
5284
5284
  },
@@ -5804,7 +5804,7 @@ var stepMetadata = {
5804
5804
  "renderVideoFromHtmlComposition": {
5805
5805
  stepType: "renderVideoFromHtmlComposition",
5806
5806
  description: "Render an HTML composition (HTML, CSS, media and seekable animation on a timeline) to a video file",
5807
- usageNotes: "- Renders an HTML composition in HyperFrames' format (https://hyperframes.heygen.com) with HyperFrames 0.8.217, in Chrome off the app's own sandbox; author for that version\n- Pass the composition as `html`, the page itself (up to 10 MB), or as `project`, the URL of a zip of a project's folder (`index.html` at its root, or the file `entryFile` names) or of one HTML file, up to 2 GiB: an app file or any public URL\n- `variables` overrides the composition's `data-composition-variables` defaults, which it reads with `window.__hyperframes.getVariables()`\n- Whatever the composition loads itself (fonts, scripts, images, video) is fetched as it renders, so each is in the project or at a public URL; a private app file needs a signed URL\n- fps 24, 30 or 60 (default 30); quality draft, standard or high (default standard); format mp4 (default), webm or mov (both keep transparency), or gif\n- An mp4, webm or mov renders in chunks of about 10 seconds side by side, so a long video takes little longer than a short one; a call finishes within about 30 minutes, and reports its progress as it goes. A gif renders whole, within 15 minutes\n- A composition HyperFrames can't render fails with HyperFrames' own reason",
5807
+ usageNotes: "- Renders an HTML composition in HyperFrames' format (https://hyperframes.heygen.com) with HyperFrames 0.9.19, in Chrome off the app's own sandbox; author for that version\n- Pass the composition as `html`, the page itself (up to 10 MB), or as `project`, the URL of a zip of a project's folder (`index.html` at its root, or the file `entryFile` names) or of one HTML file, up to 2 GiB: an app file or any public URL\n- `variables` overrides the composition's `data-composition-variables` defaults, which it reads with `window.__hyperframes.getVariables()`\n- Whatever the composition loads itself (fonts, scripts, images, video) is fetched as it renders, so each is in the project or at a public URL; a private app file needs a signed URL\n- fps 24, 30 or 60 (default 30); quality draft, standard or high (default standard); format mp4 (default), webm or mov (both keep transparency), or gif\n- An mp4, webm or mov renders in chunks of about 10 seconds side by side, so a long video takes little longer than a short one; a call finishes within about 30 minutes, and reports its progress as it goes. A gif renders whole, within 15 minutes\n- A composition HyperFrames can't render fails with HyperFrames' own reason",
5808
5808
  inputSchema: { "type": "object", "properties": { "html": { "type": "string", "description": "The composition's HTML, up to 10 MB; give this or project" }, "project": { "type": "string", "description": "URL of a zip of the project's folder, or of one self-contained HTML file; give this or html" }, "entryFile": { "type": "string", "description": "The entry HTML inside the zip, relative to its root; index.html when omitted" }, "variables": { "type": "object", "properties": {}, "required": [], "description": "Values over the composition's declared variables" }, "fps": { "enum": [24, 30, 60], "type": "number", "description": "Frames per second: 24, 30 (default) or 60" }, "quality": { "enum": ["draft", "standard", "high"], "type": "string", "description": "Encode quality: draft, standard (default) or high" }, "format": { "enum": ["mp4", "webm", "mov", "gif"], "type": "string", "description": "mp4 (default), webm or mov (both keep transparency), or gif" }, "intermediateAsset": { "type": "boolean", "description": "When true, the asset is created but hidden from the user's gallery (tagged as intermediate)" } } },
5809
5809
  outputSchema: { "type": "object", "properties": { "videoUrl": { "type": "string", "description": "URL of the rendered video" }, "durationSec": { "type": "number", "description": "Its length in seconds" }, "frames": { "type": "number", "description": "How many frames it has" } }, "required": ["videoUrl", "durationSec", "frames"] }
5810
5810
  },
@@ -6053,7 +6053,7 @@ var stepMetadata = {
6053
6053
  "transcribeAudio": {
6054
6054
  stepType: "transcribeAudio",
6055
6055
  description: "Convert an audio file to text using a transcription model.",
6056
- usageNotes: "- The default model is elevenlabs-scribe-v2: speakers, word timings, key terms and redaction, at any length.\n- The prompt field gives context (speaker names, domain) to OpenAI's and DeepInfra's models. Other models ignore it; give them `keyterms` instead.\n- Options go in `transcriptionModelOverride.config`, and support varies by model:\n - `language` \u2014 the spoken language as a code such as `en` or `es`; left out, the model detects it. Every model takes it. deepgram-nova-3 assumes `en` unless told, and also takes `multi` for speech that switches language (billed at its multilingual rate).\n - `detectLanguage: 'yes'` \u2014 detect the language instead of assuming English (deepgram-nova-3 only).\n - `includeSpeakers: 'yes'` \u2014 label each passage with a speaker. Supported by elevenlabs-scribe-v2, elevenlabs-scribe-v2-medical, deepgram-nova-3, grok-stt, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3 and gemini-3.5-transcribe; gpt-4o-transcribe-diarize always does. Scribe and Voxtral also switch `text` to a timestamped, speaker-labelled block format; Deepgram, Gemini and gpt-4o-transcribe-diarize to `Speaker N:` lines.\n - `numSpeakers` \u2014 the most people talking, to help tell them apart (Scribe).\n - `keyterms` \u2014 a list of names, products and terms the audio contains, so they're spelled right. Supported by gpt-transcribe, Scribe, deepgram-nova-3, grok-stt, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3 and gemini-3.5-transcribe. Billed as an add-on on Scribe and Deepgram.\n - `redactPii: 'yes'` \u2014 replace names, phone numbers, addresses and other personal information with a label of what was there (Scribe, deepgram-nova-3). Billed as an add-on.\n - `includeWordTimings: 'yes'` \u2014 add per-word timings and confidence inside each segment. Supported by Scribe, deepgram-nova-3, whisper-1, grok-stt, voxtral-mini-transcribe-2 and gemini-3.5-transcribe. Off by default because an hour of speech is roughly ten thousand words.\n - `subtitles: 'srt' | 'vtt'` \u2014 also return the transcript as a subtitle file, in the `subtitles` output, timed by the segments. Any model that returns segments.\n- gemini-3.5-transcribe can't take `keyterms` together with speakers, word timings or subtitles; asked for both, it's refused.\n- The `segments` output carries timings, confidence and speakers as structured data, so there is no need to slice the audio up to recover them. It is populated by Scribe, deepgram-nova-3, grok-stt, whisper-1, gpt-4o-transcribe-diarize, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3, gemini-3.5-transcribe (with word timings or subtitles) and the DeepInfra-served whisper and qwen models, and absent for models that return nothing but text (gpt-transcribe, gpt-4o-transcribe, gpt-4o-mini-transcribe, the DeepInfra-served voxtral models).\n- In a workflow, `destinationVar` receives the transcript text. Name `segmentsVar` to also keep the `segments` array. A model that returns no segments leaves that variable unset.\n- A recording of any length can be transcribed. One longer than a model takes at once (OpenAI's and DeepInfra's models, Gemini, Voxtral 3) is transcribed in pieces and joined, its timings running from the start of the recording. Speakers can't be followed from one piece to the next, so with speakers on, such a recording is refused; Scribe and Deepgram label speakers at any length.\n- `confidence` is always 0..1 with higher meaning more confident, whichever model produced it.",
6056
+ usageNotes: "- The default model is elevenlabs-scribe-v3: speakers, word timings, key terms and redaction, at any length.\n- The prompt field gives context (speaker names, domain) to OpenAI's and DeepInfra's models. Other models ignore it; give them `keyterms` instead.\n- Options go in `transcriptionModelOverride.config`, and support varies by model:\n - `language` \u2014 the spoken language as a code such as `en` or `es`; left out, the model detects it. Every model takes it. deepgram-nova-3 assumes `en` unless told, and also takes `multi` for speech that switches language (billed at its multilingual rate).\n - `detectLanguage: 'yes'` \u2014 detect the language instead of assuming English (deepgram-nova-3 only).\n - `includeSpeakers: 'yes'` \u2014 label each passage with a speaker. Supported by elevenlabs-scribe-v2, elevenlabs-scribe-v2-medical, elevenlabs-scribe-v3, deepgram-nova-3, grok-stt, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3 and gemini-3.5-transcribe; gpt-4o-transcribe-diarize always does. Scribe and Voxtral also switch `text` to a timestamped, speaker-labelled block format; Deepgram, Gemini and gpt-4o-transcribe-diarize to `Speaker N:` lines.\n - `numSpeakers` \u2014 the most people talking, to help tell them apart (Scribe).\n - `keyterms` \u2014 a list of names, products and terms the audio contains, so they're spelled right. Supported by gpt-transcribe, Scribe, deepgram-nova-3, grok-stt, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3 and gemini-3.5-transcribe. Billed as an add-on on Scribe and Deepgram.\n - `redactPii: 'yes'` \u2014 replace names, phone numbers, addresses and other personal information with a label of what was there (Scribe, deepgram-nova-3). Billed as an add-on.\n - `includeWordTimings: 'yes'` \u2014 add per-word timings and confidence inside each segment. Supported by Scribe, deepgram-nova-3, whisper-1, grok-stt, voxtral-mini-transcribe-2 and gemini-3.5-transcribe. Off by default because an hour of speech is roughly ten thousand words.\n - `subtitles: 'srt' | 'vtt'` \u2014 also return the transcript as a subtitle file, in the `subtitles` output, timed by the segments. Any model that returns segments.\n- gemini-3.5-transcribe can't take `keyterms` together with speakers, word timings or subtitles; asked for both, it's refused.\n- The `segments` output carries timings, confidence and speakers as structured data, so there is no need to slice the audio up to recover them. It is populated by Scribe, deepgram-nova-3, grok-stt, whisper-1, gpt-4o-transcribe-diarize, voxtral-mini-transcribe-2, voxtral-mini-transcribe-3, voxtral-small-transcribe-3, gemini-3.5-transcribe (with word timings or subtitles) and the DeepInfra-served whisper and qwen models, and absent for models that return nothing but text (gpt-transcribe, gpt-4o-transcribe, gpt-4o-mini-transcribe, the DeepInfra-served voxtral models).\n- In a workflow, `destinationVar` receives the transcript text. Name `segmentsVar` to also keep the `segments` array. A model that returns no segments leaves that variable unset.\n- A recording of any length can be transcribed. One longer than a model takes at once (OpenAI's and DeepInfra's models, Gemini, Voxtral 3) is transcribed in pieces and joined, its timings running from the start of the recording. Speakers can't be followed from one piece to the next, so with speakers on, such a recording is refused; Scribe and Deepgram label speakers at any length.\n- `confidence` is always 0..1 with higher meaning more confident, whichever model produced it.",
6057
6057
  inputSchema: { "type": "object", "properties": { "audioUrl": { "type": "string", "description": "URL of the audio file to transcribe" }, "prompt": { "type": "string", "description": "Optional context to improve transcription accuracy (e.g. language, speaker names, domain terms)" }, "transcriptionModelOverride": { "type": "object", "properties": { "model": { "type": "string", "description": "Audio transcription model identifier" }, "config": { "type": "object", "properties": {}, "required": [], "description": "Additional model-specific configuration" } }, "required": ["model"], "description": "Optional model configuration override. Uses the workflow's default transcription model if not specified" } }, "required": ["audioUrl", "prompt"] },
6058
6058
  outputSchema: { "type": "object", "properties": { "text": { "type": "string", "description": "The transcribed text from the audio file" }, "segments": { "type": "array", "items": { "type": "object", "properties": { "start": { "type": "number", "description": "Seconds from the start of the audio." }, "end": { "type": "number" }, "text": { "type": "string", "description": "The passage, spacing normalized." }, "confidence": { "type": "number", "description": "How sure the model was, always `0..1` with higher meaning more confident \u2014 whatever the provider reported.\n\nNormalizing this is the whole point of the type. Providers disagree on both the scale and the direction: Deepgram reports a probability in `0..1`, ElevenLabs and Whisper report a log probability in `(-inf, 0]`. Handing those through untouched would make `confidence < 0.5` mean opposite things depending on which model the caller happened to pick. Log probabilities are converted with `Math.exp`; probabilities pass through." }, "speaker": { "type": "string", "description": "Present only when diarization was asked for and the provider supports it." }, "language": { "type": "string", "description": "Present only for providers that report language per passage rather than per request. Most report it once for the whole audio \u2014 see `TranscriptionResponse.language`." }, "words": { "type": "array", "items": { "type": "object", "properties": { "text": { "type": "string" }, "start": { "type": "number", "description": "Seconds from the start of the audio." }, "end": { "type": "number" }, "confidence": { "type": "number", "description": "See the confidence convention on `TranscriptSegment`." }, "speaker": { "type": "string", "description": "Provider-assigned speaker id, present only when diarization was asked for." } }, "required": ["text"], "description": "One recognized token.\n\n`text` rather than `word` because an entry is not always a word \u2014 providers emit punctuation, spacing and non-speech audio events (laughter, footsteps) through the same channel." }, "description": "Word-level detail, when the caller asked for it (`includeWordTimings`). Off by default: an hour of speech is roughly ten thousand words, an order of magnitude more payload than the segments that contain them." } }, "required": ["start", "end", "text"], "description": "A contiguous passage of one speaker's speech, bounded in time.\n\nThe unit a review UI works in: short enough to seek to and highlight, coarse enough that an hour of audio is a few hundred entries rather than ten thousand." }, "description": "Timed passages with per-segment confidence, speaker attribution and optional word-level detail \u2014 see the step's usage notes for which models populate it. Absent for models that return nothing but text." }, "language": { "type": "string", "description": "Detected or declared language of the audio, when the model reports it." }, "durationSeconds": { "type": "number", "description": "Length of the transcribed audio, when the model reports it." }, "subtitles": { "type": "string", "description": "The transcript as an SRT or WebVTT file, when `subtitles` was asked for in the model's config and the model returned segments to time it by." } }, "required": ["text"] }
6059
6059
  },