@deepraxis/cli 0.3.0-dev.47e949a → 0.3.0-dev.6b385c0

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.
Files changed (2) hide show
  1. package/dist/praxis.js +6 -6
  2. package/package.json +1 -1
package/dist/praxis.js CHANGED
@@ -28864,7 +28864,7 @@ The default is return=minimal: success returns HTTP 201 with an empty body. Send
28864
28864
  resource: "media.generation.capabilities",
28865
28865
  method: "list",
28866
28866
  summary: "Get generation capabilities",
28867
- description: "Read what Praxis offers for each kind of generation — image, video and speech — before asking for one. Every type answers the models available for it, the model it uses when a request names none, and what each model costs per unit; the video fields beside them keep their original meaning (intents, default model, duration and aspect-ratio limits, per-library concurrency). The models come from the platform model catalogue, and an id listed here is accepted verbatim as the model field of that type’s generation request. price.amount is a decimal string with its currency; video models that charge by resolution also carry price.by_resolution, whose amount is then the cheapest tier. A model the catalogue cannot route right now is still listed, with available false, so a caller can show it as temporarily unavailable instead of hiding it. When the catalogue cannot be read the endpoint still answers: every type comes back with an empty models list and its real default_model. Generation is charged per unit and every output is stored in the library.",
28867
+ description: 'Read what Praxis offers for each kind of generation — image, video and speech — before asking for one, and read a speech model’s voices here before naming one. Every type answers the models available for it, the model it uses when a request names none, and what each model costs per unit; a speech model the platform publishes voices for also carries default_voice and voices there, each voice with an id, a name, a gender, a language and a style — so a request for "a gentle female voice" is answered by matching the style and gender rather than by guessing an identifier. Pass a voice id back verbatim as voice, or omit voice to take default_voice; a voice that is neither one of those nor a voice material of the library is refused with INVALID_PARAMS before any job is created. OpenAI voice names such as alloy and aura are not valid here. The video fields beside the types keep their original meaning (intents, default model, duration and aspect-ratio limits, per-library concurrency). The models come from the platform model catalogue, and an id listed here is accepted verbatim as the model field of that type’s generation request. price.amount is a decimal string with its currency; video models that charge by resolution also carry price.by_resolution, whose amount is then the cheapest tier. A model the catalogue cannot route right now is still listed, with available false, so a caller can show it as temporarily unavailable instead of hiding it. When the catalogue cannot be read the endpoint still answers: every type comes back with an empty models list and its real default_model, and the voices a speech model carries are absent rather than empty, because the voices come from the platform and not from the catalogue. Generation is charged per unit and every output is stored in the library.',
28868
28868
  http: "GET",
28869
28869
  path: "/media/v1/generation/capabilities",
28870
28870
  streaming: "no",
@@ -29730,7 +29730,7 @@ The default is return=minimal: success returns HTTP 201 with an empty body. Send
29730
29730
  resource: "media.speech",
29731
29731
  method: "create",
29732
29732
  summary: "Create speech",
29733
- description: "Synthesise speech from text and store the recording straight into the library. Use library_id from creation or listing; the path segment {library_id} also accepts the literal default, which resolves to the organization’s default library before the credential check, so an agent that has never listed a library can still synthesise. input is the text to speak and voice is required; speed is passed to the provider unchanged, and instructions and language are refused with OPERATION_NOT_SUPPORTED in this release. Voice cloning is not available yet, so voice is a platform voice name of the chosen model. The provider caps one request at a short stretch of text, so a long input is cut into segments at paragraph and sentence boundaries and synthesised in order, then joined into the single recording the library receives: nothing is spoken twice and nothing is dropped. Because the segments are joined byte for byte, a multi-segment input accepts only response_format mp3 (the default) or pcm; opus is available for a single-segment input. A speech job has one result however many segments it took. Synthesis is billed by the model usage it consumes, at the same prices as the model API, and the job records the charge in usage; a request refused before it starts costs nothing. The response is 202 with a queued Job whose type is speech and whose operation is generate; poll GET /media/v1/libraries/{library_id}/jobs/{job_id} until its status is completed or failed, then read the produced asset from the library — the recording carries material_type audio, audio_role voice, ai_generated true, and provenance naming this job, the text spoken and the model. Send Idempotency-Key, or idempotency_key in the body, to retry without synthesising and charging twice. Example: POST /media/v1/libraries/default/audio/speech with {\"input\":\"A red kite above a calm sea at sunrise.\",\"voice\":\"alloy\"}.",
29733
+ description: "Synthesise speech from text and store the recording straight into the library. Use library_id from creation or listing; the path segment {library_id} also accepts the literal default, which resolves to the organization’s default library before the credential check, so an agent that has never listed a library can still synthesise. input is the text to speak. voice is optional and may be omitted, in which case the model’s default voice is used; when it is given it must be one of the voices of the model being used — read them from the types.speech list of GET /media/v1/generation/capabilities, which gives each voice an id, a name, a gender, a language and a style, or pass a voice material id of this same library — and a voice that is neither is refused with INVALID_PARAMS in the voice field before any job exists, so a wrong name is free to try but reading the capabilities first is free without trying. Do not use OpenAI voice names such as alloy or aura: they are not valid here. speed is passed to the provider unchanged, and instructions and language are refused with OPERATION_NOT_SUPPORTED in this release. Cloning a voice from recordings is not available yet. The provider caps one request at a short stretch of text, so a long input is cut into segments at paragraph and sentence boundaries and synthesised in order, then joined into the single recording the library receives: nothing is spoken twice and nothing is dropped. Because the segments are joined byte for byte, a multi-segment input accepts only response_format mp3 (the default) or pcm; opus is available for a single-segment input. A speech job has one result however many segments it took. Synthesis is billed by the model usage it consumes, at the same prices as the model API, and the job records the charge in usage; a request refused before it starts costs nothing. The response is 202 with a queued Job whose type is speech and whose operation is generate; poll GET /media/v1/libraries/{library_id}/jobs/{job_id} until its status is completed or failed, then read the produced asset from the library — the recording carries material_type audio, audio_role voice, ai_generated true, and provenance naming this job, the text spoken and the model. Send Idempotency-Key, or idempotency_key in the body, to retry without synthesising and charging twice. Example: POST /media/v1/libraries/default/audio/speech with {\"input\":\"A red kite above a calm sea at sunrise.\"}, or with {\"input\":\"A red kite above a calm sea at sunrise.\",\"voice\":\"zh_female_wenroumama_uranus_bigtts\"} for a gentle female voice.",
29734
29734
  http: "POST",
29735
29735
  path: "/media/v1/libraries/{library_id}/audio/speech",
29736
29736
  streaming: "no",
@@ -29755,8 +29755,8 @@ The default is return=minimal: success returns HTTP 201 with an empty body. Send
29755
29755
  {
29756
29756
  name: "voice",
29757
29757
  type: "string",
29758
- required: true,
29759
- description: "Voice identifier or platform voice name."
29758
+ required: false,
29759
+ description: "Voice to synthesize with. Omit it to use the model’s default voice. The available values are the ids in the types.speech list of GET /media/v1/generation/capabilities, beside the voices of the model you are using, plus a voice material id of this same library. A voice that is neither is refused with INVALID_PARAMS before any job is created, so a wrong name costs nothing — read the capabilities first instead of guessing. Example ids for the platform speech model: zh_female_vv_uranus_bigtts (the default), zh_female_wenroumama_uranus_bigtts (a gentle female voice), zh_male_m191_uranus_bigtts. OpenAI voice names such as alloy are not valid here."
29760
29760
  },
29761
29761
  {
29762
29762
  name: "instructions",
@@ -57364,7 +57364,7 @@ function forbiddenHits(names, entries) {
57364
57364
  return hits;
57365
57365
  }
57366
57366
  var targetDir = (workdir, target) => `${workdir}/${target}`;
57367
- var planFileOf = (planDir, target) => `${planDir}/${target}.bin`;
57367
+ var planFileOf = (planDir, target) => `${planDir}/${target.replaceAll("/", "__")}.bin`;
57368
57368
  var STATE_USERNAME = "praxis";
57369
57369
  var STATE_PASSWORD_ENV = "TF_HTTP_PASSWORD";
57370
57370
  function stateAddress(backend, target) {
@@ -61127,7 +61127,7 @@ function startProgress(ctx, write = (text) => writeOutput({ stderr: text })) {
61127
61127
  }
61128
61128
 
61129
61129
  // src/main.ts
61130
- var version = "0.3.0-dev.47e949a";
61130
+ var version = "0.3.0-dev.6b385c0";
61131
61131
  var commands = commandTree(operations);
61132
61132
  var manual = [...workflows.map((w) => w.words), [API_WORD]];
61133
61133
  function route(argv) {
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@deepraxis/cli",
3
- "version": "0.3.0-dev.47e949a",
3
+ "version": "0.3.0-dev.6b385c0",
4
4
  "type": "module",
5
5
  "bin": {
6
6
  "praxis": "dist/praxis.js"