poddrill 0.3.1 → 0.4.0

package/.agents/skills/poddrill/SKILL.md

@@ -19,11 +19,11 @@ Do this once on the computer that will transcribe. Each step is safe to repeat.
 
 1. Install Node.js 20.19 or newer and `ffmpeg`. Use the Node `.tar.gz` archive from `https://nodejs.org/dist/v20.19.0/`, or run `npx -y node@20.19.0`. On macOS, `brew install ffmpeg`. On Linux, `sudo apt-get install -y ffmpeg`.
 
-2. Install the npm package. Run `npm install -g poddrill@0.3.1`. Do not use `@latest`. Cloning the git repository is not required. If you clone it, set `GIT_TERMINAL_PROMPT=0`.
+2. Install the npm package. Run `npm install -g poddrill@0.4.0`. Do not use `@latest`. Cloning the git repository is not required. If you clone it, set `GIT_TERMINAL_PROMPT=0`.
 
 3. Run `poddrill install-python`. This creates `~/.cache/poddrill/venv` and installs cactus-needle 3.2.0, numpy 2.5.4, and onnxruntime 1.31.0. It uses `uv` when `uv` is on PATH, and otherwise `python3 -m venv`. The npm package named cactus-needle is a different package.
 
-4. Run `poddrill fetch-weights`. This downloads `voxceleb_ECAPA512.onnx` from the pinned WeSpeaker commit, checks the byte size and SHA-256, and writes `~/.cache/poddrill/ecapa/voxceleb_ECAPA512.onnx`. WeSpeaker ECAPA-TDNN512 pretrained on VoxCeleb. VoxCeleb is used under CC BY 4.0. A later run leaves a matching file in place.
+4. Run `poddrill fetch-weights`. This downloads `voxceleb_ECAPA512.onnx` and `segmentation.onnx`, checks each byte size and SHA-256, and writes them under `~/.cache/poddrill/ecapa/` and `~/.cache/poddrill/segmentation/`. A later run leaves a matching file in place.
 
 5. Run `poddrill client doctor --client cursor --json`. Use `grok_build` or `grok_bot` in place of `cursor` when that is the client. For every check whose `state` is `bad`, run the `fix` command on that check, then run doctor again. Skip a check whose `state` is `optional`. Do not install it. Do not mention it. Do not ask the user about it. `workspace` stays `pending` until setup. Stop this loop when doctor exits 0, or when the only `bad` check is `link`.
 
@@ -53,7 +53,7 @@ Read cues only after the embedding is ready. A ready transcript can still be not
 
 ## Speaker groups
 
-An asset clusters cues that sound like one voice. That cluster is a speaker group. You assign the group a name, `quote`, `ending`, or `undetermined`. The utility does not choose the name.
+An asset clusters cues that sound like one voice. That cluster is a speaker group. The embedding stage writes the groups before that stage is ready. You assign the group a name, `quote`, `ending`, or `undetermined`. The utility does not choose the name.
 
 Label the speaker groups before you write a summary. Do not write a summary from the transcript text alone.
 
@@ -73,7 +73,7 @@ Write the WAV at `<workspace>/audio/<file>.wav`. `poddrill start` resolves the p
 
 Pass `--language <code>` on `poddrill start` when the episode is not English. The code is `en`, `de`, `fr`, `es`, `it`, `nl`, `pl`, `ja`, `ko`, or `zh`.
 
-`poddrill start` returns before the transcript exists. Poll `poddrill episode <id>` until the transcript artifact `state` is `ready` or `failed`. The id is `ep_` plus the first 16 hex characters of the sha256 of the file bytes.
+`poddrill start` returns before the transcript exists. The transcript job diarizes the episode, then transcribes it. Poll `poddrill episode <id>` until the transcript artifact `state` is `ready` or `failed`. The id is `ep_` plus the first 16 hex characters of the sha256 of the file bytes.
 
 The transcript is done when that `state` is `ready`. Exit 0 does not mean the transcript is done. If the transcript or the embedding is missing or `failed`, pass the same WAV to `poddrill start` again. The episode id does not change. `created` is false, and the failed artifact runs again.
 

package/README.md

@@ -29,7 +29,7 @@ Identify books mentioned in an episode and what it is about:
 ```
 
 ## For LLM and Agent
-The supported install is the npm package. Run `npm install -g poddrill@0.3.1`. Do not use `@latest`. Cloning this repository is not required. If you clone it, set `GIT_TERMINAL_PROMPT=0`. The agent then follows the skill shipped in the package.
+The supported install is the npm package. Run `npm install -g poddrill@0.4.0`. Do not use `@latest`. Cloning this repository is not required. If you clone it, set `GIT_TERMINAL_PROMPT=0`. The agent then follows the skill shipped in the package.
 
 ### Requirements
 
@@ -38,7 +38,9 @@ The computer that runs PodDrill needs:
 - Node.js 20.19 or newer. Install the `.tar.gz` archive from `https://nodejs.org/dist/v20.19.0/`, or run `npx -y node@20.19.0`. The published CLI is plain JavaScript. The tests in this repository use `node --experimental-strip-types` and need a newer Node.
 - `ffmpeg` on `PATH`.
 - A Python interpreter with cactus-needle 3.2.0, numpy 2.5.4, and onnxruntime 1.31.0. The npm package named cactus-needle is a different package. `poddrill install-python` creates `~/.cache/poddrill/venv` and installs those pins. It uses `uv` when `uv` is on PATH, and otherwise `python3 -m venv`. `PODDRILL_PYTHON` selects another absolute interpreter, and setup stores that path in `poddrill.json`.
-- The ECAPA weights file. `poddrill fetch-weights` downloads `voxceleb_ECAPA512.onnx` from the pinned WeSpeaker commit, checks its size and SHA-256, and writes `~/.cache/poddrill/ecapa/voxceleb_ECAPA512.onnx`. `PODDRILL_EMBED_WEIGHTS` selects another path. WeSpeaker ECAPA-TDNN512 pretrained on VoxCeleb. VoxCeleb is used under CC BY 4.0.
+- The ECAPA and segmentation weight files. `poddrill fetch-weights` downloads `voxceleb_ECAPA512.onnx` and `segmentation.onnx`, checks each size and SHA-256, and writes them under `~/.cache/poddrill/ecapa/` and `~/.cache/poddrill/segmentation/`. `PODDRILL_EMBED_WEIGHTS` and `PODDRILL_SEGMENTATION_WEIGHTS` select other paths.
+
+The transcript job diarizes the episode, then transcribes it. The embedding job runs after the transcript is ready.
 
 `poddrill client doctor` prints a `fix` command for every check that is not ok. `poddrill bootstrap` runs `install-python`, `fetch-weights`, and `setup`.
 
@@ -50,10 +52,10 @@ Grok Bot cannot write its server entry. It prints the custom server to add: the
 
 `npm test` runs the unit tests. Those tests use `node --experimental-strip-types`.
 
-`npm run test:smoke` runs a clean-home install, downloads the ECAPA file, and transcribes a short generated clip through `dist/cli.js`. It needs `python3` with `ensurepip`, `ffmpeg`, and `espeak-ng`. If one of those is missing, the test skips and names the missing tool. Run `npm run test:smoke` on a computer that has those tools. Build `dist/` with `npm run build` first.
+`npm run test:smoke` runs a clean-home install, downloads the ECAPA and segmentation files, and transcribes a short generated clip through `dist/cli.js`. It needs `python3` with `ensurepip`, `ffmpeg`, and `espeak-ng`. If one of those is missing, the test skips and names the missing tool. Run `npm run test:smoke` on a computer that has those tools. Build `dist/` with `npm run build` first.
 
 ## License
 
 MIT License. Copyright 2026 Jianyu Zhao.
 
-The ECAPA weights are separate from this license. WeSpeaker ECAPA-TDNN512 pretrained on VoxCeleb. VoxCeleb is used under CC BY 4.0.
+The ECAPA weights are separate from this license. WeSpeaker ECAPA-TDNN512 pretrained on VoxCeleb. VoxCeleb is used under CC BY 4.0. The segmentation weights are separate from this license. pyannote segmentation-3.0 is MIT.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "poddrill",
-  "version": "0.3.1",
+  "version": "0.4.0",
   "description": "Transcribe episode audio into timed text, group the voices, and store a summary.",
   "license": "MIT",
   "author": "Jianyu Zhao",