@tryhamster/gerbil 1.0.0-rc.8 → 1.0.0

package/docs/onnx-removal-plan.md

@@ -0,0 +1,339 @@
+# ONNX / transformers.js Removal Plan — Going Pure Native
+
+> Status: **PLANNING DOC**. No code is deleted by this document. It is the
+> authoritative inventory + phased plan to hand to a removal workflow.
+>
+> Branch: `feat/webgpu-engine-mobile`
+> Goal: remove the legacy ONNX Runtime / transformers.js / puppeteer-Chrome
+> inference stack so Gerbil runs **pure native** on the WebGPU engine
+> (`src/gpu/`), across text (Qwen3.5 / LFM2.5 / Gemma4), vision (Qwen3.5 ViT,
+> Gemma4 vision), embeddings (EmbeddingGemma / Qwen3-Embedding), STT (Moonshine)
+> and TTS (Kani).
+
+---
+
+## 0. TL;DR — Headline Findings
+
+- **~10,800 lines of inference-stack code become deletable** once the native
+  paths are wired in and the browser/REPL consumers are migrated. The biggest
+  single chunks: `chrome-backend.ts` (1,836), `worker.ts` (1,056),
+  `tts.ts` (1,031), `worker-entry.ts` (599), `stt.ts` (605),
+  `worker-code.generated.ts` (1.35 MB generated bundle).
+- **3 npm deps drop out:** `puppeteer-core`, `@huggingface/transformers`,
+  `onnxruntime-web`. Likely also `kokoro-js` (ONNX-backed TTS) and possibly
+  `@huggingface/hub` (only if nothing else needs HF tree/download APIs).
+- **The "all modalities are native" premise is only true inside `src/gpu/`.**
+  `WebGPUEngine` (`src/gpu/index.ts`) exposes native `generate`, `encodeImage`,
+  `describeImage`, `embed`, `speak` and `transcribe` (Moonshine). **But the
+  public surfaces — the core Node `Gerbil` class and the browser hooks — do NOT
+  yet route STT / TTS / embeddings to the native engine.** They still call ONNX.
+  **This wiring is the prerequisite for removal**, not an afterthought.
+
+### The 2-3 real decisions the user must make
+
+1. **CPU/WASM fallback (transformers.js wasm) — keep or kill?** Pure-native
+   means *WebGPU required*. Dropping the wasm path removes the only way to run
+   on machines/browsers without WebGPU (older Safari, locked-down enterprise,
+   CI without a GPU). Recommendation: **kill it** for pure-native, but this is a
+   reach tradeoff the user owns. (Detail in §3a.)
+2. **chrome-backend.ts (puppeteer headless Chrome) — remove now?** It was the
+   Node-side WebGPU path before node-dawn/native. It is already legacy. **Yes,
+   safe to remove** — but it has surprisingly wide tentacles in the CLI/REPL
+   (cache listing, process management commands). (Detail in §3b.)
+3. **Browser ONNX worker + hooks — remove once site/REPL stop using them?**
+   The chat worker (`worker.ts` + `worker-entry.ts` + the 1.35 MB
+   `worker-code.generated.ts`) and `use-embedding`/`use-speech` ONNX paths can
+   go once `use-native-engine` covers chat and native embed/speech hooks exist.
+   (Detail in §3c.)
+
+---
+
+## 1. Full Inventory — File by File
+
+Legend: **[DELETE]** = remove entirely · **[GUT]** = strip ONNX path, keep file ·
+**[KEEP]** = native, no change · **[REWIRE]** = repoint to native then clean.
+
+### 1a. Core (`src/core/`)
+
+| File | Lines | ONNX/legacy usage | Disposition |
+|---|---:|---|---|
+| `chrome-backend.ts` | 1,836 | `puppeteer-core` headless-Chrome WebGPU backend. `ChromeGPUBackend`, `getChromeCachedModels`, global browser/process management. | **[DELETE]** (after CLI/REPL detached) |
+| `stt.ts` | 605 | `WhisperSTT`: `pipeline("automatic-speech-recognition", "onnx-community/whisper-*")` via `@huggingface/transformers`, ORT wasm CDN. | **[DELETE]** → replace with native Moonshine wrapper |
+| `tts.ts` | 1,031 | `KokoroTTS` (`kokoro-js`, ONNX) + `SupertonicTTS` (`pipeline("text-to-speech")`, ONNX). | **[DELETE]** → replace with native Kani wrapper |
+| `model-compat.ts` | 79 | `patchHFCacheConfig` — patches HF cache so transformers.js loads new model_types (qwen3_5→qwen3). Pure ONNX-loader workaround. | **[DELETE]** (native loader reads safetensors directly, no patch needed) |
+| `gerbil.ts` | 2,686 | The hub. `import { AutoModelForCausalLM, AutoModelForImageTextToText, pipeline, ... } from "@huggingface/transformers"`. `effectiveBackend` switch with `onnx` / `webgpu-native` / chrome branches. `this.model` (ONNX), `this.visionModel` (ONNX), `this.chromeBackend`, `this.embedder` (ONNX pipeline). Native path = `this.webgpuEngine`. `embed()`/`embedBatch()` still ONNX (`pipeline("feature-extraction", "Xenova/all-MiniLM-L6-v2")`). STT/TTS delegate to ONNX `WhisperSTT`/`createTTS`. | **[GUT]** — collapse to native-only; remove all ONNX branches, chrome branch, `pipeline` import, model-compat call |
+
+Key ONNX surfaces inside `gerbil.ts`:
+- Imports (lines ~6-21): `AutoModelForCausalLM`, `AutoModelForImageTextToText`,
+  `pipeline as rawPipeline`, etc.
+- `effectiveBackend` switch (~386-393): three-way `onnx` / `webgpu-native` /
+  auto. After removal this becomes always-native.
+- Branches: `isBrowser && webgpu && onnx` (~440), `!isBrowser && webgpu && onnx`
+  → chrome backend (~467-486), CPU/WASM pipeline fallback (~488+), error
+  fallback to pipeline (~525-530).
+- `loadVisionModel` chrome branch (~600-620).
+- `generate`/`generateStream` dispatch: `if (this.webgpuEngine)` (native) vs
+  `else if (this.chromeBackend)` vs `else this.model.generate` (ONNX) — lines
+  ~1235-1300, ~1424-1510.
+- `embed()` (~1824) / `embedBatch()` (~1849): pure ONNX, **no native branch yet**.
+- STT/TTS methods (~2058-2310): delegate to ONNX `tts`/`stt` objects.
+- Chrome status/memory passthroughs (~1055-1101), getModelRepo ONNX maps
+  (~1026-1041).
+- Cleanup (~2388-2461): disposes webgpuEngine, chromeBackend, tts, stt.
+
+### 1b. Browser (`src/browser/`)
+
+| File | Lines | ONNX/legacy usage | Disposition |
+|---|---:|---|---|
+| `worker.ts` | 1,056 | Chat web-worker host. Builds inline worker that imports `@huggingface/transformers` from CDN, ORT wasm paths, `AutoTokenizer`/`AutoModelForCausalLM`/`AutoModelForImageTextToText`. iOS main-thread inference path. | **[DELETE]** (after chat → `use-native-engine`) |
+| `worker-entry.ts` | 599 | The real worker source compiled to the IIFE. Imports transformers v4, ORT wasm CDN, vision+text ONNX. | **[DELETE]** |
+| `worker-code.generated.ts` | ~1.35 MB | Generated IIFE bundle of `worker-entry.ts` (transformers.js fully inlined). Produced by `scripts/build-worker.mjs`. | **[DELETE]** + delete generator + build step |
+| `use-embedding.ts` | 485 | Worker that imports transformers.js from CDN, `pipeline("feature-extraction", "Xenova/all-MiniLM-L6-v2")`, ORT wasm, iOS single-thread. | **[GUT/REWIRE]** to native embed (EmbeddingGemma via `WebGPUEngine.embed`) |
+| `use-speech.ts` | 689 | TTS hook: Kokoro (`kokoro-js`) + Supertonic (`pipeline("text-to-speech")`). All ONNX. | **[REWIRE]** to native Kani |
+| `use-voice-input.ts` | 1,227 | STT+chat+TTS combo hook. Uses `createGerbilWorker` (ONNX chat worker) + ONNX STT/TTS models (`whisper-tiny.en`, `kokoro-82m`). | **[REWIRE]** to native engine + native STT/TTS |
+| `backend-selector.ts` | 245 | `selectBackend` returning `webgpu`/`wasm`/`cpu`, fallback chains, iOS wasm fallbacks. The wasm/cpu half is ONNX-fallback machinery. | **[GUT]** — drop wasm/cpu fallback (decision §3a) |
+| `use-native-engine.ts` | 217 | **[KEEP]** Already native — "no ONNX Runtime, no transformers.js." This is the template the others migrate to. | **[KEEP]** |
+| `index.ts` | — | Re-exports worker, use-embedding, use-speech, use-voice-input, backend-selector. Notes core Gerbil excluded due to chrome-backend/puppeteer. | **[REWIRE]** exports after deletions |
+
+Other browser files referencing legacy (mostly comments/guards): `preload.ts`,
+`device-guards.ts`, `download.ts`, `webgpu.ts` — touch only as needed (mostly
+ONNX wasm path hints / iOS guards). Low priority.
+
+### 1c. CLI / REPL (`src/cli/`) — see §4 for full audit
+
+| File | Lines | usage | Disposition |
+|---|---:|---|---|
+| `index.ts` | 1,289 | `--backend native|onnx|auto` flag (maps to webgpu-native/onnx). `transcribe`, `tts`, `voice` commands (ONNX STT/TTS). `chrome` admin command (`ChromeGPUBackend.getGlobalBrowserStatus`, `killAllBackends`, kill page, ps/pkill of `gerbil/chrome-cache`). transformers cache dir scanning. | **[REWIRE]** drop onnx/auto backend options → native-only; delete `chrome` command; repoint STT/TTS |
+| `repl/utils.ts` | 738 | `getChromeCachedModels`, `refreshCachedModelSizes`, transformers cache dir, `.onnx` file detection, HF `/tree/main/onnx` fetch picking q4f16/q4 ONNX files, `onnx-community/*-ONNX` model registry. | **[GUT]** remove chrome + ONNX file logic; native models load safetensors |
+| `repl/App.tsx` | 1,340 | `backend` prop ("auto"/"webgpu-native"/"onnx"), device switch webgpu/cpu, ttsModel kokoro / sttModel whisper defaults, transformers.js disposal comment. | **[REWIRE]** native-only |
+| `repl/index.tsx` | — | Chrome cleanup delay comment, backend prop. | **[GUT]** |
+| `repl/views/ModelView.tsx`, `ChatView.tsx`, `FrameworksView.tsx` | — | reference chrome-backend / onnx model ids. | **[GUT]** |
+
+### 1d. Skills / Integrations
+
+| File | usage | Disposition |
+|---|---|---|
+| `skills/builtin/transcribe.ts` (77) | Enum of whisper ONNX model ids; calls `g.transcribe`. | **[REWIRE]** to native STT model ids (Moonshine) |
+| `integrations/langchain.ts` | `GerbilEmbeddings` → `g.embed`/`g.embedBatch`. No direct ONNX, rides on core `embed()`. | **[KEEP]** (auto-native once core `embed()` is native) |
+| `integrations/ai-sdk.ts` | Embedding settings → `g.embed`. | **[KEEP]** |
+| `src/index.ts` | Exports `ChromeGPUBackend`, `getChromeCachedModels`, `WhisperSTT`, `WHISPER_MODELS`, `KokoroTTS`, `TTS_MODELS`, STT/TTS types. | **[REWIRE]** remove chrome + ONNX-class exports; export native equivalents |
+
+### 1e. Native engine (`src/gpu/`) — the target, all KEEP
+
+`index.ts` (`WebGPUEngine`: `generate`, `encodeImage`, `describeImage`, `embed`,
+`speak`, `transcribe`), `moonshine-stt.ts` / `moonshine-executor.ts` (native
+STT), `kani-tts.ts` (native TTS), `vision-executor.ts` / `vision-preprocess.ts`,
+`model-loader.ts` (safetensors), `architectures/*` (qwen2, qwen3_5, qwen3_5_vision,
+lfm2, gemma4, gemma4_vision, gemma3_encoder, moonshine, kani_tts, omnivoice).
+**No changes — this is what everything migrates onto.**
+
+### 1f. Build / deps
+
+- `scripts/build-worker.mjs` — bundles `worker-entry.ts` with transformers
+  inlined → `worker-code.generated.ts`. **[DELETE]** with the worker.
+- `package.json` scripts `build` and `pretypecheck` call `build-worker.mjs`.
+  **[EDIT]** remove that step.
+- `tsdown.config.ts` — `@huggingface/transformers` in `nodeExternals`; browser
+  build "BUNDLES transformers.js for clean DX". **[EDIT]** remove transformers
+  external + the bundling comment/path.
+- `package.json` deps to drop: `puppeteer-core`, `@huggingface/transformers`,
+  `onnxruntime-web`, `kokoro-js`; re-evaluate `@huggingface/hub`.
+
+---
+
+## 2. What Each ONNX Usage Did → Native Replacement
+
+| Modality | Old (ONNX) | New (native) | Wiring status |
+|---|---|---|---|
+| **Text gen** | transformers.js `AutoModelForCausalLM` (browser CDN + Node), chrome-backend (Node GPU) | `WebGPUEngine.generate` (qwen2/qwen3_5/lfm2/gemma4) | ✅ Wired in core `gerbil.ts` (`webgpu-native` branch) and browser `use-native-engine.ts`. ONNX text path now **DEAD** for GPU; still the CPU/WASM + chrome fallback. |
+| **Vision (VLM)** | `AutoModelForImageTextToText` (browser + chrome) | `WebGPUEngine.encodeImage` / `describeImage` (qwen3_5_vision, gemma4_vision) | ✅ Native exists. Core vision branch still has chrome fallback (~600-620) and browser worker has ONNX vision. **Partially dead.** |
+| **Embeddings** | `pipeline("feature-extraction", "Xenova/all-MiniLM-L6-v2")` (core + `use-embedding` worker) | `WebGPUEngine.embed` (EmbeddingGemma / Qwen3-Embedding) | ❌ **NOT wired** into core `embed()` or `use-embedding.ts`. Still 100% ONNX. **This is live, must be migrated, not just deleted.** |
+| **STT** | `WhisperSTT` (`pipeline("automatic-speech-recognition")`, onnx-community/whisper-*) | `MoonshineSTT` / `WebGPUEngine.transcribe` | ❌ **NOT wired** into core `transcribe()` (delegates to ONNX `WhisperSTT`). Native Moonshine sits unused in `src/gpu/`. **Live, must migrate.** |
+| **TTS** | Kokoro (`kokoro-js`) + Supertonic (`pipeline("text-to-speech")`) | `KaniTTS` / `WebGPUEngine.speak` | ❌ **NOT wired** into core `speak()` (delegates to ONNX `createTTS`). Native Kani unused outside `src/gpu/`. **Live, must migrate.** |
+
+**Dead vs deliberate-keep summary:**
+- **DEAD now (GPU mode):** ONNX text & vision generation paths in core + browser
+  worker. Safe to delete once consumers stop importing the worker.
+- **LIVE (must migrate first):** core `embed/embedBatch`, core STT (`WhisperSTT`),
+  core TTS (`createTTS`/Kokoro), and their browser hooks. These are the real
+  work — the "pure native" claim is aspirational for these three modalities at
+  the public-API layer.
+- **Deliberate keep:** all of `src/gpu/`; `use-native-engine.ts`; langchain/
+  ai-sdk integrations (ride on native `embed`).
+
+---
+
+## 3. Decision Points
+
+### 3a. CPU/WASM fallback (transformers.js wasm) — keep or remove?
+
+- **What it is:** `backend-selector.ts` wasm/cpu chains, the `pipeline()` CPU/WASM
+  branches in `gerbil.ts` (~488, ~525 fallback), iOS single-thread ORT wasm in
+  `use-embedding.ts` / worker. The *only* execution path when WebGPU is absent.
+- **Pure-native cost:** removing it = **WebGPU becomes a hard requirement.**
+  No fallback for: older Safari without WebGPU, headless CI without a GPU
+  adapter, locked-down enterprise browsers, very old Android.
+- **Recommendation:** the user wants pure native → **remove**, but ship a clear
+  capability error ("WebGPU required") rather than a silent failure. Flagged as
+  a **reach tradeoff** — quantify lost reach before committing. If any reach
+  matters, the cheap compromise is to keep `onnxruntime-web` for embeddings only
+  (smallest, most reuse-heavy) and drop everything else; but that keeps one ONNX
+  dep alive, so it half-defeats the goal.
+- **Decision owner: user.**
+
+### 3b. chrome-backend.ts (puppeteer headless Chrome) — safe to remove?
+
+- **What it was:** Node-side WebGPU via headless Chrome (puppeteer-core) before
+  node-dawn/native landed. Already deprecated; the native engine replaced it for
+  Node GPU inference.
+- **Hidden tentacles:** beyond inference it powers CLI/REPL features —
+  `getChromeCachedModels` (cache listing in `repl/utils.ts`, `index.ts`), the
+  `gerbil chrome` admin command (status/kill/ps/pkill of `gerbil/chrome-cache`),
+  and exports in `src/index.ts`. These must be detached first or builds break.
+- **Recommendation:** **YES, remove.** Lowest-risk *inference* removal, but do
+  it as a coordinated phase (detach CLI/REPL/index consumers → delete file →
+  drop `puppeteer-core`). 1,836 lines + a heavy dep gone.
+- **Decision: effectively yes; confirm no external consumer imports
+  `ChromeGPUBackend` from the package.**
+
+### 3c. Browser ONNX worker + hooks — remove once site/REPL stop using them?
+
+- **What it is:** `worker.ts` + `worker-entry.ts` + 1.35 MB generated bundle +
+  ONNX paths in `use-embedding`/`use-speech`/`use-voice-input`.
+- **Blocker:** chat must move to `use-native-engine`; embeddings/STT/TTS hooks
+  need native rewrites. The site migration doc (`docs/gerbil-site-native-
+  migration.md`, `docs/site-update-plan.md`) tracks consumer migration.
+- **Recommendation:** remove **after** native hooks exist and the site/REPL are
+  repointed. Biggest bundle-size win (drops the multi-MB inlined transformers).
+- **Decision: yes, sequenced last among browser work.**
+
+---
+
+## 4. REPL / CLI Audit
+
+The REPL/CLI is **partly native already** (it accepts `--backend native` →
+`webgpu-native` and the `webgpuEngine` path handles chat), but is riddled with
+ONNX/chrome assumptions:
+
+- **`cli/index.ts`**: every command exposes `--backend native|onnx|auto`
+  (default `auto`). `auto` + ONNX must go for native-only. The `transcribe`,
+  `tts`, and `voice` commands use ONNX STT/TTS (`whisper-tiny.en`, `kokoro-82m`)
+  — repoint to native Moonshine/Kani. The entire `chrome` admin subcommand
+  (status/kill/ps) is chrome-backend-only — **delete**. Cache listing scans the
+  transformers cache dir — switch to native cache dir.
+- **`repl/utils.ts`**: model registry is ONNX repos (`onnx-community/*-ONNX`,
+  `Phi-3-...-onnx`), cache scanning keys on `.onnx` files and the HF
+  `/tree/main/onnx` API, plus `getChromeCachedModels`. Rewrite registry to
+  native safetensors repos (`Qwen/Qwen3.5-0.8B`, etc. — the `nativeRepoMap` in
+  `gerbil.ts` already lists the mappings), scan for `.safetensors`, drop chrome.
+- **`repl/App.tsx` / `index.tsx` / views**: `backend` prop, cpu/webgpu device
+  switch, kokoro/whisper defaults, chrome cleanup delay. Collapse to native;
+  device switch becomes meaningful only if wasm fallback is kept (§3a).
+
+**Plan:** native-only REPL — remove the `onnx`/`auto` backend options (or make
+`auto` == native), delete the `chrome` command, repoint STT/TTS/embeddings model
+ids to native, rewrite cache/registry helpers to safetensors. If §3a keeps wasm,
+retain the cpu/webgpu switch; otherwise drop it.
+
+---
+
+## 5. Phased Removal Plan
+
+Ordering principle: **wire native replacements first, detach consumers, then
+delete; deps drop only after their last importer is gone.** Each phase ends with
+the no-regression suite (§6) green.
+
+### Phase 0 — Safety net (no deletion)
+- Snapshot current behavior: run `pnpm test` (vitest) + the native e2e scripts
+  (`scripts/engine/test-gerbil-e2e.mjs`, `test-vision-e2e.mjs`,
+  `test-embedding-gemma.mjs`, `test-moonshine-transcribe.mjs`,
+  `test-kani-speak.mjs`). Record baseline outputs/tokens.
+- **Risk: none.** Verifies: native paths actually work before we lean on them.
+
+### Phase 1 — Wire native embeddings / STT / TTS into the public API (ADD, don't delete)
+- Core `gerbil.ts`: make `embed`/`embedBatch` route to `WebGPUEngine.embed`;
+  make `transcribe`/`speak` route to native Moonshine/Kani (new thin wrappers
+  replacing `WhisperSTT`/`createTTS`).
+- Add native browser hooks (native embed, native speech) modeled on
+  `use-native-engine.ts`.
+- **Risk: MEDIUM** (behavioral parity — output dims, voices, languages).
+  Verify: embedding cosine-sim sanity, transcribe WER on a fixture, speak audio
+  length/sample-rate. **This phase makes the "pure native" claim true.**
+
+### Phase 2 — Remove chrome-backend (Node puppeteer GPU)
+- Detach: `cli/index.ts` (`chrome` command, cache listing), `repl/utils.ts`
+  (`getChromeCachedModels`/`refreshCachedModelSizes`), `repl` views, `src/index.ts`
+  exports, `gerbil.ts` chrome branches + passthroughs.
+- Delete `src/core/chrome-backend.ts`. Drop `puppeteer-core` from package.json.
+- **Risk: LOW–MEDIUM** (wide but mechanical; CLI admin command removed is a
+  user-visible change). Verify: build, CLI smoke (`gerbil "hi"` native), REPL
+  loads.
+
+### Phase 3 — Migrate browser chat to native, delete ONNX worker
+- Repoint site/REPL chat consumers to `use-native-engine`.
+- Delete `worker.ts`, `worker-entry.ts`, `worker-code.generated.ts`,
+  `scripts/build-worker.mjs`; remove the build/pretypecheck worker step;
+  fix `browser/index.ts` exports.
+- **Risk: MEDIUM** (iOS/Safari chat path was the worker's reason to exist;
+  confirm `use-native-engine` covers WKWebView per `docs/mobile-failure-
+  diagnosis.md`). Verify: browser e2e chat, iPad smoke.
+
+### Phase 4 — Remove ONNX embeddings / STT / TTS source
+- Delete `src/core/stt.ts`, `src/core/tts.ts`, `src/core/model-compat.ts`.
+- Gut/rewire `use-embedding.ts`, `use-speech.ts`, `use-voice-input.ts` to native.
+- Remove `gerbil.ts` ONNX imports (`AutoModel*`, `pipeline`), the `embedder`
+  ONNX field, model-compat call, getModelRepo ONNX maps.
+- Update `src/index.ts` exports (drop `WhisperSTT`/`WHISPER_MODELS`/`KokoroTTS`/
+  `TTS_MODELS` or replace with native).
+- Drop `kokoro-js` from package.json.
+- **Risk: MEDIUM.** Verify: STT/TTS/embedding suites; integration tests
+  (langchain/ai-sdk embeddings).
+
+### Phase 5 — Decide & remove CPU/WASM fallback + final dep purge (gated on §3a)
+- If removing fallback: delete wasm/cpu branches in `gerbil.ts` +
+  `backend-selector.ts` wasm chains; replace with "WebGPU required" error.
+- Remove `@huggingface/transformers` + `onnxruntime-web` from package.json and
+  `tsdown.config.ts` externals/bundling. Re-evaluate `@huggingface/hub`.
+- Native-only REPL cleanup (backend flags, device switch).
+- **Risk: HIGH** (reach loss is permanent; this is the one that needs the user's
+  explicit sign-off). Verify: full suite on a WebGPU-capable machine; confirm
+  graceful error on a no-WebGPU environment.
+
+### Net result
+~10,800 LOC removed; deps removed: `puppeteer-core`, `@huggingface/transformers`,
+`onnxruntime-web`, `kokoro-js` (+ maybe `@huggingface/hub`); the 1.35 MB worker
+bundle and its build step gone; one inference path (native) to maintain.
+
+---
+
+## 6. Subagent Execution Flow (for the removal workflow)
+
+Run as a **sequence of agents, one per phase** (phases are dependency-ordered;
+do NOT parallelize across phases — later phases assume earlier deletions). Within
+a phase, independent detach edits can be parallel sub-agents.
+
+Each agent's contract:
+1. **Read** this doc's relevant phase + the target files.
+2. **Make the phase's edits** (wire or delete per the disposition table).
+3. **Verify no regression:** `pnpm build` (TypeScript must pass — catches dangling
+   imports), `pnpm test` (vitest), and the phase-relevant native e2e scripts:
+   - Text: `scripts/engine/test-gerbil-e2e.mjs`
+   - Vision: `scripts/engine/test-vision-e2e.mjs`
+   - Embeddings: `scripts/engine/test-embedding-gemma.mjs`
+   - STT: `scripts/engine/test-moonshine-transcribe.mjs`
+   - TTS: `scripts/engine/test-kani-speak.mjs`
+4. **Grep-gate:** after a delete phase, `grep -rn "<removed symbol>" src/` must be
+   empty (e.g. `ChromeGPUBackend`, `@huggingface/transformers`, `kokoro-js`).
+5. **Report** diff summary + verification results to the orchestrator; STOP on any
+   red so a human can adjudicate (especially Phase 5 reach loss).
+
+Suggested agent lineup:
+- **Agent A (Phase 0):** baseline capture, no edits.
+- **Agent B (Phase 1):** native wiring for embed/STT/TTS — the gating change.
+- **Agent C (Phase 2):** chrome-backend removal + `puppeteer-core` drop.
+- **Agent D (Phase 3):** browser native chat migration + worker deletion.
+- **Agent E (Phase 4):** ONNX STT/TTS/embed source removal + `kokoro-js` drop.
+- **Agent F (Phase 5):** GATED — fallback decision, final dep purge, REPL cleanup.
+
+Orchestrator gates: B must pass before C–E delete the ONNX sources they replace;
+F runs only after explicit user sign-off on §3a.