Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠

Bonsai 2 2B Vision as an in-browser decision provider (frozen model + trained probe)

Claude96459ca71862
diff --git a/scripts/check-webgpu-ask.mjs b/scripts/check-webgpu-ask.mjs
index 0ecfedfe0..2507661b0 100644
--- a/scripts/check-webgpu-ask.mjs
+++ b/scripts/check-webgpu-ask.mjs
@@ -22,6 +22,7 @@
// --download download the model if it is not cached yet
// --tasks a,b run only these tasks
// --timeout-ms N per-task deadline (default 10 minutes)
+// --decoder-graph F install a replacement decoder graph in the cache first
import assert from 'node:assert/strict';
import fs from 'node:fs/promises';
import http from 'node:http';
@@ -142,6 +143,17 @@ try {
if (Date.now() - started > 60 * 60_000) throw new Error('Download did not finish within an hour');
}
}
+ if (option('--decoder-graph', '')) {
+ // Replace the cached decoder graph (same external weights) before the
+ // first chat, to check a re-exported graph through the real extension.
+ const data = (await fs.readFile(option('--decoder-graph'))).toString('base64');
+ await harness.evaluate(async ({ url, data }) => {
+ const bytes = Uint8Array.from(atob(data), c => c.charCodeAt(0));
+ const cache = await caches.open('transformers-cache');
+ await cache.put(url, new Response(bytes, { headers: { 'content-type': 'application/octet-stream', 'content-length': String(bytes.length) } }));
+ }, { url: `https://huggingface.co/${model}/resolve/main/onnx/decoder_model_merged_q2f16.onnx`, data });
+ result.decoderGraph = option('--decoder-graph');
+ }
// The provider can be selected for chat only once its files are cached.
const selected = await message('set_active_provider', { providerId: 'webgpu' });
assert.equal(selected?.ok, true, `Could not select the WebGPU provider: ${JSON.stringify(selected)}`);
diff --git a/scripts/check-webgpu-decisions.mjs b/scripts/check-webgpu-decisions.mjs
index 14ffb00b5..ed2070b0f 100644
--- a/scripts/check-webgpu-decisions.mjs
+++ b/scripts/check-webgpu-decisions.mjs
@@ -16,7 +16,8 @@
// the case where dense attention memory used to grow
// --limit N first N test requests only
// --temperature T Bonsai readout temperature (default 1)
-// --prompt NAME Bonsai decision prompt: default, strict or verify
+// --prompt NAME Bonsai decision mode: probe (default, what Settings uses),
+// readout, strict or verify
// --decoder-graph F install a replacement decoder graph in the cache first
// --return-hidden with --prompt verify, record each option's hidden state
import fs from 'node:fs/promises';
@@ -157,9 +158,16 @@ try {
result.decoderGraph = graphFile;
}
const temperature = Number(option('--temperature', 1));
- const promptVariant = option('--prompt', 'default');
+ const promptVariant = option('--prompt', 'probe');
const returnHidden = args.includes('--return-hidden');
await run('bonsai', fixture => message('webgpu_decide_fixture', { fixture: { ...fixture, temperature, promptVariant, returnHidden } }));
+ if (promptVariant === 'probe') {
+ // The decision provider path Settings uses: judge -> provider -> worker.
+ const started = Date.now();
+ const tested = await message('test_system_one', { settings: { decisionProvider: 'webgpu_chat', systemOneEnabled: true, decisionVisionMode: 'on' } });
+ result.engines.bonsai.settingsTest = { ...tested, seconds: (Date.now() - started) / 1000 };
+ console.log('settings test', JSON.stringify(result.engines.bonsai.settingsTest));
+ }
}
} catch (error) {
result.error = String(error?.stack || error);
diff --git a/scripts/models/bonsai/train_decision_probe.py b/scripts/models/bonsai/train_decision_probe.py
new file mode 100644
index 000000000..07ca18790
--- /dev/null
+++ b/scripts/models/bonsai/train_decision_probe.py
@@ -0,0 +1,185 @@
+"""Train a frozen-backbone decision probe for the Bonsai 2 Vision readout.
+
+Inputs are hardware-eval reports from scripts/check-webgpu-decisions.mjs run
+with `--prompt verify --return-hidden` (one last-position hidden state and one
+Yes-minus-No margin per option) on d1's balanced-V3 train, validation and
+final-test splits, plus the split files for the labels.
+
+Each option is scored as a * margin + w . standardize(hidden) + b_type, and a
+softmax over a question's options gives its answer, the same grouped readout
+d1 uses over its option markers. Only 2048 + a few parameters are trained;
+the regularization strength is chosen on validation, and the test split is
+scored once with the chosen model.
+
+ python scripts/models/bonsai/train_decision_probe.py --reports <dir> --data <split dir> --out probe.json
+"""
+import argparse
+import base64
+import json
+from pathlib import Path
+
+import numpy as np
+from scipy.optimize import minimize
+
+TYPES = ("choice", "noul", "score")
+
+
+def decode(text):
+ return np.frombuffer(base64.b64decode(text), dtype=np.float16).astype(np.float32)
+
+
+def load(report_path, split_path):
+ report = json.loads(Path(report_path).read_text())
+ answers = report["engines"]["bonsai"]["answers"]
+ split = json.loads(Path(split_path).read_text())
+ rows = []
+ for request in split["requests"]:
+ for entry in request["questions"]:
+ answer = answers.get(entry["id"])
+ if not answer or not answer.get("hidden") or any(h is None for h in answer["hidden"]):
+ continue
+ question = entry["question"]
+ if question["type"] == "noul":
+ names = ["no", "yes"]
+ elif question["type"] == "score":
+ names = [str(i) for i in range(len(question["criteria"]))]
+ else:
+ names = list(question["criteria"].keys())
+ label = entry["label"]
+ target = names.index(label["selected_option"]) if question["type"] != "score" else label["selected_index"]
+ rows.append(dict(entry=entry, type=question["type"], names=names, target=target,
+ margins=np.array(answer["margins"], dtype=np.float64),
+ hidden=np.stack([decode(h) for h in answer["hidden"]]).astype(np.float64)))
+ return rows
+
+
+def scores(params, row, mean, std):
+ width = mean.shape[0]
+ a, w = params[0], params[1:1 + width]
+ bias = params[1 + width + TYPES.index(row["type"])]
+ return a * row["margins"] + ((row["hidden"] - mean) / std) @ w + bias
+
+
+def objective(params, rows, mean, std, l2):
+ width = mean.shape[0]
+ loss, grad = 0.0, np.zeros_like(params)
+ for row in rows:
+ s = scores(params, row, mean, std)
+ s = s - s.max()
+ p = np.exp(s) / np.exp(s).sum()
+ loss -= np.log(p[row["target"]] + 1e-12)
+ d = p.copy()
+ d[row["target"]] -= 1
+ z = (row["hidden"] - mean) / std
+ grad[0] += d @ row["margins"]
+ grad[1:1 + width] += d @ z
+ grad[1 + width + TYPES.index(row["type"])] += d.sum()
+ n = len(rows)
+ loss = loss / n + l2 * np.sum(params[1:1 + width] ** 2)
+ grad = grad / n
+ grad[1:1 + width] += 2 * l2 * params[1:1 + width]
+ return loss, grad
+
+
+def answers_for(rows, params, mean, std):
+ out = {}
+ for row in rows:
+ s = scores(params, row, mean, std)
+ p = np.exp(s - s.max())
+ p /= p.sum()
+ q = row["entry"]["question"]
+ if q["type"] == "noul":
+ out[row["entry"]["id"]] = {"type": "noul", "noul": float(p[1])}
+ elif q["type"] == "score":
+ out[row["entry"]["id"]] = {"type": "score", "score": float(p @ np.arange(len(p))),
+ "probabilities": {n: float(v) for n, v in zip(row["names"], p)}}
+ else:
+ out[row["entry"]["id"]] = {"type": "choice", "choice": row["names"][int(p.argmax())],
+ "probabilities": {n: float(v) for n, v in zip(row["names"], p)}}
+ return out
+
+
+def metrics(rows, answers):
+ # Mirrors scripts/models/d1/decision-metrics.mjs.
+ t = dict(categorical=0, correct=0, pos=0, neg=0, tp=0, fp=0, unk=0, unk_ok=0, score_n=0, score_err=0.0, nll=0.0)
+ for row in rows:
+ entry, answer, q = row["entry"], answers[row["entry"]["id"]], row["entry"]["question"]
+ label = entry["label"]
+ if q["type"] == "score":
+ t["score_n"] += 1
+ t["score_err"] += abs(answer["score"] - label["expected_level"])
+ continue
+ t["categorical"] += 1
+ guess = answer["choice"] if q["type"] == "choice" else ("yes" if answer["noul"] >= 0.5 else "no")
+ right = guess == label["selected_option"]
+ t["correct"] += right
+ if entry.get("is_completion_claim"):
+ positives = entry.get("completion_positive_options", [])
+ if label["selected_option"] in positives:
+ t["pos"] += 1
+ t["tp"] += guess in positives
+ else:
+ t["neg"] += 1
+ t["fp"] += guess in positives
+ unknown = (entry.get("uncertainty_metadata") or {}).get("unknown_option")
+ if unknown and label["selected_option"] == unknown:
+ t["unk"] += 1
+ t["unk_ok"] += right
+ return {"categorical": f"{t['correct']}/{t['categorical']}", "categorical_accuracy": t["correct"] / max(1, t["categorical"]),
+ "completion_found": f"{t['tp']}/{t['pos']}",
+ "false_completions": f"{t['fp']}/{t['neg']}", "unknown": f"{t['unk_ok']}/{t['unk']}",
+ "score_mae": t["score_err"] / max(1, t["score_n"])}
+
+
+def fit(rows, mean, std, l2):
+ width = mean.shape[0]
+ start = np.zeros(1 + width + len(TYPES))
+ start[0] = 1.0 # begin at the zero-shot verify readout
+ result = minimize(objective, start, args=(rows, mean, std, l2), jac=True, method="L-BFGS-B", options={"maxiter": 500})
+ return result.x
+
+
+def main():
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--reports", type=Path, required=True)
+ parser.add_argument("--data", type=Path, required=True)
+ parser.add_argument("--out", type=Path, required=True)
+ args = parser.parse_args()
+ train = load(args.reports / "probe-train.json", args.data / "train.json")
+ val = load(args.reports / "probe-val.json", args.data / "validation.json")
+ test = load(args.reports / "probe-test.json", args.data / "final_test.json")
+ stacked = np.concatenate([row["hidden"] for row in train])
+ mean, std = stacked.mean(0), stacked.std(0) + 1e-3
+ width = mean.shape[0]
+ zero = np.zeros(1 + width + len(TYPES))
+ zero[0] = 1.0
+ report = {"rows": {"train": len(train), "validation": len(val), "test": len(test)},
+ "zero_shot_verify": {"validation": metrics(val, answers_for(val, zero, mean, std)),
+ "test": metrics(test, answers_for(test, zero, mean, std))},
+ "candidates": []}
+ best = None
+ for l2 in [1e-4, 1e-3, 1e-2, 3e-2, 1e-1, 3e-1, 1.0]:
+ params = fit(train, mean, std, l2)
+ val_loss, _ = objective(params, val, mean, std, 0.0)
+ val_metrics = metrics(val, answers_for(val, params, mean, std))
+ accuracy = val_metrics["categorical_accuracy"]
+ candidate = {"l2": l2, "validation_nll": float(val_loss), "validation": val_metrics,
+ "train": metrics(train, answers_for(train, params, mean, std))}
+ report["candidates"].append(candidate)
+ print(json.dumps(candidate), flush=True)
+ key = (accuracy, -val_loss)
+ if best is None or key > best[0]:
+ best = (key, l2, params)
+ _, l2, params = best
+ report["selected_l2"] = l2
+ report["test"] = metrics(test, answers_for(test, params, mean, std))
+ report["margin_weight"] = float(params[0])
+ report["type_bias"] = {t: float(params[1 + width + i]) for i, t in enumerate(TYPES)}
+ args.out.write_text(json.dumps(report, indent=2) + "\n")
+ np.savez(args.out.with_suffix(".npz"), params=params, mean=mean, std=std)
+ print("selected l2", l2, "test", json.dumps(report["test"]))
+ print("zero-shot verify test", json.dumps(report["zero_shot_verify"]["test"]))
+
+
+if __name__ == "__main__":
+ main()
diff --git a/src/chrome/src/agent/decision-config.js b/src/chrome/src/agent/decision-config.js
index eb9616e3f..ee897cc34 100644
--- a/src/chrome/src/agent/decision-config.js
+++ b/src/chrome/src/agent/decision-config.js
@@ -7,10 +7,16 @@ export const DECISION_SETTINGS_KEYS = [
D1_CONSENT_KEY,
];
+// The in-browser chat model answers decisions through its trained probe
+// (Chrome; src/chrome/src/offscreen/bonsai-decision-probe.js). Kept in sync
+// with WEBGPU_DECISION_MODEL_ID in providers/webgpu.js.
+export const WEBGPU_CHAT_DECISION_MODEL = 'webbrain-one/Ternary-Bonsai-2-2B-ONNX-WebGPU-Vision';
+
// Resolve legacy settings without rewriting credentials or opted-in features.
export function resolveDecisionConfig(stored = {}, compass = null) {
const explicitD1 = stored.decisionProvider === 'webgpu_d1' && stored.systemOneEnabled === true && stored[D1_CONSENT_KEY] === D1_CONSENT_VERSION;
- const provider = explicitD1 ? 'webgpu_d1' : compass ? 'compass' : (stored.decisionProvider || (stored.typesafeApiKey ? 'typesafe' : 'openrouter'));
+ const explicitChat = stored.decisionProvider === 'webgpu_chat' && stored.systemOneEnabled === true;
+ const provider = explicitD1 ? 'webgpu_d1' : explicitChat ? 'webgpu_chat' : compass ? 'compass' : (stored.decisionProvider || (stored.typesafeApiKey ? 'typesafe' : 'openrouter'));
if (provider === 'webgpu_d1') return {
provider, model: D1_MODEL_ID, apiKey: '', baseUrl: '', url: '', local: true,
enabled: stored.systemOneEnabled === true && stored[D1_CONSENT_KEY] === D1_CONSENT_VERSION,
@@ -20,6 +26,17 @@ export function resolveDecisionConfig(stored = {}, compass = null) {
requiresIndependentSuccessConfirmation: true,
config: { category: 'local', providerName: provider, inputCostPerMillionUsd: 0, outputCostPerMillionUsd: 0 },
};
+ // Explicitly selected, like D1: an on-device decision provider that never
+ // leaves the browser. A positive completion still needs independent proof.
+ if (provider === 'webgpu_chat') return {
+ provider, model: WEBGPU_CHAT_DECISION_MODEL, apiKey: '', baseUrl: '', url: '', local: true,
+ enabled: stored.systemOneEnabled === true,
+ doneEnabled: stored.systemOneDoneEnabled === true,
+ threshold: Number.isFinite(stored.systemOneDoneThreshold) && stored.systemOneDoneThreshold >= .5 && stored.systemOneDoneThreshold <= .99 ? stored.systemOneDoneThreshold : .9,
+ supportsVision: stored.decisionVisionMode !== 'off', experimental: true,
+ requiresIndependentSuccessConfirmation: true,
+ config: { category: 'local', providerName: provider, inputCostPerMillionUsd: 0, outputCostPerMillionUsd: 0 },
+ };
if (!['compass', 'openrouter', 'typesafe', 'local'].includes(provider)) throw new Error('Unknown decision provider.');
const defaults = { compass: DEFAULT_DECISION_MODEL, openrouter: DEFAULT_DECISION_MODEL, typesafe: 'jev-1.13.0', local: 'kev-latest' };
const model = compass ? DEFAULT_DECISION_MODEL : String(stored.decisionModel || defaults[provider]).trim();
@@ -49,6 +66,7 @@ export function resolveDecisionConfig(stored = {}, compass = null) {
}
export async function listDecisionModels(config, fetchImpl = globalThis.fetch, signal) {
+ if (config.provider === 'webgpu_chat') return [{ id: WEBGPU_CHAT_DECISION_MODEL, name: 'Ternary Bonsai 2 2B Vision (in-browser, experimental)', supportsVision: true, inputRate: 0, outputRate: 0 }];
if (config.provider === 'webgpu_d1') return [{ id: D1_MODEL_ID, name: 'D1 FP16 (experimental, on-device)', supportsVision: true, inputRate: 0, outputRate: 0 }];
return withCompletionTimeout(requestSignal => discoverDecisionModels(config, fetchImpl, requestSignal), signal, 5000);
}
diff --git a/src/chrome/src/agent/systemone-judge.js b/src/chrome/src/agent/systemone-judge.js
index 1b3b06ffb..a66a0d4c7 100644
--- a/src/chrome/src/agent/systemone-judge.js
+++ b/src/chrome/src/agent/systemone-judge.js
@@ -5,6 +5,7 @@
// interface. It never performs an action or turns a non-success into success.
import { evaluateD1 } from '../providers/d1.js';
+import { evaluateWebgpuDecision } from '../providers/webgpu-decision.js';
export const SYSTEM_ONE_API_URL = 'https://api.typesafe.ai/v1/systemone';
export const SYSTEM_ONE_MODEL = 'jev-1.13.0';
@@ -274,14 +275,26 @@ export function createSystemOneJudge({
return {
async evaluate({ apiKey, state, images, questions, signal, beforeRequest, onUsage, config, headers = {}, metadata = {} } = {}) {
const key = normalizeTypesafeApiKey(config?.apiKey ?? apiKey);
- if (!key && !['local', 'compass', 'webgpu_d1'].includes(config?.provider)) throw new Error('Decision API key is not configured.');
+ if (!key && !['local', 'compass', 'webgpu_d1', 'webgpu_chat'].includes(config?.provider)) throw new Error('Decision API key is not configured.');
const model = config?.model || SYSTEM_ONE_MODEL;
validateState(state, config?.supportsVision === true);
validateQuestions(questions);
const started = now();
- const request = requestSignal(signal, Math.max(1, Math.min(5000, Number(timeoutMs) || 5000)));
+ // The in-browser model prefills its screenshot on this GPU (about a second
+ // on a desktop card) and loads on first use after a browser restart.
+ const cap = config?.provider === 'webgpu_chat' ? 15000 : 5000;
+ const request = requestSignal(signal, Math.max(1, Math.min(cap, Number(timeoutMs) || cap)));
const retries = Math.max(0, Math.min(2, Math.floor(Number(maxRetries) || 0)));
try {
+ if (config?.provider === 'webgpu_chat') {
+ if (beforeRequest) await abortable(beforeRequest(), request.signal);
+ const result = await abortable(evaluateWebgpuDecision({ state, images, questions, config, signal: request.signal }), request.signal);
+ validateSystemOneAnswers(result.answers, questions);
+ if (result.model !== model || !Number.isInteger(result.usage?.input_tokens) || result.usage.input_tokens < 0 || result.usage?.output_tokens !== 0) throw systemOneError('JEV_INVALID_USAGE', 'Invalid in-browser decision usage/model.');
+ const metadata = { model: result.model, provider: result.provider, usage: result.usage, latencyMs: Math.max(0, now() - started), estimatedCostUsd: 0 };
+ if (onUsage) await abortable(onUsage(metadata), request.signal);
+ return { ...result, ...metadata };
+ }
if (config?.provider === 'webgpu_d1') {
if (beforeRequest) await abortable(beforeRequest(), request.signal);
const result = await abortable(evaluateD1({ state, images, questions, config, signal: request.signal }), request.signal);
diff --git a/src/chrome/src/config-transfer.js b/src/chrome/src/config-transfer.js
index 82266b412..889e10cda 100644
--- a/src/chrome/src/config-transfer.js
+++ b/src/chrome/src/config-transfer.js
@@ -236,7 +236,7 @@ function sanitizeProviders(value, { strict = false } = {}) {
}
function validSettingValue(key, value) {
- if (key === 'decisionProvider') return ['', 'openrouter', 'typesafe', 'local', 'webgpu_d1'].includes(value);
+ if (key === 'decisionProvider') return ['', 'openrouter', 'typesafe', 'local', 'webgpu_d1', 'webgpu_chat'].includes(value);
if (key === 'decisionVisionMode') return ['auto', 'on', 'off'].includes(value);
if (key === 'decisionBaseUrl') {
if (typeof value !== 'string') return false;
@@ -283,7 +283,7 @@ function normalizeSettings(source, { strict = false } = {}) {
settings.providers = sanitizeProviders(settings.providers, { strict });
if (!Object.hasOwn(source, 'decisionInputRate')) {
settings.decisionInputRate = source.decisionProvider === 'typesafe' || (!source.decisionProvider && source.typesafeApiKey)
- ? .042 : ['local', 'webgpu_d1'].includes(source.decisionProvider) ? 0 : .02;
+ ? .042 : ['local', 'webgpu_d1', 'webgpu_chat'].includes(source.decisionProvider) ? 0 : .02;
}
return settings;
}
diff --git a/src/chrome/src/offscreen/bonsai-decision-probe.js b/src/chrome/src/offscreen/bonsai-decision-probe.js
new file mode 100644
index 000000000..814cb8b50
--- /dev/null
+++ b/src/chrome/src/offscreen/bonsai-decision-probe.js
@@ -0,0 +1,40 @@
+// Decision probe for Ternary Bonsai 2 2B Vision, trained by
+// scripts/models/bonsai/train_decision_probe.py on d1's balanced-V3 train split
+// (synthetic, AI-labelled screens), regularization chosen on its validation
+// split. The backbone is frozen: each option of a question is scored from its
+// verify continuation as margin * BONSAI_DECISION_PROBE.margin + hidden . weights, where margin is
+// the Yes-minus-No logit and hidden is the decoder's last-position state
+// (decision_hidden). Standardization is folded into the weights; constant
+// offsets cancel in the softmax over a question's options.
+//
+// Validation: 110/126 categorical, 21/24 completions found, 3/48 false
+// completions. Test: 110/126 categorical, 21/24 found, 4/48 false completions.
+export const BONSAI_DECISION_PROBE = Object.freeze({
+ model: 'webbrain-one/Ternary-Bonsai-2-2B-ONNX-WebGPU-Vision',
+ hiddenSize: 2048,
+ margin: 1.3493167743628656,
+ weightsBase64: '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',
+});
+
+let weights = null;
+function probeWeights() {
+ if (!weights) {
+ const binary = atob(BONSAI_DECISION_PROBE.weightsBase64);
+ const bytes = new Uint8Array(binary.length);
+ for (let i = 0; i < binary.length; i++) bytes[i] = binary.charCodeAt(i);
+ weights = new Float32Array(bytes.buffer);
+ }
+ return weights;
+}
+
+/** Per-option scores for one question; a softmax over them is the answer. */
+export function bonsaiProbeScores(margins, hidden) {
+ const w = probeWeights();
+ return margins.map((margin, index) => {
+ const vector = hidden[index];
+ if (!vector || vector.length !== w.length) throw new Error('The decision probe needs the decoder hidden state.');
+ let dot = 0;
+ for (let i = 0; i < w.length; i++) dot += w[i] * vector[i];
+ return margin * BONSAI_DECISION_PROBE.margin + dot;
+ });
+}
diff --git a/src/chrome/src/offscreen/inference-worker.js b/src/chrome/src/offscreen/inference-worker.js
index 6f532bc95..b403f1eb4 100644
--- a/src/chrome/src/offscreen/inference-worker.js
+++ b/src/chrome/src/offscreen/inference-worker.js
@@ -17,6 +17,7 @@ import {
decisionStateText,
normalizeDecisionRequest,
} from './webgpu-decisions.js';
+import { BONSAI_DECISION_PROBE, bonsaiProbeScores } from './bonsai-decision-probe.js';
let libraryPromise = null;
let libraryVersion = null;
@@ -61,10 +62,12 @@ const WEBGPU_BONSAI2_MODEL_IDS = new Set([
// scripts/models/bonsai/chunk_gdn_prefill.py). Earlier revisions prefilled with
// an ONNX Scan that WebGPU runs one token at a time on the CPU. Installs that
// cached that graph keep it under the same URL, so replace just this small file;
-// the unchanged weight files stay cached.
+// the unchanged weight files stay cached. The 2B graph also exposes the
+// last-position hidden state as `decision_hidden` for the decision probe
+// (scripts/models/bonsai/add_decision_hidden.py); logits and caches are as before.
const BONSAI2_DECODER_GRAPH_FILE = 'onnx/decoder_model_merged_q2f16.onnx';
const BONSAI2_DECODER_GRAPH_SHA256 = new Map([
- ['webbrain-one/Ternary-Bonsai-2-2B-ONNX-WebGPU-Vision', '2768d880f4fdab46a984d697b901db6e8ac54dc9b896524cb8cb769bbb339df0'],
+ ['webbrain-one/Ternary-Bonsai-2-2B-ONNX-WebGPU-Vision', '240c07e1191911a2306faf106ee537382a1b4d5e6ecdd32b618da160db9145c1'],
['webbrain-one/Ternary-Bonsai-2-4B-ONNX-WebGPU-Vision', '52424fbb7094ed0a1036d8795e78c3da349fce1838ee6bd077523929b0bb7bb2'],
]);
// A refresh must never stall model loading on a flaky network; one attempt per
@@ -1884,7 +1887,15 @@ async function runMultimodalDecision(payload) {
try {
cache = await prefillBonsai2Cache(runtime, prefix, bonsai2PrefillSliceTokens(payload?.prefillSliceTokens), prefixLength);
const prefillMs = performance.now() - started;
- const verify = payload?.promptVariant === 'verify';
+ // `probe` (the default) scores each option's verify continuation with the
+ // trained probe; `verify` returns the raw Yes margins; `readout` and
+ // `strict` read the option codes in one pass. All but `probe` are for the
+ // diagnostic fixture path.
+ const variant = ['readout', 'default', 'strict', 'verify', 'probe'].includes(payload?.promptVariant) ? payload.promptVariant : 'probe';
+ const verify = variant === 'verify' || variant === 'probe';
+ if (variant === 'probe' && modelId !== BONSAI_DECISION_PROBE.model) {
+ throw new Error(`No decision probe is trained for ${modelId}.`);
+ }
const returnHidden = payload?.returnHidden === true;
const yesId = verify ? codeId('Yes') : null, noId = verify ? codeId('No') : null;
for (const [id, question] of request.entries) {
@@ -1898,9 +1909,16 @@ async function runMultimodalDecision(payload) {
suffixTokens += suffix.input_ids.dims.at(-1);
const output = await bonsai2ContinuationLogits(runtime, prefix, cache, suffix.input_ids, { withHidden: true });
margins.push(output.logits[yesId] - output.logits[noId]);
- if (returnHidden) hidden.push(output.hidden ? encodeFloat16Base64(output.hidden) : null);
+ if (variant === 'probe' && !output.hidden) {
+ throw new Error('In-browser decisions need the updated Ternary Bonsai 2 2B Vision decoder. Download the model again in Settings > Providers > WebGPU.');
+ }
+ hidden.push(output.hidden);
}
- answers[id] = { ...decisionAnswer(question, margins, { temperature }), ...(returnHidden ? { margins, hidden } : {}) };
+ const scores = variant === 'probe' ? bonsaiProbeScores(margins, hidden) : margins;
+ answers[id] = {
+ ...decisionAnswer(question, scores, { temperature }),
+ ...(returnHidden ? { margins, hidden: hidden.map(vector => (vector ? encodeFloat16Base64(vector) : null)) } : {}),
+ };
questionMs.push(Math.round(performance.now() - questionStarted));
continue;
}
@@ -1930,7 +1948,7 @@ async function runMultimodalDecision(payload) {
question_ms: questionMs,
total_ms: Math.round(performance.now() - started),
temperature,
- prompt_variant: payload?.promptVariant || 'default',
+ prompt_variant: variant,
},
};
} catch (error) {
diff --git a/src/chrome/src/providers/webgpu-decision.js b/src/chrome/src/providers/webgpu-decision.js
new file mode 100644
index 000000000..c483c4295
--- /dev/null
+++ b/src/chrome/src/providers/webgpu-decision.js
@@ -0,0 +1,12 @@
+import { WebGPUProvider, WEBGPU_DECISION_MODEL_ID } from './webgpu.js';
+
+// Decisions on the in-browser chat model (Ternary Bonsai 2 2B Vision and its
+// trained probe). Runs in the offscreen WebGPU worker; nothing leaves the browser.
+export async function evaluateWebgpuDecision({ state, images, questions, config, signal } = {}) {
+ if (config?.model !== WEBGPU_DECISION_MODEL_ID) throw new Error('Only Ternary Bonsai 2 2B Vision answers in-browser decisions.');
+ if (signal?.aborted) throw signal.reason;
+ const provider = new WebGPUProvider({ model: WEBGPU_DECISION_MODEL_ID });
+ const result = await provider.decide({ state, images, questions });
+ if (signal?.aborted) throw signal.reason;
+ return { model: result.model, provider: 'webgpu_chat', answers: result.answers, usage: result.usage };
+}
diff --git a/src/chrome/src/providers/webgpu.js b/src/chrome/src/providers/webgpu.js
index 6ca6749c0..d89aa7d36 100644
--- a/src/chrome/src/providers/webgpu.js
+++ b/src/chrome/src/providers/webgpu.js
@@ -202,8 +202,9 @@ export const WEBGPU_HIDDEN_MODEL_IDS = Object.freeze([
export const WEBGPU_SELECTABLE_MODEL_PRESETS = Object.freeze(WEBGPU_PICKER_MODEL_IDS
.filter(id => !WEBGPU_HIDDEN_MODEL_IDS.includes(id))
.map(id => WEBGPU_MODEL_PRESETS.find(preset => preset.id === id)));
-// Presets whose chunked decoder supports the decision readout.
-export const WEBGPU_DECISION_MODEL_IDS = new Set([WEBGPU_BONSAI2_VL_2B_MODEL_ID, WEBGPU_BONSAI2_VL_4B_MODEL_ID]);
+// The preset with a trained decision probe (src/chrome/src/offscreen/bonsai-decision-probe.js).
+export const WEBGPU_DECISION_MODEL_ID = WEBGPU_BONSAI2_VL_2B_MODEL_ID;
+export const WEBGPU_DECISION_MODEL_IDS = new Set([WEBGPU_DECISION_MODEL_ID]);
export const WEBGPU_MODEL_NOT_READY_ERROR = `${WEBGPU_COMPASS_TINY_V2_MODEL_ID} is not downloaded. Open Settings > Providers > WebGPU or Apocalypse Mode > WebGPU to download it before chatting.`;
// Chrome-only selection state. Keep this separate from the synced
// `visionModel` endpoint so enabling the fallback never overwrites a user's
@@ -495,7 +496,7 @@ export class WebGPUProvider extends WebGPUOffscreenProvider {
images,
questions,
...(Number(temperature) > 0 ? { temperature: Number(temperature) } : {}),
- ...(['strict', 'verify'].includes(promptVariant) ? { promptVariant } : {}),
+ ...(['readout', 'strict', 'verify', 'probe'].includes(promptVariant) ? { promptVariant } : {}),
...(returnHidden === true ? { returnHidden: true } : {}),
...(prefillSliceTokens == null ? {} : { prefillSliceTokens }),
});
diff --git a/src/chrome/src/ui/locales/decision-copy.mjs b/src/chrome/src/ui/locales/decision-copy.mjs
index 51a9e4a74..4b5b1b919 100644
--- a/src/chrome/src/ui/locales/decision-copy.mjs
+++ b/src/chrome/src/ui/locales/decision-copy.mjs
@@ -1,5 +1,7 @@
export const decisionEnglish = {
'st.d1.provider': 'Local decision model - Experimental',
+ 'st.decision.webgpu_chat': 'In-browser chat model (Ternary Bonsai 2 2B Vision) - Experimental',
+ 'st.decision.webgpu_chat_note': 'Answers decisions with the in-browser WebGPU chat model and a small trained readout, with no extra download. Download Ternary Bonsai 2 2B Vision under Providers first. The first decision after a browser restart loads the model and may fall back to the active model. Not a sole completion authority.',
'st.d1.warning': 'Experimental on-device decision model, about 0.95 GB download plus tokenizer. Requires hardware WebGPU and substantial GPU memory. Not a chat model or sole completion authority. CPU/WASM control operators may run. Browser support and model decisions are not universally validated.',
'st.d1.consent': 'I accept the experimental local model download from Hugging Face and its memory requirements.',
'st.d1.need_consent': 'Accept the experimental model download consent first.',
@@ -25,6 +27,8 @@ export const decisionEnglish = {
export const decisionTranslations = {
tr: {
'st.d1.provider': 'Yerel karar modeli — Deneysel',
+ 'st.decision.webgpu_chat': 'Tarayıcı içi sohbet modeli (Ternary Bonsai 2 2B Vision) — Deneysel',
+ 'st.decision.webgpu_chat_note': 'Kararları tarayıcı içi WebGPU sohbet modeli ve küçük bir eğitilmiş okuma katmanıyla verir; ek indirme gerekmez. Önce Sağlayıcılar altında Ternary Bonsai 2 2B Vision modelini indirin. Tarayıcı yeniden başladıktan sonraki ilk karar modeli yükler ve etkin modele geri düşebilir. Tek başına tamamlanma yetkilisi değildir.',
'st.d1.warning': 'Deneysel yerel karar modeli; tokenizer ile birlikte yaklaşık 0,95 GB indirme. Donanım WebGPU ve yüksek GPU belleği gerekir. Sohbet modeli veya tek başına tamamlanma yetkilisi değildir. CPU/WASM kontrol işlemleri çalışabilir. Her tarayıcı, GPU ve model kararı doğrulanmış değildir.',
'st.d1.consent': 'Hugging Face üzerinden deneysel yerel model indirmesini ve bellek gereksinimlerini kabul ediyorum.',
'st.d1.need_consent': 'Önce deneysel model indirme onayını kabul edin.',
diff --git a/src/chrome/src/ui/settings.html b/src/chrome/src/ui/settings.html
index cac927937..1e3e8829c 100644
--- a/src/chrome/src/ui/settings.html
+++ b/src/chrome/src/ui/settings.html
@@ -2261,8 +2261,9 @@
</div>
<div class="field"><label for="decision-provider" data-i18n="st.decision.provider"></label>
- <select id="decision-provider"><option value="openrouter">OpenRouter</option><option value="typesafe">TypeSafe</option><option value="local" data-i18n="st.decision.local"></option><option value="webgpu_d1" data-i18n="st.d1.provider"></option></select>
+ <select id="decision-provider"><option value="openrouter">OpenRouter</option><option value="typesafe">TypeSafe</option><option value="local" data-i18n="st.decision.local"></option><option value="webgpu_d1" data-i18n="st.d1.provider"></option><option value="webgpu_chat" data-i18n="st.decision.webgpu_chat"></option></select>
</div>
+ <p class="settings-description" id="webgpu-decision-note" hidden data-i18n="st.decision.webgpu_chat_note"></p>
<div id="d1-controls" hidden>
<p class="settings-description" data-i18n="st.d1.warning"></p>
<label><input type="checkbox" id="d1-consent"> <span data-i18n="st.d1.consent"></span></label>
diff --git a/src/chrome/src/ui/settings.js b/src/chrome/src/ui/settings.js
index 954fe5abd..2c2f812b7 100644
--- a/src/chrome/src/ui/settings.js
+++ b/src/chrome/src/ui/settings.js
@@ -1,4 +1,4 @@
-import { DECISION_SETTINGS_KEYS, DEFAULT_DECISION_MODEL, resolveDecisionConfig } from '../agent/decision-config.js';
+import { DECISION_SETTINGS_KEYS, DEFAULT_DECISION_MODEL, WEBGPU_CHAT_DECISION_MODEL, resolveDecisionConfig } from '../agent/decision-config.js';
import { D1_MODEL_ID } from '../providers/d1-config.js';
import { initD1Settings } from './d1-settings.js';
/**
@@ -239,6 +239,19 @@ const refreshD1Settings = initD1Settings(sendToBackground);
let decisionVisionSupported = false;
const decisionKeyDrafts = new Map();
let previousDecisionProvider = 'openrouter';
+// Decision providers that run locally and need no API key.
+const LOCAL_DECISION_PROVIDERS = ['local', 'webgpu_d1', 'webgpu_chat'];
+// The in-browser chat model's decision mode has a fixed model and a note.
+function syncWebgpuDecisionProvider() {
+ const selected = decisionProviderInput?.value === 'webgpu_chat';
+ const note = document.getElementById('webgpu-decision-note');
+ if (note) note.hidden = !selected;
+ if (selected) {
+ decisionModelInput.value = WEBGPU_CHAT_DECISION_MODEL;
+ decisionModelInput.readOnly = true;
+ }
+ if (systemOneApiKeyInput && selected) systemOneApiKeyInput.disabled = true;
+}
const systemOneApiKeyInput = document.getElementById('system-one-api-key');
const systemOneEnabledToggle = document.getElementById('toggle-system-one');
@@ -741,6 +754,7 @@ async function init() {
decisionDoneToggle.checked = decisionConfig.doneEnabled;
decisionThresholdInput.value = String(decisionConfig.threshold * 100);
await refreshD1Settings();
+ syncWebgpuDecisionProvider();
decisionKeyDrafts.set('typesafe', stored.typesafeApiKey || '');
decisionKeyDrafts.set('openrouter', stored.decisionApiKey || '');
decisionKeyDrafts.set('local', stored.decisionLocalApiKey || '');
@@ -1576,7 +1590,7 @@ if (btnSaveSystemOne) {
btnSaveSystemOne.addEventListener('click', async () => {
const key = normalizeTypesafeApiKey(systemOneApiKeyInput?.value);
const enabled = systemOneEnabledToggle?.checked === true;
- if (enabled && !['local', 'webgpu_d1'].includes(decisionProviderInput.value) && !isValidTypesafeApiKey(key)) {
+ if (enabled && !LOCAL_DECISION_PROVIDERS.includes(decisionProviderInput.value) && !isValidTypesafeApiKey(key)) {
showSystemOneResult('fail', t('st.system_one.need_key'));
return;
}
@@ -1589,8 +1603,8 @@ if (btnSaveSystemOne) {
try { selection = decisionSettingsDraft(); resolveDecisionConfig(selection); if (!(selection.systemOneDoneThreshold >= .5 && selection.systemOneDoneThreshold <= .99)) throw new Error('Completion threshold must be 50–99%.'); } catch (error) { showSystemOneResult('fail', error.message); return; }
await chrome.storage.local.set({
...selection,
- ...(decisionProviderInput.value === 'webgpu_d1' ? {} : decisionProviderInput.value === 'typesafe' ? { typesafeApiKey: key } : { [decisionProviderInput.value === 'local' ? 'decisionLocalApiKey' : 'decisionApiKey']: key }),
- systemOneEnabled: enabled && (['local', 'webgpu_d1'].includes(decisionProviderInput.value) || isValidTypesafeApiKey(key)),
+ ...(['webgpu_d1', 'webgpu_chat'].includes(decisionProviderInput.value) ? {} : decisionProviderInput.value === 'typesafe' ? { typesafeApiKey: key } : { [decisionProviderInput.value === 'local' ? 'decisionLocalApiKey' : 'decisionApiKey']: key }),
+ systemOneEnabled: enabled && (LOCAL_DECISION_PROVIDERS.includes(decisionProviderInput.value) || isValidTypesafeApiKey(key)),
systemOneWatchEnabled: systemOneWatchToggle?.checked === true,
systemOneCompletionEnabled: systemOneCompletionToggle?.checked === true,
systemOneFastClassifications: systemOneClassificationsToggle?.checked === true,
@@ -1639,19 +1653,20 @@ function decisionSettingsDraft() {
const card = decisionModels.find(m => m.id === decisionModelInput.value);
return { decisionProvider: decisionProviderInput.value, decisionModel: decisionModelInput.value.trim(),
decisionBaseUrl: decisionEndpointInput.value.trim(), decisionVisionMode: decisionVisionInput.value,
- decisionVisionSupported, ...(decisionProviderInput.value === 'webgpu_d1' ? {} : decisionProviderInput.value === 'typesafe' ? { typesafeApiKey: systemOneApiKeyInput.value.trim() } : { [decisionProviderInput.value === 'local' ? 'decisionLocalApiKey' : 'decisionApiKey']: systemOneApiKeyInput.value.trim() }),
+ decisionVisionSupported, ...(['webgpu_d1', 'webgpu_chat'].includes(decisionProviderInput.value) ? {} : decisionProviderInput.value === 'typesafe' ? { typesafeApiKey: systemOneApiKeyInput.value.trim() } : { [decisionProviderInput.value === 'local' ? 'decisionLocalApiKey' : 'decisionApiKey']: systemOneApiKeyInput.value.trim() }),
systemOneDoneEnabled: decisionDoneToggle.checked, systemOneDoneThreshold: Number(decisionThresholdInput.value) / 100,
- decisionInputRate: card?.inputRate ?? (decisionProviderInput.value === 'typesafe' ? .042 : ['local', 'webgpu_d1'].includes(decisionProviderInput.value) ? 0 : .02), decisionOutputRate: card?.outputRate ?? 0 };
+ decisionInputRate: card?.inputRate ?? (decisionProviderInput.value === 'typesafe' ? .042 : LOCAL_DECISION_PROVIDERS.includes(decisionProviderInput.value) ? 0 : .02), decisionOutputRate: card?.outputRate ?? 0 };
}
decisionProviderInput?.addEventListener('change', () => {
decisionKeyDrafts.set(previousDecisionProvider, systemOneApiKeyInput.value);
previousDecisionProvider = decisionProviderInput.value;
systemOneApiKeyInput.value = decisionKeyDrafts.get(previousDecisionProvider) || '';
- decisionModelInput.value = previousDecisionProvider === 'webgpu_d1' ? D1_MODEL_ID : previousDecisionProvider === 'typesafe' ? 'jev-1.13.0' : previousDecisionProvider === 'local' ? 'kev-latest' : DEFAULT_DECISION_MODEL;
+ decisionModelInput.value = previousDecisionProvider === 'webgpu_d1' ? D1_MODEL_ID : previousDecisionProvider === 'webgpu_chat' ? WEBGPU_CHAT_DECISION_MODEL : previousDecisionProvider === 'typesafe' ? 'jev-1.13.0' : previousDecisionProvider === 'local' ? 'kev-latest' : DEFAULT_DECISION_MODEL;
decisionModels = []; decisionVisionSupported = false;
document.getElementById('decision-models').replaceChildren();
document.getElementById('decision-endpoint-field').hidden = previousDecisionProvider !== 'local';
+ syncWebgpuDecisionProvider();
});
decisionModelInput?.addEventListener('input', () => {
decisionVisionSupported = decisionModels.find(m => m.id === decisionModelInput.value)?.supportsVision === true;
@@ -1671,7 +1686,7 @@ document.getElementById('btn-decision-models')?.addEventListener('click', async
btnTestSystemOne?.addEventListener('click', async () => {
const apiKey = normalizeTypesafeApiKey(systemOneApiKeyInput?.value);
- if (!apiKey && !['local', 'webgpu_d1'].includes(decisionProviderInput.value)) { showSystemOneResult('fail', t('st.system_one.need_key')); return; }
+ if (!apiKey && !LOCAL_DECISION_PROVIDERS.includes(decisionProviderInput.value)) { showSystemOneResult('fail', t('st.system_one.need_key')); return; }
btnTestSystemOne.disabled = true;
showSystemOneResult('', t('st.providers.testing'));
try {
diff --git a/src/firefox/src/agent/decision-config.js b/src/firefox/src/agent/decision-config.js
index eb9616e3f..ee897cc34 100644
--- a/src/firefox/src/agent/decision-config.js
+++ b/src/firefox/src/agent/decision-config.js
@@ -7,10 +7,16 @@ export const DECISION_SETTINGS_KEYS = [
D1_CONSENT_KEY,
];
+// The in-browser chat model answers decisions through its trained probe
+// (Chrome; src/chrome/src/offscreen/bonsai-decision-probe.js). Kept in sync
+// with WEBGPU_DECISION_MODEL_ID in providers/webgpu.js.
+export const WEBGPU_CHAT_DECISION_MODEL = 'webbrain-one/Ternary-Bonsai-2-2B-ONNX-WebGPU-Vision';
+
// Resolve legacy settings without rewriting credentials or opted-in features.
export function resolveDecisionConfig(stored = {}, compass = null) {
const explicitD1 = stored.decisionProvider === 'webgpu_d1' && stored.systemOneEnabled === true && stored[D1_CONSENT_KEY] === D1_CONSENT_VERSION;
- const provider = explicitD1 ? 'webgpu_d1' : compass ? 'compass' : (stored.decisionProvider || (stored.typesafeApiKey ? 'typesafe' : 'openrouter'));
+ const explicitChat = stored.decisionProvider === 'webgpu_chat' && stored.systemOneEnabled === true;
+ const provider = explicitD1 ? 'webgpu_d1' : explicitChat ? 'webgpu_chat' : compass ? 'compass' : (stored.decisionProvider || (stored.typesafeApiKey ? 'typesafe' : 'openrouter'));
if (provider === 'webgpu_d1') return {
provider, model: D1_MODEL_ID, apiKey: '', baseUrl: '', url: '', local: true,
enabled: stored.systemOneEnabled === true && stored[D1_CONSENT_KEY] === D1_CONSENT_VERSION,
@@ -20,6 +26,17 @@ export function resolveDecisionConfig(stored = {}, compass = null) {
requiresIndependentSuccessConfirmation: true,
config: { category: 'local', providerName: provider, inputCostPerMillionUsd: 0, outputCostPerMillionUsd: 0 },
};
+ // Explicitly selected, like D1: an on-device decision provider that never
+ // leaves the browser. A positive completion still needs independent proof.
+ if (provider === 'webgpu_chat') return {
+ provider, model: WEBGPU_CHAT_DECISION_MODEL, apiKey: '', baseUrl: '', url: '', local: true,
+ enabled: stored.systemOneEnabled === true,
+ doneEnabled: stored.systemOneDoneEnabled === true,
+ threshold: Number.isFinite(stored.systemOneDoneThreshold) && stored.systemOneDoneThreshold >= .5 && stored.systemOneDoneThreshold <= .99 ? stored.systemOneDoneThreshold : .9,
+ supportsVision: stored.decisionVisionMode !== 'off', experimental: true,
+ requiresIndependentSuccessConfirmation: true,
+ config: { category: 'local', providerName: provider, inputCostPerMillionUsd: 0, outputCostPerMillionUsd: 0 },
+ };
if (!['compass', 'openrouter', 'typesafe', 'local'].includes(provider)) throw new Error('Unknown decision provider.');
const defaults = { compass: DEFAULT_DECISION_MODEL, openrouter: DEFAULT_DECISION_MODEL, typesafe: 'jev-1.13.0', local: 'kev-latest' };
const model = compass ? DEFAULT_DECISION_MODEL : String(stored.decisionModel || defaults[provider]).trim();
@@ -49,6 +66,7 @@ export function resolveDecisionConfig(stored = {}, compass = null) {
}
export async function listDecisionModels(config, fetchImpl = globalThis.fetch, signal) {
+ if (config.provider === 'webgpu_chat') return [{ id: WEBGPU_CHAT_DECISION_MODEL, name: 'Ternary Bonsai 2 2B Vision (in-browser, experimental)', supportsVision: true, inputRate: 0, outputRate: 0 }];
if (config.provider === 'webgpu_d1') return [{ id: D1_MODEL_ID, name: 'D1 FP16 (experimental, on-device)', supportsVision: true, inputRate: 0, outputRate: 0 }];
return withCompletionTimeout(requestSignal => discoverDecisionModels(config, fetchImpl, requestSignal), signal, 5000);
}
diff --git a/src/firefox/src/agent/systemone-judge.js b/src/firefox/src/agent/systemone-judge.js
index 1b3b06ffb..a66a0d4c7 100644
--- a/src/firefox/src/agent/systemone-judge.js
+++ b/src/firefox/src/agent/systemone-judge.js
@@ -5,6 +5,7 @@
// interface. It never performs an action or turns a non-success into success.
import { evaluateD1 } from '../providers/d1.js';
+import { evaluateWebgpuDecision } from '../providers/webgpu-decision.js';
export const SYSTEM_ONE_API_URL = 'https://api.typesafe.ai/v1/systemone';
export const SYSTEM_ONE_MODEL = 'jev-1.13.0';
@@ -274,14 +275,26 @@ export function createSystemOneJudge({
return {
async evaluate({ apiKey, state, images, questions, signal, beforeRequest, onUsage, config, headers = {}, metadata = {} } = {}) {
const key = normalizeTypesafeApiKey(config?.apiKey ?? apiKey);
- if (!key && !['local', 'compass', 'webgpu_d1'].includes(config?.provider)) throw new Error('Decision API key is not configured.');
+ if (!key && !['local', 'compass', 'webgpu_d1', 'webgpu_chat'].includes(config?.provider)) throw new Error('Decision API key is not configured.');
const model = config?.model || SYSTEM_ONE_MODEL;
validateState(state, config?.supportsVision === true);
validateQuestions(questions);
const started = now();
- const request = requestSignal(signal, Math.max(1, Math.min(5000, Number(timeoutMs) || 5000)));
+ // The in-browser model prefills its screenshot on this GPU (about a second
+ // on a desktop card) and loads on first use after a browser restart.
+ const cap = config?.provider === 'webgpu_chat' ? 15000 : 5000;
+ const request = requestSignal(signal, Math.max(1, Math.min(cap, Number(timeoutMs) || cap)));
const retries = Math.max(0, Math.min(2, Math.floor(Number(maxRetries) || 0)));
try {
+ if (config?.provider === 'webgpu_chat') {
+ if (beforeRequest) await abortable(beforeRequest(), request.signal);
+ const result = await abortable(evaluateWebgpuDecision({ state, images, questions, config, signal: request.signal }), request.signal);
+ validateSystemOneAnswers(result.answers, questions);
+ if (result.model !== model || !Number.isInteger(result.usage?.input_tokens) || result.usage.input_tokens < 0 || result.usage?.output_tokens !== 0) throw systemOneError('JEV_INVALID_USAGE', 'Invalid in-browser decision usage/model.');
+ const metadata = { model: result.model, provider: result.provider, usage: result.usage, latencyMs: Math.max(0, now() - started), estimatedCostUsd: 0 };
+ if (onUsage) await abortable(onUsage(metadata), request.signal);
+ return { ...result, ...metadata };
+ }
if (config?.provider === 'webgpu_d1') {
if (beforeRequest) await abortable(beforeRequest(), request.signal);
const result = await abortable(evaluateD1({ state, images, questions, config, signal: request.signal }), request.signal);
diff --git a/src/firefox/src/config-transfer.js b/src/firefox/src/config-transfer.js
index 82266b412..889e10cda 100644
--- a/src/firefox/src/config-transfer.js
+++ b/src/firefox/src/config-transfer.js
@@ -236,7 +236,7 @@ function sanitizeProviders(value, { strict = false } = {}) {
}
function validSettingValue(key, value) {
- if (key === 'decisionProvider') return ['', 'openrouter', 'typesafe', 'local', 'webgpu_d1'].includes(value);
+ if (key === 'decisionProvider') return ['', 'openrouter', 'typesafe', 'local', 'webgpu_d1', 'webgpu_chat'].includes(value);
if (key === 'decisionVisionMode') return ['auto', 'on', 'off'].includes(value);
if (key === 'decisionBaseUrl') {
if (typeof value !== 'string') return false;
@@ -283,7 +283,7 @@ function normalizeSettings(source, { strict = false } = {}) {
settings.providers = sanitizeProviders(settings.providers, { strict });
if (!Object.hasOwn(source, 'decisionInputRate')) {
settings.decisionInputRate = source.decisionProvider === 'typesafe' || (!source.decisionProvider && source.typesafeApiKey)
- ? .042 : ['local', 'webgpu_d1'].includes(source.decisionProvider) ? 0 : .02;
+ ? .042 : ['local', 'webgpu_d1', 'webgpu_chat'].includes(source.decisionProvider) ? 0 : .02;
}
return settings;
}
diff --git a/src/firefox/src/providers/webgpu-decision.js b/src/firefox/src/providers/webgpu-decision.js
new file mode 100644
index 000000000..586da792c
--- /dev/null
+++ b/src/firefox/src/providers/webgpu-decision.js
@@ -0,0 +1,5 @@
+// Firefox has no in-browser WebGPU chat runtime, so the chat model cannot
+// answer decisions here; the decision settings never offer it.
+export async function evaluateWebgpuDecision() {
+ throw new Error('In-browser chat-model decisions are available in Chrome only.');
+}
diff --git a/src/firefox/src/ui/locales/decision-copy.mjs b/src/firefox/src/ui/locales/decision-copy.mjs
index 51a9e4a74..4b5b1b919 100644
--- a/src/firefox/src/ui/locales/decision-copy.mjs
+++ b/src/firefox/src/ui/locales/decision-copy.mjs
@@ -1,5 +1,7 @@
export const decisionEnglish = {
'st.d1.provider': 'Local decision model - Experimental',
+ 'st.decision.webgpu_chat': 'In-browser chat model (Ternary Bonsai 2 2B Vision) - Experimental',
+ 'st.decision.webgpu_chat_note': 'Answers decisions with the in-browser WebGPU chat model and a small trained readout, with no extra download. Download Ternary Bonsai 2 2B Vision under Providers first. The first decision after a browser restart loads the model and may fall back to the active model. Not a sole completion authority.',
'st.d1.warning': 'Experimental on-device decision model, about 0.95 GB download plus tokenizer. Requires hardware WebGPU and substantial GPU memory. Not a chat model or sole completion authority. CPU/WASM control operators may run. Browser support and model decisions are not universally validated.',
'st.d1.consent': 'I accept the experimental local model download from Hugging Face and its memory requirements.',
'st.d1.need_consent': 'Accept the experimental model download consent first.',
@@ -25,6 +27,8 @@ export const decisionEnglish = {
export const decisionTranslations = {
tr: {
'st.d1.provider': 'Yerel karar modeli — Deneysel',
+ 'st.decision.webgpu_chat': 'Tarayıcı içi sohbet modeli (Ternary Bonsai 2 2B Vision) — Deneysel',
+ 'st.decision.webgpu_chat_note': 'Kararları tarayıcı içi WebGPU sohbet modeli ve küçük bir eğitilmiş okuma katmanıyla verir; ek indirme gerekmez. Önce Sağlayıcılar altında Ternary Bonsai 2 2B Vision modelini indirin. Tarayıcı yeniden başladıktan sonraki ilk karar modeli yükler ve etkin modele geri düşebilir. Tek başına tamamlanma yetkilisi değildir.',
'st.d1.warning': 'Deneysel yerel karar modeli; tokenizer ile birlikte yaklaşık 0,95 GB indirme. Donanım WebGPU ve yüksek GPU belleği gerekir. Sohbet modeli veya tek başına tamamlanma yetkilisi değildir. CPU/WASM kontrol işlemleri çalışabilir. Her tarayıcı, GPU ve model kararı doğrulanmış değildir.',
'st.d1.consent': 'Hugging Face üzerinden deneysel yerel model indirmesini ve bellek gereksinimlerini kabul ediyorum.',
'st.d1.need_consent': 'Önce deneysel model indirme onayını kabul edin.',
diff --git a/test/webgpu-decisions.mjs b/test/webgpu-decisions.mjs
index 84c7e551e..a381f8833 100644
--- a/test/webgpu-decisions.mjs
+++ b/test/webgpu-decisions.mjs
@@ -101,3 +101,97 @@ test('the worker routes decisions through the offscreen host and provider', asyn
assert.match(branch, /src\/ui\/settings\.html/, 'decision fixtures are Settings-only');
assert.match(branch, /strictSecretMode/);
});
+
+test('the trained probe scores each option from its margin and hidden state', async () => {
+ const { BONSAI_DECISION_PROBE, bonsaiProbeScores } = await import('../src/chrome/src/offscreen/bonsai-decision-probe.js');
+ const { WEBGPU_DECISION_MODEL_ID } = await import('../src/chrome/src/providers/webgpu.js');
+ const { WEBGPU_CHAT_DECISION_MODEL } = await import('../src/chrome/src/agent/decision-config.js');
+ assert.equal(BONSAI_DECISION_PROBE.model, WEBGPU_DECISION_MODEL_ID);
+ assert.equal(WEBGPU_CHAT_DECISION_MODEL, WEBGPU_DECISION_MODEL_ID, 'decision-config keeps the probe model id in sync');
+ const width = BONSAI_DECISION_PROBE.hiddenSize;
+ const zero = new Float32Array(width);
+ const [a, b] = bonsaiProbeScores([2, 0], [zero, zero]);
+ assert.ok(Math.abs(a - b - 2 * BONSAI_DECISION_PROBE.margin) < 1e-9, 'with equal hidden states only the margins differ');
+ const unit = new Float32Array(width).fill(1);
+ const [c, d] = bonsaiProbeScores([0, 0], [unit, zero]);
+ assert.ok(Number.isFinite(c - d) && c !== d, 'the hidden state moves the score');
+ assert.throws(() => bonsaiProbeScores([0], [null]), /hidden state/);
+ assert.throws(() => bonsaiProbeScores([0], [new Float32Array(3)]), /hidden state/);
+});
+
+test('decisions default to the probe and refuse a decoder without decision_hidden', async () => {
+ const worker = await readFile(new URL('../src/chrome/src/offscreen/inference-worker.js', import.meta.url), 'utf8');
+ assert.match(worker, /payload\.promptVariant : 'probe';/, 'the probe is the default decision mode');
+ assert.match(worker, /variant === 'probe' && !output\.hidden\) \{\n\s*throw new Error\('In-browser decisions need the updated Ternary Bonsai 2 2B Vision decoder/);
+ assert.match(worker, /\['webbrain-one\/Ternary-Bonsai-2-2B-ONNX-WebGPU-Vision', '240c07e1191911a2306faf106ee537382a1b4d5e6ecdd32b618da160db9145c1'\]/,
+ 'the 2B decoder pin is the graph that exposes decision_hidden');
+});
+
+test('the in-browser chat model is an explicit, key-free, local decision provider that needs independent success proof', async () => {
+ for (const build of ['chrome', 'firefox']) {
+ const config = await import(`../src/${build}/src/agent/decision-config.js`);
+ const transfer = await import(`../src/${build}/src/config-transfer.js`);
+ const selected = config.resolveDecisionConfig({ decisionProvider: 'webgpu_chat', systemOneEnabled: true });
+ assert.equal(selected.provider, 'webgpu_chat');
+ assert.equal(selected.model, config.WEBGPU_CHAT_DECISION_MODEL);
+ assert.equal(selected.local, true); assert.equal(selected.apiKey, ''); assert.equal(selected.enabled, true);
+ assert.equal(selected.doneEnabled, false, 'completion outsourcing stays opt-in');
+ assert.equal(selected.requiresIndependentSuccessConfirmation, true);
+ assert.equal(selected.config.inputCostPerMillionUsd, 0);
+ assert.equal(config.resolveDecisionConfig({ decisionProvider: 'webgpu_chat', systemOneEnabled: true }, { baseUrl: 'https://api.webbrain.one/v1' }).provider, 'webgpu_chat',
+ 'an explicit local choice overrides Compass, as D1 does');
+ assert.equal(config.resolveDecisionConfig({ decisionProvider: 'webgpu_chat' }).enabled, false);
+ assert.equal((await config.listDecisionModels(selected, () => { throw new Error('no network'); })).length, 1);
+ const imported = transfer.parseConfigImport(JSON.stringify({ schema: transfer.CONFIG_SCHEMA, settings: { decisionProvider: 'webgpu_chat' } })).settings;
+ assert.equal(imported.decisionProvider, 'webgpu_chat');
+ }
+});
+
+test('the judge answers webgpu_chat decisions through the offscreen worker in Chrome and refuses them in Firefox', async () => {
+ const saved = { chrome: globalThis.chrome };
+ const sent = [];
+ const answers = { q: { type: 'noul', noul: 0.8 } };
+ globalThis.chrome = {
+ offscreen: { hasDocument: async () => true, createDocument: async () => {} },
+ runtime: {
+ lastError: null,
+ getURL: path => `chrome-extension://test/${path}`,
+ sendMessage(message, callback) {
+ sent.push(message.type);
+ const reply = message.type === 'webgpu-download-status' ? { ready: true }
+ : message.type === 'webgpu-decide' ? { ok: true, model: message.model, answers, usage: { input_tokens: 900, output_tokens: 0 } }
+ : { ok: true };
+ callback(reply);
+ },
+ },
+ storage: { local: { get: async () => ({}) } },
+ };
+ try {
+ const { resolveDecisionConfig } = await import('../src/chrome/src/agent/decision-config.js');
+ const { createSystemOneJudge } = await import('../src/chrome/src/agent/systemone-judge.js');
+ const config = resolveDecisionConfig({ decisionProvider: 'webgpu_chat', systemOneEnabled: true });
+ const result = await createSystemOneJudge({ fetchImpl: () => { throw new Error('no cloud request'); }, maxRetries: 0 })
+ .evaluate({ state: 'Page', images: [], questions: { q: { type: 'noul', instructions: 'Is it done?' } }, config });
+ assert.deepEqual(result.answers, answers);
+ assert.equal(result.usage.output_tokens, 0);
+ assert.equal(result.estimatedCostUsd, 0);
+ assert.ok(sent.includes('webgpu-decide'));
+ const firefox = await import('../src/firefox/src/agent/systemone-judge.js');
+ const firefoxConfig = (await import('../src/firefox/src/agent/decision-config.js')).resolveDecisionConfig({ decisionProvider: 'webgpu_chat', systemOneEnabled: true });
+ await assert.rejects(firefox.createSystemOneJudge({ fetchImpl: () => { throw new Error('no cloud request'); }, maxRetries: 0 })
+ .evaluate({ state: 'Page', questions: { q: { type: 'noul', instructions: 'Is it done?' } }, config: firefoxConfig }), /Chrome only/);
+ } finally {
+ if (saved.chrome === undefined) delete globalThis.chrome; else globalThis.chrome = saved.chrome;
+ }
+});
+
+test('Settings offers the chat-model decision provider in Chrome only', async () => {
+ const chromeHtml = await readFile(new URL('../src/chrome/src/ui/settings.html', import.meta.url), 'utf8');
+ const firefoxHtml = await readFile(new URL('../src/firefox/src/ui/settings.html', import.meta.url), 'utf8');
+ assert.match(chromeHtml, /<option value="webgpu_chat" data-i18n="st\.decision\.webgpu_chat"><\/option>/);
+ assert.match(chromeHtml, /id="webgpu-decision-note" hidden data-i18n="st\.decision\.webgpu_chat_note"/);
+ assert.doesNotMatch(firefoxHtml, /webgpu_chat/);
+ const { decisionEnglish, decisionTranslations } = await import('../src/chrome/src/ui/locales/decision-copy.mjs');
+ assert.ok(decisionEnglish['st.decision.webgpu_chat'] && decisionEnglish['st.decision.webgpu_chat_note']);
+ assert.ok(decisionTranslations.tr['st.decision.webgpu_chat']);
+});