Skip to content

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

Bound Clef Flash's state by UTF-8 bytes, a provable upper bound

Claude69dce0944774
diff --git a/docs/providers-and-models.md b/docs/providers-and-models.md
index fa3395854..802f262b9 100644
--- a/docs/providers-and-models.md
+++ b/docs/providers-and-models.md
@@ -90,10 +90,10 @@ provider until changed. Provider keys are stored separately.
`cloudflare/clef-flash` reads only the first ~2K tokens of a long text state
(OpenRouter's model page says Workers AI truncates it), so
-`SYSTEM_ONE_TEXT_TOKEN_BUDGETS` in `systemone-judge.js` caps the state it is sent
-at 1,500 *estimated* tokens. The estimate is deliberately high (no tokenizer
-ships with the extension): ASCII counts half a token per character, other
-characters three (the worst case for CJK and emoji), long unbroken letter/digit runs one each. A longer state is not
+`SYSTEM_ONE_TEXT_BYTE_BUDGETS` in `systemone-judge.js` caps the state it is sent
+at 1,500 UTF-8 bytes. A byte-level BPE tokenizer never produces more tokens than
+bytes, so this holds for any text without shipping a tokenizer; image blocks are
+left out because the provider counts them separately. A longer state is not
sent; the check takes the same fallback as for an unavailable decision model
(failure reason `state_too_long`) instead of judging evidence it only partly
read. In practice Clef Flash answers short decisions, and long completion
diff --git a/src/chrome/src/agent/systemone-judge.js b/src/chrome/src/agent/systemone-judge.js
index 9f8e3409d..2570b705f 100644
--- a/src/chrome/src/agent/systemone-judge.js
+++ b/src/chrome/src/agent/systemone-judge.js
@@ -132,32 +132,21 @@ function validateState(state, multimodal = false) {
// confident and wrong. A state past the budget is not sent at all; the caller
// takes the same fallback as for any unavailable decision model. Images are
// counted separately by the provider and are not part of this text.
-// The budget is in estimated tokens and sits well under the 2K the model reads
-// (see estimateTextTokens), because no tokenizer ships with the extension.
-export const SYSTEM_ONE_TEXT_TOKEN_BUDGETS = Object.freeze({ 'cloudflare/clef-flash': 1500 });
-
-// A deliberately high token estimate, so a state it passes stays under the real
-// limit for ordinary text, CJK, and identifier-heavy input alike: ASCII counts
-// half a token per character (English and JSON run near a quarter), any other
-// character three (CJK and emoji take one to three, so the worst case is
-// charged), and long unbroken runs of
-// letters and digits (hashes, base64, ids) one per character.
-export function estimateTextTokens(text) {
- let tokens = 0;
- const rest = String(text).replace(/[A-Za-z0-9+/=_-]{24,}/g, run => { tokens += run.length; return ' '; });
- for (const character of rest) tokens += character.charCodeAt(0) < 128 ? .5 : 3;
- return Math.ceil(tokens);
-}
+// The budget is in UTF-8 bytes: a byte-level BPE tokenizer never produces more
+// tokens than bytes, so the byte count is a bound that holds for any text (CJK,
+// emoji, punctuation, hashes) without shipping a tokenizer. 1,500 sits well
+// under the roughly 2K tokens the model reads.
+export const SYSTEM_ONE_TEXT_BYTE_BUDGETS = Object.freeze({ 'cloudflare/clef-flash': 1500 });
function assertStateFitsModel(model, state) {
- const budget = SYSTEM_ONE_TEXT_TOKEN_BUDGETS[model];
+ const budget = SYSTEM_ONE_TEXT_BYTE_BUDGETS[model];
if (!budget) return;
// Screenshots ride inside the state as structured image_url blocks; the
// provider counts images separately, so only those blocks' payloads are left
// out. A data URL that sits in ordinary text is text and is measured.
const text = JSON.stringify(state, (_key, value) => (value && typeof value === 'object' && value.type === 'image_url'
&& value.image_url && typeof value.image_url === 'object' ? { type: 'image_url' } : value));
- if (estimateTextTokens(text) > budget) {
+ if (new TextEncoder().encode(text).length > budget) {
throw systemOneError('JEV_STATE_EXCEEDS_MODEL_INPUT', `The decision state is longer than ${model} reads; the fallback answers instead.`);
}
}
diff --git a/src/firefox/src/agent/systemone-judge.js b/src/firefox/src/agent/systemone-judge.js
index aa8238fb1..db94c6ae1 100644
--- a/src/firefox/src/agent/systemone-judge.js
+++ b/src/firefox/src/agent/systemone-judge.js
@@ -132,32 +132,21 @@ function validateState(state, multimodal = false) {
// confident and wrong. A state past the budget is not sent at all; the caller
// takes the same fallback as for any unavailable decision model. Images are
// counted separately by the provider and are not part of this text.
-// The budget is in estimated tokens and sits well under the 2K the model reads
-// (see estimateTextTokens), because no tokenizer ships with the extension.
-export const SYSTEM_ONE_TEXT_TOKEN_BUDGETS = Object.freeze({ 'cloudflare/clef-flash': 1500 });
-
-// A deliberately high token estimate, so a state it passes stays under the real
-// limit for ordinary text, CJK, and identifier-heavy input alike: ASCII counts
-// half a token per character (English and JSON run near a quarter), any other
-// character three (CJK and emoji take one to three, so the worst case is
-// charged), and long unbroken runs of
-// letters and digits (hashes, base64, ids) one per character.
-export function estimateTextTokens(text) {
- let tokens = 0;
- const rest = String(text).replace(/[A-Za-z0-9+/=_-]{24,}/g, run => { tokens += run.length; return ' '; });
- for (const character of rest) tokens += character.charCodeAt(0) < 128 ? .5 : 3;
- return Math.ceil(tokens);
-}
+// The budget is in UTF-8 bytes: a byte-level BPE tokenizer never produces more
+// tokens than bytes, so the byte count is a bound that holds for any text (CJK,
+// emoji, punctuation, hashes) without shipping a tokenizer. 1,500 sits well
+// under the roughly 2K tokens the model reads.
+export const SYSTEM_ONE_TEXT_BYTE_BUDGETS = Object.freeze({ 'cloudflare/clef-flash': 1500 });
function assertStateFitsModel(model, state) {
- const budget = SYSTEM_ONE_TEXT_TOKEN_BUDGETS[model];
+ const budget = SYSTEM_ONE_TEXT_BYTE_BUDGETS[model];
if (!budget) return;
// Screenshots ride inside the state as structured image_url blocks; the
// provider counts images separately, so only those blocks' payloads are left
// out. A data URL that sits in ordinary text is text and is measured.
const text = JSON.stringify(state, (_key, value) => (value && typeof value === 'object' && value.type === 'image_url'
&& value.image_url && typeof value.image_url === 'object' ? { type: 'image_url' } : value));
- if (estimateTextTokens(text) > budget) {
+ if (new TextEncoder().encode(text).length > budget) {
throw systemOneError('JEV_STATE_EXCEEDS_MODEL_INPUT', `The decision state is longer than ${model} reads; the fallback answers instead.`);
}
}
diff --git a/test/systemone.mjs b/test/systemone.mjs
index 7747b5e4e..2b569e6bc 100644
--- a/test/systemone.mjs
+++ b/test/systemone.mjs
@@ -37,8 +37,8 @@ for (const build of ['chrome', 'firefox']) {
return { ok: true, json: async () => ({ model: body.model, answers: { p: { type: 'noul', noul: .9 } }, usage: { input_tokens: 10, output_tokens: 0 } }) }; } });
const one = { p: questions.p };
const config = { provider: 'openrouter', model: 'cloudflare/clef-flash', apiKey: 'test', url: 'https://example.test/v1/decisions' };
- const budget = mod.SYSTEM_ONE_TEXT_TOKEN_BUDGETS['cloudflare/clef-flash'];
- assert.ok(budget < 2000, 'the budget stays under the 2K tokens the model reads');
+ const budget = mod.SYSTEM_ONE_TEXT_BYTE_BUDGETS['cloudflare/clef-flash'];
+ assert.ok(budget < 2000, 'the budget (bytes, an upper bound on byte-level tokens) stays under the 2K tokens the model reads');
await judge.evaluate({ config, state: { evidence: 'The post was published and its page shows it. '.repeat(20) }, questions: one });
assert.equal(requests.length, 1, 'ordinary text within the budget is sent');
// Each of these stays far below the old 6,000-character cap yet can exceed 2K real tokens.
@@ -47,6 +47,7 @@ for (const build of ['chrome', 'firefox']) {
hashes: { evidence: ('a1b2c3d4e5f60718293a4b5c6d7e8f90 ').repeat(60) },
unbroken: { evidence: 'Zm9vYmFy'.repeat(250) },
emoji: { evidence: '\u{1F600}'.repeat(700) },
+ punctuation: { evidence: '\\"\\n{}[]<>|~^'.repeat(150) },
};
for (const [name, state] of Object.entries(dense)) {
assert.ok(JSON.stringify(state).length < 6000, `${name} is under the old character cap`);
@@ -64,7 +65,6 @@ for (const build of ['chrome', 'firefox']) {
assert.equal(requests.length, 2, 'a data URL in text is not sent');
await judge.evaluate({ config: { ...config, model: 'another/publisher' }, state: dense.cjk, questions: one });
assert.equal(requests.length, 3, 'other models keep the full state');
- assert.ok(mod.estimateTextTokens('x'.repeat(100)) >= 50 && mod.estimateTextTokens('æ—¥'.repeat(10)) >= 30);
});
test(`${build}: singleton Choice requests fail locally before a service call`, async () => {
let calls = 0;