Skip to content

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

Treat an OpenAI Responses output-limit error as a cut-off in the inline card

Claudec31e106bff11
diff --git a/docs/architecture.md b/docs/architecture.md
index b05e27c59..521f6a01d 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -521,7 +521,9 @@ cost allowance (90 s timeout) and returns the text. The output budget is sized t
selection (at least 2,000 tokens; Translate and Proofread grow to about two tokens per
character, up to 8,000) and never exceeds the provider's own output ceiling. If the reply
still stopped at the cap (finish reason, Anthropic `stop_reason` or Responses
-`incomplete_details`, or token usage at the budget when no reason is reported) the response
+`incomplete_details`, or token usage at the budget when no reason is reported; OpenAI
+Responses reports the limit as an error, which the background treats the same way and
+answers with whatever text the provider returned) the response
is flagged `truncated`, and the card says so and offers **Continue in side panel** instead
of passing a cut-off translation off as complete. Each call carries a job id, and a card
that is closed, retried or retargeted sends `WB_SELECTION_SHORTCUT_INLINE_CANCEL` so the
diff --git a/src/chrome/src/background.js b/src/chrome/src/background.js
index 959e48be7..dc2bafa22 100644
--- a/src/chrome/src/background.js
+++ b/src/chrome/src/background.js
@@ -73,6 +73,7 @@ import {
buildFullContextSelectionPrompt,
buildInlineSelectionMessages,
inlineAnswerWasCut,
+ inlineErrorWasCut,
inlineSelectionMaxTokens,
buildSelectionPrompt,
normalizeSelectionAction,
@@ -1782,6 +1783,14 @@ async function answerInlineSelection(msg) {
if (!text) return { ok: false, error: 'The model returned an empty answer.' };
return inlineAnswerWasCut(result, maxTokens) ? { ok: true, text, truncated: true } : { ok: true, text };
} catch (error) {
+ if (inlineErrorWasCut(error)) {
+ // The provider reported the output limit as an error. Show what it wrote, if
+ // anything, and let the card hand the rest to the side panel.
+ const partial = String(error.partialContent || '').trim();
+ return partial
+ ? { ok: true, text: partial, truncated: true }
+ : { ok: false, truncated: true, error: 'The answer was cut off before any text was produced.' };
+ }
return { ok: false, error: String(error?.message || error).slice(0, 300) };
} finally {
clearTimeout(timer);
diff --git a/src/chrome/src/content/selection-shortcut.js b/src/chrome/src/content/selection-shortcut.js
index 2111852e2..5fc33ccd1 100644
--- a/src/chrome/src/content/selection-shortcut.js
+++ b/src/chrome/src/content/selection-shortcut.js
@@ -922,12 +922,14 @@
parts.source.textContent = cardState.sourceLabel || str('autoDetect');
parts.target.value = cardState.target;
}
- parts.result.hidden = status === 'idle';
+ // A cut-off with no text has nothing to show or retry, so only the notice remains.
+ const cutOffOnly = status === 'error' && cardState.truncated;
+ parts.result.hidden = status === 'idle' || cutOffOnly;
parts.result.setAttribute('aria-busy', status === 'loading' ? 'true' : 'false');
parts.skeleton.hidden = status !== 'loading';
parts.text.hidden = status !== 'ready';
- parts.error.hidden = status !== 'error';
- parts.cut.hidden = !(status === 'ready' && cardState.truncated);
+ parts.error.hidden = status !== 'error' || cutOffOnly;
+ parts.cut.hidden = !((status === 'ready' || cutOffOnly) && cardState.truncated);
if (status === 'ready') renderAnswer(parts.text, cardState.text, kind !== 'translate');
if (status === 'error') {
parts.errorText.textContent = str('inlineFailed');
@@ -1033,10 +1035,12 @@
if (requestId !== cardRequestId || !cardState) return;
cardJobId = '';
const answer = typeof response?.text === 'string' ? response.text.trim() : '';
- if (!response?.ok || !answer) throw new Error(typeof response?.error === 'string' ? response.error : '');
+ cardState.truncated = response?.truncated === true;
+ if (!response?.ok || !answer) {
+ throw new Error(typeof response?.error === 'string' && !cardState.truncated ? response.error : '');
+ }
cardState.status = 'ready';
cardState.text = answer;
- cardState.truncated = response.truncated === true;
} catch (error) {
if (requestId !== cardRequestId || !cardState) return;
cardJobId = '';
@@ -1060,7 +1064,7 @@
// A cut-off answer is finished in the side panel, where the reply is not capped.
function continueInPanel() {
- if (!cardState || cardState.status !== 'ready') return;
+ if (!cardState?.truncated) return;
const { kind, target } = cardState;
submitSelection(kind, '', kind === 'translate' ? target : interfaceLanguage);
}
diff --git a/src/chrome/src/context-menu-storage.js b/src/chrome/src/context-menu-storage.js
index 4d4085123..199722d7e 100644
--- a/src/chrome/src/context-menu-storage.js
+++ b/src/chrome/src/context-menu-storage.js
@@ -288,6 +288,13 @@ export function inlineSelectionMaxTokens(selectionText, action, providerCeiling)
// that report none.
const INLINE_CUT_REASONS = new Set(['length', 'max_tokens', 'max_output_tokens', 'max_completion_tokens']);
+// Some providers (OpenAI Responses) report hitting the output limit as an error
+// instead of a result. Such an error counts as a cut-off, not as a failure to retry.
+export function inlineErrorWasCut(error) {
+ const reason = String(error?.incompleteReason ?? '').trim().toLowerCase();
+ return INLINE_CUT_REASONS.has(reason);
+}
+
export function inlineAnswerWasCut(result, maxTokens) {
const raw = result?.raw;
const reason = [
diff --git a/src/chrome/src/providers/openai.js b/src/chrome/src/providers/openai.js
index 1099df8c3..b37cf89a9 100644
--- a/src/chrome/src/providers/openai.js
+++ b/src/chrome/src/providers/openai.js
@@ -889,6 +889,9 @@ export class OpenAICompatibleProvider extends BaseLLMProvider {
error.isAskStreamTerminalError = !!stream && !error.isResponsesStreamFallbackSafe;
error.incomplete = true;
error.incompleteReason = reason;
+ // What the model wrote before it stopped, for callers that can show a partial answer.
+ const partial = Array.isArray(response?.output) ? this._responseText(response.output) : '';
+ if (partial) error.partialContent = partial;
return error;
}
diff --git a/src/firefox/src/background.js b/src/firefox/src/background.js
index a71f346ab..26accf549 100644
--- a/src/firefox/src/background.js
+++ b/src/firefox/src/background.js
@@ -70,6 +70,7 @@ import {
buildFullContextSelectionPrompt,
buildInlineSelectionMessages,
inlineAnswerWasCut,
+ inlineErrorWasCut,
inlineSelectionMaxTokens,
buildSelectionPrompt,
normalizeSelectionAction,
@@ -1545,6 +1546,14 @@ async function answerInlineSelection(msg) {
if (!text) return { ok: false, error: 'The model returned an empty answer.' };
return inlineAnswerWasCut(result, maxTokens) ? { ok: true, text, truncated: true } : { ok: true, text };
} catch (error) {
+ if (inlineErrorWasCut(error)) {
+ // The provider reported the output limit as an error. Show what it wrote, if
+ // anything, and let the card hand the rest to the side panel.
+ const partial = String(error.partialContent || '').trim();
+ return partial
+ ? { ok: true, text: partial, truncated: true }
+ : { ok: false, truncated: true, error: 'The answer was cut off before any text was produced.' };
+ }
return { ok: false, error: String(error?.message || error).slice(0, 300) };
} finally {
clearTimeout(timer);
diff --git a/src/firefox/src/content/selection-shortcut.js b/src/firefox/src/content/selection-shortcut.js
index 2111852e2..5fc33ccd1 100644
--- a/src/firefox/src/content/selection-shortcut.js
+++ b/src/firefox/src/content/selection-shortcut.js
@@ -922,12 +922,14 @@
parts.source.textContent = cardState.sourceLabel || str('autoDetect');
parts.target.value = cardState.target;
}
- parts.result.hidden = status === 'idle';
+ // A cut-off with no text has nothing to show or retry, so only the notice remains.
+ const cutOffOnly = status === 'error' && cardState.truncated;
+ parts.result.hidden = status === 'idle' || cutOffOnly;
parts.result.setAttribute('aria-busy', status === 'loading' ? 'true' : 'false');
parts.skeleton.hidden = status !== 'loading';
parts.text.hidden = status !== 'ready';
- parts.error.hidden = status !== 'error';
- parts.cut.hidden = !(status === 'ready' && cardState.truncated);
+ parts.error.hidden = status !== 'error' || cutOffOnly;
+ parts.cut.hidden = !((status === 'ready' || cutOffOnly) && cardState.truncated);
if (status === 'ready') renderAnswer(parts.text, cardState.text, kind !== 'translate');
if (status === 'error') {
parts.errorText.textContent = str('inlineFailed');
@@ -1033,10 +1035,12 @@
if (requestId !== cardRequestId || !cardState) return;
cardJobId = '';
const answer = typeof response?.text === 'string' ? response.text.trim() : '';
- if (!response?.ok || !answer) throw new Error(typeof response?.error === 'string' ? response.error : '');
+ cardState.truncated = response?.truncated === true;
+ if (!response?.ok || !answer) {
+ throw new Error(typeof response?.error === 'string' && !cardState.truncated ? response.error : '');
+ }
cardState.status = 'ready';
cardState.text = answer;
- cardState.truncated = response.truncated === true;
} catch (error) {
if (requestId !== cardRequestId || !cardState) return;
cardJobId = '';
@@ -1060,7 +1064,7 @@
// A cut-off answer is finished in the side panel, where the reply is not capped.
function continueInPanel() {
- if (!cardState || cardState.status !== 'ready') return;
+ if (!cardState?.truncated) return;
const { kind, target } = cardState;
submitSelection(kind, '', kind === 'translate' ? target : interfaceLanguage);
}
diff --git a/src/firefox/src/context-menu-storage.js b/src/firefox/src/context-menu-storage.js
index 9b6d62b25..e6f3ce14a 100644
--- a/src/firefox/src/context-menu-storage.js
+++ b/src/firefox/src/context-menu-storage.js
@@ -288,6 +288,13 @@ export function inlineSelectionMaxTokens(selectionText, action, providerCeiling)
// that report none.
const INLINE_CUT_REASONS = new Set(['length', 'max_tokens', 'max_output_tokens', 'max_completion_tokens']);
+// Some providers (OpenAI Responses) report hitting the output limit as an error
+// instead of a result. Such an error counts as a cut-off, not as a failure to retry.
+export function inlineErrorWasCut(error) {
+ const reason = String(error?.incompleteReason ?? '').trim().toLowerCase();
+ return INLINE_CUT_REASONS.has(reason);
+}
+
export function inlineAnswerWasCut(result, maxTokens) {
const raw = result?.raw;
const reason = [
diff --git a/src/firefox/src/providers/openai.js b/src/firefox/src/providers/openai.js
index 069729782..b1624df23 100644
--- a/src/firefox/src/providers/openai.js
+++ b/src/firefox/src/providers/openai.js
@@ -906,6 +906,9 @@ export class OpenAICompatibleProvider extends BaseLLMProvider {
error.isAskStreamTerminalError = !!stream && !error.isResponsesStreamFallbackSafe;
error.incomplete = true;
error.incompleteReason = reason;
+ // What the model wrote before it stopped, for callers that can show a partial answer.
+ const partial = Array.isArray(response?.output) ? this._responseText(response.output) : '';
+ if (partial) error.partialContent = partial;
return error;
}
diff --git a/test/fixtures/run.mjs b/test/fixtures/run.mjs
index 1e44b8ee8..6629fdc31 100644
--- a/test/fixtures/run.mjs
+++ b/test/fixtures/run.mjs
@@ -2272,6 +2272,39 @@ for (const [label, sourcePath, manualOpen] of [
});
}
+ test(`${label}: an output limit with no text left offers only the side panel, not a retry`, async (page) => {
+ await setupSelectionShortcut(page, sourcePath, { requiresManualOpen: manualOpen });
+ await page.evaluate(() => {
+ window.__selectionInlineHandler = async () => ({
+ ok: false,
+ truncated: true,
+ error: 'The answer was cut off before any text was produced.',
+ });
+ });
+ await openInlineCard(page, 'proofread');
+ await page.waitForFunction(() => window.__webbrainSelectionShortcut.getState().cardTruncated);
+ let state = await page.evaluate(() => window.__webbrainSelectionShortcut.getState());
+ if (state.cardStatus !== 'error' || state.retryRect || state.cardError || !state.continueRect || !state.cutNoticeText) {
+ throw new Error(`a cut-off without text should show the notice and the side-panel button only: ${JSON.stringify(state)}`);
+ }
+ await clickRectCenter(page, state.continueRect);
+ await page.waitForFunction(() => window.__selectionMessages.length === 1);
+ const message = await page.evaluate(() => window.__selectionMessages[0]);
+ if (message.action !== 'proofread' || !/Selected words near the viewport edge/.test(message.selectionText)) {
+ throw new Error(`continuing should hand the same action to the side panel: ${JSON.stringify(message)}`);
+ }
+
+ // A plain failure is unchanged: the reason and a retry, no side-panel button.
+ await page.evaluate(() => {
+ window.__selectionInlineHandler = async () => ({ ok: false, error: 'No provider configured' });
+ });
+ await openInlineCard(page, 'explain');
+ state = await waitForCardStatus(page, 'error');
+ if (state.cardTruncated || state.continueRect || !state.retryRect || state.cardError !== 'No provider configured') {
+ throw new Error(`an ordinary failure must keep its retry and reason: ${JSON.stringify(state)}`);
+ }
+ });
+
test(`${label}: a double click while the language bundle loads starts one card and one model call`, async (page) => {
await setupSelectionShortcut(page, sourcePath, { requiresManualOpen: manualOpen, holdLocalization: true });
const selected = await selectFixtureText(page);
diff --git a/test/run.js b/test/run.js
index fd5cfce19..2198b3f3e 100644
--- a/test/run.js
+++ b/test/run.js
@@ -694,6 +694,7 @@ const {
INLINE_SELECTION_ACTIONS: INLINE_SELECTION_ACTIONS_CH,
INLINE_SELECTION_MAX_CHARS: INLINE_SELECTION_MAX_CHARS_CH,
inlineAnswerWasCut: inlineAnswerWasCutCh,
+ inlineErrorWasCut: inlineErrorWasCutCh,
inlineSelectionMaxTokens: inlineSelectionMaxTokensCh,
buildFollowUpSelectionText: buildFollowUpSelectionTextCh,
buildInlineSelectionMessages: buildInlineSelectionMessagesCh,
@@ -715,6 +716,7 @@ const {
INLINE_SELECTION_ACTIONS: INLINE_SELECTION_ACTIONS_FX,
INLINE_SELECTION_MAX_CHARS: INLINE_SELECTION_MAX_CHARS_FX,
inlineAnswerWasCut: inlineAnswerWasCutFx,
+ inlineErrorWasCut: inlineErrorWasCutFx,
inlineSelectionMaxTokens: inlineSelectionMaxTokensFx,
buildFollowUpSelectionText: buildFollowUpSelectionTextFx,
buildInlineSelectionMessages: buildInlineSelectionMessagesFx,
@@ -51757,6 +51759,7 @@ test('inline card answers are one tool-less, bounded model call behind validated
assert.match(body, /agent\._chatWithCostAllowance\(\s*provider,\s*messages,\s*\{ maxTokens, temperature: 0, signal: controller\.signal \},\s*agent\._newCostRunState\(\),/, `${prefix}: the call is bounded, abortable and goes through the cost allowance`);
assert.match(body, /const maxTokens = inlineSelectionMaxTokens\(msg\.selectionText, msg\.action, provider\?\.maxOutputTokens\);/, `${prefix}: the budget is sized to the selection and the provider's ceiling`);
assert.match(body, /inlineAnswerWasCut\(result, maxTokens\) \? \{ ok: true, text, truncated: true \}/, `${prefix}: a capped answer is flagged instead of passing as complete`);
+ assert.match(body, /catch \(error\) \{\s*if \(inlineErrorWasCut\(error\)\) \{[\s\S]*?\{ ok: true, text: partial, truncated: true \}[\s\S]*?\{ ok: false, truncated: true, error:/, `${prefix}: a provider that reports the output limit as an error still ends in the cut-off state`);
// A card that is closed, retried or retargeted cancels the paid call it no longer needs.
assert.match(body, /const jobId = typeof msg\.jobId === 'string' && INLINE_JOB_ID_PATTERN\.test\(msg\.jobId\) \? msg\.jobId : '';\s*if \(jobId\) inlineSelectionJobs\.set\(jobId, controller\);/, `${prefix}: only well-formed job ids are tracked`);
@@ -51805,6 +51808,45 @@ test('inline card detects an answer that stopped at the output cap on every prov
}
});
+test('inline card treats a Responses output-limit error as a cut-off with any partial text', () => {
+ for (const [label, Provider, wasCutError] of [
+ ['chrome', OpenAIProviderCh, inlineErrorWasCutCh],
+ ['firefox', OpenAIProviderFx, inlineErrorWasCutFx],
+ ]) {
+ const provider = new Provider({
+ category: 'cloud',
+ providerName: 'openai',
+ baseUrl: 'https://api.openai.com/v1',
+ model: 'gpt-5.6-terra',
+ });
+ const thrown = (data) => {
+ try { provider._parseResponsesData(data); } catch (error) { return error; }
+ return null;
+ };
+ const partial = thrown({
+ status: 'incomplete',
+ incomplete_details: { reason: 'max_output_tokens' },
+ output: [{ type: 'message', content: [{ type: 'output_text', text: 'Half of the ' }, { type: 'output_text', text: 'translation' }] }],
+ });
+ assert.equal(partial?.incompleteReason, 'max_output_tokens', `${label}: the provider still raises the output limit as an error for the agent`);
+ assert.equal(partial?.partialContent, 'Half of the translation', `${label}: the error carries what the model wrote before it stopped`);
+ assert.equal(wasCutError(partial), true, `${label}: the card treats that error as a cut-off`);
+
+ const reasoningOnly = thrown({
+ status: 'incomplete',
+ incomplete_details: { reason: 'max_output_tokens' },
+ output: [{ type: 'reasoning', summary: [] }],
+ });
+ assert.equal(wasCutError(reasoningOnly), true, `${label}: a limit spent on reasoning is still a cut-off`);
+ assert.equal('partialContent' in reasoningOnly, false, `${label}: no text means no partial answer`);
+
+ assert.equal(wasCutError(thrown({ status: 'incomplete', incomplete_details: { reason: 'content_filter' } })), false, `${label}: a content filter is a failure, not a cut-off`);
+ assert.equal(wasCutError(provider._responsesIncompleteError({ incomplete_details: { reason: 'missing_response_completed' } }, { stream: true })), false, `${label}: a stream that never completed is a failure`);
+ assert.equal(wasCutError(new Error('network error')), false, `${label}: ordinary errors stay errors`);
+ assert.equal(wasCutError(null), false, `${label}: no error is not a cut-off`);
+ }
+});
+
test('inline card claims its request before waiting, cancels what it abandons and finishes cut-off answers in the panel', () => {
for (const prefix of ['src/chrome', 'src/firefox']) {
const content = fs.readFileSync(path.join(ROOT, prefix, 'src/content/selection-shortcut.js'), 'utf8');
@@ -51829,9 +51871,9 @@ test('inline card claims its request before waiting, cancels what it abandons an
assert.match(body('function cancelInline()', 'async function runInline()'), /type: INLINE_CANCEL_MESSAGE, jobId/, `${prefix}: cancellation names the job`);
// A capped answer is never shown as complete.
- assert.match(run, /cardState\.truncated = response\.truncated === true;/, `${prefix}: the card records the cut-off flag`);
- assert.match(content, /parts\.cut\.hidden = !\(status === 'ready' && cardState\.truncated\);/, `${prefix}: the notice shows for a cut-off answer only`);
- assert.match(body('function continueInPanel()', 'function resetCard()'), /submitSelection\(kind, '', kind === 'translate' \? target : interfaceLanguage\)/, `${prefix}: continuing hands the same action to the side panel`);
+ assert.match(run, /cardState\.truncated = response\?\.truncated === true;\s*if \(!response\?\.ok \|\| !answer\) \{\s*throw new Error\(typeof response\?\.error === 'string' && !cardState\.truncated \? response\.error : ''\);/, `${prefix}: the card records the cut-off flag even when no text came back`);
+ assert.match(content, /const cutOffOnly = status === 'error' && cardState\.truncated;[\s\S]*?parts\.error\.hidden = status !== 'error' \|\| cutOffOnly;\s*parts\.cut\.hidden = !\(\(status === 'ready' \|\| cutOffOnly\) && cardState\.truncated\);/, `${prefix}: the notice shows for a cut-off answer only, and a cut-off with no text drops the retry`);
+ assert.match(body('function continueInPanel()', 'function resetCard()'), /if \(!cardState\?\.truncated\) return;[\s\S]*submitSelection\(kind, '', kind === 'translate' \? target : interfaceLanguage\)/, `${prefix}: continuing is only offered for a cut-off answer and hands the same action to the side panel`);
}
});