Skip to content

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

Guard policy: a plain arithmetic answer before the calculator ran ends as partial

Barack Sokulluc0daba5e4cab
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.js
index c4ec3b583..775b0f6f3 100644
--- a/src/chrome/src/agent/agent.js
+++ b/src/chrome/src/agent/agent.js
@@ -14713,6 +14713,19 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
return calculationResultInText(summary, result, fraction);
}
+ // A plain final accepted in a literal-arithmetic task before any
+ // calculator attempt (the plan guard gives its notes first): WebBrain knows
+ // its answer was not computed. As for a premature done(success), it ends
+ // as partial with that caveat (pure-calculation-done-checks).
+ _notePrematureCalculationFinal(tabId) {
+ const policy = this._pureCalculationToolPolicy(tabId);
+ if (!policy || policy.stale || policy.attempted || !policy.names.includes('calculate')) return;
+ if (this._guardOutcome('pure-calculation-done-checks', { decided: true, tabId }) !== 'advise') return;
+ const state = this._guardRunState(tabId);
+ state.caveats.set('pure-calculation-done-checks', 'The result was not computed with the calculator.');
+ (state.outcomeDowngrades ||= new Set()).add('pure-calculation-done-checks');
+ }
+
_pureCalculationToolBlock(tabId, name, args, expectedScope) {
const policy = this._pureCalculationToolPolicy(tabId, expectedScope);
if (!policy) return null;
@@ -53437,6 +53450,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
runOptions,
);
finalResponse = this._withStandaloneUnverifiedNotice(finalResponse, standaloneGroundingGap, runOptions);
+ this._notePrematureCalculationFinal(tabId);
finalResponse = this._withCompletionCaveat(tabId, finalResponse);
messages.push(this._withResponseItems({ role: 'assistant', content: finalResponse }, result.responseItems, result.reasoningContent, provider));
onUpdate('text', { content: finalResponse, replace: true });
@@ -54819,6 +54833,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
fullText = attributedFullText;
onUpdate('text', { content: fullText, replace: true });
}
+ this._notePrematureCalculationFinal(tabId);
const caveatedFullText = this._withCompletionCaveat(tabId, fullText);
if (caveatedFullText !== fullText) {
fullText = caveatedFullText;
diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.js
index 0a5278d29..cf1f3673d 100644
--- a/src/firefox/src/agent/agent.js
+++ b/src/firefox/src/agent/agent.js
@@ -13078,6 +13078,19 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
return calculationResultInText(summary, result, fraction);
}
+ // A plain final accepted in a literal-arithmetic task before any
+ // calculator attempt (the plan guard gives its notes first): WebBrain knows
+ // its answer was not computed. As for a premature done(success), it ends
+ // as partial with that caveat (pure-calculation-done-checks).
+ _notePrematureCalculationFinal(tabId) {
+ const policy = this._pureCalculationToolPolicy(tabId);
+ if (!policy || policy.stale || policy.attempted || !policy.names.includes('calculate')) return;
+ if (this._guardOutcome('pure-calculation-done-checks', { decided: true, tabId }) !== 'advise') return;
+ const state = this._guardRunState(tabId);
+ state.caveats.set('pure-calculation-done-checks', 'The result was not computed with the calculator.');
+ (state.outcomeDowngrades ||= new Set()).add('pure-calculation-done-checks');
+ }
+
_pureCalculationToolBlock(tabId, name, args, expectedScope) {
const policy = this._pureCalculationToolPolicy(tabId, expectedScope);
if (!policy) return null;
@@ -45020,6 +45033,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
finalResponse = result.costAllowanceMessage
? `${repairedFinalContent}\n\n${result.costAllowanceMessage}`
: repairedFinalContent;
+ this._notePrematureCalculationFinal(tabId);
finalResponse = this._withCompletionCaveat(tabId, finalResponse);
messages.push(this._withResponseItems({ role: 'assistant', content: finalResponse }, result.responseItems, result.reasoningContent, provider));
onUpdate('text', { content: finalResponse });
@@ -46233,6 +46247,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
onUpdate('text_delta', { content: `\n\n${costStopMessage}` });
fullText = `${fullText}\n\n${costStopMessage}`;
}
+ this._notePrematureCalculationFinal(tabId);
const caveatedFullText = this._withCompletionCaveat(tabId, fullText);
if (caveatedFullText !== fullText) {
fullText = caveatedFullText;
diff --git a/test/pure-calculation-reply.mjs b/test/pure-calculation-reply.mjs
index 6319c8830..1e2c58348 100644
--- a/test/pure-calculation-reply.mjs
+++ b/test/pure-calculation-reply.mjs
@@ -344,6 +344,40 @@ for (const build of ['chrome', 'firefox']) {
});
}
+ for (const streaming of [false, true]) {
+ test(`${build}: ${streaming ? 'stream' : 'chat'} a plain arithmetic answer accepted before any calculation ends as partial`, async () => {
+ // The model never calls calculate. By default the plan guard's notes are
+ // given once each, then the plain answer is accepted, but not as a
+ // success: WebBrain knows the result was not computed.
+ const tab = nextTab++, updates = []; let agent, send, requests = 0;
+ agent = makeAgent(Agent, async () => {
+ requests++; send({});
+ return { content: answer };
+ });
+ agent.maxSteps = 8;
+ agent._enrichUserMessageWithCurrentPage = async (_tab, _messages, content) => {
+ send = bind(agent, tab); return { role: 'user', content };
+ };
+ agent._runPlannerIntentGate = async () => ({ proceed: true, requestKind: 'execute',
+ requiresStateChange: false, requiresSubmission: false,
+ ...agent._plannerCompletionGateFields({ read_scope: 'none', deliverables: scope.deliverables }) });
+ agent.executeTool = async (_tab, name) => {
+ if (name === 'get_accessibility_tree') return { success: true, pageContent: 'text "Owned passive counter"' };
+ assert.fail(`No tool runs in this task: ${name}`);
+ };
+ const update = (type, data) => updates.push({ type, data });
+ const options = { detachedRequestId: `plain-before-calc-${tab}`, askStreamingEnabled: false };
+ try {
+ const result = streaming ? await agent.processMessageStream(tab, ORIGINAL_CALC_CORRECTION, update, 'act', options)
+ : await agent.processMessage(tab, ORIGINAL_CALC_CORRECTION, update, 'act', [], options);
+ assert.ok(requests >= 2, 'the plain answer gets a note before it is accepted');
+ assert.equal(updates.some(e => e.type === 'tool_call' && e.data.name === 'calculate'), false);
+ assert.match(result, /The result was not computed with the calculator\./);
+ assert.ok(updates.some(e => e.type === 'run_status' && e.data?.status === 'partial'), 'the run ends as partial');
+ } finally { documents.delete(tab); }
+ });
+ }
+
for (const beforeCalculation of [false, true]) for (const change of ['none', 'no_actual_receipt', 'mixed_task', 'no_literal_task', 'quoted_task', 'provider_config', 'same_text_new_task',
'new_completion_run', 'browser_action', 'verification_debt', 'iframe_debt', 'send_obligation',
'ledger', 'pending_send', 'queued_steering', 'human_input', 'unknown_intervention', 'navigation',