Skip to content

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

Guard policy: a plain final accepted over a decided check ends as partial

Barack Sokulludb75df53312b
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.js
index 0b1dd5962..3214d6b76 100644
--- a/src/chrome/src/agent/agent.js
+++ b/src/chrome/src/agent/agent.js
@@ -5292,6 +5292,14 @@ export class Agent extends LoopDetector {
return this._guardAdvises(id, { decided }) && this._guardRunState(tabId).advised.has(id);
}
+ // A completion accepted over a check that KNOWS the task is not done
+ // (OUTCOME_DOWNGRADING_GUARDS) ends the run as partial. The done tool
+ // downgrades its own outcome; a plain final has no outcome of its own, so
+ // its run reports partial (trace status, run_status, onRunFinished).
+ _completionOutcomeDowngraded(tabId) {
+ return (this._guardRunState(tabId).outcomeDowngrades?.size || 0) > 0;
+ }
+
// Adds the "WebBrain could not verify everything" line, if any, to the
// answer of an accepted completion. _guardRunState first drops a previous
// run's caveats.
@@ -16632,7 +16640,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
}
}
if (!scheduledResume && toolResult.outcome === 'success'
- && this._guardRunState(tabId).outcomeDowngrades?.size) {
+ && this._completionOutcomeDowngraded(tabId)) {
// Accepted over a guard that knows the claim is wrong: end the run,
// but as partial so schedules and watches do not count a success.
toolResult.outcome = 'partial';
@@ -53213,6 +53221,10 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
finalResponse = this._withCompletionCaveat(tabId, finalResponse);
messages.push(this._withResponseItems({ role: 'assistant', content: finalResponse }, result.responseItems, result.reasoningContent, provider));
onUpdate('text', { content: finalResponse, replace: true });
+ if (this._completionOutcomeDowngraded(tabId)) {
+ _traceStatus = 'partial';
+ onUpdate('run_status', { status: 'partial', message: finalResponse });
+ }
break;
}
@@ -54569,6 +54581,10 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
messages.push(this._withResponseItems({ role: 'assistant', content: fullText }, responseItems, reasoningContent, provider));
this._persist(tabId);
closeTraceStep({ ok: true });
+ if (this._completionOutcomeDowngraded(tabId)) {
+ onUpdate('run_status', { status: 'partial', message: fullText });
+ return finish(fullText, 'partial');
+ }
return finish(fullText);
} catch (e) {
diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.js
index 33a3cb265..2585d0c9e 100644
--- a/src/firefox/src/agent/agent.js
+++ b/src/firefox/src/agent/agent.js
@@ -4844,6 +4844,14 @@ export class Agent extends LoopDetector {
return this._guardAdvises(id, { decided }) && this._guardRunState(tabId).advised.has(id);
}
+ // A completion accepted over a check that KNOWS the task is not done
+ // (OUTCOME_DOWNGRADING_GUARDS) ends the run as partial. The done tool
+ // downgrades its own outcome; a plain final has no outcome of its own, so
+ // its run reports partial (trace status, run_status, onRunFinished).
+ _completionOutcomeDowngraded(tabId) {
+ return (this._guardRunState(tabId).outcomeDowngrades?.size || 0) > 0;
+ }
+
// Adds the "WebBrain could not verify everything" line, if any, to the
// answer of an accepted completion. _guardRunState first drops a previous
// run's caveats.
@@ -14899,7 +14907,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
}
}
if (!scheduledResume && toolResult.outcome === 'success'
- && this._guardRunState(tabId).outcomeDowngrades?.size) {
+ && this._completionOutcomeDowngraded(tabId)) {
// Accepted over a guard that knows the claim is wrong: end the run,
// but as partial so schedules and watches do not count a success.
toolResult.outcome = 'partial';
@@ -44843,6 +44851,10 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
finalResponse = this._withCompletionCaveat(tabId, finalResponse);
messages.push(this._withResponseItems({ role: 'assistant', content: finalResponse }, result.responseItems, result.reasoningContent, provider));
onUpdate('text', { content: finalResponse });
+ if (this._completionOutcomeDowngraded(tabId)) {
+ _traceStatus = 'partial';
+ onUpdate('run_status', { status: 'partial', message: finalResponse });
+ }
break;
}
@@ -46030,6 +46042,10 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
messages.push(this._withResponseItems({ role: 'assistant', content: fullText }, responseItems, reasoningContent, provider));
this._persist(tabId);
closeTraceStep({ ok: true });
+ if (this._completionOutcomeDowngraded(tabId)) {
+ onUpdate('run_status', { status: 'partial', message: fullText });
+ return finish(fullText, 'partial');
+ }
return finish(fullText);
} catch (e) {
diff --git a/test/guard-advisory-completion.mjs b/test/guard-advisory-completion.mjs
index a1eb195f6..af24b4f75 100644
--- a/test/guard-advisory-completion.mjs
+++ b/test/guard-advisory-completion.mjs
@@ -238,6 +238,41 @@ for (const build of ['chrome', 'firefox']) {
}
}
});
+
+ test(`${label}: a plain final accepted over a check that knows the task is not done ends the run as partial`, async () => {
+ // decided: the last text went into a formatting toolbar (the check
+ // knows). Undecided: the page was not read after the last action.
+ for (const decided of [true, false]) {
+ const tab = nextTab++;
+ let n = 0;
+ const agent = loopAgent(Agent, async () => {
+ n++;
+ if (!decided && n === 1) return call('save', 'click_ax', { ref_id: 'ref_save' });
+ return { content: 'Wrote the post.' };
+ });
+ if (decided) {
+ agent._richTextToolbarGuard.hasPending = () => true;
+ agent._richTextToolbarGuard.completionAction = () => ({ name: 'type_ax', ref_id: 'ref_bold' });
+ }
+ const runStatuses = [];
+ let finished = null;
+ const onUpdate = (type, data) => { if (type === 'run_status') runStatuses.push(data?.status); };
+ const options = { detachedRequestId: `guard-${tab}`, onRunFinished: status => { finished = status; } };
+ const result = streaming
+ ? await agent.processMessageStream(tab, 'Write the post.', onUpdate, 'act', options)
+ : await agent.processMessage(tab, 'Write the post.', onUpdate, 'act', [], options);
+ assert.ok(String(result).startsWith(`Wrote the post.\n\n${CAVEAT}`), `${label}: the plain final is accepted with its caveat`);
+ if (decided) {
+ assert.match(result, /formatting toolbar instead of the editor\./);
+ assert.equal(finished, 'partial', `${label}: the run finished as a success over a check that knows it is not done`);
+ assert.deepEqual(runStatuses, ['partial'], `${label}: the scheduler did not hear the partial ending`);
+ } else {
+ assert.match(result, /The page was not checked again after the last action\./);
+ assert.equal(finished, 'done', `${label}: missing evidence alone must not downgrade the run`);
+ assert.deepEqual(runStatuses, []);
+ }
+ }
+ });
}
test(`${build}: each untraced run gets its own one refusal per guard`, async () => {