Skip to content

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

Guard policy: downgrade before the chat partial ending

Barack Sokullu247e8b00cd51
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.js
index 3214d6b76..7c1f6e93c 100644
--- a/src/chrome/src/agent/agent.js
+++ b/src/chrome/src/agent/agent.js
@@ -16630,6 +16630,14 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
return { action: 'continue' };
}
if (!scheduledResume) await cancelPendingResumes();
+ if (!scheduledResume && toolResult.outcome === 'success'
+ && 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.
+ // Before the chat ending below, so a chat run gets the same
+ // partial contract (runEnded, no auto-resume, the notice).
+ toolResult.outcome = 'partial';
+ }
if (!scheduledResume && runOptions?.scheduledRun !== true && this.chatSessions.has(tabId)
&& ['partial', 'failed'].includes(toolResult.outcome)) {
toolResult.runEnded = true;
@@ -16639,12 +16647,6 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
toolResult.summary = (toolResult.summary || partialAssistantText || '') + '\n\n' + notice;
}
}
- if (!scheduledResume && toolResult.outcome === 'success'
- && 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';
- }
onUpdate('tool_result', { name: fnName, result: toolResult });
this._doneBlockCount.delete(tabId);
rawDoneSummary = toolResult.summary || rawDoneSummary;
diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.js
index 2585d0c9e..aef4785c0 100644
--- a/src/firefox/src/agent/agent.js
+++ b/src/firefox/src/agent/agent.js
@@ -14897,6 +14897,14 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
return { action: 'continue' };
}
if (!scheduledResume) await cancelPendingResumes();
+ if (!scheduledResume && toolResult.outcome === 'success'
+ && 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.
+ // Before the chat ending below, so a chat run gets the same
+ // partial contract (runEnded, no auto-resume, the notice).
+ toolResult.outcome = 'partial';
+ }
if (!scheduledResume && runOptions?.scheduledRun !== true && this.chatSessions.has(tabId)
&& ['partial', 'failed'].includes(toolResult.outcome)) {
toolResult.runEnded = true;
@@ -14906,12 +14914,6 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
toolResult.summary = (toolResult.summary || partialAssistantText || '') + '\n\n' + notice;
}
}
- if (!scheduledResume && toolResult.outcome === 'success'
- && 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';
- }
onUpdate('tool_result', { name: fnName, result: toolResult });
this._doneBlockCount.delete(tabId);
rawDoneSummary = toolResult.summary || rawDoneSummary;
diff --git a/test/guard-advisory-completion.mjs b/test/guard-advisory-completion.mjs
index af24b4f75..e63fe4b3d 100644
--- a/test/guard-advisory-completion.mjs
+++ b/test/guard-advisory-completion.mjs
@@ -419,6 +419,39 @@ for (const build of ['chrome', 'firefox']) {
assert.equal(doneResult?.outcome, 'success', `${build}: missing evidence alone must not downgrade the outcome`);
});
+ test(`${build}: a chat run's success accepted over a decided check ends with the chat partial contract`, async () => {
+ const { chatRunEndedNotice, createChatSession } = await import(`../src/${build}/src/agent/chat-workflow.js`);
+ for (const decided of [true, false]) {
+ const h = batchAgent(Agent);
+ h.agent.chatSessions.set(h.tab, createChatSession({ threadKey: 'dom:telegram:k:123456:0' }));
+ if (decided) {
+ // Text went into a formatting toolbar: the check knows the claim is wrong.
+ h.agent._richTextToolbarGuard.completionAction = () => ({ name: 'type_ax', ref_id: 'ref_bold' });
+ } else {
+ // Only missing evidence: the page was not read again.
+ h.agent.completionInvariants.set(h.tab, { ...invariant.createCompletionInvariantState('chat'), hadAction: true,
+ verificationDebt: true, sequence: 1, lastAction: { name: 'click_ax', sequence: 1 } });
+ }
+ const done = ['done', { summary: 'Replied in the chat.', outcome: 'success' }];
+ assert.equal((await h.batch([done])).action, 'continue', `${build}: the first completion gets its note`);
+ const updates = [];
+ const accepted = await h.batch([done], updates);
+ assert.equal(accepted.action, 'return');
+ const doneResult = updates.filter(update => update.type === 'tool_result' && update.data.name === 'done').at(-1)?.data.result;
+ if (decided) {
+ assert.equal(doneResult?.outcome, 'partial', `${build}: a success over a decided check was not downgraded`);
+ assert.equal(doneResult?.runEnded, true, `${build}: a downgraded chat completion must end the chat run`);
+ assert.equal(doneResult?.autoResumeScheduled, false, `${build}: a downgraded chat completion must not auto-resume`);
+ assert.ok(String(doneResult?.summary).includes(chatRunEndedNotice()), `${build}: the chat run-ended notice is missing`);
+ } else {
+ assert.equal(doneResult?.outcome, 'success', `${build}: missing evidence alone must not downgrade a chat success`);
+ assert.equal(doneResult?.runEnded, undefined);
+ assert.equal(doneResult?.autoResumeScheduled, undefined);
+ assert.ok(!String(doneResult?.summary).includes(chatRunEndedNotice()));
+ }
+ }
+ });
+
test(`${build}: a done with no outcome is advised once, then accepted as partial`, async () => {
for (const strict of [false, true]) {
const h = batchAgent(Agent, { strict });