Skip to content

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

Carry the several-deliverables flag across a trusted Continue

Claudea3a91dbf0c8c
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.js
index cb29fcd0f..bce2f5347 100644
--- a/src/chrome/src/agent/agent.js
+++ b/src/chrome/src/agent/agent.js
@@ -40093,8 +40093,10 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
&& gateOutcome.deliverables.every(value => typeof value === 'string' && value.length <= 1000)
? [...gateOutcome.deliverables]
: [],
+ // A continuation of the same task still owes the deliverables the first run did.
multipleDeliverables: gateOutcome?.multipleDeliverables === true
- || (Array.isArray(gateOutcome?.deliverables) && gateOutcome.deliverables.length > 1),
+ || (Array.isArray(gateOutcome?.deliverables) && gateOutcome.deliverables.length > 1)
+ || (carryScopeMatches && carried?.multipleDeliverables === true),
siteWorkflowUrl: String(gateOutcome?.siteWorkflowUrl || ''),
messaging,
messagingConversationScope,
@@ -40671,6 +40673,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
scheduledResume: guard.scheduledResume === true,
conditionalSiteWorkflow: guard.conditionalSiteWorkflow,
conditionalMutationBlocked: guard.conditionalMutationBlocked === true,
+ multipleDeliverables: guard.multipleDeliverables === true || (guard.deliverables || []).length > 1,
approvedPlanAnchor: guard.approvedPlanAnchor,
approvedPlanText: guard.approvedPlanText,
taskKey: guard.taskKey,
diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.js
index b7cafdd60..5f0307043 100644
--- a/src/firefox/src/agent/agent.js
+++ b/src/firefox/src/agent/agent.js
@@ -36835,8 +36835,10 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
&& gateOutcome.deliverables.every(value => typeof value === 'string' && value.length <= 1000)
? [...gateOutcome.deliverables]
: [],
+ // A continuation of the same task still owes the deliverables the first run did.
multipleDeliverables: gateOutcome?.multipleDeliverables === true
- || (Array.isArray(gateOutcome?.deliverables) && gateOutcome.deliverables.length > 1),
+ || (Array.isArray(gateOutcome?.deliverables) && gateOutcome.deliverables.length > 1)
+ || (carryScopeMatches && carried?.multipleDeliverables === true),
siteWorkflowUrl: String(gateOutcome?.siteWorkflowUrl || ''),
messaging,
messagingConversationScope,
@@ -37413,6 +37415,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
scheduledResume: guard.scheduledResume === true,
conditionalSiteWorkflow: guard.conditionalSiteWorkflow,
conditionalMutationBlocked: guard.conditionalMutationBlocked === true,
+ multipleDeliverables: guard.multipleDeliverables === true || (guard.deliverables || []).length > 1,
approvedPlanAnchor: guard.approvedPlanAnchor,
approvedPlanText: guard.approvedPlanText,
taskKey: guard.taskKey,
diff --git a/test/post-submit-model-recovery.mjs b/test/post-submit-model-recovery.mjs
index 34c6a1b6a..e7c8ab3c5 100644
--- a/test/post-submit-model-recovery.mjs
+++ b/test/post-submit-model-recovery.mjs
@@ -165,6 +165,24 @@ for (const browser of ['chrome', 'firefox']) {
assert.deepEqual(f.operations, [], 'one verified submission cannot finish nine');
});
+ test(`${browser}: Continue keeps the task's several-deliverables flag`, async () => {
+ const f = fixture();
+ const gate = { requestKind: 'execute', requiresStateChange: true, requiresSubmission: true };
+ f.agent._startPlanExecutionGuard(f.tab, 'act', { ...gate, deliverables: [], multipleDeliverables: true }, {});
+ const first = f.agent._planExecutionGuards.get(f.tab);
+ assert.equal(first.multipleDeliverables, true);
+ first.successfulTaskToolCalls = 1;
+ first.successfulConsequentialToolCalls = 1;
+ f.agent._storeContinuationExecutionEvidence(f.tab);
+ assert.equal(f.agent._continuationExecutionEvidence.get(f.tab)?.multipleDeliverables, true, 'the first run stores the flag');
+ // The trusted continuation's own plan names no deliverables.
+ f.agent._startPlanExecutionGuard(f.tab, 'act', { ...gate, deliverables: [] }, { trustedContinuation: true });
+ assert.equal(f.agent._planExecutionGuards.get(f.tab).multipleDeliverables, true, 'the continuation restores it');
+ // A run that is not a continuation does not inherit it.
+ f.agent._startPlanExecutionGuard(f.tab, 'act', { ...gate, deliverables: [] }, {});
+ assert.equal(f.agent._planExecutionGuards.get(f.tab).multipleDeliverables, false);
+ });
+
test(`${browser}: a later self-verifying action does not make the earlier submission the latest step`, async () => {
// The submission left verification debt; scheduling a follow-up then
// succeeded and proved itself. lastAction still names the submission, but