Open-source AI browser agent for Chrome and Firefox (monorepo) ðŸ§
JavaScript ai-agentai-agentsai-toolsbrowser-aibrowser-automationbrowser-extensionchrome-aichrome-devtoolschrome-extensionchrome-extensionsfirefox-addonfirefox-addonsfirefox-extensionfirefox-extensionsgemmagemma4minimaxoffline-aioffline-llmqwen
Address PR #343 review findings on message recipient guard
Emre Sokullu5dbfa0be098b
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.jsindex 07d7193ec..7eb64006a 100644--- a/src/chrome/src/agent/agent.js+++ b/src/chrome/src/agent/agent.js@@ -17758,17 +17758,6 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d...(target?.target_kind === 'named' ? { expectedRecipients: target.recipients } : {}),});if (probe?.success === true && probe?.conclusive === true && probe.messageSend === false) return null;- // A plan that declares neither submission nor state change cannot legitimately- // dispatch a message, so an unclassifiable click under it is a read rather than- // a send. Fail open for those instead of blocking every control the classifier- // cannot resolve. A conclusive send stays blocked: a read-only plan still may- // not deliver a message without recipient authorization.- const explicitlyReadOnly = guard?.requiresSubmission === false- && guard?.requiresStateChange === false;- const conclusiveMessageSend = probe?.success === true- && probe?.conclusive === true- && probe.messageSend === true;- if (explicitlyReadOnly && !conclusiveMessageSend) return null;// Gmail reply editors can be collapsed when planning begins. Allow only a// content-verified Reply/Reply all/Forward control to open the editor;// every other unresolved click remains blocked. Recipient authorization@@ -17873,16 +17862,19 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d* "receive", approval for any identity it liked. Authorization is therefore* the intersection of what the user picked and what the page actually shows.* The bound target is always an observed candidate, never the answer text,- * and an answer matching no candidate or more than one authorizes nothing.+ * and an answer matching no candidate or only a subset of observed+ * candidates authorizes nothing. When multiple recipients are observed on+ * the page (e.g. Reply all), the answer must explicitly authorize the full+ * observed recipient set.*/- _bindClarifiedMessageRecipient(tabId, answer) {+ _bindClarifiedMessageRecipient(tabId, answer, source = 'user') {const guard = this._planExecutionGuards.get(tabId);if (!guard) return false;// Read and clear together. The consent the guard staged belongs to exactly// one clarify, the one its own error told the model to ask; every later// question is a different question and must not inherit it. Clearing on- // the failing paths too keeps a non-matching answer from leaving consent- // parked for some unrelated answer to pick up.+ // all paths (including timeout, auto, mismatch, cancellation) keeps staged+ // consent from remaining parked for an unrelated later clarify to pick up.const pending = guard.pendingRecipientAuthorization === true;const observed = Array.isArray(guard.observedRecipientCandidates)? guard.observedRecipientCandidates@@ -17890,13 +17882,14 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout dguard.pendingRecipientAuthorization = false;guard.observedRecipientCandidates = null;if (!pending || observed.length === 0) return false;+ if (!answer || source === 'timeout' || source === 'auto') return false;const normalizedAnswer = normalizeRecipientIdentity(answer);if (!normalizedAnswer) return false;const matched = observed.filter(candidate => answerNamesIdentity(normalizedAnswer,normalizeRecipientIdentity(candidate?.identity),));- if (matched.length !== 1) return false;+ if (matched.length === 0 || matched.length !== observed.length) return false;const target = normalizeMessageTarget({ target_kind: 'named', recipients: matched });if (!target) return false;guard.messaging = target;@@ -26180,6 +26173,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout dif (tabPending.size === 0) this._pendingClarifications.delete(tabId);if (response && response.cancelled) {+ this._bindClarifiedMessageRecipient(tabId, '', 'cancelled');return { success: false, cancelled: true, reason: response.reason || 'clarify cancelled' };}const answer = String(response?.answer || '').trim();@@ -26191,9 +26185,9 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d// never sees it, and the run loops until it exhausts its step budget.// A timed-out or auto-selected answer is not a confirmation and binds// nothing, matching how research escalation treats those sources below.- if (answer && source !== 'timeout' && source !== 'auto') {- this._bindClarifiedMessageRecipient(tabId, answer);- }+ // Every clarification outcome (human answer, timeout, auto) must consume+ // the staged recipient consent so a later unrelated clarify does not inherit it.+ this._bindClarifiedMessageRecipient(tabId, answer, source);const explicitResearchApproval = isResearchEscalation&& source !== 'timeout'&& source !== 'auto'diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.jsindex 6c43169b7..0373a962b 100644--- a/src/firefox/src/agent/agent.js+++ b/src/firefox/src/agent/agent.js@@ -15511,17 +15511,6 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d...(target?.target_kind === 'named' ? { expectedRecipients: target.recipients } : {}),});if (probe?.success === true && probe?.conclusive === true && probe.messageSend === false) return null;- // A plan that declares neither submission nor state change cannot legitimately- // dispatch a message, so an unclassifiable click under it is a read rather than- // a send. Fail open for those instead of blocking every control the classifier- // cannot resolve. A conclusive send stays blocked: a read-only plan still may- // not deliver a message without recipient authorization.- const explicitlyReadOnly = guard?.requiresSubmission === false- && guard?.requiresStateChange === false;- const conclusiveMessageSend = probe?.success === true- && probe?.conclusive === true- && probe.messageSend === true;- if (explicitlyReadOnly && !conclusiveMessageSend) return null;// Gmail reply editors can be collapsed when planning begins. Allow only a// content-verified Reply/Reply all/Forward control to open the editor;// every other unresolved click remains blocked. Recipient authorization@@ -15626,16 +15615,19 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d* "receive", approval for any identity it liked. Authorization is therefore* the intersection of what the user picked and what the page actually shows.* The bound target is always an observed candidate, never the answer text,- * and an answer matching no candidate or more than one authorizes nothing.+ * and an answer matching no candidate or only a subset of observed+ * candidates authorizes nothing. When multiple recipients are observed on+ * the page (e.g. Reply all), the answer must explicitly authorize the full+ * observed recipient set.*/- _bindClarifiedMessageRecipient(tabId, answer) {+ _bindClarifiedMessageRecipient(tabId, answer, source = 'user') {const guard = this._planExecutionGuards.get(tabId);if (!guard) return false;// Read and clear together. The consent the guard staged belongs to exactly// one clarify, the one its own error told the model to ask; every later// question is a different question and must not inherit it. Clearing on- // the failing paths too keeps a non-matching answer from leaving consent- // parked for some unrelated answer to pick up.+ // all paths (including timeout, auto, mismatch, cancellation) keeps staged+ // consent from remaining parked for an unrelated later clarify to pick up.const pending = guard.pendingRecipientAuthorization === true;const observed = Array.isArray(guard.observedRecipientCandidates)? guard.observedRecipientCandidates@@ -15643,13 +15635,14 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout dguard.pendingRecipientAuthorization = false;guard.observedRecipientCandidates = null;if (!pending || observed.length === 0) return false;+ if (!answer || source === 'timeout' || source === 'auto') return false;const normalizedAnswer = normalizeRecipientIdentity(answer);if (!normalizedAnswer) return false;const matched = observed.filter(candidate => answerNamesIdentity(normalizedAnswer,normalizeRecipientIdentity(candidate?.identity),));- if (matched.length !== 1) return false;+ if (matched.length === 0 || matched.length !== observed.length) return false;const target = normalizeMessageTarget({ target_kind: 'named', recipients: matched });if (!target) return false;guard.messaging = target;@@ -22944,6 +22937,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout dif (tabPending.size === 0) this._pendingClarifications.delete(tabId);if (response && response.cancelled) {+ this._bindClarifiedMessageRecipient(tabId, '', 'cancelled');return { success: false, cancelled: true, reason: response.reason || 'clarify cancelled' };}const answer = String(response?.answer || '').trim();@@ -22955,9 +22949,9 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d// never sees it, and the run loops until it exhausts its step budget.// A timed-out or auto-selected answer is not a confirmation and binds// nothing, matching how research escalation treats those sources below.- if (answer && source !== 'timeout' && source !== 'auto') {- this._bindClarifiedMessageRecipient(tabId, answer);- }+ // Every clarification outcome (human answer, timeout, auto) must consume+ // the staged recipient consent so a later unrelated clarify does not inherit it.+ this._bindClarifiedMessageRecipient(tabId, answer, source);const explicitResearchApproval = isResearchEscalation&& source !== 'timeout'&& source !== 'auto'diff --git a/test/run.js b/test/run.jsindex e1803d059..d46d5f46c 100644--- a/test/run.js+++ b/test/run.js@@ -4956,24 +4956,24 @@ test('direct-message recipient guard uses structured intent and exact active ideassert.equal(unsafe?.reasonCode, 'recipient_unverifiable_dispatch_path');}- // A plan declaring neither submission nor state change cannot legitimately- // send, so an unclassifiable click under it is a read. Blocking those made- // a draft-only Gmail task unable to operate the thread it was reading.+ // A read-only plan cannot bypass the recipient guard on an inconclusive click.+ // Planner flags describe requested intent, not DOM effects: an unclassified+ // control might be a Send button, so only a conclusively non-send action may proceed.agent._planExecutionGuards.set(tabId, {messaging: null, requiresSubmission: false, requiresStateChange: false,});probe = { success: true, conclusive: false, messageSend: null, identityCandidates: [] };- assert.equal(- await agent._messageRecipientGuardBlock(- tabId,- 'click_ax',- { ref_id: 'ref_expand_all' },- 'https://mail.google.com/mail/u/0/#inbox/thread-1',- {},- ),- null,- `${label}: read-only plan still blocked an unclassifiable thread click`,+ const readOnlyInconclusive = await agent._messageRecipientGuardBlock(+ tabId,+ 'click_ax',+ { ref_id: 'ref_expand_all' },+ 'https://mail.google.com/mail/u/0/#inbox/thread-1',+ {},);+ assert.equal(readOnlyInconclusive?.blocked, true,+ `${label}: read-only plan allowed an inconclusive click to bypass recipient guard`);+ assert.equal(readOnlyInconclusive?.reasonCode, 'message_send_classification_inconclusive',+ `${label}: inconclusive click did not fail closed`);// The same read-only plan must still not deliver an actual message.probe = {@@ -5043,14 +5043,67 @@ test('direct-message recipient guard uses structured intent and exact active ide`${label}: a later clarify inherited consent staged for an earlier one`,);+ // A timed-out or auto-selected clarify must consume staged consent without binding.+ agent._planExecutionGuards.set(tabId, clarifyGuardState([{ identity: '[email protected]', role: 'to' }]));+ assert.equal(+ agent._bindClarifiedMessageRecipient(tabId, '[email protected]', 'timeout'),+ false,+ `${label}: timed-out clarify bound a recipient`,+ );+ assert.equal(agent._planExecutionGuards.get(tabId)?.pendingRecipientAuthorization, false,+ `${label}: timed-out clarify left pendingRecipientAuthorization true`);+ assert.equal(agent._planExecutionGuards.get(tabId)?.observedRecipientCandidates, null,+ `${label}: timed-out clarify left observedRecipientCandidates intact`);+ assert.equal(+ agent._bindClarifiedMessageRecipient(tabId, '[email protected]'),+ false,+ `${label}: subsequent clarify inherited consent from timed-out clarify`,+ );++ agent._planExecutionGuards.set(tabId, clarifyGuardState([{ identity: '[email protected]', role: 'to' }]));+ assert.equal(+ agent._bindClarifiedMessageRecipient(tabId, '[email protected]', 'auto'),+ false,+ `${label}: auto-selected clarify bound a recipient`,+ );+ assert.equal(agent._planExecutionGuards.get(tabId)?.pendingRecipientAuthorization, false,+ `${label}: auto-selected clarify left pendingRecipientAuthorization true`);+ assert.equal(agent._planExecutionGuards.get(tabId)?.observedRecipientCandidates, null,+ `${label}: auto-selected clarify left observedRecipientCandidates intact`);+ assert.equal(+ agent._bindClarifiedMessageRecipient(tabId, '[email protected]'),+ false,+ `${label}: subsequent clarify inherited consent from auto clarify`,+ );++ // Multi-recipient threads (e.g. Reply all): an answer naming all observed recipients+ // authorizes the full recipient set.agent._planExecutionGuards.set(tabId, clarifyGuardState([{ identity: '[email protected]', role: 'to' },{ identity: '[email protected]', role: 'to' },]));assert.equal(agent._bindClarifiedMessageRecipient(tabId, '[email protected] and [email protected]'),+ true,+ `${label}: an answer naming all observed recipients was rejected`,+ );+ assert.deepEqual(agent._planExecutionGuards.get(tabId)?.messaging, {+ target_kind: 'named',+ recipients: [+ { identity: '[email protected]', role: 'to' },+ { identity: '[email protected]', role: 'to' },+ ],+ }, `${label}: clarify authorization did not set full recipient set`);++ // But naming only a subset of the observed recipients authorizes nothing.+ agent._planExecutionGuards.set(tabId, clarifyGuardState([+ { identity: '[email protected]', role: 'to' },+ { identity: '[email protected]', role: 'to' },+ ]));+ assert.equal(+ agent._bindClarifiedMessageRecipient(tabId, 'Send it to [email protected]'),false,- `${label}: an ambiguous answer authorized a recipient`,+ `${label}: an incomplete recipient subset authorized a send target`,);assert.equal(agent._planExecutionGuards.get(tabId)?.messaging, null);