Skip to content

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

Skip the planner only when the earlier task was read whole too

Emre Sokullua71734bda939
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.js
index 29d85a797..9ec6cb25d 100644
--- a/src/chrome/src/agent/agent.js
+++ b/src/chrome/src/agent/agent.js
@@ -16,7 +16,7 @@ import { JEV_FAST_KEYS, JEV_CLASSIFIER_THRESHOLD, JEV_BROWSER_THRESHOLD, confide
import { redactSystemOneText, wrapSystemOneData, boundedSystemOneText, systemOneStateBytes } from './systemone-evidence.js';
import { createSystemOneJudge, isSystemOneResponseContractError, systemOneFailureReason, SYSTEM_ONE_COST_PROVIDER } from './systemone-judge.js';
import { AUTO_MODE, AUTO_MODE_ASK_NOTE, REQUEST_ACTION_MODE_TOOL_NAME, acceptModeEscalation, autoModeForChoice, autoModeHardRoute, autoModePlanForChoice, carriesUntrustedText, confirmedEscalationMode, isRunMode, settleAutoModeRoute, withRoutedRuntimeMode } from './auto-mode.js';
-import { SEMANTIC_BATCH_LIMITS, SEMANTIC_DECISION_VERSION, acceptSemanticAnswer, acceptSemanticBatchAnswers, boundedSemanticData, controlEffectCommits, semanticBatchChunks, semanticBatchDecisionMessages, semanticDecisionMessages, semanticQuestion, semanticSystemOneBatchQuestions, semanticSystemOneQuestion, WORKFLOW_STATES } from './semantic-decisions.js';
+import { SEMANTIC_BATCH_LIMITS, SEMANTIC_DECISION_VERSION, acceptSemanticAnswer, acceptSemanticBatchAnswers, boundedSemanticData, controlEffectCommits, semanticDataComplete, semanticBatchChunks, semanticBatchDecisionMessages, semanticDecisionMessages, semanticQuestion, semanticSystemOneBatchQuestions, semanticSystemOneQuestion, WORKFLOW_STATES } from './semantic-decisions.js';
import { SOCIAL_PLATFORMS, socialPublicationApiPlatform, normalizePublicationContract, resolvePublicationText, publicationProgress, exactPublicationText, publicationMediaMatches, publicationContractMessages, publicationAuditMessages, parsePublicationJson, publicationAuditValidation } from './social-publish-contract.js';
import { mastodonPublicationIntentMessages, normalizeMastodonPublicationIntent, captureNativeMastodonPublication, mastodonPublicationBinding, mastodonPublicationReceipt, mastodonPublicationNavigation } from './mastodon-publication.js';
import { AGENT_TOOLS, AGENT_TOOL_NAMES, RESERVED_AGENT_TOOL_NAMES, RETIRED_AGENT_TOOL_NAMES, getToolsForMode, SYSTEM_PROMPT_BATCHING, SYSTEM_PROMPT_ASK, SYSTEM_PROMPT_ACT, SYSTEM_PROMPT_ACT_COMPACT, SYSTEM_PROMPT_ACT_MID, SYSTEM_PROMPT_ACT_LEAN, SYSTEM_PROMPT_GENERATIVE_MEDIA, SYSTEM_PROMPT_DEV_APPENDIX, SYSTEM_PROMPT_WEBMCP_ASK, SYSTEM_PROMPT_WEBMCP_ACT } from './tools.js';
@@ -9222,8 +9222,8 @@ export class Agent extends LoopDetector {
* in the same decision-model request (run_mode with plan_need). Their sure
* answers come back as `alongside: { kind: choice | null }` and are
* remembered on their own; the active model never answers them. An
- * authorizing answer among them needs the whole request: when the request
- * was cut to fit the decision model, it is not used.
+ * authorizing answer among them needs the whole request and data: when
+ * either was cut to fit the decision model, it is not used.
*/
async _semanticDecision(tabId, kind, { request = '', page = {} } = {}, { provider = this._activeProvider(tabId), signal = null, mainModel = true, retryChoices = [], alongside = [] } = {}) {
const spec = semanticQuestion(kind);
@@ -9232,9 +9232,12 @@ export class Agent extends LoopDetector {
const requestText = String(request ?? '');
const trusted = kind === 'user_turn_kind' ? requestText : requestText.slice(0, 6000);
const data = boundedSemanticData(page);
- // Keyed by the whole request, not the part a model reads: two requests
- // that share a cut-off start are different requests.
- const keyFor = question => this._sha256TextSync(JSON.stringify([SEMANTIC_DECISION_VERSION, question, requestText, data]));
+ // Whether a model reads all of it: the request is cut to 6,000 characters
+ // above, and boundedSemanticData shortens long data (an earlier task).
+ const inputComplete = trusted.length === requestText.length && semanticDataComplete(page);
+ // Keyed by the whole request, not the part a model reads, and by whether
+ // the data was cut: inputs that share a cut-off start are different.
+ const keyFor = question => this._sha256TextSync(JSON.stringify([SEMANTIC_DECISION_VERSION, question, requestText, data, inputComplete]));
const cacheKey = keyFor(kind);
const extraKinds = [...new Set(Array.isArray(alongside) ? alongside : [])].filter(extra => extra !== kind && semanticQuestion(extra));
this._semanticDecisionCache ??= new Map();
@@ -9263,9 +9266,9 @@ export class Agent extends LoopDetector {
const settings = await this._jevSettings(tabId, provider);
decisionKey = this._semanticDecisionKey(settings);
const decisionText = redactSystemOneText(JSON.stringify({ request: trusted, page: data }));
- // What the decision model reads of the request: cut to 6,000 characters
- // above, then to the decision state's budget below.
- const partialRequest = trusted.length < requestText.length || systemOneStateBytes(decisionText) > 12000;
+ // What the decision model reads: the input above, then cut to the
+ // decision state's budget below.
+ const partialInput = !inputComplete || systemOneStateBytes(decisionText) > 12000;
// Keep the existing bounded fast route, but use the full-text main
// model directly when a relationship question cannot fit in it.
const decisionFits = kind !== 'user_turn_kind' || systemOneStateBytes(decisionText) <= 12000;
@@ -9278,9 +9281,9 @@ export class Agent extends LoopDetector {
for (const extra of extraKinds) {
const extraAnswer = verdict.answers?.[extra];
const authorizing = semanticQuestion(extra).authorizing.includes(extraAnswer?.choice);
- // A part of the request cannot authorize skipping anything the rest
+ // A part of the input cannot authorize skipping anything the rest
// of it might need (plan_need: a payment after the cut).
- if (authorizing && partialRequest) continue;
+ if (authorizing && partialInput) continue;
const extraChoice = confidentChoice(extraAnswer, authorizing ? 0.85 : 0.7);
if (extraChoice) {
extraAnswers[extra] = extraChoice;
diff --git a/src/chrome/src/agent/semantic-decisions.js b/src/chrome/src/agent/semantic-decisions.js
index adb37a950..5d58d5a59 100644
--- a/src/chrome/src/agent/semantic-decisions.js
+++ b/src/chrome/src/agent/semantic-decisions.js
@@ -572,6 +572,20 @@ export function boundedSemanticData(value, depth = 0) {
return null;
}
+// Whether boundedSemanticData(value) keeps all of value: no string past 600
+// characters, no list or object past 16 entries, nothing nested too deep.
+export function semanticDataComplete(value, depth = 0) {
+ if (value == null || typeof value === 'boolean' || (typeof value === 'number' && Number.isFinite(value))) return true;
+ if (typeof value === 'string') return value.normalize('NFC').replace(/\s+/g, ' ').trim().length <= 600;
+ if (depth >= 3) return false;
+ if (Array.isArray(value)) return value.length <= 16 && value.every(item => semanticDataComplete(item, depth + 1));
+ if (typeof value === 'object') {
+ const entries = Object.entries(value);
+ return entries.length <= 16 && entries.every(([key, item]) => String(key).length <= 40 && semanticDataComplete(item, depth + 1));
+ }
+ return false;
+}
+
// The decision model's form of the question.
export function semanticSystemOneQuestion(kind) {
const spec = semanticQuestion(kind);
diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.js
index 01ec17594..61402a2a4 100644
--- a/src/firefox/src/agent/agent.js
+++ b/src/firefox/src/agent/agent.js
@@ -14,7 +14,7 @@ import { JEV_FAST_KEYS, JEV_CLASSIFIER_THRESHOLD, JEV_BROWSER_THRESHOLD, confide
import { redactSystemOneText, wrapSystemOneData, boundedSystemOneText, systemOneStateBytes } from './systemone-evidence.js';
import { createSystemOneJudge, isSystemOneResponseContractError, systemOneFailureReason, SYSTEM_ONE_COST_PROVIDER } from './systemone-judge.js';
import { AUTO_MODE, AUTO_MODE_ASK_NOTE, REQUEST_ACTION_MODE_TOOL_NAME, acceptModeEscalation, autoModeForChoice, autoModeHardRoute, autoModePlanForChoice, carriesUntrustedText, confirmedEscalationMode, isRunMode, settleAutoModeRoute, withRoutedRuntimeMode } from './auto-mode.js';
-import { SEMANTIC_BATCH_LIMITS, SEMANTIC_DECISION_VERSION, acceptSemanticAnswer, acceptSemanticBatchAnswers, boundedSemanticData, controlEffectCommits, semanticBatchChunks, semanticBatchDecisionMessages, semanticDecisionMessages, semanticQuestion, semanticSystemOneBatchQuestions, semanticSystemOneQuestion, WORKFLOW_STATES } from './semantic-decisions.js';
+import { SEMANTIC_BATCH_LIMITS, SEMANTIC_DECISION_VERSION, acceptSemanticAnswer, acceptSemanticBatchAnswers, boundedSemanticData, controlEffectCommits, semanticDataComplete, semanticBatchChunks, semanticBatchDecisionMessages, semanticDecisionMessages, semanticQuestion, semanticSystemOneBatchQuestions, semanticSystemOneQuestion, WORKFLOW_STATES } from './semantic-decisions.js';
import { firefoxBidi } from '../bidi/client.js';
import { SOCIAL_PLATFORMS, socialPublicationApiPlatform, normalizePublicationContract, resolvePublicationText, publicationProgress, exactPublicationText, publicationMediaMatches, publicationContractMessages, publicationAuditMessages, parsePublicationJson, publicationAuditValidation } from './social-publish-contract.js';
import { mastodonPublicationIntentMessages, normalizeMastodonPublicationIntent, captureNativeMastodonPublication, mastodonPublicationBinding, mastodonPublicationReceipt, mastodonPublicationNavigation } from './mastodon-publication.js';
@@ -8931,8 +8931,8 @@ export class Agent extends LoopDetector {
* in the same decision-model request (run_mode with plan_need). Their sure
* answers come back as `alongside: { kind: choice | null }` and are
* remembered on their own; the active model never answers them. An
- * authorizing answer among them needs the whole request: when the request
- * was cut to fit the decision model, it is not used.
+ * authorizing answer among them needs the whole request and data: when
+ * either was cut to fit the decision model, it is not used.
*/
async _semanticDecision(tabId, kind, { request = '', page = {} } = {}, { provider = this._activeProvider(tabId), signal = null, mainModel = true, retryChoices = [], alongside = [] } = {}) {
const spec = semanticQuestion(kind);
@@ -8941,9 +8941,12 @@ export class Agent extends LoopDetector {
const requestText = String(request ?? '');
const trusted = kind === 'user_turn_kind' ? requestText : requestText.slice(0, 6000);
const data = boundedSemanticData(page);
- // Keyed by the whole request, not the part a model reads: two requests
- // that share a cut-off start are different requests.
- const keyFor = question => this._sha256TextSync(JSON.stringify([SEMANTIC_DECISION_VERSION, question, requestText, data]));
+ // Whether a model reads all of it: the request is cut to 6,000 characters
+ // above, and boundedSemanticData shortens long data (an earlier task).
+ const inputComplete = trusted.length === requestText.length && semanticDataComplete(page);
+ // Keyed by the whole request, not the part a model reads, and by whether
+ // the data was cut: inputs that share a cut-off start are different.
+ const keyFor = question => this._sha256TextSync(JSON.stringify([SEMANTIC_DECISION_VERSION, question, requestText, data, inputComplete]));
const cacheKey = keyFor(kind);
const extraKinds = [...new Set(Array.isArray(alongside) ? alongside : [])].filter(extra => extra !== kind && semanticQuestion(extra));
this._semanticDecisionCache ??= new Map();
@@ -8972,9 +8975,9 @@ export class Agent extends LoopDetector {
const settings = await this._jevSettings(tabId, provider);
decisionKey = this._semanticDecisionKey(settings);
const decisionText = redactSystemOneText(JSON.stringify({ request: trusted, page: data }));
- // What the decision model reads of the request: cut to 6,000 characters
- // above, then to the decision state's budget below.
- const partialRequest = trusted.length < requestText.length || systemOneStateBytes(decisionText) > 12000;
+ // What the decision model reads: the input above, then cut to the
+ // decision state's budget below.
+ const partialInput = !inputComplete || systemOneStateBytes(decisionText) > 12000;
// Keep the existing bounded fast route, but use the full-text main
// model directly when a relationship question cannot fit in it.
const decisionFits = kind !== 'user_turn_kind' || systemOneStateBytes(decisionText) <= 12000;
@@ -8987,9 +8990,9 @@ export class Agent extends LoopDetector {
for (const extra of extraKinds) {
const extraAnswer = verdict.answers?.[extra];
const authorizing = semanticQuestion(extra).authorizing.includes(extraAnswer?.choice);
- // A part of the request cannot authorize skipping anything the rest
+ // A part of the input cannot authorize skipping anything the rest
// of it might need (plan_need: a payment after the cut).
- if (authorizing && partialRequest) continue;
+ if (authorizing && partialInput) continue;
const extraChoice = confidentChoice(extraAnswer, authorizing ? 0.85 : 0.7);
if (extraChoice) {
extraAnswers[extra] = extraChoice;
diff --git a/src/firefox/src/agent/semantic-decisions.js b/src/firefox/src/agent/semantic-decisions.js
index adb37a950..5d58d5a59 100644
--- a/src/firefox/src/agent/semantic-decisions.js
+++ b/src/firefox/src/agent/semantic-decisions.js
@@ -572,6 +572,20 @@ export function boundedSemanticData(value, depth = 0) {
return null;
}
+// Whether boundedSemanticData(value) keeps all of value: no string past 600
+// characters, no list or object past 16 entries, nothing nested too deep.
+export function semanticDataComplete(value, depth = 0) {
+ if (value == null || typeof value === 'boolean' || (typeof value === 'number' && Number.isFinite(value))) return true;
+ if (typeof value === 'string') return value.normalize('NFC').replace(/\s+/g, ' ').trim().length <= 600;
+ if (depth >= 3) return false;
+ if (Array.isArray(value)) return value.length <= 16 && value.every(item => semanticDataComplete(item, depth + 1));
+ if (typeof value === 'object') {
+ const entries = Object.entries(value);
+ return entries.length <= 16 && entries.every(([key, item]) => String(key).length <= 40 && semanticDataComplete(item, depth + 1));
+ }
+ return false;
+}
+
// The decision model's form of the question.
export function semanticSystemOneQuestion(kind) {
const spec = semanticQuestion(kind);
diff --git a/test/plan-when-needed.mjs b/test/plan-when-needed.mjs
index d10feedcb..242a2ca92 100644
--- a/test/plan-when-needed.mjs
+++ b/test/plan-when-needed.mjs
@@ -126,6 +126,17 @@ for (const build of BUILDS) {
assert.equal('plan' in await wide.agent._routeAutoRunMode(6, { ...question, request: 'あ'.repeat(4500) }), false);
});
+ test(`${build}: an earlier task cut to fit the decision data never skips the planner`, async () => {
+ // A short follow-up leans on the earlier task, which the data keeps to 600 characters.
+ const start = 'Read the next row and note its owner. '.repeat(20).slice(0, 600);
+ const followUp = task => ({ request: 'Continue.', page: { previous_mode: 'act', current_task: task } });
+ const { agent, requests } = decisionAgent({ run_mode: sure('act'), plan_need: sure('direct') });
+ assert.equal((await agent._routeAutoRunMode(8, followUp(start))).plan, 'direct', 'an earlier task read whole');
+ const longer = await agent._routeAutoRunMode(8, followUp(`${start} Then delete every row you read.`));
+ assert.equal(requests.length, 2, 'the longer earlier task is a new question');
+ assert.equal('plan' in longer, false);
+ });
+
test(`${build}: a remembered answer belongs to the whole request, not to its first 6,000 characters`, async () => {
const start = 'Read this paragraph. '.repeat(300).slice(0, 6000);
const { agent, requests } = decisionAgent({ run_mode: sure('act'), plan_need: sure('direct') });