neoagent 3.2.1-beta.10 → 3.2.1-beta.12

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (62) hide show
  1. package/flutter_app/lib/main_controller.dart +46 -4
  2. package/flutter_app/lib/main_models.dart +4 -2
  3. package/flutter_app/lib/main_operations.dart +0 -49
  4. package/flutter_app/lib/main_settings.dart +338 -0
  5. package/flutter_app/lib/src/backend_client.dart +19 -0
  6. package/package.json +1 -1
  7. package/server/http/routes.js +1 -0
  8. package/server/public/.last_build_id +1 -1
  9. package/server/public/flutter_bootstrap.js +1 -1
  10. package/server/public/main.dart.js +65635 -65088
  11. package/server/routes/behavior.js +80 -0
  12. package/server/routes/settings.js +4 -0
  13. package/server/services/ai/history.js +1 -1
  14. package/server/services/ai/loop/agent_engine_core.js +97 -16
  15. package/server/services/ai/loop/completion_judge.js +211 -44
  16. package/server/services/ai/loop/conversation_loop.js +6 -11
  17. package/server/services/ai/loop/messaging_delivery.js +47 -0
  18. package/server/services/ai/messagingFallback.js +2 -2
  19. package/server/services/ai/model_failure_cache.js +7 -0
  20. package/server/services/ai/systemPrompt.js +30 -128
  21. package/server/services/ai/taskAnalysis.js +47 -2
  22. package/server/services/ai/toolEvidence.js +65 -159
  23. package/server/services/ai/tools.js +60 -8
  24. package/server/services/behavior/config.js +251 -0
  25. package/server/services/behavior/defaults.js +68 -0
  26. package/server/services/behavior/delivery.js +176 -0
  27. package/server/services/behavior/index.js +43 -0
  28. package/server/services/behavior/model_client.js +35 -0
  29. package/server/services/behavior/modules/agent_identity.js +37 -0
  30. package/server/services/behavior/modules/channel_style.js +28 -0
  31. package/server/services/behavior/modules/index.js +29 -0
  32. package/server/services/behavior/modules/norms.js +101 -0
  33. package/server/services/behavior/modules/persona.js +172 -0
  34. package/server/services/behavior/modules/persona_prompt.js +238 -0
  35. package/server/services/behavior/modules/social_memory.js +86 -0
  36. package/server/services/behavior/modules/social_observability.js +94 -0
  37. package/server/services/behavior/modules/social_signals.js +41 -0
  38. package/server/services/behavior/modules/theory_of_mind.js +118 -0
  39. package/server/services/behavior/modules/turn_taking.js +237 -0
  40. package/server/services/behavior/pipeline.js +324 -0
  41. package/server/services/behavior/registry.js +78 -0
  42. package/server/services/behavior/signals.js +107 -0
  43. package/server/services/behavior/state.js +99 -0
  44. package/server/services/behavior/system_prompt.js +75 -0
  45. package/server/services/memory/manager.js +14 -33
  46. package/server/services/messaging/access_policy.js +10 -6
  47. package/server/services/messaging/automation.js +158 -27
  48. package/server/services/messaging/discord.js +11 -1
  49. package/server/services/messaging/formatting_guides.js +2 -4
  50. package/server/services/messaging/http_platforms.js +4 -3
  51. package/server/services/messaging/inbound_queue.js +74 -13
  52. package/server/services/messaging/inbound_store.js +33 -0
  53. package/server/services/messaging/manager.js +18 -0
  54. package/server/services/messaging/telegram.js +10 -1
  55. package/server/services/messaging/whatsapp.js +11 -0
  56. package/server/services/social_reach/channels/social_video.js +1 -1
  57. package/server/services/social_video/captions.js +2 -2
  58. package/server/services/social_video/service.js +194 -29
  59. package/server/services/voice/message.js +1 -1
  60. package/server/services/voice/runtime.js +2 -2
  61. package/server/utils/logger.js +19 -0
  62. package/server/services/ai/terminal_reply.js +0 -18
@@ -0,0 +1,80 @@
1
+ 'use strict';
2
+
3
+ const express = require('express');
4
+ const { requireAuth } = require('../middleware/auth');
5
+ const {
6
+ MODULE_IDS,
7
+ cloneDefaults,
8
+ getBehaviorConfig,
9
+ setBehaviorConfig,
10
+ resolveBehaviorConfig,
11
+ } = require('../services/behavior');
12
+ const {
13
+ getAgentIdFromRequest,
14
+ resolveAgentId,
15
+ } = require('../services/agents/manager');
16
+ const { invalidateSystemPromptCache } = require('../services/ai/systemPrompt');
17
+
18
+ const router = express.Router();
19
+ router.use(requireAuth);
20
+
21
+ function requestAgentId(req) {
22
+ return resolveAgentId(req.session.userId, getAgentIdFromRequest(req));
23
+ }
24
+
25
+ router.get('/', (req, res) => {
26
+ const userId = req.session.userId;
27
+ const agentId = requestAgentId(req);
28
+ res.json({
29
+ agentId,
30
+ modules: MODULE_IDS,
31
+ defaults: cloneDefaults(),
32
+ config: getBehaviorConfig(userId, agentId),
33
+ });
34
+ });
35
+
36
+ router.put('/', (req, res) => {
37
+ if (!req.body?.config || typeof req.body.config !== 'object' || Array.isArray(req.body.config)) {
38
+ return res.status(400).json({ error: 'config must be an object' });
39
+ }
40
+ const userId = req.session.userId;
41
+ const agentId = requestAgentId(req);
42
+ const config = setBehaviorConfig(userId, agentId, req.body.config);
43
+ invalidateSystemPromptCache(userId, agentId);
44
+ return res.json({ agentId, config });
45
+ });
46
+
47
+ router.get('/effective', (req, res) => {
48
+ const userId = req.session.userId;
49
+ const agentId = requestAgentId(req);
50
+ const platform = String(req.query.platform || '').trim();
51
+ const chatId = String(req.query.chatId || '').trim();
52
+ if (!platform || !chatId) {
53
+ return res.status(400).json({ error: 'platform and chatId are required' });
54
+ }
55
+ return res.json({
56
+ agentId,
57
+ config: resolveBehaviorConfig(userId, agentId, {
58
+ platform,
59
+ chatId,
60
+ isGroup: req.query.isGroup !== 'false',
61
+ }),
62
+ });
63
+ });
64
+
65
+ router.get('/diagnostics', (req, res) => {
66
+ const userId = req.session.userId;
67
+ const agentId = requestAgentId(req);
68
+ const platform = String(req.query.platform || '').trim();
69
+ const chatId = String(req.query.chatId || '').trim();
70
+ if (!platform || !chatId) {
71
+ return res.status(400).json({ error: 'platform and chatId are required' });
72
+ }
73
+ const pipeline = req.app?.locals?.behaviorPipeline;
74
+ if (!pipeline?.getDiagnostics) {
75
+ return res.status(503).json({ error: 'Behavior runtime is not initialized' });
76
+ }
77
+ return res.json(pipeline.getDiagnostics(userId, agentId, platform, chatId));
78
+ });
79
+
80
+ module.exports = router;
@@ -346,6 +346,10 @@ router.put('/', async (req, res) => {
346
346
  });
347
347
 
348
348
  tx(Object.entries(normalizedBody));
349
+ if (Object.prototype.hasOwnProperty.call(normalizedBody, 'assistant_behavior_notes')) {
350
+ const { invalidateSystemPromptCache } = require('../services/ai/systemPrompt');
351
+ invalidateSystemPromptCache(userId, agentId);
352
+ }
349
353
 
350
354
  if (Object.keys(normalizedBody).some((key) => VOICE_SETTING_KEYS.has(key))) {
351
355
  const manager = req.app?.locals?.messagingManager;
@@ -126,7 +126,7 @@ async function summarizeMessages(
126
126
  const prompt = [
127
127
  {
128
128
  role: 'system',
129
- content: 'Compress conversation context. Preserve user goals, constraints, preferences, decisions, promised follow-ups, recurring schedules, important facts, tool outcomes, and unresolved issues. Keep concrete details (names, dates, times, statuses) and avoid vague wording. Keep the same personality context, including the favorite-contact texting register when relevant. Output plain text only.'
129
+ content: 'Compress conversation context. Preserve user goals, constraints, preferences, decisions, promised follow-ups, recurring schedules, important facts, tool outcomes, unresolved issues, and any user-stated communication preferences. Keep concrete details (names, dates, times, statuses) and avoid vague wording. Output plain text only.'
130
130
  },
131
131
  {
132
132
  role: 'user',
@@ -20,6 +20,10 @@ const { summarizeCapabilityHealth } = require('../capabilityHealth');
20
20
  const { shouldAcceptTaskComplete } = require('../completion');
21
21
  const { shortenRunId, summarizeForLog } = require('../logFormat');
22
22
  const { getProviderForUser } = require('../provider_selector');
23
+ const {
24
+ recordModelFailure,
25
+ recordModelSuccess,
26
+ } = require('../model_failure_cache');
23
27
  const { runConversation } = require('./conversation_loop');
24
28
  const {
25
29
  TERMINAL_STATUSES,
@@ -638,6 +642,82 @@ class AgentEngine {
638
642
  });
639
643
  }
640
644
 
645
+ async inferStructured({
646
+ userId,
647
+ agentId = null,
648
+ modelId = null,
649
+ purpose = 'fast',
650
+ system,
651
+ prompt,
652
+ maxTokens = 400,
653
+ fallback = {},
654
+ signal = null,
655
+ }) {
656
+ const { getFailureFallbackModelId } = require('./conversation_loop');
657
+ const configuredFallbackId = getAiSettings(userId, agentId).fallback_model_id;
658
+ const failedModelIds = new Set();
659
+ let requestedModelId = modelId;
660
+ let selected = null;
661
+ let result = null;
662
+
663
+ for (let attempt = 0; attempt < 3; attempt += 1) {
664
+ selected = await getProviderForUser(
665
+ userId,
666
+ prompt,
667
+ false,
668
+ requestedModelId,
669
+ {
670
+ agentId,
671
+ signal,
672
+ selectionHint: {
673
+ purpose: purpose === 'general' ? 'general' : 'fast',
674
+ costMode: purpose === 'fast' ? 'economy' : 'balanced_auto',
675
+ },
676
+ },
677
+ );
678
+ try {
679
+ result = await this.requestStructuredJson({
680
+ provider: selected.provider,
681
+ providerName: selected.providerName,
682
+ model: selected.model,
683
+ messages: system ? [{ role: 'system', content: String(system) }] : [],
684
+ prompt: String(prompt || ''),
685
+ maxTokens,
686
+ normalize: (value) => value,
687
+ fallback,
688
+ telemetry: { userId, agentId, signal },
689
+ phase: 'behavior_inference',
690
+ });
691
+ recordModelSuccess(userId, agentId, selected.modelSelectionId);
692
+ break;
693
+ } catch (error) {
694
+ if (isAbortError(error, signal) || signal?.aborted) throw error;
695
+ recordModelFailure(userId, agentId, selected.modelSelectionId, error);
696
+ failedModelIds.add(selected.modelSelectionId);
697
+ if (attempt >= 2) throw error;
698
+ requestedModelId = await getFailureFallbackModelId(
699
+ userId,
700
+ agentId,
701
+ selected.modelSelectionId,
702
+ configuredFallbackId,
703
+ error,
704
+ signal,
705
+ failedModelIds,
706
+ );
707
+ if (!requestedModelId) throw error;
708
+ }
709
+ }
710
+
711
+ return {
712
+ parsed: result.value,
713
+ raw: result.raw,
714
+ usage: result.usage,
715
+ model: selected.model,
716
+ modelSelectionId: selected.modelSelectionId,
717
+ providerName: selected.providerName,
718
+ };
719
+ }
720
+
641
721
  async requestModelResponse({
642
722
  provider,
643
723
  providerName,
@@ -973,6 +1053,15 @@ class AgentEngine {
973
1053
  content: [
974
1054
  'Return JSON only. Distill the current thread working state. Keep it concise and concrete.',
975
1055
  'Track summary, open_commitments, unresolved_questions, referenced_entities, and last_verified_facts. Do not invent facts.',
1056
+ options.memoryScope?.scopeType === 'channel'
1057
+ ? [
1058
+ 'This is a shared room. Extract only durable room or participant facts that are safe and useful in this room.',
1059
+ 'Do not infer or store facts about the authenticated owner from another participant’s message.',
1060
+ options.context?.socialIntelligence?.message?.sender
1061
+ ? `For facts about the current speaker, use the canonical subject "${options.source}:${options.context.socialIntelligence.message.sender}".`
1062
+ : '',
1063
+ ].filter(Boolean).join(' ')
1064
+ : '',
976
1065
  buildMemoryConsolidationInstructions(new Date().toISOString()),
977
1066
  'Schema:',
978
1067
  JSON.stringify({
@@ -1049,6 +1138,14 @@ class AgentEngine {
1049
1138
  agentId: options.agentId || null,
1050
1139
  conversationId,
1051
1140
  runId,
1141
+ scope: options.memoryScope || undefined,
1142
+ metadata: options.memoryScope
1143
+ ? {
1144
+ trustLevel: 'shared_room_conversation',
1145
+ platform: options.source || null,
1146
+ chatId: options.chatId || null,
1147
+ }
1148
+ : undefined,
1052
1149
  signal: options.signal,
1053
1150
  },
1054
1151
  );
@@ -1245,22 +1342,6 @@ class AgentEngine {
1245
1342
  return options.reasoningEffort || process.env.REASONING_EFFORT || 'low';
1246
1343
  }
1247
1344
 
1248
- shouldFastCompleteVoiceReply({
1249
- options = {},
1250
- toolExecutions = [],
1251
- failedStepCount = 0,
1252
- messagingSent = false,
1253
- lastReply = '',
1254
- }) {
1255
- // Voice fast-path is structural only: short reply, no tools/failures.
1256
- // Content terminality is left to the model completion judge.
1257
- return options.latencyProfile === 'voice'
1258
- && toolExecutions.length === 0
1259
- && failedStepCount === 0
1260
- && !messagingSent
1261
- && Boolean(String(lastReply || '').trim());
1262
- }
1263
-
1264
1345
  getMessagingRetryLimit(maxIterations) {
1265
1346
  // Cap at 3: more than 3 autonomous messaging retries indicates a structural
1266
1347
  // problem (model unavailable, bad config) that more retries won't solve.
@@ -6,6 +6,7 @@ const {
6
6
  assessResearchAdequacy,
7
7
  formatResearchAdequacyGuidance,
8
8
  summarizeAvailableTools,
9
+ summarizeResearchEvidenceCatalog,
9
10
  summarizeToolExecutions,
10
11
  } = require('../toolEvidence');
11
12
 
@@ -208,7 +209,7 @@ function buildCompletionDecisionPrompt({
208
209
  const lines = [
209
210
  'Return JSON only.',
210
211
  'Decide whether this run should continue autonomously or stop now.',
211
- 'Schema: {"status":"continue|complete|blocked","reason":"short concrete reason"}',
212
+ 'Schema: {"status":"continue|complete|blocked","reason":"short concrete reason","blocker_review":{"kind":"user_input|permission|external_dependency|unavailable_capability|none","resolvable_in_run":true,"required_value":"specific missing dependency"},"research_review":{"required":true,"adequate":false,"overall_support":"primary|secondary|context|none","evidence_indexes":[1],"target_coverage":[{"target":"exact requested target","support":"primary|secondary|context|none","evidence_indexes":[1]}],"missing_targets":["exact missing target"]}}',
212
213
  'Rules:',
213
214
  '- Use "continue" whenever any safe next step remains in this same run.',
214
215
  '- Use "complete" when the requested outcome is achieved, already done, or a no-op because there is nothing matching the request to change, and the latest draft is the finished user-facing answer.',
@@ -219,8 +220,12 @@ function buildCompletionDecisionPrompt({
219
220
  '- A tool-specific API error, timeout, rate limit, or missing result inside this run is usually "continue", not "blocked", if any other available tool could still make progress.',
220
221
  '- Repeated read-only inspection that has already established the relevant object is absent or unchanged is not progress. Accept a concise complete/blocker reply instead of requiring more searching.',
221
222
  `- If completion_confidence_required is ${goalContext.effectiveCompletionConfidence} and the latest draft depends on unverified assumptions, use "continue" so the run can gather evidence, inspect state, or narrow the reply.`,
222
- '- When research intensity is light or deep, use "continue" until the required primary/source coverage and target coverage are met, or the draft is an explicit blocker naming exactly what could not be verified.',
223
+ '- When research intensity is light or deep, fill research_review from the numbered evidence entries. Do not mark complete unless every requested target has concrete support.',
223
224
  '- Search snippets, memory, and model priors are leads, not completion evidence. Prefer opened/fetched primary sources before "complete".',
225
+ '- Judge source quality semantically from the actual evidence summary. Tool names, call counts, local notes, and target words in tool arguments do not by themselves make evidence primary or prove a claim.',
226
+ '- evidence_indexes are the numeric parts of E identifiers from the Research evidence catalog. Never cite an entry that does not support the target.',
227
+ '- A real external blocker or genuinely required user input may be "blocked" even when research is incomplete. Name the missing dependency precisely; do not pretend more autonomous research can resolve user-only information.',
228
+ '- For "blocked", blocker_review.kind must identify the external dependency, resolvable_in_run must be false, and required_value must state what is missing. Otherwise choose "continue".',
224
229
  '- If the latest draft invents entities, products, people, files, results, or actions that are not supported by tool evidence, use "continue" so the run can gather evidence or rewrite into a truthful partial/blocker answer.',
225
230
  '- A polished-sounding answer is not complete if key requested targets still lack direct evidence.',
226
231
  '- If the latest draft only announces unfinished work, promises a future update, or asks the user to wait without a concrete result or blocker, use "continue" so the run keeps acting.',
@@ -229,12 +234,12 @@ function buildCompletionDecisionPrompt({
229
234
 
230
235
  if (adequacy.intensity !== 'none') {
231
236
  lines.push(
232
- `- Research intensity for this run is ${adequacy.intensity}. Current coverage: primary=${adequacy.primarySourceCount}/${adequacy.requiredPrimarySources}, secondary=${adequacy.secondarySourceCount}, targets_covered=${adequacy.coveredTargets.length}/${Math.max(adequacy.requiredTargetCoverage, adequacy.targets.length)}.`,
237
+ `- Research intensity for this run is ${adequacy.intensity}. Successful evidence candidates=${adequacy.evidenceCandidateCount}; unique calls=${adequacy.uniqueEvidenceCandidateCount}. Semantic target coverage has not been precomputed.`,
233
238
  );
234
239
  }
235
- if (adequacy.adequate === false) {
240
+ if (adequacy.structurallyReady === false) {
236
241
  lines.push(
237
- `- Research is still incomplete (${adequacy.reason}). Use "continue" unless the latest draft is an explicit blocker naming the exact missing evidence.`,
242
+ `- No source evidence is ready for review (${adequacy.reason}). Use "continue" unless the latest draft is a real external blocker.`,
238
243
  );
239
244
  }
240
245
 
@@ -255,31 +260,176 @@ function buildCompletionDecisionPrompt({
255
260
  : '',
256
261
  `Current iteration: ${iteration} of ${maxIterations}.`,
257
262
  `Available tools in this run: ${summarizeAvailableTools(tools) || 'none'}`,
258
- `Recent tool evidence:\n${summarizeToolExecutions(toolExecutions, 8) || 'none'}`,
263
+ `Recent tool evidence:\n${summarizeToolExecutions(toolExecutions, 12) || 'none'}`,
259
264
  adequacy.intensity !== 'none'
260
- ? `Research adequacy: intensity=${adequacy.intensity}; adequate=${adequacy.adequate}; reason=${adequacy.reason}`
265
+ ? `Research evidence catalog:\n${summarizeResearchEvidenceCatalog(toolExecutions) || 'none'}`
266
+ : '',
267
+ adequacy.intensity !== 'none'
268
+ ? `Research preflight: intensity=${adequacy.intensity}; structurally_ready=${adequacy.structurallyReady}; semantic_review_required=true; reason=${adequacy.reason}`
261
269
  : '',
262
270
  adequacy.targets.length
263
271
  ? `Research targets: ${adequacy.targets.join('; ')}`
264
272
  : '',
265
- adequacy.uncoveredTargets.length
266
- ? `Uncovered research targets: ${adequacy.uncoveredTargets.join('; ')}`
267
- : '',
268
273
  formatResearchAdequacyGuidance(adequacy),
269
274
  `Latest draft reply:\n${draftReply || '(empty)'}`,
270
275
  );
271
276
  return lines.filter(Boolean).join('\n');
272
277
  }
273
278
 
279
+ function normalizeEvidenceIndexes(value) {
280
+ if (!Array.isArray(value)) return [];
281
+ return [...new Set(value
282
+ .map((entry) => Number(entry))
283
+ .filter((entry) => Number.isInteger(entry) && entry > 0))]
284
+ .slice(0, 20);
285
+ }
286
+
287
+ function normalizeResearchReview(raw = null) {
288
+ const review = raw && typeof raw === 'object' && !Array.isArray(raw) ? raw : {};
289
+ const targetCoverage = Array.isArray(review.target_coverage || review.targetCoverage)
290
+ ? (review.target_coverage || review.targetCoverage)
291
+ .filter((entry) => entry && typeof entry === 'object' && !Array.isArray(entry))
292
+ .map((entry) => ({
293
+ target: String(entry.target || '').trim(),
294
+ support: ['primary', 'secondary', 'context', 'none'].includes(
295
+ String(entry.support || '').trim().toLowerCase(),
296
+ )
297
+ ? String(entry.support).trim().toLowerCase()
298
+ : 'none',
299
+ evidence_indexes: normalizeEvidenceIndexes(
300
+ entry.evidence_indexes || entry.evidenceIndexes,
301
+ ),
302
+ }))
303
+ .filter((entry) => entry.target)
304
+ .slice(0, 12)
305
+ : [];
306
+ return {
307
+ required: review.required === true,
308
+ adequate: review.adequate === true,
309
+ overall_support: ['primary', 'secondary', 'context', 'none'].includes(
310
+ String(review.overall_support || review.overallSupport || '').trim().toLowerCase(),
311
+ )
312
+ ? String(review.overall_support || review.overallSupport).trim().toLowerCase()
313
+ : 'none',
314
+ evidence_indexes: normalizeEvidenceIndexes(
315
+ review.evidence_indexes || review.evidenceIndexes,
316
+ ),
317
+ target_coverage: targetCoverage,
318
+ missing_targets: normalizeGoalCriteria(
319
+ review.missing_targets || review.missingTargets || [],
320
+ ),
321
+ };
322
+ }
323
+
274
324
  function normalizeCompletionDecision(raw, fallbackStatus = 'continue') {
275
325
  const allowed = new Set(['continue', 'complete', 'blocked']);
276
326
  const requestedStatus = String(raw.status || '').trim().toLowerCase();
327
+ const rawBlockerReview = raw.blocker_review || raw.blockerReview;
328
+ const blockerReview = rawBlockerReview
329
+ && typeof rawBlockerReview === 'object'
330
+ && !Array.isArray(rawBlockerReview)
331
+ ? rawBlockerReview
332
+ : {};
333
+ const blockerKind = String(blockerReview.kind || '').trim().toLowerCase();
277
334
  return {
278
335
  status: allowed.has(requestedStatus) ? requestedStatus : fallbackStatus,
279
336
  reason: String(raw.reason || '').trim().slice(0, 400),
337
+ blocker_review: {
338
+ kind: [
339
+ 'user_input',
340
+ 'permission',
341
+ 'external_dependency',
342
+ 'unavailable_capability',
343
+ 'none',
344
+ ].includes(blockerKind)
345
+ ? blockerKind
346
+ : 'none',
347
+ resolvable_in_run: blockerReview.resolvable_in_run !== false
348
+ && blockerReview.resolvableInRun !== false,
349
+ required_value: String(
350
+ blockerReview.required_value || blockerReview.requiredValue || '',
351
+ ).trim().slice(0, 300),
352
+ },
353
+ research_review: normalizeResearchReview(
354
+ raw.research_review || raw.researchReview,
355
+ ),
280
356
  };
281
357
  }
282
358
 
359
+ function validateResearchReview(review, researchAdequacy, toolExecutions = []) {
360
+ if (!researchAdequacy || researchAdequacy.intensity === 'none') {
361
+ return { valid: true, reason: '' };
362
+ }
363
+ if (researchAdequacy.structurallyReady === false) {
364
+ return {
365
+ valid: false,
366
+ reason: researchAdequacy.reason || 'No successful source evidence is available.',
367
+ };
368
+ }
369
+ if (!review?.required || review.adequate !== true) {
370
+ return {
371
+ valid: false,
372
+ reason: 'The completion judge did not confirm adequate research evidence.',
373
+ };
374
+ }
375
+
376
+ const isValidEvidenceIndex = (index) => {
377
+ const execution = (Array.isArray(toolExecutions) ? toolExecutions : [])[index - 1];
378
+ return Boolean(
379
+ execution
380
+ && execution.ok === true
381
+ && execution.evidenceRelevant !== false
382
+ && execution.evidenceSource !== 'messaging',
383
+ );
384
+ };
385
+ const hasValidEvidence = (indexes) => (
386
+ Array.isArray(indexes) && indexes.some(isValidEvidenceIndex)
387
+ );
388
+ const requiredTargets = Array.isArray(researchAdequacy.targets)
389
+ ? researchAdequacy.targets
390
+ : [];
391
+
392
+ if (requiredTargets.length === 0) {
393
+ const reviewIndexes = [
394
+ ...(review.evidence_indexes || []),
395
+ ...(review.target_coverage || []).flatMap((entry) => entry.evidence_indexes || []),
396
+ ];
397
+ if (!hasValidEvidence(reviewIndexes)) {
398
+ return { valid: false, reason: 'The research review cites no successful evidence entry.' };
399
+ }
400
+ if (researchAdequacy.intensity === 'deep' && review.overall_support !== 'primary') {
401
+ return { valid: false, reason: 'Deep research lacks primary source support.' };
402
+ }
403
+ if (review.overall_support === 'none') {
404
+ return { valid: false, reason: 'The research review did not classify its source support.' };
405
+ }
406
+ return { valid: true, reason: '' };
407
+ }
408
+
409
+ const coverageByTarget = new Map(
410
+ (review.target_coverage || []).map((entry) => [
411
+ String(entry.target || '').trim().toLowerCase(),
412
+ entry,
413
+ ]),
414
+ );
415
+ for (const target of requiredTargets) {
416
+ const coverage = coverageByTarget.get(String(target).trim().toLowerCase());
417
+ if (!coverage || coverage.support === 'none' || !hasValidEvidence(coverage.evidence_indexes)) {
418
+ return {
419
+ valid: false,
420
+ reason: `The research review does not cite supporting evidence for "${target}".`,
421
+ };
422
+ }
423
+ if (researchAdequacy.intensity === 'deep' && coverage.support !== 'primary') {
424
+ return {
425
+ valid: false,
426
+ reason: `Deep research still lacks primary evidence for "${target}".`,
427
+ };
428
+ }
429
+ }
430
+ return { valid: true, reason: '' };
431
+ }
432
+
283
433
  function enforceTerminalReplyDecision(decision, lastReply, options = {}) {
284
434
  // Natural-language terminality is judged by the model completion decision.
285
435
  // Runtime only enforces structural evidence contracts (research adequacy).
@@ -289,25 +439,33 @@ function enforceTerminalReplyDecision(decision, lastReply, options = {}) {
289
439
  goalContext: options.goalContext || null,
290
440
  toolExecutions: options.toolExecutions || [],
291
441
  });
292
- if (
293
- researchAdequacy
294
- && researchAdequacy.adequate === false
295
- && (decision?.status === 'complete' || decision?.status === 'blocked')
296
- ) {
297
- // Allow true blocked outcomes only when the model already chose blocked AND
298
- // research cannot progress further (no remaining targets/primary gap that
299
- // tools could still cover). Otherwise force continue for evidence gathering.
442
+ if (decision?.status === 'blocked') {
443
+ const blockerReview = decision.blocker_review;
300
444
  if (
301
- decision?.status === 'blocked'
302
- && researchAdequacy.uncoveredTargets.length === 0
303
- && researchAdequacy.primarySourceCount >= researchAdequacy.requiredPrimarySources
445
+ blockerReview
446
+ && blockerReview.kind !== 'none'
447
+ && blockerReview.resolvable_in_run === false
448
+ && blockerReview.required_value
304
449
  ) {
305
450
  return decision;
306
451
  }
307
452
  return {
453
+ ...decision,
308
454
  status: 'continue',
309
- reason: researchAdequacy.reason
310
- || 'Research evidence is still incomplete for the requested targets; continue gathering sources before finishing.',
455
+ reason: 'The completion judge did not substantiate a blocker that requires external input or state.',
456
+ };
457
+ }
458
+ if (decision?.status === 'complete') {
459
+ const reviewValidation = validateResearchReview(
460
+ decision.research_review,
461
+ researchAdequacy,
462
+ options.toolExecutions || [],
463
+ );
464
+ if (reviewValidation.valid) return decision;
465
+ return {
466
+ status: 'continue',
467
+ reason: reviewValidation.reason,
468
+ research_review: decision.research_review,
311
469
  };
312
470
  }
313
471
  return decision;
@@ -332,7 +490,7 @@ function buildChurnAssessmentPrompt({
332
490
  const lines = [
333
491
  'Return JSON only.',
334
492
  'Self-assess your current loop state — are you making genuine progress or spinning?',
335
- 'Schema: {"assessment":"progressing|churn|blocked","reason":"one short concrete sentence"}',
493
+ 'Schema: {"assessment":"progressing|churn|blocked","reason":"one short concrete sentence","blocker_kind":"user_input|permission|external_dependency|unavailable_capability|none","resolvable_in_run":true}',
336
494
  '',
337
495
  `Context: ${readOnlyCount} consecutive iteration(s) with only read/search/inspect operations — no concrete state changes yet.`,
338
496
  alreadyRead ? `Already inspected: ${alreadyRead}.` : '',
@@ -343,44 +501,39 @@ function buildChurnAssessmentPrompt({
343
501
  `Iteration: ${iteration}`,
344
502
  `Recent tool evidence:\n${summarizeToolExecutions(toolExecutions, 6) || 'none'}`,
345
503
  adequacy.intensity !== 'none'
346
- ? `Research adequacy: intensity=${adequacy.intensity}; adequate=${adequacy.adequate}; covered=${adequacy.coveredTargets.length}/${Math.max(adequacy.requiredTargetCoverage, adequacy.targets.length)}; primary=${adequacy.primarySourceCount}/${adequacy.requiredPrimarySources}.`
504
+ ? `Research preflight: intensity=${adequacy.intensity}; evidence_candidates=${adequacy.evidenceCandidateCount}; semantic target coverage must be judged from the evidence summaries.`
347
505
  : '',
348
- adequacy.uncoveredTargets.length
349
- ? `Still uncovered research targets: ${adequacy.uncoveredTargets.join('; ')}.`
506
+ adequacy.targets.length
507
+ ? `Requested research targets: ${adequacy.targets.join('; ')}.`
350
508
  : '',
351
509
  '',
352
510
  'Assessment rules:',
353
511
  '- "progressing": You are systematically gathering necessary context and the next concrete action is already determined — you know exactly what to do next.',
354
512
  '- "churn": You are re-reading/re-searching information already in context, or exploring without a clear next concrete step. Accept the nudge and act.',
355
513
  '- "blocked": No concrete action is available in this run. You have all the evidence needed to deliver a truthful final answer or a specific external blocker.',
514
+ '- Use "blocked" only with a non-none blocker_kind and resolvable_in_run=false. Partial evidence, a failed attempt, or uncertainty is not by itself an external blocker.',
356
515
  '- For multi-target research, keep "progressing" while uncovered targets remain and a fresh primary source can still be opened. Do not mark "blocked" just because you have partial notes.',
357
516
  '- Re-querying the same snippet source for an already covered target is "churn". Opening a different primary source for an uncovered target is "progressing".',
358
- adequacy.adequate === false
359
- ? '- Research adequacy is currently incomplete. Prefer "progressing" if a new primary source for an uncovered target is still available; use "churn" only for repeated identical reads; do not use "blocked" unless tools cannot reach remaining targets.'
517
+ adequacy.structurallyReady === false
518
+ ? '- No successful source evidence exists yet. Prefer "progressing" if a concrete source-gathering step remains, but use "blocked" for a genuine external dependency or required user input.'
360
519
  : '',
361
520
  ];
362
521
  return lines.filter(Boolean).join('\n');
363
522
  }
364
523
 
365
- function enforceChurnAssessment(assessment, options = {}) {
524
+ function enforceChurnAssessment(assessment) {
366
525
  const normalized = normalizeChurnAssessment(assessment);
367
- const researchAdequacy = options.researchAdequacy
368
- || assessResearchAdequacy({
369
- analysis: options.analysis || null,
370
- goalContext: options.goalContext || null,
371
- toolExecutions: options.toolExecutions || [],
372
- });
373
526
  if (
374
- researchAdequacy
375
- && researchAdequacy.adequate === false
376
- && researchAdequacy.intensity !== 'none'
377
- && normalized.assessment === 'blocked'
378
- && (researchAdequacy.uncoveredTargets?.length > 0 || researchAdequacy.primarySourceCount < researchAdequacy.requiredPrimarySources)
527
+ normalized.assessment === 'blocked'
528
+ && (
529
+ normalized.blocker_kind === 'none'
530
+ || normalized.resolvable_in_run !== false
531
+ )
379
532
  ) {
380
533
  return {
381
- assessment: 'progressing',
382
- reason: researchAdequacy.reason
383
- || 'Research targets remain uncovered; keep gathering primary sources instead of force-finishing.',
534
+ ...normalized,
535
+ assessment: 'churn',
536
+ reason: 'The churn assessment did not identify a concrete external blocker.',
384
537
  };
385
538
  }
386
539
  return normalized;
@@ -389,9 +542,21 @@ function enforceChurnAssessment(assessment, options = {}) {
389
542
  function normalizeChurnAssessment(raw) {
390
543
  const allowed = new Set(['progressing', 'churn', 'blocked']);
391
544
  const assessment = String(raw?.assessment || '').trim().toLowerCase();
545
+ const blockerKind = String(raw?.blocker_kind || raw?.blockerKind || '').trim().toLowerCase();
392
546
  return {
393
547
  assessment: allowed.has(assessment) ? assessment : 'churn',
394
548
  reason: String(raw?.reason || '').trim().slice(0, 300),
549
+ blocker_kind: [
550
+ 'user_input',
551
+ 'permission',
552
+ 'external_dependency',
553
+ 'unavailable_capability',
554
+ 'none',
555
+ ].includes(blockerKind)
556
+ ? blockerKind
557
+ : 'none',
558
+ resolvable_in_run: raw?.resolvable_in_run !== false
559
+ && raw?.resolvableInRun !== false,
395
560
  };
396
561
  }
397
562
 
@@ -406,6 +571,8 @@ module.exports = {
406
571
  enforceChurnAssessment,
407
572
  normalizeChurnAssessment,
408
573
  normalizeCompletionDecision,
574
+ normalizeResearchReview,
409
575
  normalizeGoalContract,
410
576
  resolveRunGoalContext,
577
+ validateResearchReview,
411
578
  };