fix(chat): the model was never being shown the conversation

.ai/BOOTSTRAP.md

@@ -50,7 +50,7 @@ Full validation before release: `npm run release:preflight`.
 - Shared packages: @claw/shared-auth, @claw/shared-constants, @claw/shared-entitlements, @claw/shared-rabbitmq, @claw/shared-types, @claw/shared-utilities
 - Events: 178 on `claw.events`
 - Permissions: 51 · Env vars: 351
-- API endpoints: 643 · Frontend pages: 143
+- API endpoints: 644 · Frontend pages: 143
 
 This file is generated. To change it, edit the renderer + policy sources and run
 `npm run knowledge:build`.

.ai/manifests/api-endpoints.json

@@ -2331,6 +2331,11 @@
         "route": "/health",
         "source": "apps/claw-routing-service/src/modules/health/controllers/health.controller.ts"
       },
+      {
+        "method": "GET",
+        "route": "/internal/router-models/context-window/:provider/:model",
+        "source": "apps/claw-routing-service/src/modules/router-models/controllers/model-context-window-internal.controller.ts"
+      },
       {
         "method": "GET",
         "route": "/internal/router-models/costs/:provider/:model",
@@ -3253,5 +3258,5 @@
     ]
   },
   "generated": true,
-  "total": 643
+  "total": 644
 }

.ai/manifests/governance.json

@@ -252,7 +252,7 @@
   ],
   "skills": [
     {
-      "bytes": 19868,
+      "bytes": 20103,
       "file": "skills/00-index.md"
     },
     {
@@ -339,6 +339,10 @@
       "bytes": 4776,
       "file": "skills/add-workspace-connector.md"
     },
+    {
+      "bytes": 8787,
+      "file": "skills/audit-conversational-context.md"
+    },
     {
       "bytes": 5494,
       "file": "skills/backend-architecture-review.md"

.ai/manifests/hashes.json

@@ -1,24 +1,24 @@
 {
   "generated": true,
   "hashes": {
-    ".ai/BOOTSTRAP.md": "67f2767a",
-    ".ai/manifests/api-endpoints.json": "c1aec4ca",
+    ".ai/BOOTSTRAP.md": "c2216623",
+    ".ai/manifests/api-endpoints.json": "995a14e1",
     ".ai/manifests/data-ownership.json": "501d74dc",
     ".ai/manifests/docker-services.json": "255da189",
     ".ai/manifests/environment-variables.json": "33602cbe",
     ".ai/manifests/event-graph.json": "c1cd05bc",
     ".ai/manifests/frontend-routes.json": "a52b7ece",
-    ".ai/manifests/governance.json": "ef960bf2",
-    ".ai/manifests/i18n.json": "5321b991",
+    ".ai/manifests/governance.json": "f71a269b",
+    ".ai/manifests/i18n.json": "3befb756",
     ".ai/manifests/nginx-routes.json": "53c65be1",
     ".ai/manifests/packages.json": "cc928a4a",
     ".ai/manifests/permissions.json": "c87c2c04",
     ".ai/manifests/ports.json": "6b737943",
     ".ai/manifests/prisma-models.json": "ac082a39",
     ".ai/manifests/rabbitmq-events.json": "bcaeb411",
     ".ai/manifests/repository.json": "047d93d0",
-    ".ai/manifests/services.json": "ee33ea47",
-    ".ai/manifests/tests.json": "17a5bc52",
+    ".ai/manifests/services.json": "9c92d2c4",
+    ".ai/manifests/tests.json": "4b6be224",
     ".ai/manifests/workspace-dependency-graph.json": "80e5438b",
     ".ai/manifests/workspaces.json": "b5133e99",
     ".ai/packs/README.md": "2e64753b",
@@ -36,7 +36,7 @@
     "apps/claw-agent-service/AGENTS.md": "8d90d82d",
     "apps/claw-audit-service/AGENTS.md": "2342ed5d",
     "apps/claw-auth-service/AGENTS.md": "6ff5cb9c",
-    "apps/claw-chat-service/AGENTS.md": "4a134984",
+    "apps/claw-chat-service/AGENTS.md": "ed858ddb",
     "apps/claw-client-logs-service/AGENTS.md": "7d329099",
     "apps/claw-connector-service/AGENTS.md": "45f37482",
     "apps/claw-file-generation-service/AGENTS.md": "f247a3e4",
@@ -45,11 +45,11 @@
     "apps/claw-health-service/AGENTS.md": "5ebb83b5",
     "apps/claw-image-service/AGENTS.md": "ba82000d",
     "apps/claw-llamacpp-service/AGENTS.md": "5514cb48",
-    "apps/claw-memory-service/AGENTS.md": "5357edef",
+    "apps/claw-memory-service/AGENTS.md": "9ddbf165",
     "apps/claw-ollama-service/AGENTS.md": "13684e8e",
     "apps/claw-payment-service/AGENTS.md": "e1e9991f",
     "apps/claw-research-service/AGENTS.md": "c6c46e1f",
-    "apps/claw-routing-service/AGENTS.md": "dced4a0e",
+    "apps/claw-routing-service/AGENTS.md": "302993c4",
     "apps/claw-server-logs-service/AGENTS.md": "fc1c971a",
     "apps/claw-workspace-service/AGENTS.md": "bf6bb5c0",
     "packages/shared-auth/AGENTS.md": "049afff2",

.ai/manifests/i18n.json

@@ -1,5 +1,5 @@
 {
-  "approxKeyCount": 5001,
+  "approxKeyCount": 5015,
   "generated": true,
   "locales": [
     "ar",

.ai/manifests/services.json

@@ -135,7 +135,7 @@
         "FileDeliveryRecord",
         "MessageAttachment"
       ],
-      "testFiles": 117,
+      "testFiles": 125,
       "testRunner": "jest"
     },
     {
@@ -320,7 +320,7 @@
         "MemoryUsage",
         "WorkspaceObjectEmbedding"
       ],
-      "testFiles": 12,
+      "testFiles": 14,
       "testRunner": "jest"
     },
     {
@@ -421,7 +421,7 @@
     {
       "database": "postgresql",
       "dir": "apps/claw-routing-service",
-      "endpointCount": 69,
+      "endpointCount": 70,
       "internalDeps": [
         "@claw/shared-constants",
         "@claw/shared-entitlements",

.ai/manifests/tests.json

@@ -38,7 +38,7 @@
     },
     "claw-chat-service": {
       "runner": "jest",
-      "testFiles": 117
+      "testFiles": 125
     },
     "claw-client-logs-service": {
       "runner": "jest",
@@ -74,7 +74,7 @@
     },
     "claw-memory-service": {
       "runner": "jest",
-      "testFiles": 12
+      "testFiles": 14
     },
     "claw-ollama-service": {
       "runner": "jest",
@@ -102,5 +102,5 @@
     }
   },
   "generated": true,
-  "total": 1109
+  "total": 1119
 }

apps/claw-chat-service/AGENTS.md

@@ -21,7 +21,7 @@ npm run dev
 - Database: postgresql
 - Prisma models: ChatMessage, ChatMessageContextReceipt, ChatShare, ChatShareMessage, ChatShareMessageAsset, ChatThread, FileDeliveryRecord, MessageAttachment
 - API endpoints: 47 (see `.ai/manifests/api-endpoints.json`)
-- Test files: 117 (jest)
+- Test files: 125 (jest)
 - Depends on: @claw/shared-constants, @claw/shared-entitlements, @claw/shared-rabbitmq, @claw/shared-types, @claw/shared-utilities
 
 ## Before editing

apps/claw-chat-service/prisma/migrations/20260830120000_add_cross_thread_context/migration.sql

@@ -0,0 +1,13 @@
+-- ADR-087 — cross-thread retrieval.
+--
+-- Default FALSE, and that is the whole privacy posture in one word: reaching
+-- into a user's other conversations is opt-in. Every existing thread therefore
+-- keeps behaving exactly as it does today after this migration runs.
+ALTER TABLE "chat_threads"
+  ADD COLUMN "use_cross_thread_context" BOOLEAN NOT NULL DEFAULT false;
+
+-- Stage 2 of retrieval reads recent messages for a set of candidate threads.
+-- The existing single-column thread_id index cannot serve the ORDER BY, so
+-- without this the read degrades to a scan as a user's history grows.
+CREATE INDEX IF NOT EXISTS "chat_messages_thread_id_created_at_idx"
+  ON "chat_messages" ("thread_id", "created_at");

apps/claw-chat-service/prisma/schema.prisma

@@ -52,6 +52,12 @@ model ChatThread {
   // === Integration V2 — per-thread memory/context switches ===
   useMemory          Boolean     @default(true) @map("use_memory")
   useContext         Boolean     @default(true) @map("use_context")
+  // === ADR-087 — cross-thread retrieval ===
+  // OFF by default, and deliberately so. Silently reaching into a user's other
+  // conversations is a privacy decision, not a quality tweak: it must be asked
+  // for. When off, nothing from another thread can enter the prompt except a
+  // durable memory the user's own memory preferences already allow.
+  useCrossThreadContext Boolean  @default(false) @map("use_cross_thread_context")
   createdAt          DateTime    @default(now()) @map("created_at")
   updatedAt          DateTime    @updatedAt @map("updated_at")
 
@@ -106,6 +112,9 @@ model ChatMessage {
 
   @@index([threadId])
   @@index([createdAt])
+  // Cross-thread retrieval reads a user's recent messages across many threads.
+  // Without this the stage-2 read is a scan of every message in the table.
+  @@index([threadId, createdAt])
   @@map("chat_messages")
 }
 

apps/claw-chat-service/src/__tests__/consensus-execution.manager.spec.ts

@@ -3,6 +3,11 @@ import { ConsensusExecutionManager } from '../modules/chat-messages/managers/con
 import type { ParallelModelTarget } from '../modules/chat-messages/types/parallel.types';
 import type { AssembledContext } from '../modules/chat-messages/types/context.types';
 import { createFakePaygAccessControl } from '../modules/chat-messages/__tests__/helpers/fake-payg-access-control.helper';
+import {
+  disabledCrossThreadResult,
+  emptyConversationManifest,
+  fallbackModelTokenBudget,
+} from '../modules/chat-messages/utilities/assembled-context.utility';
 
 jest.spyOn(AppConfig, 'get').mockReturnValue({
   CHAT_DATABASE_URL: 'postgresql://test:test@localhost:5432/test',
@@ -58,6 +63,9 @@ describe('ConsensusExecutionManager', () => {
     fileContents: [],
     workspaceCitations: [],
     tokenBudget: 4096,
+    modelBudget: fallbackModelTokenBudget(),
+    conversationManifest: emptyConversationManifest(),
+    crossThread: disabledCrossThreadResult(),
     researchEvidence: [],
     researchRunId: null,
     researchWarnings: [],

apps/claw-chat-service/src/__tests__/escalation-chain.manager.spec.ts

@@ -3,6 +3,11 @@ import { EscalationChainManager } from '../modules/chat-messages/managers/escala
 import { EscalationChainStatus } from '../common/enums/escalation-chain-status.enum';
 import type { EscalationChainStep } from '../modules/chat-messages/types/escalation-chain.types';
 import type { AssembledContext } from '../modules/chat-messages/types/context.types';
+import {
+  disabledCrossThreadResult,
+  emptyConversationManifest,
+  fallbackModelTokenBudget,
+} from '../modules/chat-messages/utilities/assembled-context.utility';
 
 jest.spyOn(AppConfig, 'get').mockReturnValue({
   CHAT_DATABASE_URL: 'postgresql://test:test@localhost:5432/test',
@@ -62,6 +67,9 @@ describe('EscalationChainManager', () => {
     fileContents: [],
     workspaceCitations: [],
     tokenBudget: 4096,
+    modelBudget: fallbackModelTokenBudget(),
+    conversationManifest: emptyConversationManifest(),
+    crossThread: disabledCrossThreadResult(),
     researchEvidence: [],
     researchRunId: null,
     researchWarnings: [],

apps/claw-chat-service/src/__tests__/parallel-execution.manager.spec.ts

@@ -5,6 +5,11 @@ import type {
   ParallelModelTarget,
 } from '../modules/chat-messages/types/parallel.types';
 import type { AssembledContext } from '../modules/chat-messages/types/context.types';
+import {
+  disabledCrossThreadResult,
+  emptyConversationManifest,
+  fallbackModelTokenBudget,
+} from '../modules/chat-messages/utilities/assembled-context.utility';
 
 jest.spyOn(AppConfig, 'get').mockReturnValue({
   CHAT_DATABASE_URL: 'postgresql://test:test@localhost:5432/test',
@@ -97,6 +102,9 @@ describe('ParallelExecutionManager', () => {
     fileContents: [],
     workspaceCitations: [],
     tokenBudget: 4096,
+    modelBudget: fallbackModelTokenBudget(),
+    conversationManifest: emptyConversationManifest(),
+    crossThread: disabledCrossThreadResult(),
     researchEvidence: [],
     researchRunId: null,
     researchWarnings: [],

apps/claw-chat-service/src/app/config/app.config.ts

@@ -26,6 +26,10 @@ const appConfigSchema = z.object({
   OLLAMA_SERVICE_URL: z.string().min(1).default('http://ollama-service:4008'),
   LLAMACPP_SERVICE_URL: z.string().min(1).default('http://llamacpp-service:4017'),
   CONNECTOR_SERVICE_URL: z.string().min(1).default('http://connector-service:4003'),
+  // Read for one thing only: the selected model's real context window, which
+  // chat-service must never keep its own copy of. See ModelContextWindowClient
+  // and ADR-086.
+  ROUTING_SERVICE_URL: z.string().min(1).default('https://routing-service:4004'),
   MEMORY_SERVICE_URL: z.string().min(1).default('http://memory-service:4005'),
   FILE_SERVICE_URL: z.string().min(1).default('http://file-service:4006'),
   IMAGE_SERVICE_URL: z.string().min(1).default('http://image-service:4012'),

apps/claw-chat-service/src/common/constants/execution.constants.ts

@@ -26,8 +26,28 @@ export const IMAGE_PROVIDER_PREFIX = 'IMAGE_';
 
 export const FILE_GENERATION_PROVIDER = 'FILE_GENERATION';
 
+/**
+ * DEPRECATED as a context rule. Kept only as the Runtime V2 tool-trail window.
+ *
+ * This was the whole conversational memory of the product: `assemble` sliced
+ * the thread to the last twenty messages and everything downstream cut further.
+ * Conversational history is now selected by ContextComposerManager against a
+ * real token budget — see ADR-086. Do not reintroduce a message-count rule.
+ */
 export const THREAD_CONTEXT_LIMIT = 20;
 
+/**
+ * How many rows are read from the database for one generation.
+ *
+ * Not a context rule — a read cap. The composer decides what of this reaches
+ * the model. It was 20, applied at the query, so a hundred-message thread had
+ * eighty messages that no amount of budget could recover: they were never
+ * loaded. Four hundred rows of a single indexed thread is a cheap read and
+ * covers a very long conversation; beyond it, hierarchical summaries (Batch 2)
+ * take over rather than an ever-larger SELECT.
+ */
+export const THREAD_HISTORY_FETCH_LIMIT = 400;
+
 export const MEMORY_FETCH_LIMIT = 20;
 
 /**

apps/claw-chat-service/src/common/constants/index.ts

@@ -21,6 +21,7 @@ export {
   OLLAMA_PROVIDER,
   PROVIDER_BASE_URLS,
   THREAD_CONTEXT_LIMIT,
+  THREAD_HISTORY_FETCH_LIMIT,
   VIDEO_MIME_PREFIX,
   WORKSPACE_CONTEXT_LIMIT,
 } from './execution.constants';

apps/claw-chat-service/src/common/utilities/receipt-from-context.utility.ts

@@ -4,6 +4,7 @@ import {
   type MemorySensitivity,
   type MemoryType,
   type RetrievalBundle,
+  type RetrievalConversationSummary,
   RetrievalReason,
 } from '@claw/shared-types';
 import type { AssembledContext } from '../../modules/chat-messages/types/context.types';
@@ -45,12 +46,59 @@ export function receiptFromAssembledContext(
     memories,
     packItems,
     assemblyOrder: [
+      ...(context.conversationManifest?.includedMessageIds ?? []).map((id) => `message:${id}`),
+      ...(context.crossThread?.selections ?? []).map(
+        (selection) => `prior-message:${selection.messageId}`,
+      ),
       ...memories.map((m) => `memory:${m.id}`),
       ...packItems.map((p) => `pack:${p.id}`),
     ],
     tokenBudget: context.tokenBudget,
     tokenBudgetUsed,
-    retrievalLatencyMs: 0,
-    warnings: [],
+    retrievalLatencyMs: context.conversationManifest?.retrievalMs ?? 0,
+    warnings: context.conversationManifest?.warnings ?? [],
+    conversation: conversationSummary(context),
+  };
+}
+
+/**
+ * The conversational half of the receipt.
+ *
+ * This is the part that answers "why did the AI forget this?". Everything here
+ * is derived from the composer's own decision record, not recomputed, so the
+ * receipt cannot disagree with what was actually sent.
+ */
+function conversationSummary(context: AssembledContext): RetrievalConversationSummary | undefined {
+  // A receipt is a diagnostic artifact. It must never be the reason a
+  // generation fails, so a context built by a path that predates the manifest
+  // yields no conversation summary rather than a thrown TypeError.
+  const manifest = context.conversationManifest;
+  if (manifest === undefined || manifest === null) {
+    return undefined;
+  }
+  const omissionReasons: Record<string, string> = {};
+  for (const omitted of manifest.omitted) {
+    omissionReasons[omitted.messageId] = omitted.reason;
+  }
+  return {
+    totalThreadMessages: manifest.totalThreadMessages,
+    includedMessageIds: manifest.includedMessageIds,
+    includedTurnCount: manifest.includedTurnCount,
+    omittedMessageIds: manifest.omitted.map((omitted) => omitted.messageId),
+    omissionReasons,
+    estimatedInputTokens: manifest.estimatedInputTokens,
+    contextWindowTokens: manifest.budget.contextWindowTokens,
+    reservedOutputTokens: manifest.budget.reservedOutputTokens,
+    availableInputTokens: manifest.budget.availableInputTokens,
+    contextWindowSource: manifest.budget.source,
+    referenceSignals: manifest.referenceSignal.signals,
+    priorThreadsSearched: context.crossThread?.searchedThreadIds ?? [],
+    priorThreadsUsed: context.crossThread?.usedThreadIds ?? [],
+    priorMessageIds: (context.crossThread?.selections ?? []).map(
+      (selection) => selection.messageId,
+    ),
+    crossThreadSkipReason: context.crossThread?.skippedReason ?? null,
+    retrievalMs: manifest.retrievalMs,
+    selectionMs: manifest.selectionMs,
   };
 }