talon-agent 4.6.0 → 5.0.0

package/README.md

@@ -119,9 +119,9 @@ index.ts                    Composition root
   |   +-- bus/              Typed pub-sub spine + event journal
   |   +-- vfs/              The talon:// namespace (~/.talon/ns), FUSE-backed
   |   +-- mesh/             Device mesh: presence, exec/fs channel, teleport
-  |   +-- soul/             Soul kernel — associative recall over memory
-  |   +-- scripts/          Run-to-completion execution of saved scripts
-  |   +-- scripting/        WASM-sandboxed Lua runner for trigger scripts
+  |   +-- memory/           Typed memory store policy: core view, import, taps
+  |   +-- scripts/          Run-to-completion execution of saved scripts +
+  |   |                     the WASM-sandboxed Lua runner for trigger scripts
   |   +-- daemon/           Start / stop / restart, pidfile, discovery
   |   +-- plugin/           Plugin loader, registry, hot-reload
   |   +-- update/           Self-update for git-checkout deployments

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "talon-agent",
-  "version": "4.6.0",
+  "version": "5.0.0",
   "description": "Multi-frontend AI agent with full tool access, streaming, cron jobs, and plugin system",
   "author": "Dylan Neve",
   "license": "MIT",
@@ -80,10 +80,11 @@
     "test:watch": "vitest",
     "test:coverage": "vitest run --coverage",
     "typecheck": "tsc --noEmit",
-    "lint": "oxlint src/",
+    "lint": "oxlint --max-warnings 0 src/",
     "depcruise": "node scripts/check-architecture.mjs",
     "ratchets": "node scripts/check-ratchets.mjs",
     "function-size": "node scripts/check-function-size.mjs",
+    "tree": "node scripts/check-tree.mjs",
     "knip": "knip",
     "format": "prettier --write src/ prompts/",
     "format:check": "prettier --check src/ prompts/",

package/prompts/README.md

@@ -95,7 +95,7 @@ frontend, so it never belongs in the frontend `.md` files:
 The delivery TOOL NAMES are per-frontend (`end_turn`/`send`/`react` on
 telegram & discord, `end_turn`/`send_message`/`react` on native,
 `end_turn`/`send_message` on teams) and are injected into the templates
-by `src/backend/shared/delivery-contract.ts`, which is also where the
+by `src/backend/runtime/prompt/delivery-contract.ts`, which is also where the
 frontend-aware `[FLOW VIOLATION]` reminder and the first-turn nudge
 (appended to the turn-0 user message) are built.
 
@@ -109,7 +109,7 @@ contradictions between backends.
 `heartbeat.md` and `dream.md` are not part of the chat system prompt — they
 are standalone prompts for the background heartbeat and memory-consolidation
 (dream) agents, loaded by `src/core/background/heartbeat.ts` /
-`src/core/background/dream.ts`.
+`src/core/background/dream/index.ts`.
 
 ## Token budget
 

package/prompts/system/memory-recall.md

@@ -59,3 +59,19 @@ memory instead is what crowds a memory file with snapshots that were true for
 an hour.
 
 Memory updates should usually be quiet unless the user asks about them.
+
+### Remembering
+
+When the `remember`, `recall`, and `forget` tools are available, they are the
+canonical store — use them in preference to editing memory files by hand.
+
+Call `remember` as you learn something durable: a stated preference, a
+correction to something you had wrong, a fact about a person or a project the
+user would expect you to still know next week. One claim per call.
+
+Never store secrets or credentials, live operational status, or anything the
+user asked you to keep off the record — and never store a claim as fact when
+you are only guessing.
+
+If a claim is already on file in a different wording, replace it rather than
+adding a second copy; `forget` needs a reason, and drops are recoverable.

package/src/app.ts

@@ -10,8 +10,11 @@ import { getFrontends } from "./core/config/index.js";
 import { startUploadCleanup, stopUploadCleanup } from "./util/workspace.js";
 import { flushDatabase } from "./storage/db.js";
 import { getActiveCount, stopAllTurns } from "./core/engine/dispatcher.js";
-import { startPulseTimer, stopPulseTimer } from "./core/background/pulse.js";
-import { stopPlanAlerts } from "./core/background/plan-alerts.js";
+import {
+  startPulseTimer,
+  stopPulseTimer,
+} from "./core/background/pulse/pulse.js";
+import { stopPlanAlerts } from "./core/background/pulse/plan-alerts.js";
 import {
   startHeartbeatTimer,
   stopHeartbeatTimer,
@@ -21,7 +24,7 @@ import {
   startCronTimer,
   stopCronTimer,
   runStartupCatchup,
-} from "./core/background/cron.js";
+} from "./core/background/cron/scheduler.js";
 import { shutdownTriggers } from "./core/background/triggers/index.js";
 import { pruneSettledTriggers } from "./storage/triggers.js";
 import { startWatchdog, stopWatchdog } from "./util/watchdog.js";
@@ -51,6 +54,11 @@ import {
   writePidRecord,
   removePidRecordIfOwnedBy,
 } from "./core/daemon/pidfile.js";
+import {
+  recordBootMetrics,
+  startResourceSampler,
+  stopResourceSampler,
+} from "./core/daemon/resource-sampler.js";
 
 const { config } = await bootPhase("bootstrap", () => bootstrap());
 
@@ -202,6 +210,7 @@ async function gracefulShutdown(signal: string): Promise<void> {
   });
   await shutdownStep("triggers", shutdownTriggers);
   await shutdownStep("watchdog", stopWatchdog);
+  await shutdownStep("resource sampler", stopResourceSampler);
   await shutdownStep("upload cleanup", stopUploadCleanup);
   await shutdownStep("mcp hub", async () => {
     const { shutdownHub } = await import("./core/mcp-hub/index.js");
@@ -311,7 +320,13 @@ async function main(): Promise<void> {
   await bootPhase("frontends start", () =>
     Promise.all(blockingFrontends.map((frontend) => frontend.start())),
   );
-  log("bot", `Ready in ${bootReport()}`);
+  // Phase 0 accounting (docs/ts-migration-plan.md): the boot is over the
+  // moment the frontends are listening, so the totals are folded into the
+  // metrics store here, from the same uptime figure the log line prints.
+  const bootMs = Math.round(process.uptime() * 1000);
+  recordBootMetrics(bootMs);
+  startResourceSampler();
+  log("bot", `Ready in ${bootReport(bootMs)}`);
   for (const frontend of stdinFrontends) {
     void frontend
       .start()

package/src/backend/claude-sdk/handler.ts

@@ -6,7 +6,7 @@
  * (session expired / context overflow / model fallback via
  * `applyRetryDecisionStream`) and the post-result watchdog. The phases
  * after the stream — accounting, the trailing-prose contract, the result
- * events — are the shared ones in `backend/shared/turn-phases.ts`.
+ * events — are the shared ones in `backend/runtime/turn/turn-phases.ts`.
  *
  * The exported async generator `runChatTurn` is what the factory wires
  * onto `ChatBackend.runChatTurn` — no wrapper, no callback shim.
@@ -34,12 +34,13 @@ import {
 } from "../../core/agent-runtime/events.js";
 import type { ChatRunParams } from "../../core/agent-runtime/capabilities.js";
 import { makeBareModelRef } from "../../core/agent-runtime/model-ref.js";
-import { applyRetryDecisionStream } from "../shared/handle-retry.js";
+import { applyRetryDecisionStream } from "../runtime/turn/handle-retry.js";
 import { getConfig } from "./state.js";
 import { buildSdkOptions, getActiveFrontends } from "./options.js";
 import { waitForMcpServersReady } from "./mcp-ready.js";
 import { invalidatePlanUsage } from "./plan-usage.js";
-import { frontendsForChat } from "../shared/frontends.js";
+import { frontendsForChat } from "../runtime/frontends.js";
+import { rollUpTurnCache } from "../runtime/cache/cache-metrics.js";
 import {
   createStreamState,
   isSystemInit,
@@ -47,10 +48,12 @@ import {
   isAssistant,
   isResult,
   isRateLimitEvent,
+  isCompactBoundary,
   isUserMessage,
   extractToolResults,
   processStreamDelta,
   processAssistantMessage,
+  processCompactBoundary,
   processResultMessage,
   type StreamState,
 } from "./stream.js";
@@ -74,7 +77,7 @@ import {
   enforceTrailingProse,
   buildResultEvents,
   turnUsageSnapshot,
-} from "../shared/index.js";
+} from "../runtime/index.js";
 
 // ── Post-result watchdog ────────────────────────────────────────────────────
 // The SDK's PostToolBatch hook is the canonical loop-terminator — it returns
@@ -334,6 +337,12 @@ async function* consumeSdkStream(
       yield* translateToolResults(ctx, message);
       continue;
     }
+    // The SDK compacted the session mid-turn. Measured, not acted on: the
+    // turn carries on exactly as before.
+    if (isCompactBoundary(message)) {
+      processCompactBoundary(message, ctx.chatId);
+      continue;
+    }
     // The turn just moved the plan's usage — drop the cached windows so
     // the next /status reads them again instead of showing pre-turn
     // figures.
@@ -399,6 +408,7 @@ function buildTurnPrompt(
     senderHandle: params.senderHandle,
     isGroup: params.isGroup,
     messageId: params.messageId,
+    retrievedMemory: params.retrievedMemory,
   });
   if (frontend && previousTurns === 0 && !internal.flowRetries) {
     prompt += `\n\n${buildFirstTurnReminder(frontend)}`;
@@ -445,12 +455,23 @@ function accountFailedClaudeTurn(
 /**
  * The aggregate `cache=NN%` can't distinguish a turn that reused the
  * previous turn's prefix from one that re-wrote it — see
- * shared/cache-telemetry.ts. A lookback overflow only *predicts* a miss,
+ * runtime/cache/cache-telemetry.ts. A lookback overflow only *predicts* a miss,
  * so warn when this turn's verdict proves the previous turn's overflow
  * cost a prefix re-write, then record this turn's overflow for the next.
+ *
+ * Also the rollup seam: the verdict goes into `cache.first_request.*` (and
+ * `cache.session_start.*` on a chat's first turn) here rather than in
+ * `runChatTurn`, which is at its size limit. `turnsIncludingThis` is the
+ * session's turn count with this turn already counted — `incrementTurns`
+ * has run by the time this is called, so 1 means "this was the first".
  */
-function reportCacheVerdict(chatId: string, state: StreamState): void {
+function reportCacheVerdict(
+  chatId: string,
+  state: StreamState,
+  turnsIncludingThis: number,
+): void {
   if (state.cacheStats) {
+    rollUpTurnCache(chatId, state.cacheStats, turnsIncludingThis);
     const overflow = priorLookbackOverflow(chatId);
     if (
       overflow !== undefined &&
@@ -647,7 +668,7 @@ export async function* runChatTurn(
   incrementTurns(chatId);
 
   state.allResponseText += state.currentBlockText;
-  reportCacheVerdict(chatId, state);
+  reportCacheVerdict(chatId, state, session.turns);
 
   const usage = turnUsageSnapshot(state);
   log(

package/src/backend/claude-sdk/one-shot.ts

@@ -19,7 +19,7 @@ import { log, logWarn } from "../../util/log.js";
 import { ALLOWED_TOOLS_BACKGROUND } from "../../core/constants.js";
 import { EFFORT_MAP } from "./constants.js";
 import { buildMcpServers, buildPluginMcpServers } from "./options.js";
-import { warnIfBelowCacheMinimum } from "../shared/cache-telemetry.js";
+import { warnIfBelowCacheMinimum } from "../runtime/cache/cache-telemetry.js";
 
 const DEFAULT_SUBPROCESS_KILL_GRACE_MS = 5 * 1000;
 

package/src/backend/claude-sdk/options.ts

@@ -15,7 +15,7 @@ import type {
   HookCallback,
   HookJSONOutput,
 } from "@anthropic-ai/claude-agent-sdk";
-import type { PreparedSystemPrompt } from "../shared/system-prompt.js";
+import type { PreparedSystemPrompt } from "../runtime/prompt/system-prompt.js";
 import { getSession } from "../../storage/sessions.js";
 import { getChatSettings } from "../../storage/chat-settings.js";
 import { resolveModelId } from "../../core/models/catalog.js";
@@ -25,11 +25,12 @@ import {
   pluginHubUrl,
   hubPluginServerNames,
 } from "../../core/mcp-hub/index.js";
-import { nonTerminalFrontends, frontendsForChat } from "../shared/frontends.js";
 import {
-  noteToolFingerprint,
-  toolFingerprint,
-} from "../shared/cache-telemetry.js";
+  nonTerminalFrontends,
+  frontendsForChat,
+} from "../runtime/frontends.js";
+import { toolFingerprint } from "../runtime/cache/cache-telemetry.js";
+import { reportToolFingerprint } from "../runtime/cache/cache-metrics.js";
 import { log, logError } from "../../util/log.js";
 import { getConfig, getBridgePort } from "./state.js";
 import { ALLOWED_TOOLS_CHAT, EFFORT_MAP } from "./constants.js";
@@ -395,7 +396,7 @@ export function buildSdkOptions(
   // mid-session invalidates the system prompt and every cached message after
   // it — the most expensive cache event there is, and one the aggregate
   // hit-rate can't show. Plugin-provided MCP servers are the mutable part.
-  noteToolFingerprint(
+  reportToolFingerprint(
     chatId,
     toolFingerprint(builtinTools, Object.keys(mcpServers)),
   );

package/src/backend/claude-sdk/stream.ts

@@ -7,6 +7,7 @@
  */
 
 import type {
+  SDKCompactBoundaryMessage,
   SDKMessage,
   SDKSystemMessage,
   SDKPartialAssistantMessage,
@@ -22,7 +23,8 @@ import { checkModelDrift } from "./model-drift.js";
 import {
   turnCacheStats,
   type TurnCacheStats,
-} from "../shared/cache-telemetry.js";
+} from "../runtime/cache/cache-telemetry.js";
+import { recordCompactBoundary } from "../runtime/cache/cache-metrics.js";
 
 // ── Stream state accumulator ────────────────────────────────────────────────
 
@@ -46,7 +48,7 @@ export type StreamState = {
    * `usage.iterations`. Undefined when the provider reported none — the
    * aggregate totals above cannot distinguish a turn that read the previous
    * turn's prefix from one that re-wrote it, and that distinction is the
-   * whole cost signal (see shared/cache-telemetry.ts).
+   * whole cost signal (see runtime/cache/cache-telemetry.ts).
    */
   cacheStats: TurnCacheStats | undefined;
   lastStreamUpdate: number;
@@ -156,6 +158,18 @@ export function isRateLimitEvent(msg: SDKMessage): boolean {
   return msg.type === "rate_limit_event";
 }
 
+/**
+ * Emitted when the SDK compacts the session — the transcript before the
+ * boundary is replaced by a summary, so every cached block after the system
+ * prompt is gone and the next turn re-writes it. Shares `type: 'system'`
+ * with the init message, so the subtype is what tells them apart.
+ */
+export function isCompactBoundary(
+  msg: SDKMessage,
+): msg is SDKCompactBoundaryMessage {
+  return msg.type === "system" && msg.subtype === "compact_boundary";
+}
+
 // ── Message processors ──────────────────────────────────────────────────────
 
 /** Output of `processStreamDelta` when the throttle interval has elapsed. */
@@ -334,6 +348,19 @@ export function extractToolResults(msg: SDKUserMessage): ToolResultInfo[] {
   return results;
 }
 
+/**
+ * Record a compaction boundary. Observational only: the turn proceeds
+ * exactly as it did before this message was recognised — what changes is
+ * that `session.compacted.*` now says how often the SDK compacts and how
+ * much it reclaims (docs/cache-economics.md, PR A step 3).
+ */
+export function processCompactBoundary(
+  msg: SDKCompactBoundaryMessage,
+  chatId: string,
+): void {
+  recordCompactBoundary(chatId, msg.compact_metadata);
+}
+
 /** Context fill of the last API iteration + whether the turn's first request hit the cache. */
 function readContextFill(
   state: StreamState,
@@ -450,4 +477,4 @@ export function processResultMessage(
 export {
   normalizeForDedupe,
   isDuplicateOfDelivered,
-} from "../shared/delivered-text.js";
+} from "../runtime/turn/delivered-text.js";

package/src/backend/claude-sdk/warm.ts

@@ -11,7 +11,7 @@ import { getSession } from "../../storage/sessions.js";
 import { log, logWarn } from "../../util/log.js";
 import { getConfig } from "./state.js";
 import { buildSdkOptions } from "./options.js";
-import { prepareSystemPrompt } from "../shared/index.js";
+import { prepareSystemPrompt } from "../runtime/index.js";
 
 export async function warmSession(chatId: string): Promise<void> {
   // Guard against being called before initAgent()

package/src/backend/codex/constants.ts

@@ -2,7 +2,7 @@
  * Codex backend constants.
  */
 
-import { buildDeliveryContract } from "../shared/delivery-contract.js";
+import { buildDeliveryContract } from "../runtime/prompt/delivery-contract.js";
 
 /**
  * System-prompt suffix appended to the user-configured system prompt.

package/src/backend/codex/factory.ts

@@ -10,8 +10,8 @@ import { registerBackend } from "../../core/agent-runtime/backend-registry.js";
 import { codexDoctorChecks } from "./doctor.js";
 import type { BackendFactory } from "../../core/agent-runtime/backend-registry.js";
 import { log } from "../../util/log.js";
-import { handlerToEvents } from "../shared/handler-to-events.js";
-import { interruptChatTurn } from "../shared/turn-interrupt.js";
+import { handlerToEvents } from "../runtime/turn/handler-to-events.js";
+import { interruptChatTurn } from "../runtime/turn/turn-interrupt.js";
 import {
   composeBackend,
   type ChatBackend,

package/src/backend/codex/handler/events.ts

@@ -16,7 +16,7 @@ import {
   recordToolUse,
   recordToolCall,
   type StreamState,
-} from "../../shared/index.js";
+} from "../../runtime/index.js";
 
 export interface HandleEventContext {
   state: StreamState;

package/src/backend/codex/handler/message.ts

@@ -7,11 +7,14 @@
  * `codex.resumeThread(id)`, the rollout-JSONL live/settle usage accounting
  * (`rollout-accounting.ts`), and the ChatGPT-OAuth model-mismatch recovery
  * ladder. The post-stream phases are the shared ones in
- * `backend/shared/turn-phases.ts`.
+ * `backend/runtime/turn/turn-phases.ts`.
  */
 
 import type { Thread, Usage } from "@openai/codex-sdk";
-import type { QueryParams, QueryResult } from "../../shared/handler-types.js";
+import type {
+  QueryParams,
+  QueryResult,
+} from "../../runtime/turn/handler-types.js";
 import {
   getSession,
   incrementTurns,
@@ -37,7 +40,7 @@ import {
   nameSessionFromFirstMessage,
   finishCallbackTurn,
   type StreamState,
-} from "../../shared/index.js";
+} from "../../runtime/index.js";
 
 import {
   codexSystemPromptSuffix,
@@ -48,7 +51,7 @@ import {
 import {
   frontendsForChat,
   nonTerminalFrontends,
-} from "../../shared/frontends.js";
+} from "../../runtime/frontends.js";
 import { getState } from "../state.js";
 import { ensureCodex, getCodexAuthInfo } from "../init.js";
 import {
@@ -213,6 +216,23 @@ async function maybeFallbackForChatGptMismatch(
  * ChatGPT-OAuth account is swapped pre-emptively rather than letting the
  * first turn fail.
  */
+/**
+ * The turn's user prompt: the shared framing every backend emits, plus
+ * whatever this turn's memory retrieval produced. `formatUserPrompt` is
+ * the one place `retrievedMemory` is rendered — see
+ * `backend/runtime/prompt/prompt-format.ts`.
+ */
+function buildTurnPrompt(params: QueryParams): string {
+  return formatUserPrompt({
+    text: params.text,
+    senderName: params.senderName ?? "user",
+    senderHandle: params.senderHandle,
+    isGroup: params.isGroup,
+    messageId: params.messageId,
+    retrievedMemory: params.retrievedMemory,
+  });
+}
+
 function resolveCodexModel(chatId: string, requested: string | undefined) {
   const authInfo = getCodexAuthInfo();
   const authAwareDefault =
@@ -409,7 +429,7 @@ export async function handleMessage(
   }
   const codex = ensureCodex(params.chatId);
 
-  const { chatId, text, senderName, senderHandle, isGroup, messageId } = params;
+  const { chatId, text, senderName, isGroup } = params;
   const t0 = Date.now();
   const session = getSession(chatId);
   const previousTurns = session.turns;
@@ -432,13 +452,7 @@ export async function handleMessage(
     sessionEpoch: session.createdAt,
   });
 
-  const prompt = formatUserPrompt({
-    text,
-    senderName: senderName ?? "user",
-    senderHandle,
-    isGroup,
-    messageId,
-  });
+  const prompt = buildTurnPrompt(params);
 
   log("agent", `[${chatId}] <- (${text.length} chars)`);
   traceMessage(chatId, "in", text, { senderName, isGroup });

package/src/backend/codex/handler/rollout-accounting.ts

@@ -19,7 +19,7 @@ import {
   pushLiveUsage,
   recordTokens,
   type StreamState,
-} from "../../shared/index.js";
+} from "../../runtime/index.js";
 import { CODEX_LIVE_POLL_INTERVAL_MS } from "../constants.js";
 import {
   readLastRolloutSnapshot,

package/src/backend/codex/init.ts

@@ -25,7 +25,7 @@ import { Codex } from "@openai/codex-sdk";
 import type { TalonConfig } from "../../core/config/index.js";
 import type { FrontendName } from "../../core/agent-runtime/backend-registry.js";
 import { log, logWarn } from "../../util/log.js";
-import { nonTerminalFrontends } from "../shared/frontends.js";
+import { nonTerminalFrontends } from "../runtime/frontends.js";
 import { getState } from "./state.js";
 import { asCodexConfig, buildCodexMcpServers } from "./mcp-config.js";
 import { detectCodexAuth, type CodexAuthInfo } from "./auth.js";

package/src/backend/codex/mcp-config.ts

@@ -30,7 +30,7 @@ import {
   hubPluginServerNames,
 } from "../../core/mcp-hub/index.js";
 import type { ToolExclusionConfig } from "../../core/tools/mcp-env.js";
-import { frontendsForChat } from "../shared/frontends.js";
+import { frontendsForChat } from "../runtime/frontends.js";
 
 /**
  * AppToolApproval values accepted by Codex's `mcp_servers.<name>` table.

package/src/backend/codex/one-shot.ts

@@ -19,7 +19,7 @@
 
 import type { OneShotAgentParams, OneShotUsage } from "../../core/types.js";
 import { log, logWarn } from "../../util/log.js";
-import { appendBackendSuffix } from "../shared/index.js";
+import { appendBackendSuffix } from "../runtime/index.js";
 import { ensureCodex, getCodexAuthInfo } from "./init.js";
 import {
   CODEX_SYSTEM_PROMPT_SUFFIX,

package/src/backend/kilo/handler/message.ts

@@ -3,7 +3,10 @@
  * Kilo's server bindings.
  */
 
-import type { QueryParams, QueryResult } from "../../shared/handler-types.js";
+import type {
+  QueryParams,
+  QueryResult,
+} from "../../runtime/turn/handler-types.js";
 import { runRemoteChatTurn } from "../../remote-server/chat-turn.js";
 import {
   ensureServer,

package/src/backend/openai-agents/constants.ts

@@ -10,7 +10,7 @@
  * orchestration) with MCP servers wired from the plugin system.
  */
 
-import { buildDeliveryContract } from "../shared/delivery-contract.js";
+import { buildDeliveryContract } from "../runtime/prompt/delivery-contract.js";
 
 /**
  * Agents-specific addendum to the shared tool-only delivery contract:

package/src/backend/openai-agents/factory.ts

@@ -13,8 +13,8 @@ import { registerBackend } from "../../core/agent-runtime/backend-registry.js";
 import { openAIAgentsDoctorChecks } from "./doctor.js";
 import type { BackendFactory } from "../../core/agent-runtime/backend-registry.js";
 import { log } from "../../util/log.js";
-import { handlerToEvents } from "../shared/handler-to-events.js";
-import { interruptChatTurn } from "../shared/turn-interrupt.js";
+import { handlerToEvents } from "../runtime/turn/handler-to-events.js";
+import { interruptChatTurn } from "../runtime/turn/turn-interrupt.js";
 import {
   composeBackend,
   type ChatBackend,

package/src/backend/openai-agents/handler/events.ts

@@ -9,7 +9,7 @@ import {
   createStreamState,
   recordToolUse,
   recordToolCall,
-} from "../../shared/index.js";
+} from "../../runtime/index.js";
 
 export interface HandleRunItemContext {
   state: ReturnType<typeof createStreamState>;

package/src/backend/openai-agents/handler/message.ts

@@ -6,11 +6,14 @@
  * MCP bundle, iterating the `StreamedRunResult` (see `events.ts`), and
  * `.cancel()` on terminator. The post-stream phases — accounting, the
  * trailing-prose flow-violation retry, the result — are the shared ones in
- * `backend/shared/turn-phases.ts`.
+ * `backend/runtime/turn/turn-phases.ts`.
  */
 
 import { Agent, run } from "@openai/agents";
-import type { QueryParams, QueryResult } from "../../shared/handler-types.js";
+import type {
+  QueryParams,
+  QueryResult,
+} from "../../runtime/turn/handler-types.js";
 import { getSession, incrementTurns } from "../../../storage/sessions.js";
 import { getChatSettings } from "../../../storage/chat-settings.js";
 import { log, logError, logWarn } from "../../../util/log.js";
@@ -33,7 +36,7 @@ import {
   enforceTrailingProse,
   finishCallbackTurn,
   type StreamState,
-} from "../../shared/index.js";
+} from "../../runtime/index.js";
 
 import {
   buildOpenAiAgentsSuffix,
@@ -43,7 +46,7 @@ import {
 } from "../constants.js";
 import { getState, getOrCreateSession } from "../state.js";
 import { getActiveFrontends } from "../init.js";
-import { frontendsForChat } from "../../shared/frontends.js";
+import { frontendsForChat } from "../../runtime/frontends.js";
 import { getOrCreateBundle } from "../mcp-pool.js";
 import { OPENAI_AGENTS_BUILTIN_TOOLS } from "../builtins.js";
 import { activeAborts } from "./state.js";
@@ -105,6 +108,7 @@ function buildTurnPrompt(
     senderHandle: params.senderHandle,
     isGroup: params.isGroup,
     messageId: params.messageId,
+    retrievedMemory: params.retrievedMemory,
   });
   if (frontend && previousTurns === 0 && !isRetry) {
     prompt += `\n\n${buildFirstTurnReminder(frontend)}`;

package/src/backend/openai-agents/init.ts

@@ -46,7 +46,7 @@ import { setDefaultOpenAIClient, setOpenAIAPI } from "@openai/agents";
 import type { TalonConfig } from "../../core/config/index.js";
 import type { FrontendName } from "../../core/agent-runtime/backend-registry.js";
 import { log, logWarn } from "../../util/log.js";
-import { nonTerminalFrontends } from "../shared/frontends.js";
+import { nonTerminalFrontends } from "../runtime/frontends.js";
 import { getState } from "./state.js";
 import { startDiscovery } from "./discovery.js";
 

package/src/backend/openai-agents/mcp-pool.ts

@@ -31,7 +31,7 @@ import {
   hubPluginServerNames,
 } from "../../core/mcp-hub/index.js";
 import type { ToolExclusionConfig } from "../../core/tools/mcp-env.js";
-import { frontendsForChat } from "../shared/frontends.js";
+import { frontendsForChat } from "../runtime/frontends.js";
 import { log, logWarn } from "../../util/log.js";
 
 /**

package/src/backend/opencode/handler/message.ts

@@ -3,7 +3,10 @@
  * bound to OpenCode's server bindings.
  */
 
-import type { QueryParams, QueryResult } from "../../shared/handler-types.js";
+import type {
+  QueryParams,
+  QueryResult,
+} from "../../runtime/turn/handler-types.js";
 import { runRemoteChatTurn } from "../../remote-server/chat-turn.js";
 import {
   ensureServer,