talon-agent 4.6.1 → 5.0.0

package/src/backend/remote-server/chat-turn.ts

@@ -5,7 +5,7 @@
  * Resolves the active model against the server's catalog, makes sure the
  * server, session, and this chat's MCP servers exist, builds the prompt
  * pair, drives the turn (`./turn.ts`), then runs the shared post-turn
- * phases (`backend/shared/turn-phases.ts`) with the one family-specific
+ * phases (`backend/runtime/turn/turn-phases.ts`) with the one family-specific
  * step in between: the usage fallback from the session summary.
  *
  * Kilo and OpenCode ran byte-for-byte copies of this (modulo the backend
@@ -18,8 +18,14 @@ import { getSession, incrementTurns } from "../../storage/sessions.js";
 import { getChatSettings } from "../../storage/chat-settings.js";
 import { log, logError } from "../../util/log.js";
 import { traceMessage } from "../../util/trace.js";
-import type { QueryParams, QueryResult } from "../shared/handler-types.js";
-import { frontendsForChat, nonTerminalFrontends } from "../shared/frontends.js";
+import type {
+  QueryParams,
+  QueryResult,
+} from "../runtime/turn/handler-types.js";
+import {
+  frontendsForChat,
+  nonTerminalFrontends,
+} from "../runtime/frontends.js";
 import {
   createStreamState,
   recordTokens,
@@ -33,7 +39,7 @@ import {
   accountFailedTurn,
   nameSessionFromFirstMessage,
   finishCallbackTurn,
-} from "../shared/index.js";
+} from "../runtime/index.js";
 import type { RemoteAgentClient } from "./client.js";
 import type { RemoteServerBindings } from "./server-bindings.js";
 import { getTurnSummary, type RemoteSessionClient } from "./session-helpers.js";
@@ -68,16 +74,7 @@ export async function runRemoteChatTurn<TClient extends RemoteAgentClient>(
   const config = bindings.getConfig();
   if (!config) throw new Error(`${label} agent not initialized`);
 
-  const {
-    chatId,
-    text,
-    senderName,
-    senderHandle,
-    isGroup,
-    messageId,
-    onTextBlock,
-    onToolUse,
-  } = params;
+  const { chatId, text, senderName, onTextBlock, onToolUse } = params;
   const t0 = Date.now();
   const session = getSession(chatId);
   const previousTurns = session.turns;
@@ -108,14 +105,8 @@ export async function runRemoteChatTurn<TClient extends RemoteAgentClient>(
     pluginMcpServerNames,
   );
 
-  // Build the prompt (time tag + sender + msg_id reference).
-  const prompt = formatUserPrompt({
-    text,
-    senderName: senderName ?? "user",
-    senderHandle,
-    isGroup,
-    messageId,
-  });
+  // Build the prompt (time tag + sender + msg_id + recalled memory).
+  const prompt = buildTurnPrompt(params);
 
   // Per-session frozen prompt + this backend's delivery suffix.
   const { text: systemPrompt } = prepareSystemPrompt({
@@ -130,7 +121,7 @@ export async function runRemoteChatTurn<TClient extends RemoteAgentClient>(
   });
 
   log("agent", `[${chatId}] <- (${text.length} chars)`);
-  traceMessage(chatId, "in", text, { senderName, isGroup });
+  traceMessage(chatId, "in", text, { senderName, isGroup: params.isGroup });
 
   // Bind the stream state to the chat so token mutators mirror counts
   // into the live-turn overlay — /status updates while the turn runs.
@@ -268,6 +259,23 @@ export async function runRemoteChatTurn<TClient extends RemoteAgentClient>(
  * session summaries can race on cancellation, so a failure leaves the
  * counts at zero.
  */
+/**
+ * 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,
+  });
+}
+
 async function fillUsageFromSummary(
   oc: RemoteSessionClient,
   sessionId: string,

package/src/backend/remote-server/events.ts

@@ -33,7 +33,7 @@ import {
   recordToolUse,
   recordToolCall,
   type StreamState,
-} from "../shared/index.js";
+} from "../runtime/index.js";
 import {
   extractAssistantUsage,
   type RemoteAssistantInfo,
@@ -68,7 +68,7 @@ export interface EventProcessingContext {
   chatId: string;
   /** Session id we're scoped to — events for other sessions are dropped. */
   sessionId: string;
-  /** Stream state accumulator (shared/). */
+  /** Stream state accumulator (runtime/turn/). */
   state: StreamState;
   /** Tool-call ids we've already fired callbacks for. */
   seenToolCallIds: Set<string>;
@@ -164,7 +164,7 @@ export async function processStreamEvent(
  * lands so token/context stats are live mid-turn instead of appearing
  * only at the post-loop accounting. `recordTokens` also mirrors the
  * counts into the chat's live-turn overlay when the state is bound to a
- * chat (see `shared/stream-state.ts: pushLiveUsage`).
+ * chat (see `runtime/turn/stream-state.ts: pushLiveUsage`).
  *
  * Values are the message's totals (not deltas), matching what the
  * post-loop `extractAssistantUsage(lastAssistant.info)` reads — so the

package/src/backend/remote-server/factory.ts

@@ -28,9 +28,12 @@ import type { OneShotAgentParams, OneShotUsage } from "../../core/types.js";
 import type { TalonConfig } from "../../core/config/index.js";
 import { binaryOnPath } from "../../util/binary-on-path.js";
 import { log } from "../../util/log.js";
-import { handlerToEvents } from "../shared/handler-to-events.js";
-import type { QueryParams, QueryResult } from "../shared/handler-types.js";
-import { interruptChatTurn } from "../shared/turn-interrupt.js";
+import { handlerToEvents } from "../runtime/turn/handler-to-events.js";
+import type {
+  QueryParams,
+  QueryResult,
+} from "../runtime/turn/handler-types.js";
+import { interruptChatTurn } from "../runtime/turn/turn-interrupt.js";
 import type { RemoteModelProvider } from "./model-catalog/provider.js";
 import type { RemoteSessionSnapshot } from "./session-helpers.js";
 

package/src/backend/remote-server/index.ts

@@ -28,7 +28,7 @@
  * What's NOT here (intentionally):
  *
  *   - Tool definitions, frontend prompt format — those are backend-
- *     agnostic and live in `core/` and `backend/shared/`.
+ *     agnostic and live in `core/` and `backend/runtime/`.
  *
  * This barrel exposes the helper layer for tests and the conformance
  * suite; the bindings modules import from the concrete files directly.

package/src/backend/remote-server/mcp.ts

@@ -50,7 +50,7 @@ import { ALL_TOOLS, nativeTools } from "../../core/tools/index.js";
 import {
   frontendForChatId,
   nonTerminalFrontends,
-} from "../shared/frontends.js";
+} from "../runtime/frontends.js";
 
 // ── Constants ───────────────────────────────────────────────────────────────
 

package/src/backend/remote-server/one-shot.ts

@@ -41,7 +41,7 @@ import {
   type RemoteSessionClient,
   type RemoteAssistantInfo,
 } from "./session-helpers.js";
-import { appendBackendSuffix, sleep } from "../shared/index.js";
+import { appendBackendSuffix, sleep } from "../runtime/index.js";
 import { buildPermissionRuleset } from "./sessions.js";
 
 // ── Client surface ──────────────────────────────────────────────────────────

package/src/backend/remote-server/server-bindings.ts

@@ -25,7 +25,7 @@ import type { FrontendName } from "../../core/agent-runtime/backend-registry.js"
 import {
   buildDeliveryContract,
   type DeliveryMode,
-} from "../shared/delivery-contract.js";
+} from "../runtime/prompt/delivery-contract.js";
 import type { RemoteAgentClient } from "./client.js";
 import { createRemoteServerState, errMsg } from "./state.js";
 import { ensureRemoteServer, stopRemoteServer } from "./lifecycle.js";

package/src/backend/remote-server/turn.ts

@@ -13,7 +13,7 @@
  */
 
 import { logWarn } from "../../util/log.js";
-import { createStreamState, recordTokens, sleep } from "../shared/index.js";
+import { createStreamState, recordTokens, sleep } from "../runtime/index.js";
 import { processStreamEvent, finalizePartsIntoState } from "./events.js";
 import { findLastAssistantMessage } from "./messages.js";
 import {

package/src/backend/runtime/cache/cache-metrics.ts

@@ -0,0 +1,126 @@
+/**
+ * Prompt-cache rollups — the per-turn verdict, the tool fingerprint and the
+ * SDK's compaction boundaries, folded into counters an operator can read.
+ *
+ * `cache-telemetry.ts` next door decides WHAT a turn did (verdict,
+ * fingerprint, lookback risk) and is pure apart from its own per-chat maps.
+ * This module is the side-effecting half: it writes to the metrics store and
+ * logs, so the telemetry stays testable without a store.
+ *
+ * Why these numbers exist (docs/cache-economics.md, PR A): every decision in
+ * PR B ("one prefix per frontend") and PR C ("compact when the cache has gone
+ * cold") needs a before/after figure, and nothing here changes a prompt byte
+ * or a turn's behaviour — measurement only.
+ *
+ *   - `cache.first_request.{hit,miss,none}` — did the previous turn's prefix
+ *     survive? The turn's first request is the only one that can say, and it
+ *     is the number that tracks cost.
+ *   - `cache.first_request.{read,write}_tokens` — how big the prefix is, so a
+ *     miss can be priced.
+ *   - `cache.session_start.{hit,miss,none}` — the same verdict restricted to a
+ *     chat's FIRST turn. A hit there means another chat had already warmed an
+ *     identical prefix: the cross-chat sharing signal PR B is aiming at.
+ *   - `cache.tool_fingerprint.changed` — tools render before the system
+ *     prompt, so a change invalidates everything after it.
+ *   - `session.compacted.{manual,auto}` + `session.compact.{pre,post}_tokens`
+ *     — how often the SDK compacts and how much it reclaims, the baseline
+ *     PR C's idle-compaction policy is measured against.
+ */
+
+import { log } from "../../../util/log.js";
+import {
+  incrementCounter,
+  noteCacheVerdict,
+  recordHistogram,
+} from "../../../storage/metrics.js";
+import {
+  crossTurnVerdict,
+  fingerprintHash,
+  hasToolFingerprint,
+  noteToolFingerprint,
+  type CrossTurnVerdict,
+  type TurnCacheStats,
+} from "./cache-telemetry.js";
+
+/**
+ * Roll one finished turn's cache behaviour up into the metrics store and
+ * remember the verdict for `/status`.
+ *
+ * `turnsIncludingThis` is the chat's turn count with this turn already
+ * counted, so 1 means the turn was the session's first — the only turn whose
+ * verdict says anything about cross-chat prefix sharing.
+ */
+export function rollUpTurnCache(
+  chatId: string,
+  stats: TurnCacheStats,
+  turnsIncludingThis: number,
+): CrossTurnVerdict {
+  const verdict = crossTurnVerdict(stats);
+  incrementCounter(`cache.first_request.${verdict}`);
+  // Recorded on every turn, zeros included: the average is then "tokens per
+  // turn", which is what a miss costs, rather than an average over the
+  // turns that happened to be interesting.
+  recordHistogram("cache.first_request.read_tokens", stats.firstRead);
+  recordHistogram("cache.first_request.write_tokens", stats.firstWrite);
+  if (turnsIncludingThis <= 1) {
+    incrementCounter(`cache.session_start.${verdict}`);
+  }
+  noteCacheVerdict(chatId, verdict);
+  return verdict;
+}
+
+/**
+ * Record a chat's tool set: one info line the first time a chat is seen (so
+ * two chats on the same frontend can be compared by eye — identical hash +
+ * identical static prompt ⇒ shared cache), and a counter whenever it changes
+ * mid-process. `noteToolFingerprint` still owns the warning it already logs
+ * for a mid-session change.
+ */
+export function reportToolFingerprint(
+  chatId: string,
+  fingerprint: readonly string[],
+): void {
+  const seen = hasToolFingerprint(chatId);
+  const changed = noteToolFingerprint(chatId, fingerprint);
+  if (!seen) {
+    log(
+      "agent",
+      `[${chatId}] tool fingerprint ${fingerprintHash(fingerprint)} ` +
+        `(${fingerprint.length} tools)`,
+    );
+  }
+  if (changed) incrementCounter("cache.tool_fingerprint.changed");
+}
+
+/** The SDK's compaction metadata, narrowed to the fields that are measured. */
+export type CompactBoundary = {
+  trigger: "manual" | "auto";
+  pre_tokens: number;
+  post_tokens?: number;
+};
+
+/**
+ * Record a compaction boundary the SDK reported mid-stream. Purely
+ * observational — the turn carries on exactly as it did before, this just
+ * stops the event being dropped on the floor.
+ */
+export function recordCompactBoundary(
+  chatId: string,
+  meta: CompactBoundary,
+): void {
+  const trigger = meta.trigger === "manual" ? "manual" : "auto";
+  incrementCounter(`session.compacted.${trigger}`);
+  const pre = Number.isFinite(meta.pre_tokens) ? meta.pre_tokens : 0;
+  const post =
+    typeof meta.post_tokens === "number" && Number.isFinite(meta.post_tokens)
+      ? meta.post_tokens
+      : undefined;
+  recordHistogram("session.compact.pre_tokens", pre);
+  if (post !== undefined) recordHistogram("session.compact.post_tokens", post);
+  log(
+    "agent",
+    `[${chatId}] context compacted (${trigger}): ${pre} tokens` +
+      (post !== undefined ? ` -> ${post}` : "") +
+      ` — the prefix after the boundary is new, so the next turn re-writes it`,
+  );
+}

package/src/backend/{shared → runtime/cache}/cache-telemetry.ts

@@ -36,7 +36,9 @@
  * did.
  */
 
-import { logWarn } from "../../util/log.js";
+import { createHash } from "node:crypto";
+
+import { logWarn } from "../../../util/log.js";
 
 // ── Per-turn cache stats ────────────────────────────────────────────────────
 
@@ -244,7 +246,7 @@ export function cacheMinimumTokens(model: string): number | undefined {
   return best?.min;
 }
 
-/** Cheap tokenizer-free estimate (~4 chars/token), matching soul/projector. */
+/** Cheap tokenizer-free estimate (~4 chars/token), the house heuristic. */
 function estimateTokens(text: string): number {
   return Math.ceil(text.length / 4);
 }
@@ -307,6 +309,24 @@ export function toolFingerprint(
   ].sort();
 }
 
+/**
+ * Short, stable digest of a tool set — the thing to compare between two
+ * chats. Two chats whose fingerprints (and static prompt) match render a
+ * byte-identical prefix and therefore share the cache; two that don't,
+ * can't, however similar they look (docs/cache-economics.md, PR B).
+ */
+export function fingerprintHash(fingerprint: readonly string[]): string {
+  return createHash("sha256")
+    .update(fingerprint.join("\n"))
+    .digest("hex")
+    .slice(0, 12);
+}
+
+/** True once a fingerprint has been recorded for this chat in this process. */
+export function hasToolFingerprint(chatId: string): boolean {
+  return lastToolSets.has(chatId);
+}
+
 /**
  * Record this turn's tool set for a chat and warn when it differs from the
  * previous turn's. Returns true when a change was detected.

package/src/backend/{shared → runtime}/index.ts

@@ -1,30 +1,36 @@
 /**
- * Shared backend framework — barrel re-export.
+ * Backend runtime library — barrel re-export.
  *
  * Helpers used by every concrete backend (`claude-sdk`, `codex`,
  * `kilo`, `opencode`, `openai-agents`) to keep behaviour aligned and
- * avoid copy-paste drift.
+ * avoid copy-paste drift. The modules sit in three groups — `turn/`
+ * (what a turn does once the SDK loop is running), `prompt/` (the text
+ * handed to the model) and `cache/` (prompt-cache telemetry) — with the
+ * cross-cutting vocabulary (`usage`, `metrics`, `frontends`, `sleep`)
+ * at the root beside this barrel.
  *
  * What's here:
- *   - `delivered-text` — scratchpad/dedup primitives.
- *   - `delivery-contract` — per-backend response-flow contract
+ *   - `turn/delivered-text` — scratchpad/dedup primitives.
+ *   - `prompt/delivery-contract` — per-backend response-flow contract
  *     (rendered from prompts/system templates), frontend-aware
  *     flow-violation reminder, first-turn nudge.
- *   - `flow-violation` — flow-violation detection + reminder text.
+ *   - `turn/flow-violation` — flow-violation detection + reminder text.
  *   - `metrics` — the shared metric vocabulary (tool calls, per-turn
  *     rollups, flow violations) with backend dimensions.
- *   - `prompt-format` — user-prompt formatter ([time] [Name] [msg_id:N]).
+ *   - `prompt/prompt-format` — user-prompt formatter
+ *     ([time] [Name] [msg_id:N]).
  *   - `frontends` — config `frontend` → messaging-frontend list.
  *   - `extractSessionName` — re-exported from `util/session-name` so the
  *     backends keep one import site; the helper itself is frontend-neutral.
  *   - `usage` — cache-hit % + log summarisers.
- *   - `system-prompt` — per-session prompt snapshots + backend suffix
- *     (assembly itself lives in `core/prompt/`).
- *   - `model-retry` — session-expiry / context-overflow / fallback decisions.
- *   - `stream-state` — backend-agnostic accumulator for stream loops.
- *   - `turn-interrupt` — user-driven mid-turn interrupt registry (the
+ *   - `prompt/system-prompt` — per-session prompt snapshots + backend
+ *     suffix (assembly itself lives in `core/prompt/`).
+ *   - `turn/model-retry` — session-expiry / context-overflow / fallback
+ *     decisions.
+ *   - `turn/stream-state` — backend-agnostic accumulator for stream loops.
+ *   - `turn/turn-interrupt` — user-driven mid-turn interrupt registry (the
  *     shared `ChatBackend.interruptChatTurn` for callback backends).
- *   - `turn-phases` — the post-stream phases (accounting, session name,
+ *   - `turn/turn-phases` — the post-stream phases (accounting, session name,
  *     trailing-prose contract, result tail) every handler runs.
  *
  * What's NOT here (intentionally):
@@ -34,17 +40,17 @@
  *     (the spawn/env contract they share is `core/tools/mcp-env.ts`).
  */
 
-export { captureDeliveredText } from "./delivered-text.js";
+export { captureDeliveredText } from "./turn/delivered-text.js";
 
-export { registerTurnInterrupt } from "./turn-interrupt.js";
+export { registerTurnInterrupt } from "./turn/turn-interrupt.js";
 
-export { formatUserPrompt } from "./prompt-format.js";
+export { formatUserPrompt } from "./prompt/prompt-format.js";
 
 export {
   buildDeliveryContract,
   buildFlowViolationReminder,
   buildFirstTurnReminder,
-} from "./delivery-contract.js";
+} from "./prompt/delivery-contract.js";
 
 export { summarizeUsage } from "./usage.js";
 
@@ -57,9 +63,12 @@ export {
   priorLookbackOverflow,
   noteLookbackRisk,
   CACHE_LOOKBACK_BLOCKS,
-} from "./cache-telemetry.js";
+} from "./cache/cache-telemetry.js";
 
-export { prepareSystemPrompt, appendBackendSuffix } from "./system-prompt.js";
+export {
+  prepareSystemPrompt,
+  appendBackendSuffix,
+} from "./prompt/system-prompt.js";
 
 export {
   createStreamState,
@@ -71,19 +80,19 @@ export {
   pushLiveUsage,
   finalizeResponseText,
   type StreamState,
-} from "./stream-state.js";
+} from "./turn/stream-state.js";
 
 export {
   routeDelivery,
   buildDeliveryFailureReminder,
   TextBlockDeliveryError,
-} from "./delivery.js";
+} from "./turn/delivery.js";
 
 export { sleep } from "./sleep.js";
 
 export { recordToolCall } from "./metrics.js";
 
-export { applyRetryDecision } from "./handle-retry.js";
+export { applyRetryDecision } from "./turn/handle-retry.js";
 
 export {
   accountTurn,
@@ -92,6 +101,6 @@ export {
   enforceTrailingProse,
   finishCallbackTurn,
   turnUsageSnapshot,
-} from "./turn-phases.js";
+} from "./turn/turn-phases.js";
 
-export { buildResultEvents } from "./result-events.js";
+export { buildResultEvents } from "./turn/result-events.js";

package/src/backend/{shared → runtime/prompt}/delivery-contract.ts

@@ -34,7 +34,7 @@
  * logic-adjacent strings, like log messages.
  */
 
-import { loadSystemTemplate } from "../../core/prompt/templates.js";
+import { loadSystemTemplate } from "../../../core/prompt/templates.js";
 
 // ── Types ───────────────────────────────────────────────────────────────────
 

package/src/backend/{shared → runtime/prompt}/prompt-format.ts

@@ -15,7 +15,7 @@
  *   DM (no msg_id): "[2026-05-15 11:01:23] actual text"
  */
 
-import { formatFullDatetime } from "../../util/time.js";
+import { formatFullDatetime } from "../../../util/time.js";
 
 // ── Public API ──────────────────────────────────────────────────────────────
 
@@ -38,8 +38,33 @@ export type PromptFormatInputs = {
   messageId?: number | string;
   /** When true, omit the leading `[YYYY-MM-DD HH:MM:SS]` tag. */
   omitTimeTag?: boolean;
+  /**
+   * Memory retrieved for THIS turn (`core/memory/turn-retrieval.ts`),
+   * already ranked, trust-filtered and budgeted. Appended after the
+   * message text under a verify-first header.
+   *
+   * This is the **only** place retrieved memory enters a prompt, and
+   * every backend reaches it through this one helper — which is what
+   * keeps #639's divergence (two backends reading the field, four
+   * silently dropping it) from coming back. It belongs to the user
+   * turn: it must never reach `prepareSystemPrompt()`, a prompt
+   * addition, or a backend `system` field, or it would break the
+   * per-session frozen prompt (plan §3.6).
+   *
+   * Absent or blank → the returned prompt is BYTE-IDENTICAL to what
+   * this helper produced before the field existed.
+   */
+  retrievedMemory?: string;
 };
 
+/**
+ * Header on the injected block. "Verify before relying on it" is
+ * deliberate: retrieval is a bm25 guess, not a fact, and the model
+ * should treat a recalled line as a lead rather than as truth.
+ */
+export const RECALLED_MEMORY_HEADER =
+  "[Recalled from memory — verify before relying on it]";
+
 /**
  * Format a user prompt for the AI backend.
  *
@@ -49,6 +74,13 @@ export type PromptFormatInputs = {
  * is on the other end.
  */
 export function formatUserPrompt(inputs: PromptFormatInputs): string {
+  return withRecalledMemory(formatMessageLine(inputs), inputs.retrievedMemory);
+}
+
+// ── Helpers ─────────────────────────────────────────────────────────────────
+
+/** The message line itself — time tag, sender label, `msg_id`, text. */
+function formatMessageLine(inputs: PromptFormatInputs): string {
   const timeTag = inputs.omitTimeTag ? "" : `[${formatFullDatetime()}]`;
   const msgIdHint =
     inputs.messageId !== undefined ? ` [msg_id:${inputs.messageId}]` : "";
@@ -68,7 +100,19 @@ export function formatUserPrompt(inputs: PromptFormatInputs): string {
   return joinNonEmpty(timeTag, inputs.text);
 }
 
-// ── Helpers ─────────────────────────────────────────────────────────────────
+/**
+ * Append the retrieved-memory block AFTER the message, never before:
+ * the user's own words stay the first thing the model reads, and the
+ * recalled lines read as an annotation on them.
+ */
+function withRecalledMemory(
+  prompt: string,
+  memory: string | undefined,
+): string {
+  const block = memory?.trim();
+  if (!block) return prompt;
+  return `${prompt}\n\n${RECALLED_MEMORY_HEADER}\n${block}`;
+}
 
 function joinNonEmpty(...parts: string[]): string {
   return parts

package/src/backend/{shared → runtime/prompt}/system-prompt.ts

@@ -39,10 +39,10 @@ import {
   joinSystemPromptParts,
   type SystemPromptParts,
   type TalonConfig,
-} from "../../core/config/index.js";
-import { getPluginPromptAdditions } from "../../core/plugin/index.js";
-import { onPromptInputsChanged } from "../../core/prompt/invalidation.js";
-import { frontendForChatId, nonTerminalFrontends } from "./frontends.js";
+} from "../../../core/config/index.js";
+import { getPluginPromptAdditions } from "../../../core/plugin/index.js";
+import { onPromptInputsChanged } from "../../../core/prompt/invalidation.js";
+import { frontendForChatId, nonTerminalFrontends } from "../frontends.js";
 
 // ── Types ───────────────────────────────────────────────────────────────────
 

package/src/backend/{shared → runtime/turn}/delivered-text.ts

@@ -25,7 +25,7 @@
  * whether to surface a flow violation.
  */
 
-import { stripMcpPrefix } from "../../core/tools/index.js";
+import { stripMcpPrefix } from "../../../core/tools/index.js";
 
 // ── Constants ───────────────────────────────────────────────────────────────
 

package/src/backend/{shared → runtime/turn}/delivery.ts

@@ -35,8 +35,8 @@
 
 import type { StreamState } from "./stream-state.js";
 import { undeliveredResponseText } from "./stream-state.js";
-import { logWarn } from "../../util/log.js";
-import { incrementCounter } from "../../storage/metrics.js";
+import { logWarn } from "../../../util/log.js";
+import { incrementCounter } from "../../../storage/metrics.js";
 
 /** Route the delivery decision selected. */
 export type DeliveryRoute =

package/src/backend/{shared → runtime/turn}/handle-retry.ts

@@ -23,12 +23,12 @@
  */
 
 import type { QueryParams, QueryResult } from "./handler-types.js";
-import { classify, type TalonError } from "../../core/errors.js";
-import { logWarn } from "../../util/log.js";
-import { incrementCounter } from "../../storage/metrics.js";
-import { resetSession } from "../../storage/sessions.js";
+import { classify, type TalonError } from "../../../core/errors.js";
+import { logWarn } from "../../../util/log.js";
+import { incrementCounter } from "../../../storage/metrics.js";
+import { resetSession } from "../../../storage/sessions.js";
 import { classifyRetry } from "./model-retry.js";
-import type { AgentEvent } from "../../core/agent-runtime/events.js";
+import type { AgentEvent } from "../../../core/agent-runtime/events.js";
 
 /** Inputs for `applyRetryDecision`. */
 export interface ApplyRetryDecisionInputs {