talon-agent 3.7.0 → 3.8.1

package/src/util/config.ts

@@ -5,6 +5,7 @@ import { dirs, files as pathFiles } from "./paths.js";
 import { hardenTalonPermissions } from "./harden.js";
 import { setTimezone } from "./time.js";
 import { BACKEND_IDS } from "../core/agent-runtime/model-ref.js";
+import { REASONING_LEVEL_ORDER } from "../core/models/reasoning-levels.js";
 import {
   assembleSystemPrompt,
   joinSystemPromptParts,
@@ -28,6 +29,20 @@ const BACKEND_ID_ENUM = [...BACKEND_IDS] as [
   ...(typeof BACKEND_IDS)[number][],
 ];
 
+/**
+ * Reasoning-effort literal source for the background-agent knobs
+ * (`heartbeatEffort` / `dreamEffort`). Same trick as `BACKEND_ID_ENUM`:
+ * reuse the single source of truth (`REASONING_LEVEL_ORDER`) so adding a
+ * level to the vocabulary doesn't need a second edit here.
+ *
+ * There is no `"adaptive"` member — leaving the field unset IS adaptive
+ * (the backend/model default), matching how per-chat `effort` behaves.
+ */
+const REASONING_EFFORT_ENUM = [...REASONING_LEVEL_ORDER] as [
+  (typeof REASONING_LEVEL_ORDER)[number],
+  ...(typeof REASONING_LEVEL_ORDER)[number][],
+];
+
 // ── Config schema ───────────────────────────────────────────────────────────
 
 /** Path-based Talon plugin (loaded as a Node module). */
@@ -241,23 +256,6 @@ const mem0SettingsSchema = z.object({
   userId: z.string().min(1).optional(),
 });
 
-/**
- * Memory pre-retrieval (Phase B) — automatic per-turn palace retrieval.
- * Ships INERT: the flag gates wiring a real retriever into the dispatcher.
- * With `enabled: false` (default) prompts are byte-identical to previous
- * behavior. See docs in core/memory/retrieval.ts for the trust policy.
- */
-const memoryPreRetrievalSchema = z.object({
-  /** Master switch. Default off — plumbing merges disabled. */
-  enabled: z.boolean().default(false),
-  /** Maximum retrieved items injected per turn. */
-  maxResults: z.number().int().min(1).max(10).default(3),
-  /** Hard cap on injected characters, provenance labels included. */
-  maxChars: z.number().int().min(200).max(20000).default(3000),
-  /** In groups, filter private-user drawers before injection. */
-  groupPrivateUserFiltering: z.boolean().default(true),
-});
-
 const configSchema = z.object({
   frontend: z.union([frontendEnum, z.array(frontendEnum)]).default("telegram"),
   botToken: z.string().optional(),
@@ -317,6 +315,15 @@ const configSchema = z.object({
    */
   backendDefaults: z.record(z.string(), z.string()).optional(),
   dreamModel: z.string().optional(), // Model used for background memory consolidation (defaults to main model)
+  /**
+   * Reasoning effort for the dream / memory-consolidation agent. Unset =
+   * the backend/model default. Pair with `dreamModel` when you want a
+   * cheap model that still thinks hard (or an expensive one that doesn't).
+   *
+   * Honoured by backends with a reasoning knob (Claude SDK, Codex);
+   * silently ignored by Kilo / OpenCode, which have none.
+   */
+  dreamEffort: z.enum(REASONING_EFFORT_ENUM).optional(),
   maxMessageLength: z.number().int().min(100).default(4000),
   concurrency: z.number().int().min(1).max(20).default(1),
   apiId: z.number().int().optional(),
@@ -327,8 +334,6 @@ const configSchema = z.object({
   pulseIntervalMs: z.number().int().min(60000).default(300000),
   /** Background memory-consolidation (dream) runs. Mirrors `pulse`/`heartbeat`. */
   dream: z.boolean().default(true),
-  /** Memory pre-retrieval (Phase B). Optional; absent = disabled. */
-  memoryPreRetrieval: memoryPreRetrievalSchema.optional(),
   /**
    * Periodic background agent (default: on, hourly). Advances open
    * goals, runs user-defined maintenance, and proactively messages
@@ -339,6 +344,16 @@ const configSchema = z.object({
   heartbeat: z.boolean().default(true),
   heartbeatIntervalMinutes: z.number().int().min(5).default(60),
   heartbeatModel: z.string().optional(), // Model for heartbeat agent (defaults to main model)
+  /**
+   * Reasoning effort for the heartbeat agent. Unset = the backend/model
+   * default. Pair with `heartbeatModel` — e.g. `"high"` so unattended
+   * goal work reasons harder than a chat turn, or `"low"` to keep hourly
+   * runs cheap.
+   *
+   * Honoured by backends with a reasoning knob (Claude SDK, Codex);
+   * silently ignored by Kilo / OpenCode, which have none.
+   */
+  heartbeatEffort: z.enum(REASONING_EFFORT_ENUM).optional(),
   braveApiKey: z.string().optional(),
   /**
    * Codex-specific OpenAI API key. Prefer this, CODEX_API_KEY, or

package/src/core/memory/retrieval.ts

@@ -1,92 +0,0 @@
-/**
- * Memory pre-retrieval (Phase B) — injectable retriever boundary.
- *
- * Phase A made `memory.md` the compact always-loaded layer; Phase B closes the
- * recall gap by handing each chat turn a small, relevant palace slice without
- * depending on the model to call `mempalace_search` itself. This module owns
- * the CORE side of that boundary: a typed retriever dependency the Weaver can
- * call before `runChatTurn(...)`, returning plain data
- * (`RetrievedMemory | undefined`) that backends fold into the live user
- * prompt via `formatPromptWithRetrievedMemory`.
- *
- * Deliberate constraints (see docs/memory-phase-b-pre-retrieval.md):
- *
- *   - The retriever receives a small context object and returns plain data.
- *     It never imports backend handlers, and retrieval output never enters
- *     `prepareSystemPrompt()` / frozen prompt snapshots — cache safety first.
- *   - Fail closed: a broken retriever must never block chat delivery. The
- *     Weaver catches errors, logs one warning, and runs the turn without
- *     injected memory.
- *   - The PRODUCTION retriever is intentionally absent in this first pass.
- *     There is no clean in-process MemPalace call surface yet (the plugin is
- *     an MCP server, not a typed module), and shelling out per message from
- *     the dispatch hot path is explicitly forbidden. Until a deliberate
- *     bridge exists, deployments get `noopMemoryRetriever` — plumbing lands
- *     inert, prompts stay byte-identical.
- *   - Trust-aware injection (#373): when a real adapter lands, only items
- *     with `trustLevel` in `AUTO_INJECT_TRUST_LEVELS` may be auto-injected.
- *     `user_claim` / `group_chat` content stays pull-only, or a poisoned
- *     drawer becomes a persistent prompt injection in every session.
- */
-
-import type {
-  RetrievedMemory,
-  RetrievedMemoryTrustLevel,
-} from "../agent-runtime/capabilities.js";
-
-// ── Retriever contract ──────────────────────────────────────────────────────
-
-/** Context handed to a retriever for one chat turn. */
-export type MemoryRetrievalContext = {
-  runKind: "chat";
-  chatId: string;
-  /** The raw incoming message text (retrievers trim/cap it themselves). */
-  text: string;
-  senderName: string;
-  isGroup?: boolean;
-};
-
-/**
- * A memory retriever: small context in, bounded plain data out.
- * `undefined` means "inject nothing" — the turn proceeds unchanged.
- */
-export type MemoryRetriever = (
-  context: MemoryRetrievalContext,
-) => Promise<RetrievedMemory | undefined>;
-
-// ── Trust policy (#373) ─────────────────────────────────────────────────────
-
-/**
- * Trust levels eligible for automatic injection. Anything else (including an
- * absent `trustLevel`) must be filtered by adapters before returning items.
- */
-export const AUTO_INJECT_TRUST_LEVELS: ReadonlySet<RetrievedMemoryTrustLevel> =
-  new Set(["dylan_direct", "bot_inferred", "heartbeat_synthesis"]);
-
-/** True when an item's trust level allows automatic injection. */
-export function isAutoInjectTrusted(
-  trustLevel: RetrievedMemoryTrustLevel | undefined,
-): boolean {
-  return trustLevel !== undefined && AUTO_INJECT_TRUST_LEVELS.has(trustLevel);
-}
-
-/**
- * Enforce the #373 trust policy on a retrieval result: drop every item that
- * is not explicitly auto-inject trusted. Adapters should call this as their
- * last step, and the Weaver's `prefetchMemory` applies it again to whatever
- * a retriever returns — a buggy adapter cannot leak low-trust items into
- * the prompt.
- */
-export function filterAutoInjectable(
-  memory: RetrievedMemory | undefined,
-): RetrievedMemory | undefined {
-  if (!memory) return undefined;
-  const items = memory.items.filter((i) => isAutoInjectTrusted(i.trustLevel));
-  if (items.length === 0) return undefined;
-  return { ...memory, items };
-}
-
-// ── Default (inert) retriever ───────────────────────────────────────────────
-
-/** The no-op retriever: Phase B plumbing present, injection disabled. */
-export const noopMemoryRetriever: MemoryRetriever = async () => undefined;

package/src/core/weaver/memory-prefetch.ts

@@ -1,65 +0,0 @@
-/**
- * Memory prefetch (Phase B) — optional pre-retrieval of palace memory
- * for live user messages, strictly fail-closed: a broken palace must
- * never block chat delivery. The result is dynamic turn context passed
- * through ChatRunParams; it never touches the frozen prompt.
- *
- * The #373 trust policy is enforced HERE, not just in adapters: whatever a
- * retriever returns is passed through `filterAutoInjectable` before it can
- * reach the prompt, so a buggy or future adapter cannot leak `user_claim` /
- * `group_chat` items into auto-injection.
- */
-
-import type { RetrievedMemory } from "../agent-runtime/capabilities.js";
-import type { MemoryRetriever } from "../memory/retrieval.js";
-import { filterAutoInjectable } from "../memory/retrieval.js";
-import { logDebug, logWarn } from "../../util/log.js";
-
-export type PrefetchMemoryInput = {
-  chatId: string;
-  text: string;
-  senderName: string;
-  isGroup: boolean;
-  /** Request id for log correlation. */
-  reqId: string;
-};
-
-export async function prefetchMemory(
-  retrieve: MemoryRetriever,
-  input: PrefetchMemoryInput,
-): Promise<RetrievedMemory | undefined> {
-  try {
-    const raw = await retrieve({
-      runKind: "chat",
-      chatId: input.chatId,
-      text: input.text,
-      senderName: input.senderName,
-      isGroup: input.isGroup,
-    });
-    const retrieved = filterAutoInjectable(raw);
-    const droppedCount =
-      (raw?.items.length ?? 0) - (retrieved?.items.length ?? 0);
-    if (droppedCount > 0) {
-      logWarn(
-        "dispatcher",
-        `[${input.reqId}] memory pre-retrieval: dropped ${droppedCount} ` +
-          `low-trust item(s) the retriever failed to filter (#373)`,
-      );
-    }
-    if (retrieved) {
-      logDebug(
-        "dispatcher",
-        `[${input.reqId}] memory pre-retrieval: ${retrieved.items.length} item(s), ` +
-          `${retrieved.items.reduce((n, i) => n + i.text.length, 0)} chars`,
-      );
-    }
-    return retrieved;
-  } catch (err) {
-    logWarn(
-      "dispatcher",
-      `[${input.reqId}] memory pre-retrieval failed (running turn without it): ` +
-        (err instanceof Error ? err.message : String(err)),
-    );
-    return undefined;
-  }
-}