@animalabs/connectome-host 0.7.3 → 0.7.4

package/src/recipe.ts

@@ -95,6 +95,11 @@ export interface RecipeStrategy {
   /** Override wording for the witnessed-record instruction ({targetTokens}
    *  substituted). */
   witnessedInstruction?: string;
+  /** Identity reminder appended to every compression/merge instruction.
+   *  For agents in multi-resident channels (and older models especially):
+   *  names the agent and directs attribution so pure-witness chunks don't
+   *  flip the summarizer into another speaker's identity. */
+  identityReminder?: string;
 }
 
 export interface RecipeAgent {
@@ -421,10 +426,19 @@ export interface RecipeModules {
   /**
    * Lesson retrieval-injection (requires `lessons`). OPT-IN — defaults to off
    * and is deliberately not part of the standard recipe: it injects
-   * context-dependent content into every compile and spends two Haiku calls
-   * per turn. Enable only for agents that actually curate a lesson library.
+   * context-dependent content into every compile and spends up to two
+   * configured retrieval-model calls. Enable only for agents that actually
+   * curate a lesson library.
    */
-  retrieval?: boolean | { model?: string; maxInjected?: number };
+  retrieval?: boolean | {
+    model?: string;
+    maxInjected?: number;
+    /**
+     * Optional OpenAI Responses/Codex reasoning effort for both retrieval calls.
+     * Requires an explicit retrieval model.
+     */
+    reasoningEffort?: 'none' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'max';
+  };
   wake?: boolean | import('@animalabs/agent-framework').GateConfig;
   workspace?: boolean | { mounts: RecipeWorkspaceMount[]; configMount?: boolean };
   /**
@@ -718,7 +732,7 @@ export const DEFAULT_RECIPE: Recipe = {
   },
   modules: {
     // subagents + lessons + retrieval deliberately omitted — all opt-in only
-    // (retrieval additionally adds per-turn context churn + Haiku costs);
+    // (retrieval additionally adds per-turn context churn + retrieval-model costs);
     // see the RecipeModules field docs.
     wake: true,
     workspace: true,
@@ -1255,6 +1269,43 @@ export function validateRecipe(raw: unknown): Recipe {
       }
     }
 
+    // Validate retrieval provider reasoning when configured.
+    const retrieval = mods.retrieval;
+    if (retrieval !== undefined && typeof retrieval !== 'boolean') {
+      if (!retrieval || typeof retrieval !== 'object' || Array.isArray(retrieval)) {
+        throw new Error('Recipe modules.retrieval must be a boolean or object.');
+      }
+      const retrievalConfig = retrieval as Record<string, unknown>;
+      if (retrievalConfig.reasoningContext !== undefined) {
+        throw new Error(
+          'modules.retrieval.reasoningContext is not supported: retrieval model calls ' +
+          'are independent one-shot requests with no earlier reasoning items.',
+        );
+      }
+      const efforts = ['none', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'];
+      if (retrievalConfig.reasoningEffort !== undefined &&
+          (typeof retrievalConfig.reasoningEffort !== 'string'
+            || !efforts.includes(retrievalConfig.reasoningEffort))) {
+        throw new Error(`Invalid modules.retrieval.reasoningEffort ${JSON.stringify(retrievalConfig.reasoningEffort)}.`);
+      }
+      if (retrievalConfig.reasoningEffort !== undefined
+          && agent.provider !== 'openai-responses'
+          && agent.provider !== 'openai-codex') {
+        throw new Error(
+          'modules.retrieval.reasoningEffort requires agent.provider ' +
+          '"openai-responses" or "openai-codex".',
+        );
+      }
+      if (retrievalConfig.reasoningEffort !== undefined
+          && (typeof retrievalConfig.model !== 'string'
+            || !retrievalConfig.model.trim())) {
+        throw new Error(
+          'modules.retrieval.model must be a non-empty string when ' +
+          'modules.retrieval.reasoningEffort is configured.',
+        );
+      }
+    }
+
     // Validate ttsRelay if present — url + token are load-bearing, and a
     // recipe that names the module but can't reach a relay should fail at
     // load, not silently stream nowhere.

package/src/retrieval-config.ts

@@ -0,0 +1,39 @@
+import type { Membrane } from '@animalabs/membrane';
+import type { RecipeAgent, RecipeModules } from './recipe.js';
+import type { RetrievalModuleConfig } from './modules/retrieval-module.js';
+
+type RetrievalRecipeConfig = Exclude<RecipeModules['retrieval'], boolean | undefined>;
+
+/** Translate the recipe's retrieval block into the module's runtime config. */
+export function buildRetrievalModuleConfig(
+  membrane: Membrane,
+  retrieval: RecipeModules['retrieval'],
+  provider: RecipeAgent['provider'] = 'anthropic',
+): RetrievalModuleConfig {
+  const config: RetrievalRecipeConfig = typeof retrieval === 'object' ? retrieval : {};
+  if (config.reasoningEffort
+      && provider !== 'openai-responses'
+      && provider !== 'openai-codex') {
+    throw new Error(
+      'modules.retrieval.reasoningEffort requires agent.provider ' +
+      '"openai-responses" or "openai-codex".',
+    );
+  }
+  if (config.reasoningEffort
+      && (typeof config.model !== 'string' || !config.model.trim())) {
+    throw new Error(
+      'modules.retrieval.model must be a non-empty string when ' +
+      'modules.retrieval.reasoningEffort is configured.',
+    );
+  }
+  const retrievalReasoning = config.reasoningEffort
+    ? { effort: config.reasoningEffort }
+    : undefined;
+
+  return {
+    membrane,
+    retrievalModel: config.model,
+    retrievalReasoning,
+    maxInjectedLessons: config.maxInjected,
+  };
+}

package/src/strategies/frontdesk-strategy.ts

@@ -1,30 +1,16 @@
 import { AutobiographicalStrategy } from '@animalabs/context-manager';
 import type {
   AutobiographicalConfig,
+  Chunk,
   ContextEntry,
   MessageStoreView,
   ContextLogView,
   TokenBudget,
   StoredMessage,
-  SummaryEntry,
 } from '@animalabs/context-manager';
 import type { ContentBlock } from '@animalabs/membrane';
 import { formatZonedTime, resolveTimeZone } from '@animalabs/agent-framework';
 
-// Structural mirror of AutobiographicalStrategy's internal Chunk.
-// Kept inline because @animalabs/context-manager does not currently export it.
-interface Chunk {
-  index: number;
-  startIndex: number;
-  endIndex: number;
-  messages: StoredMessage[];
-  tokens: number;
-  compressed: boolean;
-  diary?: string;
-  summaryId?: string;
-  phaseType?: string;
-}
-
 export type FrontdeskStrategyOptions = Partial<AutobiographicalConfig> & { timeZone?: string };
 
 /**
@@ -35,10 +21,22 @@ export type FrontdeskStrategyOptions = Partial<AutobiographicalConfig> & { timeZ
  *  1. Provenance wrapping — prepends a `[zulip · #channel · topic · @user · HH:MM · msg-id]`
  *     header to each MCPL-originated entry so the agent knows the message came from a
  *     channel and which reply path to use.
- *  2. Topic-aware compression — chunk boundaries prefer Zulip-topic transitions and the
- *     compression prompt instructs per-topic structure.
- *  3. Question/mention salience — unanswered user questions and @mentions are preserved
- *     verbatim longer (both during compression and during L1 selection under budget).
+ *  2. Topic-aware compression — chunk boundaries close at Zulip-topic transitions
+ *     (via the base chunker's `chunkBoundaryHint` seam) and the compression prompt
+ *     instructs per-topic structure.
+ *  3. Question/mention salience — unanswered user questions and @mentions are named
+ *     verbatim in the compression prompt so summaries preserve them.
+ *
+ * History note: through conhost 0.7.x this class forked the whole of
+ * `rebuildChunks` for feature 2 — written against a pre-chunk-persistence
+ * base, silently bypassing chunk records and the fail-closed orphan guard —
+ * and biased the hierarchical renderer's L1 selection for feature 3.
+ * Frontdesk agents now ride the adaptive path (see framework-strategy.ts
+ * defaults): chunking goes through the base implementation with a boundary
+ * hint, and salient content survives through the compression prompt rather
+ * than selection-order bias. Stores created by the fork carry no chunk
+ * records; context-manager's `migrateChunkRecords` backfills them from L1
+ * `sourceIds` on first load.
  */
 export class FrontdeskStrategy extends AutobiographicalStrategy {
   override readonly name: string = 'frontdesk';
@@ -162,77 +160,13 @@ export class FrontdeskStrategy extends AutobiographicalStrategy {
   // ==========================================================================
 
   /**
-   * Override rebuildChunks to additionally close chunk boundaries at Zulip-topic
-   * transitions. Falls back to base size/count behaviour when topic metadata is absent.
+   * Close chunk boundaries at Zulip-topic transitions, so summaries of
+   * unrelated topics are not fused. Rides the base chunker (context-manager
+   * ≥0.6.3): record persistence, minimum-size and tool-pairing guards all
+   * apply to hinted closes.
    */
-  protected override rebuildChunks(store: MessageStoreView): void {
-    const messagesToChunk = this.getCompressibleMessages(store);
-
-    const existingCompressed = new Map<string, Chunk>();
-    for (const chunk of this.chunks as unknown as Chunk[]) {
-      if (chunk.compressed) {
-        existingCompressed.set(this.chunkKey(chunk as never), chunk);
-      }
-    }
-
-    this.chunks = [];
-    this.compressionQueue = [];
-
-    let currentChunk: StoredMessage[] = [];
-    let currentTokens = 0;
-    let chunkFilteredStart = 0;
-    const MIN_CHUNK = 4;
-
-    const push = (startIdx: number, endIdx: number, msgs: StoredMessage[], tokens: number) => {
-      const chunk = this.createChunk(
-        this.chunks.length,
-        startIdx,
-        endIdx,
-        msgs,
-        tokens,
-        existingCompressed as never,
-      );
-      this.chunks.push(chunk);
-      if (!chunk.compressed) this.compressionQueue.push(chunk.index);
-    };
-
-    for (let i = 0; i < messagesToChunk.length; i++) {
-      const msg = messagesToChunk[i];
-      let msgTokens = store.estimateTokens(msg);
-      if (this.config.attachmentsIgnoreSize) {
-        msgTokens = this.estimateTextOnlyTokens(msg);
-      }
-
-      // Topic boundary: close current chunk BEFORE adding msg when topic changes
-      // and the chunk has at least MIN_CHUNK messages. This keeps summaries of
-      // unrelated topics from being merged.
-      if (
-        currentChunk.length >= MIN_CHUNK &&
-        this.isTopicBoundary(currentChunk[currentChunk.length - 1], msg)
-      ) {
-        push(chunkFilteredStart, i, currentChunk, currentTokens);
-        currentChunk = [];
-        currentTokens = 0;
-        chunkFilteredStart = i;
-      }
-
-      currentChunk.push(msg);
-      currentTokens += msgTokens;
-
-      const shouldClose =
-        currentTokens >= this.config.targetChunkTokens && currentChunk.length >= MIN_CHUNK;
-
-      if (shouldClose) {
-        push(chunkFilteredStart, i + 1, currentChunk, currentTokens);
-        currentChunk = [];
-        currentTokens = 0;
-        chunkFilteredStart = i + 1;
-      }
-    }
-
-    if (currentChunk.length >= MIN_CHUNK) {
-      push(chunkFilteredStart, messagesToChunk.length, currentChunk, currentTokens);
-    }
+  protected override chunkBoundaryHint(prev: StoredMessage, next: StoredMessage): boolean {
+    return this.isTopicBoundary(prev, next);
   }
 
   protected isTopicBoundary(prev: StoredMessage, curr: StoredMessage): boolean {
@@ -255,6 +189,11 @@ export class FrontdeskStrategy extends AutobiographicalStrategy {
   // ==========================================================================
 
   protected override getCompressionInstruction(chunk: Chunk, targetTokens: number): string {
+    // Witnessed chunks keep the base treatment (recipe-configurable
+    // witnessed prompt); the old fork predated it and steamrolled it.
+    if (this.chunkIsWitnessed(chunk)) {
+      return super.getCompressionInstruction(chunk, targetTokens);
+    }
     const topics = new Set<string>();
     for (const m of chunk.messages) {
       const t = this.extractTopicKey(m);
@@ -289,43 +228,13 @@ export class FrontdeskStrategy extends AutobiographicalStrategy {
   }
 
   // ==========================================================================
-  // Feature 3b: Salience-biased L1 selection
-  // ==========================================================================
-
-  protected override selectL1Summaries(
-    shownL1: SummaryEntry[],
-    budget: number,
-    maxTokens: number,
-  ): { selected: SummaryEntry[]; tokensUsed: number } {
-    if (shownL1.length === 0) return { selected: [], tokensUsed: 0 };
-
-    const isSalient = (s: SummaryEntry): boolean =>
-      s.sourceIds.some((id) => this.salientSourceIds.has(id));
-
-    const salient: SummaryEntry[] = [];
-    const routine: SummaryEntry[] = [];
-    for (const s of shownL1) {
-      (isSalient(s) ? salient : routine).push(s);
-    }
-
-    const selected: SummaryEntry[] = [];
-    let used = 0;
-
-    for (const group of [salient, routine]) {
-      for (const s of group) {
-        if (used + s.tokens > budget) break;
-        if (used + s.tokens > maxTokens) break;
-        selected.push(s);
-        used += s.tokens;
-      }
-    }
-
-    return { selected, tokensUsed: used };
-  }
-
-  // ==========================================================================
-  // Salience tracking (shared by 3a and 3b)
+  // Salience tracking (feeds the compression instruction)
   // ==========================================================================
+  // The pre-adaptive frontdesk also overrode the hierarchical renderer's
+  // selectL1Summaries to emit salient L1s first under budget pressure. Under
+  // adaptive resolution the picker selects a coverage frontier — emission
+  // order is not budget-competitive — so that bias is retired; salient
+  // content survives because getCompressionInstruction names it verbatim.
 
   /**
    * Recompute which user messages are "unanswered questions or mentions":