feat: add lane config option for embedded agent isolation

README.md

@@ -278,6 +278,10 @@ docker exec -it hermes-memory hermes
 | `pipeline.l2MinIntervalSeconds` | `900` | Minimum interval between two L2 passes within the same session |
 | `recall.timeoutMs` | `5000` | Recall timeout; on timeout, skip injection without blocking the conversation |
 | `extraction.enableDedup` | `true` | L1 vector dedup / conflict detection |
+| `extraction.model` | *(default model)* | LLM model for L1 extraction (format: `provider/model`) |
+| `extraction.lane` | *(main lane)* | Global lane for extraction runs (e.g. `"cron"` to use `"cron-nested"` lane) |
+| `persona.model` | *(default model)* | LLM model for L2 scene extraction & L3 persona generation |
+| `persona.lane` | *(main lane)* | Global lane for persona/scene runs |
 | `capture.excludeAgents` | `[]` | Glob patterns to exclude specific agents (e.g. `bench-judge-*`) |
 | `capture.l0l1RetentionDays` | `0` | Local retention days for L0 / L1 files; `0` = never clean up |
 | `offload.mildOffloadRatio` | `0.5` | Mild compression trigger ratio (of context window) |

src/core/persona/persona-generator.ts

@@ -36,6 +36,8 @@ export class PersonaGenerator {
     config: unknown;
     model?: string;
     backupCount?: number;
+    /** Global lane name for the embedded agent run. */
+    lane?: string;
     logger?: Logger;
     /** Plugin instance ID for metric reporting (optional) */
     instanceId?: string;
@@ -54,6 +56,7 @@ export class PersonaGenerator {
     this.runner = opts.llmRunner ?? new CleanContextRunner({
       config: opts.config,
       modelRef: opts.model,
+      lane: opts.lane,
       enableTools: true,
       logger: opts.logger,
     });

src/core/record/l1-dedup.ts

@@ -68,10 +68,12 @@ export async function batchDedup(params: {
   conflictRecallTopK?: number;
   /** Override embedding timeout for capture-path calls (milliseconds) */
   embeddingTimeoutMs?: number;
+  /** Global lane name for the embedded agent run. */
+  lane?: string;
   /** Host-neutral LLM runner — when provided, used instead of CleanContextRunner. */
   llmRunner?: LLMRunner;
 }): Promise<DedupDecision[]> {
-  const { memories, config, logger, model, vectorStore, embeddingService, llmRunner } = params;
+  const { memories, config, logger, model, vectorStore, embeddingService, llmRunner, lane } = params;
   const topK = params.conflictRecallTopK ?? 5;
 
   if (memories.length === 0) {
@@ -139,7 +141,7 @@ export async function batchDedup(params: {
   }
 
   // Phase 2: Batch LLM judgment
-  return runLlmJudgment(matches, memories, config, logger, model, llmRunner);
+  return runLlmJudgment(matches, memories, config, logger, model, lane, llmRunner);
 }
 
 /**
@@ -151,6 +153,7 @@ async function runLlmJudgment(
   config: unknown,
   logger: Logger | undefined,
   model: string | undefined,
+  lane: string | undefined,
   llmRunner?: LLMRunner,
 ): Promise<DedupDecision[]> {
   logger?.debug?.(`${TAG} Running batch conflict detection for ${memories.length} memories`);
@@ -172,6 +175,7 @@ async function runLlmJudgment(
       const runner = new CleanContextRunner({
         config,
         modelRef: model,
+        lane,
         enableTools: false,
         logger,
       });

src/core/record/l1-extractor.ts

@@ -98,6 +98,8 @@ export async function extractL1Memories(params: {
     model?: string;
     /** Previous scene name for continuity */
     previousSceneName?: string;
+    /** Global lane name for the embedded agent run. */
+    lane?: string;
     /** Vector store for cosine similarity candidate recall */
     vectorStore?: IMemoryStore;
     /** Embedding service for computing query vectors */
@@ -165,6 +167,7 @@ export async function extractL1Memories(params: {
       config,
       logger,
       model: options.model,
+      lane: options.lane,
       llmRunner: options.llmRunner,
     });
     logger?.debug?.(`${TAG} LLM detected ${scenes.length} scene(s)`);
@@ -307,6 +310,8 @@ async function callLlmExtraction(params: {
   config: unknown;
   logger?: Logger;
   model?: string;
+  /** Global lane name for the embedded agent run. */
+  lane?: string;
   /** Host-neutral LLM runner — when provided, used instead of CleanContextRunner. */
   llmRunner?: LLMRunner;
 }): Promise<SceneSegment[]> {
@@ -338,6 +343,7 @@ async function callLlmExtraction(params: {
     const runner = new CleanContextRunner({
       config,
       modelRef: model,
+      lane: params.lane,
       enableTools: false,
       logger,
     });

src/core/scene/scene-extractor.ts

@@ -54,6 +54,11 @@ export interface SceneExtractorOptions {
   logger?: ExtractorLogger;
   /** Plugin instance ID for metric reporting (optional) */
   instanceId?: string;
+  /**
+   * Global lane name for the embedded agent run.
+   * @default undefined → OpenClaw defaults to "main"
+   */
+  lane?: string;
   /**
    * Host-neutral LLM runner. When provided, used instead of creating
    * a CleanContextRunner (decouples from OpenClaw runtime).
@@ -106,6 +111,7 @@ export class SceneExtractor {
     this.runner = opts.llmRunner ?? new CleanContextRunner({
       config: opts.config,
       modelRef: opts.model,
+      lane: opts.lane,
       enableTools: true,
       logger: opts.logger,
     });

src/utils/clean-context-runner.ts

@@ -285,6 +285,18 @@ export interface CleanContextRunnerOptions {
   agentRuntime?: EmbeddedAgentRuntimeLike;
   /** Allow the LLM to use tools (read_file, write_to_file, etc). Default: false */
   enableTools?: boolean;
+  /**
+   * Global lane name for the embedded agent run.
+   *
+   * OpenClaw routes embedded runs through a session lane (one-at-a-time per session)
+   * and a global lane (shared parallelism pool). By default the global lane is `"main"`,
+   * which is also used by the primary chat agent. Setting this to `"cron"` (→ `"cron-nested"`)
+   * or another named lane prevents long-running extraction calls from competing with
+   * interactive chat turns.
+   *
+   * @default undefined → OpenClaw defaults to `"main"`
+   */
+  lane?: string;
   /** Logger instance for detailed tracing */
   logger?: RunnerLogger;
 }
@@ -459,6 +471,9 @@ export class CleanContextRunner {
         // providers (qwencode) reject with "[] is too short - 'tools'".
         // Instead rely on cleanConfig.tools.allow to restrict the tool set
         // to a minimal read-only tool (when enableTools=false).
+        // Pass lane to use a dedicated global lane (e.g. "cron" → "cron-nested")
+        // instead of the default "main" lane shared with interactive sessions.
+        lane: this.options.lane,
         disableTools: false,
         extraSystemPrompt: effectiveSystemPrompt,
         streamParams: {