@kolisachint/hoocode-agent 0.4.71 → 0.4.73

package/dist/core/routing/local-inference.d.ts

@@ -2,7 +2,7 @@
  * Local-inference routing.
  *
  * Optional, opt-in routing of certain non-critical work (conversation
- * compaction, and large read/bash tool-result compression) to a local
+ * compaction, and large bash tool-result compression) to a local
  * "executor" model running on an OpenAI-compatible endpoint (for example an
  * MLX server), while the primary model handles all planning, reasoning, edits,
  * and tool-call synthesis.
@@ -21,10 +21,23 @@ import type { ModelRegistry } from "../model-registry.js";
 export type TurnKind = "primary" | "summarization" | "tool-result";
 export type RoutingMode = "primary-only" | "executor-for-summarization" | "executor-for-tool-results" | "shadow-executor";
 export declare const ROUTING_MODES: readonly RoutingMode[];
-/** Tool names whose output is worth compressing (validated). Others pass through. */
+/**
+ * Tool names whose output is worth compressing (validated). Others pass
+ * through. Only `bash` qualifies: its verbose output is mostly low-value noise
+ * around a few load-bearing facts. `read` was measured to compress ~0% on real
+ * source code (every line is a keep-line) and was removed. Fact-list tools
+ * (grep/find/ls) were never compressible (every line is a distinct fact).
+ */
 export declare const COMPRESSIBLE_TOOLS: Set<string>;
-/** Default minimum tool-result size (bytes) before compression is attempted. */
-export declare const DEFAULT_TOOL_RESULT_MIN_BYTES = 2048;
+/**
+ * Global size band (bytes) for local-inference routing. Applies to BOTH
+ * tool-result compression and compaction summarization. Inputs below the
+ * minimum are not worth offloading; inputs above the maximum are slow and risk
+ * GPU OOM on small machines, so they fall back to the primary model. Tunable
+ * per-machine via `minBytes`/`maxBytes` in the executor config block.
+ */
+export declare const DEFAULT_MIN_BYTES = 2048;
+export declare const DEFAULT_MAX_BYTES = 8192;
 /** Optional local server the harness manages for the executor. */
 export interface ExecutorServerConfig {
     /** Command to launch (default: "mlx_lm.server"). */
@@ -42,8 +55,10 @@ export interface ExecutorServerConfig {
 export interface ExecutorConfig {
     provider: string;
     model: string;
-    /** Minimum tool-result size (bytes) before compression is attempted. */
-    toolResultMinBytes?: number;
+    /** Minimum input size (bytes) before local inference is attempted. */
+    minBytes?: number;
+    /** Maximum input size (bytes); larger inputs fall back to the primary model. */
+    maxBytes?: number;
     /** When set, the harness spawns/health-checks/stops this local server. */
     server?: ExecutorServerConfig;
 }
@@ -74,7 +89,8 @@ export declare class LocalInferenceRouter {
     private readonly mode;
     private readonly executorConfig;
     private readonly executor;
-    private readonly toolResultMinBytes;
+    private readonly minBytes;
+    private readonly maxBytes;
     private constructor();
     static create(opts: {
         mode: RoutingMode;
@@ -92,8 +108,22 @@ export declare class LocalInferenceRouter {
      * executor is exercised separately for measurement).
      */
     selectModel(turnKind: TurnKind, primary: Model<Api>): Model<Api>;
+    /** True when an input size falls within the configured local-inference band. */
+    withinSizeBand(bytes: number): boolean;
+    /** The configured size band, for logging/diagnostics. */
+    getSizeBand(): {
+        minBytes: number;
+        maxBytes: number;
+    };
     /** Whether a given tool's result should be compressed via the executor. */
     shouldCompressToolResult(toolName: string, contentBytes: number): boolean;
+    /**
+     * Whether to route summarization to the executor for a conversation of the
+     * given serialized size. Requires summarization routing active, the executor
+     * available, and the size within the band (oversized conversations fall back
+     * to the primary to avoid slow local runs and GPU OOM).
+     */
+    shouldRouteSummarization(bytes: number): boolean;
     getExecutorModel(): Model<Api> | undefined;
 }
 //# sourceMappingURL=local-inference.d.ts.map

package/dist/core/routing/local-inference.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"local-inference.d.ts","sourceRoot":"","sources":["../../../src/core/routing/local-inference.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,MAAM,yBAAyB,CAAC;AAC1D,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,sBAAsB,CAAC;AAE1D,4EAA4E;AAC5E,MAAM,MAAM,QAAQ,GAAG,SAAS,GAAG,eAAe,GAAG,aAAa,CAAC;AAEnE,MAAM,MAAM,WAAW,GACpB,cAAc,GACd,4BAA4B,GAC5B,2BAA2B,GAC3B,iBAAiB,CAAC;AAErB,eAAO,MAAM,aAAa,EAAE,SAAS,WAAW,EAKtC,CAAC;AAEX,qFAAqF;AACrF,eAAO,MAAM,kBAAkB,aAA4B,CAAC;AAE5D,gFAAgF;AAChF,eAAO,MAAM,6BAA6B,OAAO,CAAC;AAElD,kEAAkE;AAClE,MAAM,WAAW,oBAAoB;IACpC,oDAAoD;IACpD,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,iDAAiD;IACjD,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC;IAChB,2FAA2F;IAC3F,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,sFAAsF;IACtF,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,kFAAkF;IAClF,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,6DAA6D;AAC7D,MAAM,WAAW,cAAc;IAC9B,QAAQ,EAAE,MAAM,CAAC;IACjB,KAAK,EAAE,MAAM,CAAC;IACd,wEAAwE;IACxE,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,0EAA0E;IAC1E,MAAM,CAAC,EAAE,oBAAoB,CAAC;CAC9B;AAED,sCAAsC;AACtC,MAAM,WAAW,aAAa;IAC7B,IAAI,CAAC,EAAE,WAAW,CAAC;IACnB,QAAQ,CAAC,EAAE,cAAc,CAAC;CAC1B;AAMD;;;;;;;GAOG;AACH,wBAAgB,kBAAkB,CAAC,IAAI,EAAE;IACxC,UAAU,CAAC,EAAE,OAAO,CAAC;IACrB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,UAAU,CAAC,EAAE,WAAW,CAAC;CACzB,GAAG,WAAW,CASd;AAED;;;;GAIG;AACH,qBAAa,oBAAoB;IAChC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAc;IACnC,OAAO,CAAC,QAAQ,CAAC,cAAc,CAA6B;IAC5D,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAyB;IAClD,OAAO,CAAC,QAAQ,CAAC,kBAAkB,CAAS;IAE5C,OAAO,eASN;IAED,MAAM,CAAC,MAAM,CAAC,IAAI,EAAE;QACnB,IAAI,EAAE,WAAW,CAAC;QAClB,MAAM,EAAE,aAAa,GAAG,SAAS,CAAC;QAClC,QAAQ,EAAE,aAAa,CAAC;KACxB,GAAG,oBAAoB,CAOvB;IAED,OAAO,IAAI,WAAW,CAErB;IAED,gFAAgF;IAChF,mBAAmB,IAAI,OAAO,CAE7B;IAED,iBAAiB,IAAI,cAAc,GAAG,SAAS,CAE9C;IAED;;;;;OAKG;IACH,WAAW,CAAC,QAAQ,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC,GAAG,CAAC,CAW/D;IAED,2EAA2E;IAC3E,wBAAwB,CAAC,QAAQ,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,GAAG,OAAO,CAKxE;IAED,gBAAgB,IAAI,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CAEzC;CACD","sourcesContent":["/**\n * Local-inference routing.\n *\n * Optional, opt-in routing of certain non-critical work (conversation\n * compaction, and large read/bash tool-result compression) to a local\n * \"executor\" model running on an OpenAI-compatible endpoint (for example an\n * MLX server), while the primary model handles all planning, reasoning, edits,\n * and tool-call synthesis.\n *\n * Everything here is INERT unless explicitly enabled via the\n * `--enable-local-inference` flag or the `HOOCODE_ROUTING_MODE` env var. On any\n * executor resolution/availability problem the caller falls back to the primary\n * model (compaction) or the raw tool result (tool-result compression). The\n * router never throws into the agent loop.\n *\n * Design and validation: see docs/local-executor-routing.md.\n */\n\nimport type { Api, Model } from \"@kolisachint/hoocode-ai\";\nimport type { ModelRegistry } from \"../model-registry.js\";\n\n/** Work that may be routed to the executor instead of the primary model. */\nexport type TurnKind = \"primary\" | \"summarization\" | \"tool-result\";\n\nexport type RoutingMode =\n\t| \"primary-only\"\n\t| \"executor-for-summarization\"\n\t| \"executor-for-tool-results\"\n\t| \"shadow-executor\";\n\nexport const ROUTING_MODES: readonly RoutingMode[] = [\n\t\"primary-only\",\n\t\"executor-for-summarization\",\n\t\"executor-for-tool-results\",\n\t\"shadow-executor\",\n] as const;\n\n/** Tool names whose output is worth compressing (validated). Others pass through. */\nexport const COMPRESSIBLE_TOOLS = new Set([\"read\", \"bash\"]);\n\n/** Default minimum tool-result size (bytes) before compression is attempted. */\nexport const DEFAULT_TOOL_RESULT_MIN_BYTES = 2048;\n\n/** Optional local server the harness manages for the executor. */\nexport interface ExecutorServerConfig {\n\t/** Command to launch (default: \"mlx_lm.server\"). */\n\tcommand?: string;\n\t/** Extra args appended to the launch command. */\n\targs?: string[];\n\t/** Host to health-check and bind (default: derived from executor baseUrl or 127.0.0.1). */\n\thost?: string;\n\t/** Port to health-check and bind (default: derived from executor baseUrl or 8080). */\n\tport?: number;\n\t/** Max milliseconds to wait for the server to become healthy (default: 30000). */\n\tstartupTimeoutMs?: number;\n}\n\n/** Executor model reference as configured in models.json. */\nexport interface ExecutorConfig {\n\tprovider: string;\n\tmodel: string;\n\t/** Minimum tool-result size (bytes) before compression is attempted. */\n\ttoolResultMinBytes?: number;\n\t/** When set, the harness spawns/health-checks/stops this local server. */\n\tserver?: ExecutorServerConfig;\n}\n\n/** `routing` block in models.json. */\nexport interface RoutingConfig {\n\tmode?: RoutingMode;\n\texecutor?: ExecutorConfig;\n}\n\nfunction isRoutingMode(value: unknown): value is RoutingMode {\n\treturn typeof value === \"string\" && (ROUTING_MODES as readonly string[]).includes(value);\n}\n\n/**\n * Resolve the effective routing mode from CLI flag, env var, and config.\n *\n * Activation requires either the flag or the env var; config alone never\n * activates routing (decision: explicit opt-in only). When activated without an\n * explicit mode, defaults to `executor-for-summarization` (the lowest-risk\n * mode). When not activated, always `primary-only`.\n */\nexport function resolveRoutingMode(opts: {\n\tenableFlag?: boolean;\n\tenvMode?: string;\n\tconfigMode?: RoutingMode;\n}): RoutingMode {\n\tconst envMode = opts.envMode?.trim();\n\tconst envActivates = envMode !== undefined && envMode !== \"\" && envMode !== \"primary-only\";\n\tconst activated = opts.enableFlag === true || envActivates;\n\tif (!activated) return \"primary-only\";\n\n\tif (envMode && isRoutingMode(envMode)) return envMode;\n\tif (opts.configMode && isRoutingMode(opts.configMode)) return opts.configMode;\n\treturn \"executor-for-summarization\";\n}\n\n/**\n * Router that decides, per turn kind, whether to use the executor model and\n * resolves it from the registry. Holds no mutable state beyond the resolved\n * executor model.\n */\nexport class LocalInferenceRouter {\n\tprivate readonly mode: RoutingMode;\n\tprivate readonly executorConfig: ExecutorConfig | undefined;\n\tprivate readonly executor: Model<Api> | undefined;\n\tprivate readonly toolResultMinBytes: number;\n\n\tprivate constructor(\n\t\tmode: RoutingMode,\n\t\texecutorConfig: ExecutorConfig | undefined,\n\t\texecutor: Model<Api> | undefined,\n\t) {\n\t\tthis.mode = mode;\n\t\tthis.executorConfig = executorConfig;\n\t\tthis.executor = executor;\n\t\tthis.toolResultMinBytes = executorConfig?.toolResultMinBytes ?? DEFAULT_TOOL_RESULT_MIN_BYTES;\n\t}\n\n\tstatic create(opts: {\n\t\tmode: RoutingMode;\n\t\tconfig: RoutingConfig | undefined;\n\t\tregistry: ModelRegistry;\n\t}): LocalInferenceRouter {\n\t\tconst executorConfig = opts.config?.executor;\n\t\tlet executor: Model<Api> | undefined;\n\t\tif (opts.mode !== \"primary-only\" && executorConfig) {\n\t\t\texecutor = opts.registry.find(executorConfig.provider, executorConfig.model);\n\t\t}\n\t\treturn new LocalInferenceRouter(opts.mode, executorConfig, executor);\n\t}\n\n\tgetMode(): RoutingMode {\n\t\treturn this.mode;\n\t}\n\n\t/** True when routing is active and an executor model is resolved and usable. */\n\tisExecutorAvailable(): boolean {\n\t\treturn this.mode !== \"primary-only\" && this.executor !== undefined;\n\t}\n\n\tgetExecutorConfig(): ExecutorConfig | undefined {\n\t\treturn this.executorConfig;\n\t}\n\n\t/**\n\t * Pick the model to use for a turn. Returns the executor when the mode routes\n\t * that turn kind and the executor is available; otherwise returns the primary\n\t * model. `shadow-executor` always returns the primary for the live path (the\n\t * executor is exercised separately for measurement).\n\t */\n\tselectModel(turnKind: TurnKind, primary: Model<Api>): Model<Api> {\n\t\tif (!this.isExecutorAvailable() || !this.executor) return primary;\n\t\tswitch (this.mode) {\n\t\t\tcase \"executor-for-summarization\":\n\t\t\t\treturn turnKind === \"summarization\" ? this.executor : primary;\n\t\t\tcase \"executor-for-tool-results\":\n\t\t\t\treturn turnKind === \"tool-result\" ? this.executor : primary;\n\t\t\tcase \"shadow-executor\":\n\t\t\tcase \"primary-only\":\n\t\t\t\treturn primary;\n\t\t}\n\t}\n\n\t/** Whether a given tool's result should be compressed via the executor. */\n\tshouldCompressToolResult(toolName: string, contentBytes: number): boolean {\n\t\tif (this.mode !== \"executor-for-tool-results\") return false;\n\t\tif (!this.isExecutorAvailable()) return false;\n\t\tif (!COMPRESSIBLE_TOOLS.has(toolName)) return false;\n\t\treturn contentBytes >= this.toolResultMinBytes;\n\t}\n\n\tgetExecutorModel(): Model<Api> | undefined {\n\t\treturn this.executor;\n\t}\n}\n"]}
+{"version":3,"file":"local-inference.d.ts","sourceRoot":"","sources":["../../../src/core/routing/local-inference.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,MAAM,yBAAyB,CAAC;AAC1D,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,sBAAsB,CAAC;AAE1D,4EAA4E;AAC5E,MAAM,MAAM,QAAQ,GAAG,SAAS,GAAG,eAAe,GAAG,aAAa,CAAC;AAEnE,MAAM,MAAM,WAAW,GACpB,cAAc,GACd,4BAA4B,GAC5B,2BAA2B,GAC3B,iBAAiB,CAAC;AAErB,eAAO,MAAM,aAAa,EAAE,SAAS,WAAW,EAKtC,CAAC;AAEX;;;;;;GAMG;AACH,eAAO,MAAM,kBAAkB,aAAoB,CAAC;AAEpD;;;;;;GAMG;AACH,eAAO,MAAM,iBAAiB,OAAO,CAAC;AACtC,eAAO,MAAM,iBAAiB,OAAO,CAAC;AAEtC,kEAAkE;AAClE,MAAM,WAAW,oBAAoB;IACpC,oDAAoD;IACpD,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,iDAAiD;IACjD,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC;IAChB,2FAA2F;IAC3F,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,sFAAsF;IACtF,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,kFAAkF;IAClF,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,6DAA6D;AAC7D,MAAM,WAAW,cAAc;IAC9B,QAAQ,EAAE,MAAM,CAAC;IACjB,KAAK,EAAE,MAAM,CAAC;IACd,sEAAsE;IACtE,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,gFAAgF;IAChF,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,0EAA0E;IAC1E,MAAM,CAAC,EAAE,oBAAoB,CAAC;CAC9B;AAED,sCAAsC;AACtC,MAAM,WAAW,aAAa;IAC7B,IAAI,CAAC,EAAE,WAAW,CAAC;IACnB,QAAQ,CAAC,EAAE,cAAc,CAAC;CAC1B;AAMD;;;;;;;GAOG;AACH,wBAAgB,kBAAkB,CAAC,IAAI,EAAE;IACxC,UAAU,CAAC,EAAE,OAAO,CAAC;IACrB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,UAAU,CAAC,EAAE,WAAW,CAAC;CACzB,GAAG,WAAW,CASd;AAED;;;;GAIG;AACH,qBAAa,oBAAoB;IAChC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAc;IACnC,OAAO,CAAC,QAAQ,CAAC,cAAc,CAA6B;IAC5D,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAyB;IAClD,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAS;IAClC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAS;IAElC,OAAO,eAUN;IAED,MAAM,CAAC,MAAM,CAAC,IAAI,EAAE;QACnB,IAAI,EAAE,WAAW,CAAC;QAClB,MAAM,EAAE,aAAa,GAAG,SAAS,CAAC;QAClC,QAAQ,EAAE,aAAa,CAAC;KACxB,GAAG,oBAAoB,CAOvB;IAED,OAAO,IAAI,WAAW,CAErB;IAED,gFAAgF;IAChF,mBAAmB,IAAI,OAAO,CAE7B;IAED,iBAAiB,IAAI,cAAc,GAAG,SAAS,CAE9C;IAED;;;;;OAKG;IACH,WAAW,CAAC,QAAQ,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC,GAAG,CAAC,CAW/D;IAED,gFAAgF;IAChF,cAAc,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAErC;IAED,yDAAyD;IACzD,WAAW,IAAI;QAAE,QAAQ,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAEpD;IAED,2EAA2E;IAC3E,wBAAwB,CAAC,QAAQ,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,GAAG,OAAO,CAKxE;IAED;;;;;OAKG;IACH,wBAAwB,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAI/C;IAED,gBAAgB,IAAI,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CAEzC;CACD","sourcesContent":["/**\n * Local-inference routing.\n *\n * Optional, opt-in routing of certain non-critical work (conversation\n * compaction, and large bash tool-result compression) to a local\n * \"executor\" model running on an OpenAI-compatible endpoint (for example an\n * MLX server), while the primary model handles all planning, reasoning, edits,\n * and tool-call synthesis.\n *\n * Everything here is INERT unless explicitly enabled via the\n * `--enable-local-inference` flag or the `HOOCODE_ROUTING_MODE` env var. On any\n * executor resolution/availability problem the caller falls back to the primary\n * model (compaction) or the raw tool result (tool-result compression). The\n * router never throws into the agent loop.\n *\n * Design and validation: see docs/local-executor-routing.md.\n */\n\nimport type { Api, Model } from \"@kolisachint/hoocode-ai\";\nimport type { ModelRegistry } from \"../model-registry.js\";\n\n/** Work that may be routed to the executor instead of the primary model. */\nexport type TurnKind = \"primary\" | \"summarization\" | \"tool-result\";\n\nexport type RoutingMode =\n\t| \"primary-only\"\n\t| \"executor-for-summarization\"\n\t| \"executor-for-tool-results\"\n\t| \"shadow-executor\";\n\nexport const ROUTING_MODES: readonly RoutingMode[] = [\n\t\"primary-only\",\n\t\"executor-for-summarization\",\n\t\"executor-for-tool-results\",\n\t\"shadow-executor\",\n] as const;\n\n/**\n * Tool names whose output is worth compressing (validated). Others pass\n * through. Only `bash` qualifies: its verbose output is mostly low-value noise\n * around a few load-bearing facts. `read` was measured to compress ~0% on real\n * source code (every line is a keep-line) and was removed. Fact-list tools\n * (grep/find/ls) were never compressible (every line is a distinct fact).\n */\nexport const COMPRESSIBLE_TOOLS = new Set([\"bash\"]);\n\n/**\n * Global size band (bytes) for local-inference routing. Applies to BOTH\n * tool-result compression and compaction summarization. Inputs below the\n * minimum are not worth offloading; inputs above the maximum are slow and risk\n * GPU OOM on small machines, so they fall back to the primary model. Tunable\n * per-machine via `minBytes`/`maxBytes` in the executor config block.\n */\nexport const DEFAULT_MIN_BYTES = 2048;\nexport const DEFAULT_MAX_BYTES = 8192;\n\n/** Optional local server the harness manages for the executor. */\nexport interface ExecutorServerConfig {\n\t/** Command to launch (default: \"mlx_lm.server\"). */\n\tcommand?: string;\n\t/** Extra args appended to the launch command. */\n\targs?: string[];\n\t/** Host to health-check and bind (default: derived from executor baseUrl or 127.0.0.1). */\n\thost?: string;\n\t/** Port to health-check and bind (default: derived from executor baseUrl or 8080). */\n\tport?: number;\n\t/** Max milliseconds to wait for the server to become healthy (default: 30000). */\n\tstartupTimeoutMs?: number;\n}\n\n/** Executor model reference as configured in models.json. */\nexport interface ExecutorConfig {\n\tprovider: string;\n\tmodel: string;\n\t/** Minimum input size (bytes) before local inference is attempted. */\n\tminBytes?: number;\n\t/** Maximum input size (bytes); larger inputs fall back to the primary model. */\n\tmaxBytes?: number;\n\t/** When set, the harness spawns/health-checks/stops this local server. */\n\tserver?: ExecutorServerConfig;\n}\n\n/** `routing` block in models.json. */\nexport interface RoutingConfig {\n\tmode?: RoutingMode;\n\texecutor?: ExecutorConfig;\n}\n\nfunction isRoutingMode(value: unknown): value is RoutingMode {\n\treturn typeof value === \"string\" && (ROUTING_MODES as readonly string[]).includes(value);\n}\n\n/**\n * Resolve the effective routing mode from CLI flag, env var, and config.\n *\n * Activation requires either the flag or the env var; config alone never\n * activates routing (decision: explicit opt-in only). When activated without an\n * explicit mode, defaults to `executor-for-summarization` (the lowest-risk\n * mode). When not activated, always `primary-only`.\n */\nexport function resolveRoutingMode(opts: {\n\tenableFlag?: boolean;\n\tenvMode?: string;\n\tconfigMode?: RoutingMode;\n}): RoutingMode {\n\tconst envMode = opts.envMode?.trim();\n\tconst envActivates = envMode !== undefined && envMode !== \"\" && envMode !== \"primary-only\";\n\tconst activated = opts.enableFlag === true || envActivates;\n\tif (!activated) return \"primary-only\";\n\n\tif (envMode && isRoutingMode(envMode)) return envMode;\n\tif (opts.configMode && isRoutingMode(opts.configMode)) return opts.configMode;\n\treturn \"executor-for-summarization\";\n}\n\n/**\n * Router that decides, per turn kind, whether to use the executor model and\n * resolves it from the registry. Holds no mutable state beyond the resolved\n * executor model.\n */\nexport class LocalInferenceRouter {\n\tprivate readonly mode: RoutingMode;\n\tprivate readonly executorConfig: ExecutorConfig | undefined;\n\tprivate readonly executor: Model<Api> | undefined;\n\tprivate readonly minBytes: number;\n\tprivate readonly maxBytes: number;\n\n\tprivate constructor(\n\t\tmode: RoutingMode,\n\t\texecutorConfig: ExecutorConfig | undefined,\n\t\texecutor: Model<Api> | undefined,\n\t) {\n\t\tthis.mode = mode;\n\t\tthis.executorConfig = executorConfig;\n\t\tthis.executor = executor;\n\t\tthis.minBytes = executorConfig?.minBytes ?? DEFAULT_MIN_BYTES;\n\t\tthis.maxBytes = executorConfig?.maxBytes ?? DEFAULT_MAX_BYTES;\n\t}\n\n\tstatic create(opts: {\n\t\tmode: RoutingMode;\n\t\tconfig: RoutingConfig | undefined;\n\t\tregistry: ModelRegistry;\n\t}): LocalInferenceRouter {\n\t\tconst executorConfig = opts.config?.executor;\n\t\tlet executor: Model<Api> | undefined;\n\t\tif (opts.mode !== \"primary-only\" && executorConfig) {\n\t\t\texecutor = opts.registry.find(executorConfig.provider, executorConfig.model);\n\t\t}\n\t\treturn new LocalInferenceRouter(opts.mode, executorConfig, executor);\n\t}\n\n\tgetMode(): RoutingMode {\n\t\treturn this.mode;\n\t}\n\n\t/** True when routing is active and an executor model is resolved and usable. */\n\tisExecutorAvailable(): boolean {\n\t\treturn this.mode !== \"primary-only\" && this.executor !== undefined;\n\t}\n\n\tgetExecutorConfig(): ExecutorConfig | undefined {\n\t\treturn this.executorConfig;\n\t}\n\n\t/**\n\t * Pick the model to use for a turn. Returns the executor when the mode routes\n\t * that turn kind and the executor is available; otherwise returns the primary\n\t * model. `shadow-executor` always returns the primary for the live path (the\n\t * executor is exercised separately for measurement).\n\t */\n\tselectModel(turnKind: TurnKind, primary: Model<Api>): Model<Api> {\n\t\tif (!this.isExecutorAvailable() || !this.executor) return primary;\n\t\tswitch (this.mode) {\n\t\t\tcase \"executor-for-summarization\":\n\t\t\t\treturn turnKind === \"summarization\" ? this.executor : primary;\n\t\t\tcase \"executor-for-tool-results\":\n\t\t\t\treturn turnKind === \"tool-result\" ? this.executor : primary;\n\t\t\tcase \"shadow-executor\":\n\t\t\tcase \"primary-only\":\n\t\t\t\treturn primary;\n\t\t}\n\t}\n\n\t/** True when an input size falls within the configured local-inference band. */\n\twithinSizeBand(bytes: number): boolean {\n\t\treturn bytes >= this.minBytes && bytes <= this.maxBytes;\n\t}\n\n\t/** The configured size band, for logging/diagnostics. */\n\tgetSizeBand(): { minBytes: number; maxBytes: number } {\n\t\treturn { minBytes: this.minBytes, maxBytes: this.maxBytes };\n\t}\n\n\t/** Whether a given tool's result should be compressed via the executor. */\n\tshouldCompressToolResult(toolName: string, contentBytes: number): boolean {\n\t\tif (this.mode !== \"executor-for-tool-results\") return false;\n\t\tif (!this.isExecutorAvailable()) return false;\n\t\tif (!COMPRESSIBLE_TOOLS.has(toolName)) return false;\n\t\treturn this.withinSizeBand(contentBytes);\n\t}\n\n\t/**\n\t * Whether to route summarization to the executor for a conversation of the\n\t * given serialized size. Requires summarization routing active, the executor\n\t * available, and the size within the band (oversized conversations fall back\n\t * to the primary to avoid slow local runs and GPU OOM).\n\t */\n\tshouldRouteSummarization(bytes: number): boolean {\n\t\tif (this.mode !== \"executor-for-summarization\") return false;\n\t\tif (!this.isExecutorAvailable()) return false;\n\t\treturn this.withinSizeBand(bytes);\n\t}\n\n\tgetExecutorModel(): Model<Api> | undefined {\n\t\treturn this.executor;\n\t}\n}\n"]}

package/dist/core/routing/local-inference.js

@@ -2,7 +2,7 @@
  * Local-inference routing.
  *
  * Optional, opt-in routing of certain non-critical work (conversation
- * compaction, and large read/bash tool-result compression) to a local
+ * compaction, and large bash tool-result compression) to a local
  * "executor" model running on an OpenAI-compatible endpoint (for example an
  * MLX server), while the primary model handles all planning, reasoning, edits,
  * and tool-call synthesis.
@@ -21,10 +21,23 @@ export const ROUTING_MODES = [
     "executor-for-tool-results",
     "shadow-executor",
 ];
-/** Tool names whose output is worth compressing (validated). Others pass through. */
-export const COMPRESSIBLE_TOOLS = new Set(["read", "bash"]);
-/** Default minimum tool-result size (bytes) before compression is attempted. */
-export const DEFAULT_TOOL_RESULT_MIN_BYTES = 2048;
+/**
+ * Tool names whose output is worth compressing (validated). Others pass
+ * through. Only `bash` qualifies: its verbose output is mostly low-value noise
+ * around a few load-bearing facts. `read` was measured to compress ~0% on real
+ * source code (every line is a keep-line) and was removed. Fact-list tools
+ * (grep/find/ls) were never compressible (every line is a distinct fact).
+ */
+export const COMPRESSIBLE_TOOLS = new Set(["bash"]);
+/**
+ * Global size band (bytes) for local-inference routing. Applies to BOTH
+ * tool-result compression and compaction summarization. Inputs below the
+ * minimum are not worth offloading; inputs above the maximum are slow and risk
+ * GPU OOM on small machines, so they fall back to the primary model. Tunable
+ * per-machine via `minBytes`/`maxBytes` in the executor config block.
+ */
+export const DEFAULT_MIN_BYTES = 2048;
+export const DEFAULT_MAX_BYTES = 8192;
 function isRoutingMode(value) {
     return typeof value === "string" && ROUTING_MODES.includes(value);
 }
@@ -57,12 +70,14 @@ export class LocalInferenceRouter {
     mode;
     executorConfig;
     executor;
-    toolResultMinBytes;
+    minBytes;
+    maxBytes;
     constructor(mode, executorConfig, executor) {
         this.mode = mode;
         this.executorConfig = executorConfig;
         this.executor = executor;
-        this.toolResultMinBytes = executorConfig?.toolResultMinBytes ?? DEFAULT_TOOL_RESULT_MIN_BYTES;
+        this.minBytes = executorConfig?.minBytes ?? DEFAULT_MIN_BYTES;
+        this.maxBytes = executorConfig?.maxBytes ?? DEFAULT_MAX_BYTES;
     }
     static create(opts) {
         const executorConfig = opts.config?.executor;
@@ -101,6 +116,14 @@ export class LocalInferenceRouter {
                 return primary;
         }
     }
+    /** True when an input size falls within the configured local-inference band. */
+    withinSizeBand(bytes) {
+        return bytes >= this.minBytes && bytes <= this.maxBytes;
+    }
+    /** The configured size band, for logging/diagnostics. */
+    getSizeBand() {
+        return { minBytes: this.minBytes, maxBytes: this.maxBytes };
+    }
     /** Whether a given tool's result should be compressed via the executor. */
     shouldCompressToolResult(toolName, contentBytes) {
         if (this.mode !== "executor-for-tool-results")
@@ -109,7 +132,20 @@ export class LocalInferenceRouter {
             return false;
         if (!COMPRESSIBLE_TOOLS.has(toolName))
             return false;
-        return contentBytes >= this.toolResultMinBytes;
+        return this.withinSizeBand(contentBytes);
+    }
+    /**
+     * Whether to route summarization to the executor for a conversation of the
+     * given serialized size. Requires summarization routing active, the executor
+     * available, and the size within the band (oversized conversations fall back
+     * to the primary to avoid slow local runs and GPU OOM).
+     */
+    shouldRouteSummarization(bytes) {
+        if (this.mode !== "executor-for-summarization")
+            return false;
+        if (!this.isExecutorAvailable())
+            return false;
+        return this.withinSizeBand(bytes);
     }
     getExecutorModel() {
         return this.executor;

package/dist/core/routing/local-inference.js.map

@@ -1 +1 @@
-{"version":3,"file":"local-inference.js","sourceRoot":"","sources":["../../../src/core/routing/local-inference.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAcH,MAAM,CAAC,MAAM,aAAa,GAA2B;IACpD,cAAc;IACd,4BAA4B;IAC5B,2BAA2B;IAC3B,iBAAiB;CACR,CAAC;AAEX,qFAAqF;AACrF,MAAM,CAAC,MAAM,kBAAkB,GAAG,IAAI,GAAG,CAAC,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,CAAC;AAE5D,gFAAgF;AAChF,MAAM,CAAC,MAAM,6BAA6B,GAAG,IAAI,CAAC;AAgClD,SAAS,aAAa,CAAC,KAAc,EAAwB;IAC5D,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAK,aAAmC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;AAAA,CACzF;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,kBAAkB,CAAC,IAIlC,EAAe;IACf,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,EAAE,IAAI,EAAE,CAAC;IACrC,MAAM,YAAY,GAAG,OAAO,KAAK,SAAS,IAAI,OAAO,KAAK,EAAE,IAAI,OAAO,KAAK,cAAc,CAAC;IAC3F,MAAM,SAAS,GAAG,IAAI,CAAC,UAAU,KAAK,IAAI,IAAI,YAAY,CAAC;IAC3D,IAAI,CAAC,SAAS;QAAE,OAAO,cAAc,CAAC;IAEtC,IAAI,OAAO,IAAI,aAAa,CAAC,OAAO,CAAC;QAAE,OAAO,OAAO,CAAC;IACtD,IAAI,IAAI,CAAC,UAAU,IAAI,aAAa,CAAC,IAAI,CAAC,UAAU,CAAC;QAAE,OAAO,IAAI,CAAC,UAAU,CAAC;IAC9E,OAAO,4BAA4B,CAAC;AAAA,CACpC;AAED;;;;GAIG;AACH,MAAM,OAAO,oBAAoB;IACf,IAAI,CAAc;IAClB,cAAc,CAA6B;IAC3C,QAAQ,CAAyB;IACjC,kBAAkB,CAAS;IAE5C,YACC,IAAiB,EACjB,cAA0C,EAC1C,QAAgC,EAC/B;QACD,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,CAAC,cAAc,GAAG,cAAc,CAAC;QACrC,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;QACzB,IAAI,CAAC,kBAAkB,GAAG,cAAc,EAAE,kBAAkB,IAAI,6BAA6B,CAAC;IAAA,CAC9F;IAED,MAAM,CAAC,MAAM,CAAC,IAIb,EAAwB;QACxB,MAAM,cAAc,GAAG,IAAI,CAAC,MAAM,EAAE,QAAQ,CAAC;QAC7C,IAAI,QAAgC,CAAC;QACrC,IAAI,IAAI,CAAC,IAAI,KAAK,cAAc,IAAI,cAAc,EAAE,CAAC;YACpD,QAAQ,GAAG,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,cAAc,CAAC,QAAQ,EAAE,cAAc,CAAC,KAAK,CAAC,CAAC;QAC9E,CAAC;QACD,OAAO,IAAI,oBAAoB,CAAC,IAAI,CAAC,IAAI,EAAE,cAAc,EAAE,QAAQ,CAAC,CAAC;IAAA,CACrE;IAED,OAAO,GAAgB;QACtB,OAAO,IAAI,CAAC,IAAI,CAAC;IAAA,CACjB;IAED,gFAAgF;IAChF,mBAAmB,GAAY;QAC9B,OAAO,IAAI,CAAC,IAAI,KAAK,cAAc,IAAI,IAAI,CAAC,QAAQ,KAAK,SAAS,CAAC;IAAA,CACnE;IAED,iBAAiB,GAA+B;QAC/C,OAAO,IAAI,CAAC,cAAc,CAAC;IAAA,CAC3B;IAED;;;;;OAKG;IACH,WAAW,CAAC,QAAkB,EAAE,OAAmB,EAAc;QAChE,IAAI,CAAC,IAAI,CAAC,mBAAmB,EAAE,IAAI,CAAC,IAAI,CAAC,QAAQ;YAAE,OAAO,OAAO,CAAC;QAClE,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;YACnB,KAAK,4BAA4B;gBAChC,OAAO,QAAQ,KAAK,eAAe,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC;YAC/D,KAAK,2BAA2B;gBAC/B,OAAO,QAAQ,KAAK,aAAa,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC;YAC7D,KAAK,iBAAiB,CAAC;YACvB,KAAK,cAAc;gBAClB,OAAO,OAAO,CAAC;QACjB,CAAC;IAAA,CACD;IAED,2EAA2E;IAC3E,wBAAwB,CAAC,QAAgB,EAAE,YAAoB,EAAW;QACzE,IAAI,IAAI,CAAC,IAAI,KAAK,2BAA2B;YAAE,OAAO,KAAK,CAAC;QAC5D,IAAI,CAAC,IAAI,CAAC,mBAAmB,EAAE;YAAE,OAAO,KAAK,CAAC;QAC9C,IAAI,CAAC,kBAAkB,CAAC,GAAG,CAAC,QAAQ,CAAC;YAAE,OAAO,KAAK,CAAC;QACpD,OAAO,YAAY,IAAI,IAAI,CAAC,kBAAkB,CAAC;IAAA,CAC/C;IAED,gBAAgB,GAA2B;QAC1C,OAAO,IAAI,CAAC,QAAQ,CAAC;IAAA,CACrB;CACD","sourcesContent":["/**\n * Local-inference routing.\n *\n * Optional, opt-in routing of certain non-critical work (conversation\n * compaction, and large read/bash tool-result compression) to a local\n * \"executor\" model running on an OpenAI-compatible endpoint (for example an\n * MLX server), while the primary model handles all planning, reasoning, edits,\n * and tool-call synthesis.\n *\n * Everything here is INERT unless explicitly enabled via the\n * `--enable-local-inference` flag or the `HOOCODE_ROUTING_MODE` env var. On any\n * executor resolution/availability problem the caller falls back to the primary\n * model (compaction) or the raw tool result (tool-result compression). The\n * router never throws into the agent loop.\n *\n * Design and validation: see docs/local-executor-routing.md.\n */\n\nimport type { Api, Model } from \"@kolisachint/hoocode-ai\";\nimport type { ModelRegistry } from \"../model-registry.js\";\n\n/** Work that may be routed to the executor instead of the primary model. */\nexport type TurnKind = \"primary\" | \"summarization\" | \"tool-result\";\n\nexport type RoutingMode =\n\t| \"primary-only\"\n\t| \"executor-for-summarization\"\n\t| \"executor-for-tool-results\"\n\t| \"shadow-executor\";\n\nexport const ROUTING_MODES: readonly RoutingMode[] = [\n\t\"primary-only\",\n\t\"executor-for-summarization\",\n\t\"executor-for-tool-results\",\n\t\"shadow-executor\",\n] as const;\n\n/** Tool names whose output is worth compressing (validated). Others pass through. */\nexport const COMPRESSIBLE_TOOLS = new Set([\"read\", \"bash\"]);\n\n/** Default minimum tool-result size (bytes) before compression is attempted. */\nexport const DEFAULT_TOOL_RESULT_MIN_BYTES = 2048;\n\n/** Optional local server the harness manages for the executor. */\nexport interface ExecutorServerConfig {\n\t/** Command to launch (default: \"mlx_lm.server\"). */\n\tcommand?: string;\n\t/** Extra args appended to the launch command. */\n\targs?: string[];\n\t/** Host to health-check and bind (default: derived from executor baseUrl or 127.0.0.1). */\n\thost?: string;\n\t/** Port to health-check and bind (default: derived from executor baseUrl or 8080). */\n\tport?: number;\n\t/** Max milliseconds to wait for the server to become healthy (default: 30000). */\n\tstartupTimeoutMs?: number;\n}\n\n/** Executor model reference as configured in models.json. */\nexport interface ExecutorConfig {\n\tprovider: string;\n\tmodel: string;\n\t/** Minimum tool-result size (bytes) before compression is attempted. */\n\ttoolResultMinBytes?: number;\n\t/** When set, the harness spawns/health-checks/stops this local server. */\n\tserver?: ExecutorServerConfig;\n}\n\n/** `routing` block in models.json. */\nexport interface RoutingConfig {\n\tmode?: RoutingMode;\n\texecutor?: ExecutorConfig;\n}\n\nfunction isRoutingMode(value: unknown): value is RoutingMode {\n\treturn typeof value === \"string\" && (ROUTING_MODES as readonly string[]).includes(value);\n}\n\n/**\n * Resolve the effective routing mode from CLI flag, env var, and config.\n *\n * Activation requires either the flag or the env var; config alone never\n * activates routing (decision: explicit opt-in only). When activated without an\n * explicit mode, defaults to `executor-for-summarization` (the lowest-risk\n * mode). When not activated, always `primary-only`.\n */\nexport function resolveRoutingMode(opts: {\n\tenableFlag?: boolean;\n\tenvMode?: string;\n\tconfigMode?: RoutingMode;\n}): RoutingMode {\n\tconst envMode = opts.envMode?.trim();\n\tconst envActivates = envMode !== undefined && envMode !== \"\" && envMode !== \"primary-only\";\n\tconst activated = opts.enableFlag === true || envActivates;\n\tif (!activated) return \"primary-only\";\n\n\tif (envMode && isRoutingMode(envMode)) return envMode;\n\tif (opts.configMode && isRoutingMode(opts.configMode)) return opts.configMode;\n\treturn \"executor-for-summarization\";\n}\n\n/**\n * Router that decides, per turn kind, whether to use the executor model and\n * resolves it from the registry. Holds no mutable state beyond the resolved\n * executor model.\n */\nexport class LocalInferenceRouter {\n\tprivate readonly mode: RoutingMode;\n\tprivate readonly executorConfig: ExecutorConfig | undefined;\n\tprivate readonly executor: Model<Api> | undefined;\n\tprivate readonly toolResultMinBytes: number;\n\n\tprivate constructor(\n\t\tmode: RoutingMode,\n\t\texecutorConfig: ExecutorConfig | undefined,\n\t\texecutor: Model<Api> | undefined,\n\t) {\n\t\tthis.mode = mode;\n\t\tthis.executorConfig = executorConfig;\n\t\tthis.executor = executor;\n\t\tthis.toolResultMinBytes = executorConfig?.toolResultMinBytes ?? DEFAULT_TOOL_RESULT_MIN_BYTES;\n\t}\n\n\tstatic create(opts: {\n\t\tmode: RoutingMode;\n\t\tconfig: RoutingConfig | undefined;\n\t\tregistry: ModelRegistry;\n\t}): LocalInferenceRouter {\n\t\tconst executorConfig = opts.config?.executor;\n\t\tlet executor: Model<Api> | undefined;\n\t\tif (opts.mode !== \"primary-only\" && executorConfig) {\n\t\t\texecutor = opts.registry.find(executorConfig.provider, executorConfig.model);\n\t\t}\n\t\treturn new LocalInferenceRouter(opts.mode, executorConfig, executor);\n\t}\n\n\tgetMode(): RoutingMode {\n\t\treturn this.mode;\n\t}\n\n\t/** True when routing is active and an executor model is resolved and usable. */\n\tisExecutorAvailable(): boolean {\n\t\treturn this.mode !== \"primary-only\" && this.executor !== undefined;\n\t}\n\n\tgetExecutorConfig(): ExecutorConfig | undefined {\n\t\treturn this.executorConfig;\n\t}\n\n\t/**\n\t * Pick the model to use for a turn. Returns the executor when the mode routes\n\t * that turn kind and the executor is available; otherwise returns the primary\n\t * model. `shadow-executor` always returns the primary for the live path (the\n\t * executor is exercised separately for measurement).\n\t */\n\tselectModel(turnKind: TurnKind, primary: Model<Api>): Model<Api> {\n\t\tif (!this.isExecutorAvailable() || !this.executor) return primary;\n\t\tswitch (this.mode) {\n\t\t\tcase \"executor-for-summarization\":\n\t\t\t\treturn turnKind === \"summarization\" ? this.executor : primary;\n\t\t\tcase \"executor-for-tool-results\":\n\t\t\t\treturn turnKind === \"tool-result\" ? this.executor : primary;\n\t\t\tcase \"shadow-executor\":\n\t\t\tcase \"primary-only\":\n\t\t\t\treturn primary;\n\t\t}\n\t}\n\n\t/** Whether a given tool's result should be compressed via the executor. */\n\tshouldCompressToolResult(toolName: string, contentBytes: number): boolean {\n\t\tif (this.mode !== \"executor-for-tool-results\") return false;\n\t\tif (!this.isExecutorAvailable()) return false;\n\t\tif (!COMPRESSIBLE_TOOLS.has(toolName)) return false;\n\t\treturn contentBytes >= this.toolResultMinBytes;\n\t}\n\n\tgetExecutorModel(): Model<Api> | undefined {\n\t\treturn this.executor;\n\t}\n}\n"]}
+{"version":3,"file":"local-inference.js","sourceRoot":"","sources":["../../../src/core/routing/local-inference.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAcH,MAAM,CAAC,MAAM,aAAa,GAA2B;IACpD,cAAc;IACd,4BAA4B;IAC5B,2BAA2B;IAC3B,iBAAiB;CACR,CAAC;AAEX;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAG,IAAI,GAAG,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;AAEpD;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,iBAAiB,GAAG,IAAI,CAAC;AACtC,MAAM,CAAC,MAAM,iBAAiB,GAAG,IAAI,CAAC;AAkCtC,SAAS,aAAa,CAAC,KAAc,EAAwB;IAC5D,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAK,aAAmC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;AAAA,CACzF;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,kBAAkB,CAAC,IAIlC,EAAe;IACf,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,EAAE,IAAI,EAAE,CAAC;IACrC,MAAM,YAAY,GAAG,OAAO,KAAK,SAAS,IAAI,OAAO,KAAK,EAAE,IAAI,OAAO,KAAK,cAAc,CAAC;IAC3F,MAAM,SAAS,GAAG,IAAI,CAAC,UAAU,KAAK,IAAI,IAAI,YAAY,CAAC;IAC3D,IAAI,CAAC,SAAS;QAAE,OAAO,cAAc,CAAC;IAEtC,IAAI,OAAO,IAAI,aAAa,CAAC,OAAO,CAAC;QAAE,OAAO,OAAO,CAAC;IACtD,IAAI,IAAI,CAAC,UAAU,IAAI,aAAa,CAAC,IAAI,CAAC,UAAU,CAAC;QAAE,OAAO,IAAI,CAAC,UAAU,CAAC;IAC9E,OAAO,4BAA4B,CAAC;AAAA,CACpC;AAED;;;;GAIG;AACH,MAAM,OAAO,oBAAoB;IACf,IAAI,CAAc;IAClB,cAAc,CAA6B;IAC3C,QAAQ,CAAyB;IACjC,QAAQ,CAAS;IACjB,QAAQ,CAAS;IAElC,YACC,IAAiB,EACjB,cAA0C,EAC1C,QAAgC,EAC/B;QACD,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,CAAC,cAAc,GAAG,cAAc,CAAC;QACrC,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;QACzB,IAAI,CAAC,QAAQ,GAAG,cAAc,EAAE,QAAQ,IAAI,iBAAiB,CAAC;QAC9D,IAAI,CAAC,QAAQ,GAAG,cAAc,EAAE,QAAQ,IAAI,iBAAiB,CAAC;IAAA,CAC9D;IAED,MAAM,CAAC,MAAM,CAAC,IAIb,EAAwB;QACxB,MAAM,cAAc,GAAG,IAAI,CAAC,MAAM,EAAE,QAAQ,CAAC;QAC7C,IAAI,QAAgC,CAAC;QACrC,IAAI,IAAI,CAAC,IAAI,KAAK,cAAc,IAAI,cAAc,EAAE,CAAC;YACpD,QAAQ,GAAG,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,cAAc,CAAC,QAAQ,EAAE,cAAc,CAAC,KAAK,CAAC,CAAC;QAC9E,CAAC;QACD,OAAO,IAAI,oBAAoB,CAAC,IAAI,CAAC,IAAI,EAAE,cAAc,EAAE,QAAQ,CAAC,CAAC;IAAA,CACrE;IAED,OAAO,GAAgB;QACtB,OAAO,IAAI,CAAC,IAAI,CAAC;IAAA,CACjB;IAED,gFAAgF;IAChF,mBAAmB,GAAY;QAC9B,OAAO,IAAI,CAAC,IAAI,KAAK,cAAc,IAAI,IAAI,CAAC,QAAQ,KAAK,SAAS,CAAC;IAAA,CACnE;IAED,iBAAiB,GAA+B;QAC/C,OAAO,IAAI,CAAC,cAAc,CAAC;IAAA,CAC3B;IAED;;;;;OAKG;IACH,WAAW,CAAC,QAAkB,EAAE,OAAmB,EAAc;QAChE,IAAI,CAAC,IAAI,CAAC,mBAAmB,EAAE,IAAI,CAAC,IAAI,CAAC,QAAQ;YAAE,OAAO,OAAO,CAAC;QAClE,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;YACnB,KAAK,4BAA4B;gBAChC,OAAO,QAAQ,KAAK,eAAe,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC;YAC/D,KAAK,2BAA2B;gBAC/B,OAAO,QAAQ,KAAK,aAAa,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC;YAC7D,KAAK,iBAAiB,CAAC;YACvB,KAAK,cAAc;gBAClB,OAAO,OAAO,CAAC;QACjB,CAAC;IAAA,CACD;IAED,gFAAgF;IAChF,cAAc,CAAC,KAAa,EAAW;QACtC,OAAO,KAAK,IAAI,IAAI,CAAC,QAAQ,IAAI,KAAK,IAAI,IAAI,CAAC,QAAQ,CAAC;IAAA,CACxD;IAED,yDAAyD;IACzD,WAAW,GAA2C;QACrD,OAAO,EAAE,QAAQ,EAAE,IAAI,CAAC,QAAQ,EAAE,QAAQ,EAAE,IAAI,CAAC,QAAQ,EAAE,CAAC;IAAA,CAC5D;IAED,2EAA2E;IAC3E,wBAAwB,CAAC,QAAgB,EAAE,YAAoB,EAAW;QACzE,IAAI,IAAI,CAAC,IAAI,KAAK,2BAA2B;YAAE,OAAO,KAAK,CAAC;QAC5D,IAAI,CAAC,IAAI,CAAC,mBAAmB,EAAE;YAAE,OAAO,KAAK,CAAC;QAC9C,IAAI,CAAC,kBAAkB,CAAC,GAAG,CAAC,QAAQ,CAAC;YAAE,OAAO,KAAK,CAAC;QACpD,OAAO,IAAI,CAAC,cAAc,CAAC,YAAY,CAAC,CAAC;IAAA,CACzC;IAED;;;;;OAKG;IACH,wBAAwB,CAAC,KAAa,EAAW;QAChD,IAAI,IAAI,CAAC,IAAI,KAAK,4BAA4B;YAAE,OAAO,KAAK,CAAC;QAC7D,IAAI,CAAC,IAAI,CAAC,mBAAmB,EAAE;YAAE,OAAO,KAAK,CAAC;QAC9C,OAAO,IAAI,CAAC,cAAc,CAAC,KAAK,CAAC,CAAC;IAAA,CAClC;IAED,gBAAgB,GAA2B;QAC1C,OAAO,IAAI,CAAC,QAAQ,CAAC;IAAA,CACrB;CACD","sourcesContent":["/**\n * Local-inference routing.\n *\n * Optional, opt-in routing of certain non-critical work (conversation\n * compaction, and large bash tool-result compression) to a local\n * \"executor\" model running on an OpenAI-compatible endpoint (for example an\n * MLX server), while the primary model handles all planning, reasoning, edits,\n * and tool-call synthesis.\n *\n * Everything here is INERT unless explicitly enabled via the\n * `--enable-local-inference` flag or the `HOOCODE_ROUTING_MODE` env var. On any\n * executor resolution/availability problem the caller falls back to the primary\n * model (compaction) or the raw tool result (tool-result compression). The\n * router never throws into the agent loop.\n *\n * Design and validation: see docs/local-executor-routing.md.\n */\n\nimport type { Api, Model } from \"@kolisachint/hoocode-ai\";\nimport type { ModelRegistry } from \"../model-registry.js\";\n\n/** Work that may be routed to the executor instead of the primary model. */\nexport type TurnKind = \"primary\" | \"summarization\" | \"tool-result\";\n\nexport type RoutingMode =\n\t| \"primary-only\"\n\t| \"executor-for-summarization\"\n\t| \"executor-for-tool-results\"\n\t| \"shadow-executor\";\n\nexport const ROUTING_MODES: readonly RoutingMode[] = [\n\t\"primary-only\",\n\t\"executor-for-summarization\",\n\t\"executor-for-tool-results\",\n\t\"shadow-executor\",\n] as const;\n\n/**\n * Tool names whose output is worth compressing (validated). Others pass\n * through. Only `bash` qualifies: its verbose output is mostly low-value noise\n * around a few load-bearing facts. `read` was measured to compress ~0% on real\n * source code (every line is a keep-line) and was removed. Fact-list tools\n * (grep/find/ls) were never compressible (every line is a distinct fact).\n */\nexport const COMPRESSIBLE_TOOLS = new Set([\"bash\"]);\n\n/**\n * Global size band (bytes) for local-inference routing. Applies to BOTH\n * tool-result compression and compaction summarization. Inputs below the\n * minimum are not worth offloading; inputs above the maximum are slow and risk\n * GPU OOM on small machines, so they fall back to the primary model. Tunable\n * per-machine via `minBytes`/`maxBytes` in the executor config block.\n */\nexport const DEFAULT_MIN_BYTES = 2048;\nexport const DEFAULT_MAX_BYTES = 8192;\n\n/** Optional local server the harness manages for the executor. */\nexport interface ExecutorServerConfig {\n\t/** Command to launch (default: \"mlx_lm.server\"). */\n\tcommand?: string;\n\t/** Extra args appended to the launch command. */\n\targs?: string[];\n\t/** Host to health-check and bind (default: derived from executor baseUrl or 127.0.0.1). */\n\thost?: string;\n\t/** Port to health-check and bind (default: derived from executor baseUrl or 8080). */\n\tport?: number;\n\t/** Max milliseconds to wait for the server to become healthy (default: 30000). */\n\tstartupTimeoutMs?: number;\n}\n\n/** Executor model reference as configured in models.json. */\nexport interface ExecutorConfig {\n\tprovider: string;\n\tmodel: string;\n\t/** Minimum input size (bytes) before local inference is attempted. */\n\tminBytes?: number;\n\t/** Maximum input size (bytes); larger inputs fall back to the primary model. */\n\tmaxBytes?: number;\n\t/** When set, the harness spawns/health-checks/stops this local server. */\n\tserver?: ExecutorServerConfig;\n}\n\n/** `routing` block in models.json. */\nexport interface RoutingConfig {\n\tmode?: RoutingMode;\n\texecutor?: ExecutorConfig;\n}\n\nfunction isRoutingMode(value: unknown): value is RoutingMode {\n\treturn typeof value === \"string\" && (ROUTING_MODES as readonly string[]).includes(value);\n}\n\n/**\n * Resolve the effective routing mode from CLI flag, env var, and config.\n *\n * Activation requires either the flag or the env var; config alone never\n * activates routing (decision: explicit opt-in only). When activated without an\n * explicit mode, defaults to `executor-for-summarization` (the lowest-risk\n * mode). When not activated, always `primary-only`.\n */\nexport function resolveRoutingMode(opts: {\n\tenableFlag?: boolean;\n\tenvMode?: string;\n\tconfigMode?: RoutingMode;\n}): RoutingMode {\n\tconst envMode = opts.envMode?.trim();\n\tconst envActivates = envMode !== undefined && envMode !== \"\" && envMode !== \"primary-only\";\n\tconst activated = opts.enableFlag === true || envActivates;\n\tif (!activated) return \"primary-only\";\n\n\tif (envMode && isRoutingMode(envMode)) return envMode;\n\tif (opts.configMode && isRoutingMode(opts.configMode)) return opts.configMode;\n\treturn \"executor-for-summarization\";\n}\n\n/**\n * Router that decides, per turn kind, whether to use the executor model and\n * resolves it from the registry. Holds no mutable state beyond the resolved\n * executor model.\n */\nexport class LocalInferenceRouter {\n\tprivate readonly mode: RoutingMode;\n\tprivate readonly executorConfig: ExecutorConfig | undefined;\n\tprivate readonly executor: Model<Api> | undefined;\n\tprivate readonly minBytes: number;\n\tprivate readonly maxBytes: number;\n\n\tprivate constructor(\n\t\tmode: RoutingMode,\n\t\texecutorConfig: ExecutorConfig | undefined,\n\t\texecutor: Model<Api> | undefined,\n\t) {\n\t\tthis.mode = mode;\n\t\tthis.executorConfig = executorConfig;\n\t\tthis.executor = executor;\n\t\tthis.minBytes = executorConfig?.minBytes ?? DEFAULT_MIN_BYTES;\n\t\tthis.maxBytes = executorConfig?.maxBytes ?? DEFAULT_MAX_BYTES;\n\t}\n\n\tstatic create(opts: {\n\t\tmode: RoutingMode;\n\t\tconfig: RoutingConfig | undefined;\n\t\tregistry: ModelRegistry;\n\t}): LocalInferenceRouter {\n\t\tconst executorConfig = opts.config?.executor;\n\t\tlet executor: Model<Api> | undefined;\n\t\tif (opts.mode !== \"primary-only\" && executorConfig) {\n\t\t\texecutor = opts.registry.find(executorConfig.provider, executorConfig.model);\n\t\t}\n\t\treturn new LocalInferenceRouter(opts.mode, executorConfig, executor);\n\t}\n\n\tgetMode(): RoutingMode {\n\t\treturn this.mode;\n\t}\n\n\t/** True when routing is active and an executor model is resolved and usable. */\n\tisExecutorAvailable(): boolean {\n\t\treturn this.mode !== \"primary-only\" && this.executor !== undefined;\n\t}\n\n\tgetExecutorConfig(): ExecutorConfig | undefined {\n\t\treturn this.executorConfig;\n\t}\n\n\t/**\n\t * Pick the model to use for a turn. Returns the executor when the mode routes\n\t * that turn kind and the executor is available; otherwise returns the primary\n\t * model. `shadow-executor` always returns the primary for the live path (the\n\t * executor is exercised separately for measurement).\n\t */\n\tselectModel(turnKind: TurnKind, primary: Model<Api>): Model<Api> {\n\t\tif (!this.isExecutorAvailable() || !this.executor) return primary;\n\t\tswitch (this.mode) {\n\t\t\tcase \"executor-for-summarization\":\n\t\t\t\treturn turnKind === \"summarization\" ? this.executor : primary;\n\t\t\tcase \"executor-for-tool-results\":\n\t\t\t\treturn turnKind === \"tool-result\" ? this.executor : primary;\n\t\t\tcase \"shadow-executor\":\n\t\t\tcase \"primary-only\":\n\t\t\t\treturn primary;\n\t\t}\n\t}\n\n\t/** True when an input size falls within the configured local-inference band. */\n\twithinSizeBand(bytes: number): boolean {\n\t\treturn bytes >= this.minBytes && bytes <= this.maxBytes;\n\t}\n\n\t/** The configured size band, for logging/diagnostics. */\n\tgetSizeBand(): { minBytes: number; maxBytes: number } {\n\t\treturn { minBytes: this.minBytes, maxBytes: this.maxBytes };\n\t}\n\n\t/** Whether a given tool's result should be compressed via the executor. */\n\tshouldCompressToolResult(toolName: string, contentBytes: number): boolean {\n\t\tif (this.mode !== \"executor-for-tool-results\") return false;\n\t\tif (!this.isExecutorAvailable()) return false;\n\t\tif (!COMPRESSIBLE_TOOLS.has(toolName)) return false;\n\t\treturn this.withinSizeBand(contentBytes);\n\t}\n\n\t/**\n\t * Whether to route summarization to the executor for a conversation of the\n\t * given serialized size. Requires summarization routing active, the executor\n\t * available, and the size within the band (oversized conversations fall back\n\t * to the primary to avoid slow local runs and GPU OOM).\n\t */\n\tshouldRouteSummarization(bytes: number): boolean {\n\t\tif (this.mode !== \"executor-for-summarization\") return false;\n\t\tif (!this.isExecutorAvailable()) return false;\n\t\treturn this.withinSizeBand(bytes);\n\t}\n\n\tgetExecutorModel(): Model<Api> | undefined {\n\t\treturn this.executor;\n\t}\n}\n"]}

package/dist/core/routing/tool-result-prompts.d.ts

@@ -2,13 +2,14 @@
  * Validated per-tool extractive compression prompts.
  *
  * These prompts were validated against Qwen3-4B (see
- * docs/local-executor-routing.md). Only `read` and `bash` outputs compress
- * safely (95%/94% size reduction at 100% critical-fact retention) because their
- * verbose output is mostly low-value noise around a few load-bearing facts.
- * Fact-list outputs (grep/find/ls) are intentionally excluded: every line is a
- * distinct fact, so compression drops matches.
+ * docs/local-executor-routing.md). Only `bash` output compresses safely: its
+ * verbose output is mostly low-value noise (progress/passing lines) around a
+ * few load-bearing facts (errors, counts, exit codes). `read` was removed after
+ * measurement showed ~0% reduction on real source code (every line is a
+ * keep-line). Fact-list outputs (grep/find/ls) are intentionally excluded:
+ * every line is a distinct fact, so compression drops matches.
  *
- * The prompts are extractive (keep identifiers, drop only redundant filler),
+ * The prompt is extractive (keep identifiers, drop only redundant filler),
  * not abstractive (rewrite in prose), which is what made retention reliable.
  */
 export declare const TOOL_RESULT_SYSTEM_PROMPT: string;

package/dist/core/routing/tool-result-prompts.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"tool-result-prompts.d.ts","sourceRoot":"","sources":["../../../src/core/routing/tool-result-prompts.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAEH,eAAO,MAAM,yBAAyB,QAEkD,CAAC;AAkBzF,0FAA0F;AAC1F,wBAAgB,mBAAmB,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CAExE;AAED,gEAAgE;AAChE,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CAI1F;AAED;;;;GAIG;AACH,wBAAgB,cAAc,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAKnD","sourcesContent":["/**\n * Validated per-tool extractive compression prompts.\n *\n * These prompts were validated against Qwen3-4B (see\n * docs/local-executor-routing.md). Only `read` and `bash` outputs compress\n * safely (95%/94% size reduction at 100% critical-fact retention) because their\n * verbose output is mostly low-value noise around a few load-bearing facts.\n * Fact-list outputs (grep/find/ls) are intentionally excluded: every line is a\n * distinct fact, so compression drops matches.\n *\n * The prompts are extractive (keep identifiers, drop only redundant filler),\n * not abstractive (rewrite in prose), which is what made retention reliable.\n */\n\nexport const TOOL_RESULT_SYSTEM_PROMPT =\n\t\"You compress tool output for another AI to consume. Be extractive: keep exact identifiers, \" +\n\t\"never paraphrase facts, and never invent anything. Output only the compressed result.\";\n\nconst READ_PROMPT =\n\t\"Compress this file read. Output the path EXACTLY as given, then ONLY declaration lines as \" +\n\t\"`LINE: code` for: imports, class/function/const declarations WITH their values, throw/status-check \" +\n\t\"lines, and TODO comments. Drop plain function-body statements and blank/comment-only lines. Keep \" +\n\t\"every line number, identifier, string literal, and numeric value exactly. No prose.\";\n\nconst BASH_PROMPT =\n\t\"Compress this command output. Keep ONLY: the command, every error/warning with its file:line:col \" +\n\t\"and code, and any final counts/timings/exit code. Drop progress bars, info lines, and passing/OK \" +\n\t\"lines. Keep all numbers and paths exactly. No prose.\";\n\nconst TOOL_RESULT_PROMPTS: Record<string, string> = {\n\tread: READ_PROMPT,\n\tbash: BASH_PROMPT,\n};\n\n/** Get the extractive compression prompt for a tool, or undefined if not compressible. */\nexport function getToolResultPrompt(toolName: string): string | undefined {\n\treturn TOOL_RESULT_PROMPTS[toolName];\n}\n\n/** Build the full prompt text for compressing a tool result. */\nexport function buildToolResultPrompt(toolName: string, output: string): string | undefined {\n\tconst instruction = getToolResultPrompt(toolName);\n\tif (!instruction) return undefined;\n\treturn `${instruction}\\n\\n${output}`;\n}\n\n/**\n * Remove reasoning-model `<think>...</think>` blocks (including empty ones that\n * Qwen3 emits even under `/no_think`) and trim the result. Reasoning models\n * leave these tags in the text stream; they must not leak into context.\n */\nexport function stripThinkTags(text: string): string {\n\treturn text\n\t\t.replace(/<think>[\\s\\S]*?<\\/think>/g, \"\")\n\t\t.replace(/<\\/?think>/g, \"\")\n\t\t.trim();\n}\n"]}
+{"version":3,"file":"tool-result-prompts.d.ts","sourceRoot":"","sources":["../../../src/core/routing/tool-result-prompts.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,eAAO,MAAM,yBAAyB,QAEkD,CAAC;AAWzF,0FAA0F;AAC1F,wBAAgB,mBAAmB,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CAExE;AAED,gEAAgE;AAChE,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CAI1F;AAED;;;;GAIG;AACH,wBAAgB,cAAc,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAKnD","sourcesContent":["/**\n * Validated per-tool extractive compression prompts.\n *\n * These prompts were validated against Qwen3-4B (see\n * docs/local-executor-routing.md). Only `bash` output compresses safely: its\n * verbose output is mostly low-value noise (progress/passing lines) around a\n * few load-bearing facts (errors, counts, exit codes). `read` was removed after\n * measurement showed ~0% reduction on real source code (every line is a\n * keep-line). Fact-list outputs (grep/find/ls) are intentionally excluded:\n * every line is a distinct fact, so compression drops matches.\n *\n * The prompt is extractive (keep identifiers, drop only redundant filler),\n * not abstractive (rewrite in prose), which is what made retention reliable.\n */\n\nexport const TOOL_RESULT_SYSTEM_PROMPT =\n\t\"You compress tool output for another AI to consume. Be extractive: keep exact identifiers, \" +\n\t\"never paraphrase facts, and never invent anything. Output only the compressed result.\";\n\nconst BASH_PROMPT =\n\t\"Compress this command output. Keep ONLY: the command, every error/warning with its file:line:col \" +\n\t\"and code, and any final counts/timings/exit code. Drop progress bars, info lines, and passing/OK \" +\n\t\"lines. Keep all numbers and paths exactly. No prose.\";\n\nconst TOOL_RESULT_PROMPTS: Record<string, string> = {\n\tbash: BASH_PROMPT,\n};\n\n/** Get the extractive compression prompt for a tool, or undefined if not compressible. */\nexport function getToolResultPrompt(toolName: string): string | undefined {\n\treturn TOOL_RESULT_PROMPTS[toolName];\n}\n\n/** Build the full prompt text for compressing a tool result. */\nexport function buildToolResultPrompt(toolName: string, output: string): string | undefined {\n\tconst instruction = getToolResultPrompt(toolName);\n\tif (!instruction) return undefined;\n\treturn `${instruction}\\n\\n${output}`;\n}\n\n/**\n * Remove reasoning-model `<think>...</think>` blocks (including empty ones that\n * Qwen3 emits even under `/no_think`) and trim the result. Reasoning models\n * leave these tags in the text stream; they must not leak into context.\n */\nexport function stripThinkTags(text: string): string {\n\treturn text\n\t\t.replace(/<think>[\\s\\S]*?<\\/think>/g, \"\")\n\t\t.replace(/<\\/?think>/g, \"\")\n\t\t.trim();\n}\n"]}

package/dist/core/routing/tool-result-prompts.js

@@ -2,26 +2,22 @@
  * Validated per-tool extractive compression prompts.
  *
  * These prompts were validated against Qwen3-4B (see
- * docs/local-executor-routing.md). Only `read` and `bash` outputs compress
- * safely (95%/94% size reduction at 100% critical-fact retention) because their
- * verbose output is mostly low-value noise around a few load-bearing facts.
- * Fact-list outputs (grep/find/ls) are intentionally excluded: every line is a
- * distinct fact, so compression drops matches.
+ * docs/local-executor-routing.md). Only `bash` output compresses safely: its
+ * verbose output is mostly low-value noise (progress/passing lines) around a
+ * few load-bearing facts (errors, counts, exit codes). `read` was removed after
+ * measurement showed ~0% reduction on real source code (every line is a
+ * keep-line). Fact-list outputs (grep/find/ls) are intentionally excluded:
+ * every line is a distinct fact, so compression drops matches.
  *
- * The prompts are extractive (keep identifiers, drop only redundant filler),
+ * The prompt is extractive (keep identifiers, drop only redundant filler),
  * not abstractive (rewrite in prose), which is what made retention reliable.
  */
 export const TOOL_RESULT_SYSTEM_PROMPT = "You compress tool output for another AI to consume. Be extractive: keep exact identifiers, " +
     "never paraphrase facts, and never invent anything. Output only the compressed result.";
-const READ_PROMPT = "Compress this file read. Output the path EXACTLY as given, then ONLY declaration lines as " +
-    "`LINE: code` for: imports, class/function/const declarations WITH their values, throw/status-check " +
-    "lines, and TODO comments. Drop plain function-body statements and blank/comment-only lines. Keep " +
-    "every line number, identifier, string literal, and numeric value exactly. No prose.";
 const BASH_PROMPT = "Compress this command output. Keep ONLY: the command, every error/warning with its file:line:col " +
     "and code, and any final counts/timings/exit code. Drop progress bars, info lines, and passing/OK " +
     "lines. Keep all numbers and paths exactly. No prose.";
 const TOOL_RESULT_PROMPTS = {
-    read: READ_PROMPT,
     bash: BASH_PROMPT,
 };
 /** Get the extractive compression prompt for a tool, or undefined if not compressible. */

package/dist/core/routing/tool-result-prompts.js.map

@@ -1 +1 @@
-{"version":3,"file":"tool-result-prompts.js","sourceRoot":"","sources":["../../../src/core/routing/tool-result-prompts.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAEH,MAAM,CAAC,MAAM,yBAAyB,GACrC,6FAA6F;IAC7F,uFAAuF,CAAC;AAEzF,MAAM,WAAW,GAChB,4FAA4F;IAC5F,qGAAqG;IACrG,mGAAmG;IACnG,qFAAqF,CAAC;AAEvF,MAAM,WAAW,GAChB,mGAAmG;IACnG,mGAAmG;IACnG,sDAAsD,CAAC;AAExD,MAAM,mBAAmB,GAA2B;IACnD,IAAI,EAAE,WAAW;IACjB,IAAI,EAAE,WAAW;CACjB,CAAC;AAEF,0FAA0F;AAC1F,MAAM,UAAU,mBAAmB,CAAC,QAAgB,EAAsB;IACzE,OAAO,mBAAmB,CAAC,QAAQ,CAAC,CAAC;AAAA,CACrC;AAED,gEAAgE;AAChE,MAAM,UAAU,qBAAqB,CAAC,QAAgB,EAAE,MAAc,EAAsB;IAC3F,MAAM,WAAW,GAAG,mBAAmB,CAAC,QAAQ,CAAC,CAAC;IAClD,IAAI,CAAC,WAAW;QAAE,OAAO,SAAS,CAAC;IACnC,OAAO,GAAG,WAAW,OAAO,MAAM,EAAE,CAAC;AAAA,CACrC;AAED;;;;GAIG;AACH,MAAM,UAAU,cAAc,CAAC,IAAY,EAAU;IACpD,OAAO,IAAI;SACT,OAAO,CAAC,2BAA2B,EAAE,EAAE,CAAC;SACxC,OAAO,CAAC,aAAa,EAAE,EAAE,CAAC;SAC1B,IAAI,EAAE,CAAC;AAAA,CACT","sourcesContent":["/**\n * Validated per-tool extractive compression prompts.\n *\n * These prompts were validated against Qwen3-4B (see\n * docs/local-executor-routing.md). Only `read` and `bash` outputs compress\n * safely (95%/94% size reduction at 100% critical-fact retention) because their\n * verbose output is mostly low-value noise around a few load-bearing facts.\n * Fact-list outputs (grep/find/ls) are intentionally excluded: every line is a\n * distinct fact, so compression drops matches.\n *\n * The prompts are extractive (keep identifiers, drop only redundant filler),\n * not abstractive (rewrite in prose), which is what made retention reliable.\n */\n\nexport const TOOL_RESULT_SYSTEM_PROMPT =\n\t\"You compress tool output for another AI to consume. Be extractive: keep exact identifiers, \" +\n\t\"never paraphrase facts, and never invent anything. Output only the compressed result.\";\n\nconst READ_PROMPT =\n\t\"Compress this file read. Output the path EXACTLY as given, then ONLY declaration lines as \" +\n\t\"`LINE: code` for: imports, class/function/const declarations WITH their values, throw/status-check \" +\n\t\"lines, and TODO comments. Drop plain function-body statements and blank/comment-only lines. Keep \" +\n\t\"every line number, identifier, string literal, and numeric value exactly. No prose.\";\n\nconst BASH_PROMPT =\n\t\"Compress this command output. Keep ONLY: the command, every error/warning with its file:line:col \" +\n\t\"and code, and any final counts/timings/exit code. Drop progress bars, info lines, and passing/OK \" +\n\t\"lines. Keep all numbers and paths exactly. No prose.\";\n\nconst TOOL_RESULT_PROMPTS: Record<string, string> = {\n\tread: READ_PROMPT,\n\tbash: BASH_PROMPT,\n};\n\n/** Get the extractive compression prompt for a tool, or undefined if not compressible. */\nexport function getToolResultPrompt(toolName: string): string | undefined {\n\treturn TOOL_RESULT_PROMPTS[toolName];\n}\n\n/** Build the full prompt text for compressing a tool result. */\nexport function buildToolResultPrompt(toolName: string, output: string): string | undefined {\n\tconst instruction = getToolResultPrompt(toolName);\n\tif (!instruction) return undefined;\n\treturn `${instruction}\\n\\n${output}`;\n}\n\n/**\n * Remove reasoning-model `<think>...</think>` blocks (including empty ones that\n * Qwen3 emits even under `/no_think`) and trim the result. Reasoning models\n * leave these tags in the text stream; they must not leak into context.\n */\nexport function stripThinkTags(text: string): string {\n\treturn text\n\t\t.replace(/<think>[\\s\\S]*?<\\/think>/g, \"\")\n\t\t.replace(/<\\/?think>/g, \"\")\n\t\t.trim();\n}\n"]}
+{"version":3,"file":"tool-result-prompts.js","sourceRoot":"","sources":["../../../src/core/routing/tool-result-prompts.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,MAAM,CAAC,MAAM,yBAAyB,GACrC,6FAA6F;IAC7F,uFAAuF,CAAC;AAEzF,MAAM,WAAW,GAChB,mGAAmG;IACnG,mGAAmG;IACnG,sDAAsD,CAAC;AAExD,MAAM,mBAAmB,GAA2B;IACnD,IAAI,EAAE,WAAW;CACjB,CAAC;AAEF,0FAA0F;AAC1F,MAAM,UAAU,mBAAmB,CAAC,QAAgB,EAAsB;IACzE,OAAO,mBAAmB,CAAC,QAAQ,CAAC,CAAC;AAAA,CACrC;AAED,gEAAgE;AAChE,MAAM,UAAU,qBAAqB,CAAC,QAAgB,EAAE,MAAc,EAAsB;IAC3F,MAAM,WAAW,GAAG,mBAAmB,CAAC,QAAQ,CAAC,CAAC;IAClD,IAAI,CAAC,WAAW;QAAE,OAAO,SAAS,CAAC;IACnC,OAAO,GAAG,WAAW,OAAO,MAAM,EAAE,CAAC;AAAA,CACrC;AAED;;;;GAIG;AACH,MAAM,UAAU,cAAc,CAAC,IAAY,EAAU;IACpD,OAAO,IAAI;SACT,OAAO,CAAC,2BAA2B,EAAE,EAAE,CAAC;SACxC,OAAO,CAAC,aAAa,EAAE,EAAE,CAAC;SAC1B,IAAI,EAAE,CAAC;AAAA,CACT","sourcesContent":["/**\n * Validated per-tool extractive compression prompts.\n *\n * These prompts were validated against Qwen3-4B (see\n * docs/local-executor-routing.md). Only `bash` output compresses safely: its\n * verbose output is mostly low-value noise (progress/passing lines) around a\n * few load-bearing facts (errors, counts, exit codes). `read` was removed after\n * measurement showed ~0% reduction on real source code (every line is a\n * keep-line). Fact-list outputs (grep/find/ls) are intentionally excluded:\n * every line is a distinct fact, so compression drops matches.\n *\n * The prompt is extractive (keep identifiers, drop only redundant filler),\n * not abstractive (rewrite in prose), which is what made retention reliable.\n */\n\nexport const TOOL_RESULT_SYSTEM_PROMPT =\n\t\"You compress tool output for another AI to consume. Be extractive: keep exact identifiers, \" +\n\t\"never paraphrase facts, and never invent anything. Output only the compressed result.\";\n\nconst BASH_PROMPT =\n\t\"Compress this command output. Keep ONLY: the command, every error/warning with its file:line:col \" +\n\t\"and code, and any final counts/timings/exit code. Drop progress bars, info lines, and passing/OK \" +\n\t\"lines. Keep all numbers and paths exactly. No prose.\";\n\nconst TOOL_RESULT_PROMPTS: Record<string, string> = {\n\tbash: BASH_PROMPT,\n};\n\n/** Get the extractive compression prompt for a tool, or undefined if not compressible. */\nexport function getToolResultPrompt(toolName: string): string | undefined {\n\treturn TOOL_RESULT_PROMPTS[toolName];\n}\n\n/** Build the full prompt text for compressing a tool result. */\nexport function buildToolResultPrompt(toolName: string, output: string): string | undefined {\n\tconst instruction = getToolResultPrompt(toolName);\n\tif (!instruction) return undefined;\n\treturn `${instruction}\\n\\n${output}`;\n}\n\n/**\n * Remove reasoning-model `<think>...</think>` blocks (including empty ones that\n * Qwen3 emits even under `/no_think`) and trim the result. Reasoning models\n * leave these tags in the text stream; they must not leak into context.\n */\nexport function stripThinkTags(text: string): string {\n\treturn text\n\t\t.replace(/<think>[\\s\\S]*?<\\/think>/g, \"\")\n\t\t.replace(/<\\/?think>/g, \"\")\n\t\t.trim();\n}\n"]}

package/examples/extensions/custom-provider-anthropic/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@kolisachint/hoocode-extension-custom-provider-anthropic",
 	"private": true,
-	"version": "0.2.68",
+	"version": "0.2.70",
 	"type": "module",
 	"engines": {
 		"bun": ">=1.0.0"

package/examples/extensions/custom-provider-gitlab-duo/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@kolisachint/hoocode-extension-custom-provider-gitlab-duo",
 	"private": true,
-	"version": "0.2.68",
+	"version": "0.2.70",
 	"type": "module",
 	"engines": {
 		"bun": ">=1.0.0"

package/examples/extensions/sandbox/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@kolisachint/hoocode-extension-sandbox",
 	"private": true,
-	"version": "0.2.68",
+	"version": "0.2.70",
 	"type": "module",
 	"engines": {
 		"bun": ">=1.0.0"

package/examples/extensions/with-deps/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@kolisachint/hoocode-extension-with-deps",
 	"private": true,
-	"version": "0.2.68",
+	"version": "0.2.70",
 	"type": "module",
 	"engines": {
 		"bun": ">=1.0.0"

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@kolisachint/hoocode-agent",
-	"version": "0.4.71",
+	"version": "0.4.73",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"type": "module",
 	"hoocodeConfig": {
@@ -50,9 +50,9 @@
 		"prepublishOnly": "npm run clean && npm run build"
 	},
 	"dependencies": {
-		"@kolisachint/hoocode-agent-core": "^0.4.71",
-		"@kolisachint/hoocode-ai": "^0.4.71",
-		"@kolisachint/hoocode-tui": "^0.4.71",
+		"@kolisachint/hoocode-agent-core": "^0.4.73",
+		"@kolisachint/hoocode-ai": "^0.4.73",
+		"@kolisachint/hoocode-tui": "^0.4.73",
 		"@silvia-odwyer/photon-node": "^0.3.4",
 		"chalk": "^5.5.0",
 		"cli-highlight": "^2.1.11",