@kolisachint/hoocode-agent 0.4.74 → 0.4.76

package/templates/agents/plan.md

@@ -15,7 +15,7 @@ description: |
   Cost: Low–medium (read-only)
   Isolation: Can run in parallel with explore tasks
 tools: read, grep, find, ls
-model: inherit
+model: standard
 background: false
 ---
 You are a planning agent running inside hoocode. You research the codebase and

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

@@ -1,128 +0,0 @@
-/**
- * Local-inference routing.
- *
- * Optional, opt-in routing of certain non-critical work (conversation
- * 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.
- *
- * Everything here is INERT unless explicitly enabled via the
- * `--enable-local-inference` flag or the `HOOCODE_ROUTING_MODE` env var. On any
- * executor resolution/availability problem the caller falls back to the primary
- * model (compaction) or the raw tool result (tool-result compression). The
- * router never throws into the agent loop.
- *
- * Design and validation: see docs/local-executor-routing.md.
- */
-import type { Api, Model } from "@kolisachint/hoocode-ai";
-import type { ModelRegistry } from "../model-registry.js";
-/** Work that may be routed to the executor instead of the primary model. */
-export type TurnKind = "primary" | "summarization" | "tool-result";
-export type RoutingMode = "primary-only" | "executor-for-summarization" | "executor-for-tool-results";
-export declare const ROUTING_MODES: readonly RoutingMode[];
-/**
- * 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>;
-/** Optional local server the harness manages for the executor. */
-export interface ExecutorServerConfig {
-    /** Command to launch (default: "mlx_lm.server"). */
-    command?: string;
-    /** Extra args appended to the launch command. */
-    args?: string[];
-    /** Host to health-check and bind (default: derived from executor baseUrl or 127.0.0.1). */
-    host?: string;
-    /** Port to health-check and bind (default: derived from executor baseUrl or 8080). */
-    port?: number;
-    /** Max milliseconds to wait for the server to become healthy (default: 30000). */
-    startupTimeoutMs?: number;
-}
-/**
- * Executor model reference as configured in models.json.
- *
- * The size band (`minBytes`/`maxBytes`) is wired entirely from config — there
- * are no built-in byte defaults. When a bound is omitted it is not applied:
- * `minBytes` defaults to 0 (no lower gate) and `maxBytes` to unbounded. Set
- * both in models.json to gate which inputs route to the executor; on local
- * hardware a `maxBytes` guards against slow runs / GPU OOM, while a hosted or
- * large-memory executor can leave it unset (or high) to offload large inputs.
- */
-export interface ExecutorConfig {
-    provider: string;
-    model: string;
-    /** Minimum input size (bytes) before local inference is attempted. Omitted = 0 (no lower gate). */
-    minBytes?: number;
-    /** Maximum input size (bytes); larger inputs fall back to the primary model. Omitted = unbounded. */
-    maxBytes?: number;
-    /** When set, the harness spawns/health-checks/stops this local server. */
-    server?: ExecutorServerConfig;
-}
-/** `routing` block in models.json. */
-export interface RoutingConfig {
-    mode?: RoutingMode;
-    executor?: ExecutorConfig;
-}
-/**
- * Resolve the effective routing mode from CLI flag, env var, and config.
- *
- * Activation requires either the flag or the env var; config alone never
- * activates routing (decision: explicit opt-in only). When activated without an
- * explicit mode, defaults to `executor-for-summarization` (the lowest-risk
- * mode). When not activated, always `primary-only`.
- */
-export declare function resolveRoutingMode(opts: {
-    enableFlag?: boolean;
-    envMode?: string;
-    configMode?: RoutingMode;
-}): RoutingMode;
-/**
- * Router that decides, per turn kind, whether to use the executor model and
- * resolves it from the registry. Holds no mutable state beyond the resolved
- * executor model.
- */
-export declare class LocalInferenceRouter {
-    private readonly mode;
-    private readonly executorConfig;
-    private readonly executor;
-    private readonly minBytes;
-    private readonly maxBytes;
-    private constructor();
-    static create(opts: {
-        mode: RoutingMode;
-        config: RoutingConfig | undefined;
-        registry: ModelRegistry;
-    }): LocalInferenceRouter;
-    getMode(): RoutingMode;
-    /** True when routing is active and an executor model is resolved and usable. */
-    isExecutorAvailable(): boolean;
-    getExecutorConfig(): ExecutorConfig | undefined;
-    /**
-     * Pick the model to use for a turn. Returns the executor when the mode routes
-     * that turn kind and the executor is available; otherwise returns the primary
-     * model.
-     */
-    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 +0,0 @@
-{"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,GAAG,cAAc,GAAG,4BAA4B,GAAG,2BAA2B,CAAC;AAEtG,eAAO,MAAM,aAAa,EAAE,SAAS,WAAW,EAItC,CAAC;AAEX;;;;;;GAMG;AACH,eAAO,MAAM,kBAAkB,aAAoB,CAAC;AAEpD,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;;;;;;;;;GASG;AACH,MAAM,WAAW,cAAc;IAC9B,QAAQ,EAAE,MAAM,CAAC;IACjB,KAAK,EAAE,MAAM,CAAC;IACd,mGAAmG;IACnG,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,qGAAqG;IACrG,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,eAaN;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;;;;OAIG;IACH,WAAW,CAAC,QAAQ,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC,GAAG,CAAC,CAU/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 = \"primary-only\" | \"executor-for-summarization\" | \"executor-for-tool-results\";\n\nexport const ROUTING_MODES: readonly RoutingMode[] = [\n\t\"primary-only\",\n\t\"executor-for-summarization\",\n\t\"executor-for-tool-results\",\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/** 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/**\n * Executor model reference as configured in models.json.\n *\n * The size band (`minBytes`/`maxBytes`) is wired entirely from config — there\n * are no built-in byte defaults. When a bound is omitted it is not applied:\n * `minBytes` defaults to 0 (no lower gate) and `maxBytes` to unbounded. Set\n * both in models.json to gate which inputs route to the executor; on local\n * hardware a `maxBytes` guards against slow runs / GPU OOM, while a hosted or\n * large-memory executor can leave it unset (or high) to offload large inputs.\n */\nexport interface ExecutorConfig {\n\tprovider: string;\n\tmodel: string;\n\t/** Minimum input size (bytes) before local inference is attempted. Omitted = 0 (no lower gate). */\n\tminBytes?: number;\n\t/** Maximum input size (bytes); larger inputs fall back to the primary model. Omitted = unbounded. */\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\t// No built-in byte defaults: an omitted bound is simply not applied\n\t\t// (min 0 = no lower gate, max +Infinity = unbounded). The band is wired\n\t\t// entirely from the models.json executor config.\n\t\tthis.minBytes = executorConfig?.minBytes ?? 0;\n\t\tthis.maxBytes = executorConfig?.maxBytes ?? Number.POSITIVE_INFINITY;\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.\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 \"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

@@ -1,145 +0,0 @@
-/**
- * Local-inference routing.
- *
- * Optional, opt-in routing of certain non-critical work (conversation
- * 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.
- *
- * Everything here is INERT unless explicitly enabled via the
- * `--enable-local-inference` flag or the `HOOCODE_ROUTING_MODE` env var. On any
- * executor resolution/availability problem the caller falls back to the primary
- * model (compaction) or the raw tool result (tool-result compression). The
- * router never throws into the agent loop.
- *
- * Design and validation: see docs/local-executor-routing.md.
- */
-export const ROUTING_MODES = [
-    "primary-only",
-    "executor-for-summarization",
-    "executor-for-tool-results",
-];
-/**
- * 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"]);
-function isRoutingMode(value) {
-    return typeof value === "string" && ROUTING_MODES.includes(value);
-}
-/**
- * Resolve the effective routing mode from CLI flag, env var, and config.
- *
- * Activation requires either the flag or the env var; config alone never
- * activates routing (decision: explicit opt-in only). When activated without an
- * explicit mode, defaults to `executor-for-summarization` (the lowest-risk
- * mode). When not activated, always `primary-only`.
- */
-export function resolveRoutingMode(opts) {
-    const envMode = opts.envMode?.trim();
-    const envActivates = envMode !== undefined && envMode !== "" && envMode !== "primary-only";
-    const activated = opts.enableFlag === true || envActivates;
-    if (!activated)
-        return "primary-only";
-    if (envMode && isRoutingMode(envMode))
-        return envMode;
-    if (opts.configMode && isRoutingMode(opts.configMode))
-        return opts.configMode;
-    return "executor-for-summarization";
-}
-/**
- * Router that decides, per turn kind, whether to use the executor model and
- * resolves it from the registry. Holds no mutable state beyond the resolved
- * executor model.
- */
-export class LocalInferenceRouter {
-    mode;
-    executorConfig;
-    executor;
-    minBytes;
-    maxBytes;
-    constructor(mode, executorConfig, executor) {
-        this.mode = mode;
-        this.executorConfig = executorConfig;
-        this.executor = executor;
-        // No built-in byte defaults: an omitted bound is simply not applied
-        // (min 0 = no lower gate, max +Infinity = unbounded). The band is wired
-        // entirely from the models.json executor config.
-        this.minBytes = executorConfig?.minBytes ?? 0;
-        this.maxBytes = executorConfig?.maxBytes ?? Number.POSITIVE_INFINITY;
-    }
-    static create(opts) {
-        const executorConfig = opts.config?.executor;
-        let executor;
-        if (opts.mode !== "primary-only" && executorConfig) {
-            executor = opts.registry.find(executorConfig.provider, executorConfig.model);
-        }
-        return new LocalInferenceRouter(opts.mode, executorConfig, executor);
-    }
-    getMode() {
-        return this.mode;
-    }
-    /** True when routing is active and an executor model is resolved and usable. */
-    isExecutorAvailable() {
-        return this.mode !== "primary-only" && this.executor !== undefined;
-    }
-    getExecutorConfig() {
-        return this.executorConfig;
-    }
-    /**
-     * Pick the model to use for a turn. Returns the executor when the mode routes
-     * that turn kind and the executor is available; otherwise returns the primary
-     * model.
-     */
-    selectModel(turnKind, primary) {
-        if (!this.isExecutorAvailable() || !this.executor)
-            return primary;
-        switch (this.mode) {
-            case "executor-for-summarization":
-                return turnKind === "summarization" ? this.executor : primary;
-            case "executor-for-tool-results":
-                return turnKind === "tool-result" ? this.executor : primary;
-            case "primary-only":
-                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")
-            return false;
-        if (!this.isExecutorAvailable())
-            return false;
-        if (!COMPRESSIBLE_TOOLS.has(toolName))
-            return false;
-        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;
-    }
-}
-//# sourceMappingURL=local-inference.js.map

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

@@ -1 +0,0 @@
-{"version":3,"file":"local-inference.js","sourceRoot":"","sources":["../../../src/core/routing/local-inference.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAUH,MAAM,CAAC,MAAM,aAAa,GAA2B;IACpD,cAAc;IACd,4BAA4B;IAC5B,2BAA2B;CAClB,CAAC;AAEX;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAG,IAAI,GAAG,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;AA2CpD,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,oEAAoE;QACpE,wEAAwE;QACxE,iDAAiD;QACjD,IAAI,CAAC,QAAQ,GAAG,cAAc,EAAE,QAAQ,IAAI,CAAC,CAAC;QAC9C,IAAI,CAAC,QAAQ,GAAG,cAAc,EAAE,QAAQ,IAAI,MAAM,CAAC,iBAAiB,CAAC;IAAA,CACrE;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;;;;OAIG;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,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 = \"primary-only\" | \"executor-for-summarization\" | \"executor-for-tool-results\";\n\nexport const ROUTING_MODES: readonly RoutingMode[] = [\n\t\"primary-only\",\n\t\"executor-for-summarization\",\n\t\"executor-for-tool-results\",\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/** 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/**\n * Executor model reference as configured in models.json.\n *\n * The size band (`minBytes`/`maxBytes`) is wired entirely from config — there\n * are no built-in byte defaults. When a bound is omitted it is not applied:\n * `minBytes` defaults to 0 (no lower gate) and `maxBytes` to unbounded. Set\n * both in models.json to gate which inputs route to the executor; on local\n * hardware a `maxBytes` guards against slow runs / GPU OOM, while a hosted or\n * large-memory executor can leave it unset (or high) to offload large inputs.\n */\nexport interface ExecutorConfig {\n\tprovider: string;\n\tmodel: string;\n\t/** Minimum input size (bytes) before local inference is attempted. Omitted = 0 (no lower gate). */\n\tminBytes?: number;\n\t/** Maximum input size (bytes); larger inputs fall back to the primary model. Omitted = unbounded. */\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\t// No built-in byte defaults: an omitted bound is simply not applied\n\t\t// (min 0 = no lower gate, max +Infinity = unbounded). The band is wired\n\t\t// entirely from the models.json executor config.\n\t\tthis.minBytes = executorConfig?.minBytes ?? 0;\n\t\tthis.maxBytes = executorConfig?.maxBytes ?? Number.POSITIVE_INFINITY;\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.\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 \"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/metrics.d.ts

@@ -1,27 +0,0 @@
-/**
- * Local-inference routing metrics and diagnostic logging.
- *
- * Routing is an optimization, so failures must never surface as errors in the
- * agent UI. This module provides a single, quiet sink for routing diagnostics
- * and per-turn metrics. Output is gated behind HOOCODE_ROUTING_DEBUG so normal
- * sessions stay silent (decision: fall back, log silently, never hard-fail).
- */
-import type { TurnKind } from "./local-inference.js";
-/** Record that an executor turn failed and the caller fell back to primary/raw. */
-export declare function logLocalInferenceFallback(turnKind: TurnKind, error: unknown): void;
-/** Per-turn routing metrics for cost/perf comparison. */
-export interface RoutingMetrics {
-    turnKind: TurnKind;
-    provider: string;
-    model: string;
-    inputTokens?: number;
-    outputTokens?: number;
-    latencyMs: number;
-    /** Bytes before/after for tool-result compression. */
-    bytesBefore?: number;
-    bytesAfter?: number;
-    fallback: boolean;
-}
-/** Record per-turn routing metrics (quiet unless HOOCODE_ROUTING_DEBUG). */
-export declare function logRoutingMetrics(m: RoutingMetrics): void;
-//# sourceMappingURL=metrics.d.ts.map

package/dist/core/routing/metrics.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"metrics.d.ts","sourceRoot":"","sources":["../../../src/core/routing/metrics.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,sBAAsB,CAAC;AAOrD,mFAAmF;AACnF,wBAAgB,yBAAyB,CAAC,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,OAAO,GAAG,IAAI,CAIlF;AAED,yDAAyD;AACzD,MAAM,WAAW,cAAc;IAC9B,QAAQ,EAAE,QAAQ,CAAC;IACnB,QAAQ,EAAE,MAAM,CAAC;IACjB,KAAK,EAAE,MAAM,CAAC;IACd,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,SAAS,EAAE,MAAM,CAAC;IAClB,sDAAsD;IACtD,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,QAAQ,EAAE,OAAO,CAAC;CAClB;AAED,4EAA4E;AAC5E,wBAAgB,iBAAiB,CAAC,CAAC,EAAE,cAAc,GAAG,IAAI,CAYzD","sourcesContent":["/**\n * Local-inference routing metrics and diagnostic logging.\n *\n * Routing is an optimization, so failures must never surface as errors in the\n * agent UI. This module provides a single, quiet sink for routing diagnostics\n * and per-turn metrics. Output is gated behind HOOCODE_ROUTING_DEBUG so normal\n * sessions stay silent (decision: fall back, log silently, never hard-fail).\n */\n\nimport type { TurnKind } from \"./local-inference.js\";\n\nfunction debugEnabled(): boolean {\n\tconst v = process.env.HOOCODE_ROUTING_DEBUG;\n\treturn v === \"1\" || v === \"true\" || v === \"yes\";\n}\n\n/** Record that an executor turn failed and the caller fell back to primary/raw. */\nexport function logLocalInferenceFallback(turnKind: TurnKind, error: unknown): void {\n\tif (!debugEnabled()) return;\n\tconst message = error instanceof Error ? error.message : String(error);\n\tconsole.warn(`[local-inference] ${turnKind} fell back to primary: ${message}`);\n}\n\n/** Per-turn routing metrics for cost/perf comparison. */\nexport interface RoutingMetrics {\n\tturnKind: TurnKind;\n\tprovider: string;\n\tmodel: string;\n\tinputTokens?: number;\n\toutputTokens?: number;\n\tlatencyMs: number;\n\t/** Bytes before/after for tool-result compression. */\n\tbytesBefore?: number;\n\tbytesAfter?: number;\n\tfallback: boolean;\n}\n\n/** Record per-turn routing metrics (quiet unless HOOCODE_ROUTING_DEBUG). */\nexport function logRoutingMetrics(m: RoutingMetrics): void {\n\tif (!debugEnabled()) return;\n\tconst parts = [\n\t\t`turn=${m.turnKind}`,\n\t\t`model=${m.provider}/${m.model}`,\n\t\t`latency=${m.latencyMs}ms`,\n\t\tm.inputTokens !== undefined ? `in=${m.inputTokens}` : undefined,\n\t\tm.outputTokens !== undefined ? `out=${m.outputTokens}` : undefined,\n\t\tm.bytesBefore !== undefined && m.bytesAfter !== undefined ? `bytes=${m.bytesBefore}->${m.bytesAfter}` : undefined,\n\t\tm.fallback ? \"fallback=true\" : undefined,\n\t].filter((p): p is string => p !== undefined);\n\tconsole.error(`[local-inference] ${parts.join(\" \")}`);\n}\n"]}

package/dist/core/routing/metrics.js

@@ -1,35 +0,0 @@
-/**
- * Local-inference routing metrics and diagnostic logging.
- *
- * Routing is an optimization, so failures must never surface as errors in the
- * agent UI. This module provides a single, quiet sink for routing diagnostics
- * and per-turn metrics. Output is gated behind HOOCODE_ROUTING_DEBUG so normal
- * sessions stay silent (decision: fall back, log silently, never hard-fail).
- */
-function debugEnabled() {
-    const v = process.env.HOOCODE_ROUTING_DEBUG;
-    return v === "1" || v === "true" || v === "yes";
-}
-/** Record that an executor turn failed and the caller fell back to primary/raw. */
-export function logLocalInferenceFallback(turnKind, error) {
-    if (!debugEnabled())
-        return;
-    const message = error instanceof Error ? error.message : String(error);
-    console.warn(`[local-inference] ${turnKind} fell back to primary: ${message}`);
-}
-/** Record per-turn routing metrics (quiet unless HOOCODE_ROUTING_DEBUG). */
-export function logRoutingMetrics(m) {
-    if (!debugEnabled())
-        return;
-    const parts = [
-        `turn=${m.turnKind}`,
-        `model=${m.provider}/${m.model}`,
-        `latency=${m.latencyMs}ms`,
-        m.inputTokens !== undefined ? `in=${m.inputTokens}` : undefined,
-        m.outputTokens !== undefined ? `out=${m.outputTokens}` : undefined,
-        m.bytesBefore !== undefined && m.bytesAfter !== undefined ? `bytes=${m.bytesBefore}->${m.bytesAfter}` : undefined,
-        m.fallback ? "fallback=true" : undefined,
-    ].filter((p) => p !== undefined);
-    console.error(`[local-inference] ${parts.join(" ")}`);
-}
-//# sourceMappingURL=metrics.js.map

package/dist/core/routing/metrics.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"metrics.js","sourceRoot":"","sources":["../../../src/core/routing/metrics.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH,SAAS,YAAY,GAAY;IAChC,MAAM,CAAC,GAAG,OAAO,CAAC,GAAG,CAAC,qBAAqB,CAAC;IAC5C,OAAO,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK,MAAM,IAAI,CAAC,KAAK,KAAK,CAAC;AAAA,CAChD;AAED,mFAAmF;AACnF,MAAM,UAAU,yBAAyB,CAAC,QAAkB,EAAE,KAAc,EAAQ;IACnF,IAAI,CAAC,YAAY,EAAE;QAAE,OAAO;IAC5B,MAAM,OAAO,GAAG,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;IACvE,OAAO,CAAC,IAAI,CAAC,qBAAqB,QAAQ,0BAA0B,OAAO,EAAE,CAAC,CAAC;AAAA,CAC/E;AAgBD,4EAA4E;AAC5E,MAAM,UAAU,iBAAiB,CAAC,CAAiB,EAAQ;IAC1D,IAAI,CAAC,YAAY,EAAE;QAAE,OAAO;IAC5B,MAAM,KAAK,GAAG;QACb,QAAQ,CAAC,CAAC,QAAQ,EAAE;QACpB,SAAS,CAAC,CAAC,QAAQ,IAAI,CAAC,CAAC,KAAK,EAAE;QAChC,WAAW,CAAC,CAAC,SAAS,IAAI;QAC1B,CAAC,CAAC,WAAW,KAAK,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC,SAAS;QAC/D,CAAC,CAAC,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,YAAY,EAAE,CAAC,CAAC,CAAC,SAAS;QAClE,CAAC,CAAC,WAAW,KAAK,SAAS,IAAI,CAAC,CAAC,UAAU,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,WAAW,KAAK,CAAC,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC,SAAS;QACjH,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC,CAAC,SAAS;KACxC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAe,EAAE,CAAC,CAAC,KAAK,SAAS,CAAC,CAAC;IAC9C,OAAO,CAAC,KAAK,CAAC,qBAAqB,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;AAAA,CACtD","sourcesContent":["/**\n * Local-inference routing metrics and diagnostic logging.\n *\n * Routing is an optimization, so failures must never surface as errors in the\n * agent UI. This module provides a single, quiet sink for routing diagnostics\n * and per-turn metrics. Output is gated behind HOOCODE_ROUTING_DEBUG so normal\n * sessions stay silent (decision: fall back, log silently, never hard-fail).\n */\n\nimport type { TurnKind } from \"./local-inference.js\";\n\nfunction debugEnabled(): boolean {\n\tconst v = process.env.HOOCODE_ROUTING_DEBUG;\n\treturn v === \"1\" || v === \"true\" || v === \"yes\";\n}\n\n/** Record that an executor turn failed and the caller fell back to primary/raw. */\nexport function logLocalInferenceFallback(turnKind: TurnKind, error: unknown): void {\n\tif (!debugEnabled()) return;\n\tconst message = error instanceof Error ? error.message : String(error);\n\tconsole.warn(`[local-inference] ${turnKind} fell back to primary: ${message}`);\n}\n\n/** Per-turn routing metrics for cost/perf comparison. */\nexport interface RoutingMetrics {\n\tturnKind: TurnKind;\n\tprovider: string;\n\tmodel: string;\n\tinputTokens?: number;\n\toutputTokens?: number;\n\tlatencyMs: number;\n\t/** Bytes before/after for tool-result compression. */\n\tbytesBefore?: number;\n\tbytesAfter?: number;\n\tfallback: boolean;\n}\n\n/** Record per-turn routing metrics (quiet unless HOOCODE_ROUTING_DEBUG). */\nexport function logRoutingMetrics(m: RoutingMetrics): void {\n\tif (!debugEnabled()) return;\n\tconst parts = [\n\t\t`turn=${m.turnKind}`,\n\t\t`model=${m.provider}/${m.model}`,\n\t\t`latency=${m.latencyMs}ms`,\n\t\tm.inputTokens !== undefined ? `in=${m.inputTokens}` : undefined,\n\t\tm.outputTokens !== undefined ? `out=${m.outputTokens}` : undefined,\n\t\tm.bytesBefore !== undefined && m.bytesAfter !== undefined ? `bytes=${m.bytesBefore}->${m.bytesAfter}` : undefined,\n\t\tm.fallback ? \"fallback=true\" : undefined,\n\t].filter((p): p is string => p !== undefined);\n\tconsole.error(`[local-inference] ${parts.join(\" \")}`);\n}\n"]}

package/dist/core/routing/mlx-server.d.ts

@@ -1,43 +0,0 @@
-/**
- * Local executor server lifecycle manager.
- *
- * When `routing.executor.server` is configured AND local-inference routing is
- * active, the harness can spawn a local OpenAI-compatible server (for example
- * `mlx_lm.server`), wait for it to become healthy, and stop it on shutdown.
- *
- * This is best-effort: if the server fails to start or never becomes healthy,
- * `ensureStarted()` resolves to false and the caller degrades to the primary
- * model (routing falls back). It never throws into the agent loop.
- */
-import type { ExecutorServerConfig } from "./local-inference.js";
-export interface MlxServerOptions {
-    config: ExecutorServerConfig;
-    /** Executor model id, passed as --model. */
-    modelId: string;
-    /** Executor baseUrl, used to derive host/port and health endpoint. */
-    baseUrl?: string;
-}
-export declare class MlxServerManager {
-    private readonly command;
-    private readonly args;
-    private readonly host;
-    private readonly port;
-    private readonly startupTimeoutMs;
-    private readonly modelId;
-    private child;
-    private startPromise;
-    constructor(opts: MlxServerOptions);
-    private healthUrl;
-    private isHealthy;
-    /**
-     * Ensure the server is running and healthy. Idempotent: concurrent callers
-     * share one start attempt. Returns false on any failure (caller degrades to
-     * primary). If a server is already healthy (user-started), reuses it without
-     * spawning.
-     */
-    ensureStarted(signal?: AbortSignal): Promise<boolean>;
-    private start;
-    /** Stop the spawned server (no-op if we reused an external one). */
-    stop(): void;
-}
-//# sourceMappingURL=mlx-server.d.ts.map

package/dist/core/routing/mlx-server.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"mlx-server.d.ts","sourceRoot":"","sources":["../../../src/core/routing/mlx-server.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAGH,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,sBAAsB,CAAC;AAoBjE,MAAM,WAAW,gBAAgB;IAChC,MAAM,EAAE,oBAAoB,CAAC;IAC7B,4CAA4C;IAC5C,OAAO,EAAE,MAAM,CAAC;IAChB,sEAAsE;IACtE,OAAO,CAAC,EAAE,MAAM,CAAC;CACjB;AAED,qBAAa,gBAAgB;IAC5B,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAS;IACjC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAW;IAChC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAS;IAC9B,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAS;IAC9B,OAAO,CAAC,QAAQ,CAAC,gBAAgB,CAAS;IAC1C,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAS;IACjC,OAAO,CAAC,KAAK,CAA2B;IACxC,OAAO,CAAC,YAAY,CAA+B;IAEnD,YAAY,IAAI,EAAE,gBAAgB,EAQjC;IAED,OAAO,CAAC,SAAS;YAIH,SAAS;IASvB;;;;;OAKG;IACG,aAAa,CAAC,MAAM,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,OAAO,CAAC,CAI1D;YAEa,KAAK;IAmCnB,oEAAoE;IACpE,IAAI,IAAI,IAAI,CAMX;CACD","sourcesContent":["/**\n * Local executor server lifecycle manager.\n *\n * When `routing.executor.server` is configured AND local-inference routing is\n * active, the harness can spawn a local OpenAI-compatible server (for example\n * `mlx_lm.server`), wait for it to become healthy, and stop it on shutdown.\n *\n * This is best-effort: if the server fails to start or never becomes healthy,\n * `ensureStarted()` resolves to false and the caller degrades to the primary\n * model (routing falls back). It never throws into the agent loop.\n */\n\nimport { type ChildProcess, spawn } from \"node:child_process\";\nimport type { ExecutorServerConfig } from \"./local-inference.js\";\nimport { logLocalInferenceFallback } from \"./metrics.js\";\n\nconst DEFAULT_COMMAND = \"mlx_lm.server\";\nconst DEFAULT_HOST = \"127.0.0.1\";\nconst DEFAULT_PORT = 8080;\nconst DEFAULT_STARTUP_TIMEOUT_MS = 30_000;\nconst HEALTH_POLL_INTERVAL_MS = 500;\n\n/** Derive host/port from a baseUrl like \"http://127.0.0.1:8080/v1\". */\nfunction parseBaseUrl(baseUrl: string | undefined): { host?: string; port?: number } {\n\tif (!baseUrl) return {};\n\ttry {\n\t\tconst u = new URL(baseUrl);\n\t\treturn { host: u.hostname, port: u.port ? Number(u.port) : undefined };\n\t} catch {\n\t\treturn {};\n\t}\n}\n\nexport interface MlxServerOptions {\n\tconfig: ExecutorServerConfig;\n\t/** Executor model id, passed as --model. */\n\tmodelId: string;\n\t/** Executor baseUrl, used to derive host/port and health endpoint. */\n\tbaseUrl?: string;\n}\n\nexport class MlxServerManager {\n\tprivate readonly command: string;\n\tprivate readonly args: string[];\n\tprivate readonly host: string;\n\tprivate readonly port: number;\n\tprivate readonly startupTimeoutMs: number;\n\tprivate readonly modelId: string;\n\tprivate child: ChildProcess | undefined;\n\tprivate startPromise: Promise<boolean> | undefined;\n\n\tconstructor(opts: MlxServerOptions) {\n\t\tconst fromUrl = parseBaseUrl(opts.baseUrl);\n\t\tthis.command = opts.config.command ?? DEFAULT_COMMAND;\n\t\tthis.args = opts.config.args ?? [];\n\t\tthis.host = opts.config.host ?? fromUrl.host ?? DEFAULT_HOST;\n\t\tthis.port = opts.config.port ?? fromUrl.port ?? DEFAULT_PORT;\n\t\tthis.startupTimeoutMs = opts.config.startupTimeoutMs ?? DEFAULT_STARTUP_TIMEOUT_MS;\n\t\tthis.modelId = opts.modelId;\n\t}\n\n\tprivate healthUrl(): string {\n\t\treturn `http://${this.host}:${this.port}/v1/models`;\n\t}\n\n\tprivate async isHealthy(signal?: AbortSignal): Promise<boolean> {\n\t\ttry {\n\t\t\tconst res = await fetch(this.healthUrl(), { method: \"GET\", signal });\n\t\t\treturn res.ok;\n\t\t} catch {\n\t\t\treturn false;\n\t\t}\n\t}\n\n\t/**\n\t * Ensure the server is running and healthy. Idempotent: concurrent callers\n\t * share one start attempt. Returns false on any failure (caller degrades to\n\t * primary). If a server is already healthy (user-started), reuses it without\n\t * spawning.\n\t */\n\tasync ensureStarted(signal?: AbortSignal): Promise<boolean> {\n\t\tif (this.startPromise) return this.startPromise;\n\t\tthis.startPromise = this.start(signal);\n\t\treturn this.startPromise;\n\t}\n\n\tprivate async start(signal?: AbortSignal): Promise<boolean> {\n\t\t// Reuse an already-running server (e.g. user-started) without spawning.\n\t\tif (await this.isHealthy(signal)) return true;\n\n\t\ttry {\n\t\t\tthis.child = spawn(\n\t\t\t\tthis.command,\n\t\t\t\t[...this.args, \"--model\", this.modelId, \"--host\", this.host, \"--port\", String(this.port)],\n\t\t\t\t{\n\t\t\t\t\tstdio: \"ignore\",\n\t\t\t\t\tdetached: false,\n\t\t\t\t},\n\t\t\t);\n\t\t\tthis.child.on(\"error\", (error) => {\n\t\t\t\tlogLocalInferenceFallback(\"primary\", error);\n\t\t\t\tthis.child = undefined;\n\t\t\t});\n\t\t} catch (error) {\n\t\t\tlogLocalInferenceFallback(\"primary\", error);\n\t\t\tthis.child = undefined;\n\t\t\treturn false;\n\t\t}\n\n\t\tconst deadline = Date.now() + this.startupTimeoutMs;\n\t\twhile (Date.now() < deadline) {\n\t\t\tif (signal?.aborted) return false;\n\t\t\tif (!this.child) return false; // spawn errored\n\t\t\tif (await this.isHealthy(signal)) return true;\n\t\t\tawait new Promise((r) => setTimeout(r, HEALTH_POLL_INTERVAL_MS));\n\t\t}\n\t\t// Timed out: stop whatever we spawned and degrade.\n\t\tthis.stop();\n\t\treturn false;\n\t}\n\n\t/** Stop the spawned server (no-op if we reused an external one). */\n\tstop(): void {\n\t\tif (this.child && !this.child.killed) {\n\t\t\tthis.child.kill(\"SIGTERM\");\n\t\t}\n\t\tthis.child = undefined;\n\t\tthis.startPromise = undefined;\n\t}\n}\n"]}

package/dist/core/routing/mlx-server.js

@@ -1,115 +0,0 @@
-/**
- * Local executor server lifecycle manager.
- *
- * When `routing.executor.server` is configured AND local-inference routing is
- * active, the harness can spawn a local OpenAI-compatible server (for example
- * `mlx_lm.server`), wait for it to become healthy, and stop it on shutdown.
- *
- * This is best-effort: if the server fails to start or never becomes healthy,
- * `ensureStarted()` resolves to false and the caller degrades to the primary
- * model (routing falls back). It never throws into the agent loop.
- */
-import { spawn } from "node:child_process";
-import { logLocalInferenceFallback } from "./metrics.js";
-const DEFAULT_COMMAND = "mlx_lm.server";
-const DEFAULT_HOST = "127.0.0.1";
-const DEFAULT_PORT = 8080;
-const DEFAULT_STARTUP_TIMEOUT_MS = 30_000;
-const HEALTH_POLL_INTERVAL_MS = 500;
-/** Derive host/port from a baseUrl like "http://127.0.0.1:8080/v1". */
-function parseBaseUrl(baseUrl) {
-    if (!baseUrl)
-        return {};
-    try {
-        const u = new URL(baseUrl);
-        return { host: u.hostname, port: u.port ? Number(u.port) : undefined };
-    }
-    catch {
-        return {};
-    }
-}
-export class MlxServerManager {
-    command;
-    args;
-    host;
-    port;
-    startupTimeoutMs;
-    modelId;
-    child;
-    startPromise;
-    constructor(opts) {
-        const fromUrl = parseBaseUrl(opts.baseUrl);
-        this.command = opts.config.command ?? DEFAULT_COMMAND;
-        this.args = opts.config.args ?? [];
-        this.host = opts.config.host ?? fromUrl.host ?? DEFAULT_HOST;
-        this.port = opts.config.port ?? fromUrl.port ?? DEFAULT_PORT;
-        this.startupTimeoutMs = opts.config.startupTimeoutMs ?? DEFAULT_STARTUP_TIMEOUT_MS;
-        this.modelId = opts.modelId;
-    }
-    healthUrl() {
-        return `http://${this.host}:${this.port}/v1/models`;
-    }
-    async isHealthy(signal) {
-        try {
-            const res = await fetch(this.healthUrl(), { method: "GET", signal });
-            return res.ok;
-        }
-        catch {
-            return false;
-        }
-    }
-    /**
-     * Ensure the server is running and healthy. Idempotent: concurrent callers
-     * share one start attempt. Returns false on any failure (caller degrades to
-     * primary). If a server is already healthy (user-started), reuses it without
-     * spawning.
-     */
-    async ensureStarted(signal) {
-        if (this.startPromise)
-            return this.startPromise;
-        this.startPromise = this.start(signal);
-        return this.startPromise;
-    }
-    async start(signal) {
-        // Reuse an already-running server (e.g. user-started) without spawning.
-        if (await this.isHealthy(signal))
-            return true;
-        try {
-            this.child = spawn(this.command, [...this.args, "--model", this.modelId, "--host", this.host, "--port", String(this.port)], {
-                stdio: "ignore",
-                detached: false,
-            });
-            this.child.on("error", (error) => {
-                logLocalInferenceFallback("primary", error);
-                this.child = undefined;
-            });
-        }
-        catch (error) {
-            logLocalInferenceFallback("primary", error);
-            this.child = undefined;
-            return false;
-        }
-        const deadline = Date.now() + this.startupTimeoutMs;
-        while (Date.now() < deadline) {
-            if (signal?.aborted)
-                return false;
-            if (!this.child)
-                return false; // spawn errored
-            if (await this.isHealthy(signal))
-                return true;
-            await new Promise((r) => setTimeout(r, HEALTH_POLL_INTERVAL_MS));
-        }
-        // Timed out: stop whatever we spawned and degrade.
-        this.stop();
-        return false;
-    }
-    /** Stop the spawned server (no-op if we reused an external one). */
-    stop() {
-        if (this.child && !this.child.killed) {
-            this.child.kill("SIGTERM");
-        }
-        this.child = undefined;
-        this.startPromise = undefined;
-    }
-}
-//# sourceMappingURL=mlx-server.js.map

package/dist/core/routing/mlx-server.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"mlx-server.js","sourceRoot":"","sources":["../../../src/core/routing/mlx-server.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAqB,KAAK,EAAE,MAAM,oBAAoB,CAAC;AAE9D,OAAO,EAAE,yBAAyB,EAAE,MAAM,cAAc,CAAC;AAEzD,MAAM,eAAe,GAAG,eAAe,CAAC;AACxC,MAAM,YAAY,GAAG,WAAW,CAAC;AACjC,MAAM,YAAY,GAAG,IAAI,CAAC;AAC1B,MAAM,0BAA0B,GAAG,MAAM,CAAC;AAC1C,MAAM,uBAAuB,GAAG,GAAG,CAAC;AAEpC,uEAAuE;AACvE,SAAS,YAAY,CAAC,OAA2B,EAAoC;IACpF,IAAI,CAAC,OAAO;QAAE,OAAO,EAAE,CAAC;IACxB,IAAI,CAAC;QACJ,MAAM,CAAC,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,CAAC;QAC3B,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,QAAQ,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,SAAS,EAAE,CAAC;IACxE,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,EAAE,CAAC;IACX,CAAC;AAAA,CACD;AAUD,MAAM,OAAO,gBAAgB;IACX,OAAO,CAAS;IAChB,IAAI,CAAW;IACf,IAAI,CAAS;IACb,IAAI,CAAS;IACb,gBAAgB,CAAS;IACzB,OAAO,CAAS;IACzB,KAAK,CAA2B;IAChC,YAAY,CAA+B;IAEnD,YAAY,IAAsB,EAAE;QACnC,MAAM,OAAO,GAAG,YAAY,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;QAC3C,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,MAAM,CAAC,OAAO,IAAI,eAAe,CAAC;QACtD,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC,MAAM,CAAC,IAAI,IAAI,EAAE,CAAC;QACnC,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC,MAAM,CAAC,IAAI,IAAI,OAAO,CAAC,IAAI,IAAI,YAAY,CAAC;QAC7D,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC,MAAM,CAAC,IAAI,IAAI,OAAO,CAAC,IAAI,IAAI,YAAY,CAAC;QAC7D,IAAI,CAAC,gBAAgB,GAAG,IAAI,CAAC,MAAM,CAAC,gBAAgB,IAAI,0BAA0B,CAAC;QACnF,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC;IAAA,CAC5B;IAEO,SAAS,GAAW;QAC3B,OAAO,UAAU,IAAI,CAAC,IAAI,IAAI,IAAI,CAAC,IAAI,YAAY,CAAC;IAAA,CACpD;IAEO,KAAK,CAAC,SAAS,CAAC,MAAoB,EAAoB;QAC/D,IAAI,CAAC;YACJ,MAAM,GAAG,GAAG,MAAM,KAAK,CAAC,IAAI,CAAC,SAAS,EAAE,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC,CAAC;YACrE,OAAO,GAAG,CAAC,EAAE,CAAC;QACf,CAAC;QAAC,MAAM,CAAC;YACR,OAAO,KAAK,CAAC;QACd,CAAC;IAAA,CACD;IAED;;;;;OAKG;IACH,KAAK,CAAC,aAAa,CAAC,MAAoB,EAAoB;QAC3D,IAAI,IAAI,CAAC,YAAY;YAAE,OAAO,IAAI,CAAC,YAAY,CAAC;QAChD,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;QACvC,OAAO,IAAI,CAAC,YAAY,CAAC;IAAA,CACzB;IAEO,KAAK,CAAC,KAAK,CAAC,MAAoB,EAAoB;QAC3D,wEAAwE;QACxE,IAAI,MAAM,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC;YAAE,OAAO,IAAI,CAAC;QAE9C,IAAI,CAAC;YACJ,IAAI,CAAC,KAAK,GAAG,KAAK,CACjB,IAAI,CAAC,OAAO,EACZ,CAAC,GAAG,IAAI,CAAC,IAAI,EAAE,SAAS,EAAE,IAAI,CAAC,OAAO,EAAE,QAAQ,EAAE,IAAI,CAAC,IAAI,EAAE,QAAQ,EAAE,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,EACzF;gBACC,KAAK,EAAE,QAAQ;gBACf,QAAQ,EAAE,KAAK;aACf,CACD,CAAC;YACF,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,OAAO,EAAE,CAAC,KAAK,EAAE,EAAE,CAAC;gBACjC,yBAAyB,CAAC,SAAS,EAAE,KAAK,CAAC,CAAC;gBAC5C,IAAI,CAAC,KAAK,GAAG,SAAS,CAAC;YAAA,CACvB,CAAC,CAAC;QACJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,yBAAyB,CAAC,SAAS,EAAE,KAAK,CAAC,CAAC;YAC5C,IAAI,CAAC,KAAK,GAAG,SAAS,CAAC;YACvB,OAAO,KAAK,CAAC;QACd,CAAC;QAED,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,IAAI,CAAC,gBAAgB,CAAC;QACpD,OAAO,IAAI,CAAC,GAAG,EAAE,GAAG,QAAQ,EAAE,CAAC;YAC9B,IAAI,MAAM,EAAE,OAAO;gBAAE,OAAO,KAAK,CAAC;YAClC,IAAI,CAAC,IAAI,CAAC,KAAK;gBAAE,OAAO,KAAK,CAAC,CAAC,gBAAgB;YAC/C,IAAI,MAAM,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC;gBAAE,OAAO,IAAI,CAAC;YAC9C,MAAM,IAAI,OAAO,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,UAAU,CAAC,CAAC,EAAE,uBAAuB,CAAC,CAAC,CAAC;QAClE,CAAC;QACD,mDAAmD;QACnD,IAAI,CAAC,IAAI,EAAE,CAAC;QACZ,OAAO,KAAK,CAAC;IAAA,CACb;IAED,oEAAoE;IACpE,IAAI,GAAS;QACZ,IAAI,IAAI,CAAC,KAAK,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC;YACtC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;QAC5B,CAAC;QACD,IAAI,CAAC,KAAK,GAAG,SAAS,CAAC;QACvB,IAAI,CAAC,YAAY,GAAG,SAAS,CAAC;IAAA,CAC9B;CACD","sourcesContent":["/**\n * Local executor server lifecycle manager.\n *\n * When `routing.executor.server` is configured AND local-inference routing is\n * active, the harness can spawn a local OpenAI-compatible server (for example\n * `mlx_lm.server`), wait for it to become healthy, and stop it on shutdown.\n *\n * This is best-effort: if the server fails to start or never becomes healthy,\n * `ensureStarted()` resolves to false and the caller degrades to the primary\n * model (routing falls back). It never throws into the agent loop.\n */\n\nimport { type ChildProcess, spawn } from \"node:child_process\";\nimport type { ExecutorServerConfig } from \"./local-inference.js\";\nimport { logLocalInferenceFallback } from \"./metrics.js\";\n\nconst DEFAULT_COMMAND = \"mlx_lm.server\";\nconst DEFAULT_HOST = \"127.0.0.1\";\nconst DEFAULT_PORT = 8080;\nconst DEFAULT_STARTUP_TIMEOUT_MS = 30_000;\nconst HEALTH_POLL_INTERVAL_MS = 500;\n\n/** Derive host/port from a baseUrl like \"http://127.0.0.1:8080/v1\". */\nfunction parseBaseUrl(baseUrl: string | undefined): { host?: string; port?: number } {\n\tif (!baseUrl) return {};\n\ttry {\n\t\tconst u = new URL(baseUrl);\n\t\treturn { host: u.hostname, port: u.port ? Number(u.port) : undefined };\n\t} catch {\n\t\treturn {};\n\t}\n}\n\nexport interface MlxServerOptions {\n\tconfig: ExecutorServerConfig;\n\t/** Executor model id, passed as --model. */\n\tmodelId: string;\n\t/** Executor baseUrl, used to derive host/port and health endpoint. */\n\tbaseUrl?: string;\n}\n\nexport class MlxServerManager {\n\tprivate readonly command: string;\n\tprivate readonly args: string[];\n\tprivate readonly host: string;\n\tprivate readonly port: number;\n\tprivate readonly startupTimeoutMs: number;\n\tprivate readonly modelId: string;\n\tprivate child: ChildProcess | undefined;\n\tprivate startPromise: Promise<boolean> | undefined;\n\n\tconstructor(opts: MlxServerOptions) {\n\t\tconst fromUrl = parseBaseUrl(opts.baseUrl);\n\t\tthis.command = opts.config.command ?? DEFAULT_COMMAND;\n\t\tthis.args = opts.config.args ?? [];\n\t\tthis.host = opts.config.host ?? fromUrl.host ?? DEFAULT_HOST;\n\t\tthis.port = opts.config.port ?? fromUrl.port ?? DEFAULT_PORT;\n\t\tthis.startupTimeoutMs = opts.config.startupTimeoutMs ?? DEFAULT_STARTUP_TIMEOUT_MS;\n\t\tthis.modelId = opts.modelId;\n\t}\n\n\tprivate healthUrl(): string {\n\t\treturn `http://${this.host}:${this.port}/v1/models`;\n\t}\n\n\tprivate async isHealthy(signal?: AbortSignal): Promise<boolean> {\n\t\ttry {\n\t\t\tconst res = await fetch(this.healthUrl(), { method: \"GET\", signal });\n\t\t\treturn res.ok;\n\t\t} catch {\n\t\t\treturn false;\n\t\t}\n\t}\n\n\t/**\n\t * Ensure the server is running and healthy. Idempotent: concurrent callers\n\t * share one start attempt. Returns false on any failure (caller degrades to\n\t * primary). If a server is already healthy (user-started), reuses it without\n\t * spawning.\n\t */\n\tasync ensureStarted(signal?: AbortSignal): Promise<boolean> {\n\t\tif (this.startPromise) return this.startPromise;\n\t\tthis.startPromise = this.start(signal);\n\t\treturn this.startPromise;\n\t}\n\n\tprivate async start(signal?: AbortSignal): Promise<boolean> {\n\t\t// Reuse an already-running server (e.g. user-started) without spawning.\n\t\tif (await this.isHealthy(signal)) return true;\n\n\t\ttry {\n\t\t\tthis.child = spawn(\n\t\t\t\tthis.command,\n\t\t\t\t[...this.args, \"--model\", this.modelId, \"--host\", this.host, \"--port\", String(this.port)],\n\t\t\t\t{\n\t\t\t\t\tstdio: \"ignore\",\n\t\t\t\t\tdetached: false,\n\t\t\t\t},\n\t\t\t);\n\t\t\tthis.child.on(\"error\", (error) => {\n\t\t\t\tlogLocalInferenceFallback(\"primary\", error);\n\t\t\t\tthis.child = undefined;\n\t\t\t});\n\t\t} catch (error) {\n\t\t\tlogLocalInferenceFallback(\"primary\", error);\n\t\t\tthis.child = undefined;\n\t\t\treturn false;\n\t\t}\n\n\t\tconst deadline = Date.now() + this.startupTimeoutMs;\n\t\twhile (Date.now() < deadline) {\n\t\t\tif (signal?.aborted) return false;\n\t\t\tif (!this.child) return false; // spawn errored\n\t\t\tif (await this.isHealthy(signal)) return true;\n\t\t\tawait new Promise((r) => setTimeout(r, HEALTH_POLL_INTERVAL_MS));\n\t\t}\n\t\t// Timed out: stop whatever we spawned and degrade.\n\t\tthis.stop();\n\t\treturn false;\n\t}\n\n\t/** Stop the spawned server (no-op if we reused an external one). */\n\tstop(): void {\n\t\tif (this.child && !this.child.killed) {\n\t\t\tthis.child.kill(\"SIGTERM\");\n\t\t}\n\t\tthis.child = undefined;\n\t\tthis.startPromise = undefined;\n\t}\n}\n"]}

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

@@ -1,26 +0,0 @@
-/**
- * Validated per-tool extractive compression prompts.
- *
- * These prompts were validated against Qwen3-4B (see
- * 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 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;
-/** Get the extractive compression prompt for a tool, or undefined if not compressible. */
-export declare function getToolResultPrompt(toolName: string): string | undefined;
-/** Build the full prompt text for compressing a tool result. */
-export declare function buildToolResultPrompt(toolName: string, output: string): string | undefined;
-/**
- * Remove reasoning-model `<think>...</think>` blocks (including empty ones that
- * Qwen3 emits even under `/no_think`) and trim the result. Reasoning models
- * leave these tags in the text stream; they must not leak into context.
- */
-export declare function stripThinkTags(text: string): string;
-//# sourceMappingURL=tool-result-prompts.d.ts.map