@kolisachint/hoocode-agent 0.4.70 → 0.4.72

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

@@ -0,0 +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;;;;;;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/metrics.d.ts

@@ -0,0 +1,27 @@
+/**
+ * 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

@@ -0,0 +1 @@
+{"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

@@ -0,0 +1,35 @@
+/**
+ * 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

@@ -0,0 +1 @@
+{"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

@@ -0,0 +1,43 @@
+/**
+ * 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

@@ -0,0 +1 @@
+{"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

@@ -0,0 +1,115 @@
+/**
+ * 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

@@ -0,0 +1 @@
+{"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

@@ -0,0 +1,26 @@
+/**
+ * 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

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

@@ -0,0 +1 @@
+{"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

@@ -0,0 +1,45 @@
+/**
+ * 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 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 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 = {
+    bash: BASH_PROMPT,
+};
+/** Get the extractive compression prompt for a tool, or undefined if not compressible. */
+export function getToolResultPrompt(toolName) {
+    return TOOL_RESULT_PROMPTS[toolName];
+}
+/** Build the full prompt text for compressing a tool result. */
+export function buildToolResultPrompt(toolName, output) {
+    const instruction = getToolResultPrompt(toolName);
+    if (!instruction)
+        return undefined;
+    return `${instruction}\n\n${output}`;
+}
+/**
+ * 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 function stripThinkTags(text) {
+    return text
+        .replace(/<think>[\s\S]*?<\/think>/g, "")
+        .replace(/<\/?think>/g, "")
+        .trim();
+}
+//# sourceMappingURL=tool-result-prompts.js.map

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

@@ -0,0 +1 @@
+{"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/dist/core/sdk.d.ts

@@ -57,6 +57,12 @@ export interface CreateAgentSessionOptions {
     settingsManager?: SettingsManager;
     /** Session start event metadata for extension runtime startup. */
     sessionStartEvent?: SessionStartEvent;
+    /**
+     * Master gate for local-inference routing (compaction + tool-result
+     * compression via a local executor model). Off by default. Routing also
+     * requires a `routing` block in models.json. See docs/local-executor-routing.md.
+     */
+    enableLocalInference?: boolean;
 }
 /** Result from createAgentSession */
 export interface CreateAgentSessionResult {

package/dist/core/sdk.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"sdk.d.ts","sourceRoot":"","sources":["../../src/core/sdk.ts"],"names":[],"mappings":"AACA,OAAO,EAA4B,KAAK,aAAa,EAAE,MAAM,iCAAiC,CAAC;AAC/F,OAAO,EAAoC,KAAK,KAAK,EAAgB,MAAM,yBAAyB,CAAC;AAErG,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAElD,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAEhD,OAAO,KAAK,EAAmB,oBAAoB,EAAE,iBAAiB,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AAEtH,OAAO,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAC;AAEpD,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AAE3D,OAAO,EAAwB,cAAc,EAAE,MAAM,sBAAsB,CAAC;AAC5E,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AAIxD,OAAO,EACN,cAAc,EACd,iBAAiB,EACjB,cAAc,EACd,cAAc,EACd,cAAc,EACd,YAAY,EACZ,mBAAmB,EACnB,cAAc,EACd,eAAe,EAEf,qBAAqB,EACrB,MAAM,kBAAkB,CAAC;AAE1B,MAAM,WAAW,yBAAyB;IACzC,4EAA4E;IAC5E,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,yDAAyD;IACzD,QAAQ,CAAC,EAAE,MAAM,CAAC;IAElB,oFAAoF;IACpF,WAAW,CAAC,EAAE,WAAW,CAAC;IAC1B,uFAAuF;IACvF,aAAa,CAAC,EAAE,aAAa,CAAC;IAE9B,iEAAiE;IACjE,KAAK,CAAC,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;IACnB,4FAA4F;IAC5F,aAAa,CAAC,EAAE,aAAa,CAAC;IAC9B,gEAAgE;IAChE,YAAY,CAAC,EAAE,KAAK,CAAC;QAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;QAAC,aAAa,CAAC,EAAE,aAAa,CAAA;KAAE,CAAC,CAAC;IAE3E;;;;;;OAMG;IACH,OAAO,CAAC,EAAE,KAAK,GAAG,SAAS,CAAC;IAC5B;;;;;;OAMG;IACH,KAAK,CAAC,EAAE,MAAM,EAAE,CAAC;IACjB;;;OAGG;IACH,eAAe,CAAC,EAAE,MAAM,EAAE,CAAC;IAC3B,gEAAgE;IAChE,WAAW,CAAC,EAAE,cAAc,EAAE,CAAC;IAE/B,oEAAoE;IACpE,cAAc,CAAC,EAAE,cAAc,CAAC;IAEhC,2DAA2D;IAC3D,cAAc,CAAC,EAAE,cAAc,CAAC;IAEhC,uEAAuE;IACvE,eAAe,CAAC,EAAE,eAAe,CAAC;IAClC,kEAAkE;IAClE,iBAAiB,CAAC,EAAE,iBAAiB,CAAC;CACtC;AAED,qCAAqC;AACrC,MAAM,WAAW,wBAAwB;IACxC,0BAA0B;IAC1B,OAAO,EAAE,YAAY,CAAC;IACtB,mEAAmE;IACnE,gBAAgB,EAAE,oBAAoB,CAAC;IACvC,wEAAwE;IACxE,oBAAoB,CAAC,EAAE,MAAM,CAAC;CAC9B;AAID,YAAY,EAAE,eAAe,EAAE,WAAW,EAAE,MAAM,wBAAwB,CAAC;AAC3E,OAAO,EAAE,aAAa,EAAE,qBAAqB,EAAE,iBAAiB,EAAE,MAAM,qBAAqB,CAAC;AAC9F,cAAc,4BAA4B,CAAC;AAC3C,YAAY,EACX,YAAY,EACZ,uBAAuB,EACvB,gBAAgB,EAChB,gBAAgB,EAChB,gBAAgB,EAChB,kBAAkB,EAClB,cAAc,GACd,MAAM,uBAAuB,CAAC;AAC/B,YAAY,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AAC5D,YAAY,EAAE,KAAK,EAAE,MAAM,aAAa,CAAC;AACzC,YAAY,EAAE,IAAI,EAAE,MAAM,kBAAkB,CAAC;AAE7C,OAAO,EACN,qBAAqB,EAErB,iBAAiB,EACjB,mBAAmB,EACnB,cAAc,EACd,cAAc,EACd,cAAc,EACd,eAAe,EACf,cAAc,EACd,cAAc,EACd,YAAY,GACZ,CAAC;AAsCF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,wBAAsB,kBAAkB,CAAC,OAAO,GAAE,yBAA8B,GAAG,OAAO,CAAC,wBAAwB,CAAC,CAqOnH","sourcesContent":["import { join } from \"node:path\";\nimport { Agent, type AgentMessage, type ThinkingLevel } from \"@kolisachint/hoocode-agent-core\";\nimport { clampThinkingLevel, type Message, type Model, streamSimple } from \"@kolisachint/hoocode-ai\";\nimport { getAgentDir } from \"../config.js\";\nimport { AgentSession } from \"./agent-session.js\";\nimport { formatNoModelsAvailableMessage } from \"./auth-guidance.js\";\nimport { AuthStorage } from \"./auth-storage.js\";\nimport { DEFAULT_THINKING_LEVEL } from \"./defaults.js\";\nimport type { ExtensionRunner, LoadExtensionsResult, SessionStartEvent, ToolDefinition } from \"./extensions/index.js\";\nimport { convertToLlm, createBackgroundPlaceholderText, createBackgroundTaskMessage } from \"./messages.js\";\nimport { ModelRegistry } from \"./model-registry.js\";\nimport { findInitialModel } from \"./model-resolver.js\";\nimport type { ResourceLoader } from \"./resource-loader.js\";\nimport { DefaultResourceLoader } from \"./resource-loader.js\";\nimport { getDefaultSessionDir, SessionManager } from \"./session-manager.js\";\nimport { SettingsManager } from \"./settings-manager.js\";\nimport { peekSubagentPool } from \"./subagent-pool-instance.js\";\nimport { isInstallTelemetryEnabled } from \"./telemetry.js\";\nimport { time } from \"./timings.js\";\nimport {\n\tcreateBashTool,\n\tcreateCodingTools,\n\tcreateEditTool,\n\tcreateFindTool,\n\tcreateGrepTool,\n\tcreateLsTool,\n\tcreateReadOnlyTools,\n\tcreateReadTool,\n\tcreateWriteTool,\n\ttype ToolName,\n\twithFileMutationQueue,\n} from \"./tools/index.js\";\n\nexport interface CreateAgentSessionOptions {\n\t/** Working directory for project-local discovery. Default: process.cwd() */\n\tcwd?: string;\n\t/** Global config directory. Default: ~/.hoocode/agent */\n\tagentDir?: string;\n\n\t/** Auth storage for credentials. Default: AuthStorage.create(agentDir/auth.json) */\n\tauthStorage?: AuthStorage;\n\t/** Model registry. Default: ModelRegistry.create(authStorage, agentDir/models.json) */\n\tmodelRegistry?: ModelRegistry;\n\n\t/** Model to use. Default: from settings, else first available */\n\tmodel?: Model<any>;\n\t/** Thinking level. Default: from settings, else 'medium' (clamped to model capabilities) */\n\tthinkingLevel?: ThinkingLevel;\n\t/** Models available for cycling (Ctrl+P in interactive mode) */\n\tscopedModels?: Array<{ model: Model<any>; thinkingLevel?: ThinkingLevel }>;\n\n\t/**\n\t * Optional default tool suppression mode when no explicit allowlist is provided.\n\t *\n\t * - \"all\": start with no tools enabled\n\t * - \"builtin\": disable the default built-in tools (read, bash, edit, write)\n\t *   but keep extension/custom tools enabled\n\t */\n\tnoTools?: \"all\" | \"builtin\";\n\t/**\n\t * Optional allowlist of tool names.\n\t *\n\t * When omitted, hoocode enables the default built-in tools (read, bash, edit, write)\n\t * and leaves extension/custom tools enabled unless `noTools` changes that default.\n\t * When provided, only the listed tool names are enabled.\n\t */\n\ttools?: string[];\n\t/**\n\t * Optional denylist of tool names, subtracted from whatever set is otherwise\n\t * enabled (allowlist or default). Applied to built-in, extension, and custom tools.\n\t */\n\tdisallowedTools?: string[];\n\t/** Custom tools to register (in addition to built-in tools). */\n\tcustomTools?: ToolDefinition[];\n\n\t/** Resource loader. When omitted, DefaultResourceLoader is used. */\n\tresourceLoader?: ResourceLoader;\n\n\t/** Session manager. Default: SessionManager.create(cwd) */\n\tsessionManager?: SessionManager;\n\n\t/** Settings manager. Default: SettingsManager.create(cwd, agentDir) */\n\tsettingsManager?: SettingsManager;\n\t/** Session start event metadata for extension runtime startup. */\n\tsessionStartEvent?: SessionStartEvent;\n}\n\n/** Result from createAgentSession */\nexport interface CreateAgentSessionResult {\n\t/** The created session */\n\tsession: AgentSession;\n\t/** Extensions result (for UI context setup in interactive mode) */\n\textensionsResult: LoadExtensionsResult;\n\t/** Warning if session was restored with a different model than saved */\n\tmodelFallbackMessage?: string;\n}\n\n// Re-exports\n\nexport type { AgentDefinition, AgentSource } from \"./agent-frontmatter.js\";\nexport { AgentRegistry, formatAgentsForPrompt, loadAgentRegistry } from \"./agent-registry.js\";\nexport * from \"./agent-session-runtime.js\";\nexport type {\n\tExtensionAPI,\n\tExtensionCommandContext,\n\tExtensionContext,\n\tExtensionFactory,\n\tSlashCommandInfo,\n\tSlashCommandSource,\n\tToolDefinition,\n} from \"./extensions/index.js\";\nexport type { PromptTemplate } from \"./prompt-templates.js\";\nexport type { Skill } from \"./skills.js\";\nexport type { Tool } from \"./tools/index.js\";\n\nexport {\n\twithFileMutationQueue,\n\t// Tool factories (for custom cwd)\n\tcreateCodingTools,\n\tcreateReadOnlyTools,\n\tcreateReadTool,\n\tcreateBashTool,\n\tcreateEditTool,\n\tcreateWriteTool,\n\tcreateGrepTool,\n\tcreateFindTool,\n\tcreateLsTool,\n};\n\n// Helper Functions\n\nfunction getDefaultAgentDir(): string {\n\treturn getAgentDir();\n}\n\nfunction getAttributionHeaders(\n\tmodel: Model<any>,\n\tsettingsManager: SettingsManager,\n): Record<string, string> | undefined {\n\tif (!isInstallTelemetryEnabled(settingsManager)) {\n\t\treturn undefined;\n\t}\n\n\tif (model.provider === \"openrouter\" || model.baseUrl.includes(\"openrouter.ai\")) {\n\t\treturn {\n\t\t\t\"HTTP-Referer\": \"https://github.com/kolisachint/hoocode\",\n\t\t\t\"X-OpenRouter-Title\": \"hoocode\",\n\t\t\t\"X-OpenRouter-Categories\": \"cli-agent\",\n\t\t};\n\t}\n\n\tif (\n\t\tmodel.provider === \"cloudflare-workers-ai\" ||\n\t\tmodel.provider === \"cloudflare-ai-gateway\" ||\n\t\tmodel.baseUrl.includes(\"api.cloudflare.com\") ||\n\t\tmodel.baseUrl.includes(\"gateway.ai.cloudflare.com\")\n\t) {\n\t\treturn {\n\t\t\t\"User-Agent\": \"hoocode\",\n\t\t};\n\t}\n\n\treturn undefined;\n}\n\n/**\n * Create an AgentSession with the specified options.\n *\n * @example\n * ```typescript\n * // Minimal - uses defaults\n * const { session } = await createAgentSession();\n *\n * // With explicit model\n * import { getModel } from '@kolisachint/hoocode-ai';\n * const { session } = await createAgentSession({\n *   model: getModel('anthropic', 'claude-opus-4-5'),\n *   thinkingLevel: 'high',\n * });\n *\n * // Continue previous session\n * const { session, modelFallbackMessage } = await createAgentSession({\n *   continueSession: true,\n * });\n *\n * // Full control\n * const loader = new DefaultResourceLoader({\n *   cwd: process.cwd(),\n *   agentDir: getAgentDir(),\n *   settingsManager: SettingsManager.create(),\n * });\n * await loader.reload();\n * const { session } = await createAgentSession({\n *   model: myModel,\n *   tools: [readTool, bashTool],\n *   resourceLoader: loader,\n *   sessionManager: SessionManager.inMemory(),\n * });\n * ```\n */\nexport async function createAgentSession(options: CreateAgentSessionOptions = {}): Promise<CreateAgentSessionResult> {\n\tconst cwd = options.cwd ?? options.sessionManager?.getCwd() ?? process.cwd();\n\tconst agentDir = options.agentDir ?? getDefaultAgentDir();\n\tlet resourceLoader = options.resourceLoader;\n\n\t// Use provided or create AuthStorage and ModelRegistry\n\tconst authPath = options.agentDir ? join(agentDir, \"auth.json\") : undefined;\n\tconst modelsPath = options.agentDir ? join(agentDir, \"models.json\") : undefined;\n\tconst authStorage = options.authStorage ?? AuthStorage.create(authPath);\n\tconst modelRegistry = options.modelRegistry ?? ModelRegistry.create(authStorage, modelsPath);\n\n\tconst settingsManager = options.settingsManager ?? SettingsManager.create(cwd, agentDir);\n\tconst sessionManager = options.sessionManager ?? SessionManager.create(cwd, getDefaultSessionDir(cwd, agentDir));\n\n\tif (!resourceLoader) {\n\t\tresourceLoader = new DefaultResourceLoader({ cwd, agentDir, settingsManager });\n\t\tawait resourceLoader.reload();\n\t\ttime(\"resourceLoader.reload\");\n\t}\n\n\t// Check if session has existing data to restore\n\tconst existingSession = sessionManager.buildSessionContext();\n\tconst hasExistingSession = existingSession.messages.length > 0;\n\tconst hasThinkingEntry = sessionManager.getBranch().some((entry) => entry.type === \"thinking_level_change\");\n\n\tlet model = options.model;\n\tlet modelFallbackMessage: string | undefined;\n\n\t// If session has data, try to restore model from it\n\tif (!model && hasExistingSession && existingSession.model) {\n\t\tconst restoredModel = modelRegistry.find(existingSession.model.provider, existingSession.model.modelId);\n\t\tif (restoredModel && modelRegistry.hasConfiguredAuth(restoredModel)) {\n\t\t\tmodel = restoredModel;\n\t\t}\n\t\tif (!model) {\n\t\t\tmodelFallbackMessage = `Could not restore model ${existingSession.model.provider}/${existingSession.model.modelId}`;\n\t\t}\n\t}\n\n\t// If still no model, use findInitialModel (checks settings default, then provider defaults)\n\tif (!model) {\n\t\tconst result = await findInitialModel({\n\t\t\tscopedModels: [],\n\t\t\tisContinuing: hasExistingSession,\n\t\t\tdefaultProvider: settingsManager.getDefaultProvider(),\n\t\t\tdefaultModelId: settingsManager.getDefaultModel(),\n\t\t\tdefaultThinkingLevel: settingsManager.getDefaultThinkingLevel(),\n\t\t\tmodelRegistry,\n\t\t});\n\t\tmodel = result.model;\n\t\tif (!model) {\n\t\t\tmodelFallbackMessage = formatNoModelsAvailableMessage();\n\t\t} else if (modelFallbackMessage) {\n\t\t\tmodelFallbackMessage += `. Using ${model.provider}/${model.id}`;\n\t\t}\n\t}\n\n\tlet thinkingLevel = options.thinkingLevel;\n\n\t// If session has data, restore thinking level from it\n\tif (thinkingLevel === undefined && hasExistingSession) {\n\t\tthinkingLevel = hasThinkingEntry\n\t\t\t? (existingSession.thinkingLevel as ThinkingLevel)\n\t\t\t: (settingsManager.getDefaultThinkingLevel() ?? DEFAULT_THINKING_LEVEL);\n\t}\n\n\t// Fall back to settings default\n\tif (thinkingLevel === undefined) {\n\t\tthinkingLevel = settingsManager.getDefaultThinkingLevel() ?? DEFAULT_THINKING_LEVEL;\n\t}\n\n\t// Clamp to model capabilities\n\tif (!model) {\n\t\tthinkingLevel = \"off\";\n\t} else {\n\t\tthinkingLevel = clampThinkingLevel(model, thinkingLevel) as ThinkingLevel;\n\t}\n\n\tconst defaultActiveToolNames: ToolName[] = [\"read\", \"bash\", \"edit\", \"write\", \"grep\", \"find\", \"ls\"];\n\tconst allowedToolNames = options.tools ?? (options.noTools === \"all\" ? [] : undefined);\n\tconst initialActiveToolNames: string[] = options.tools\n\t\t? [...options.tools]\n\t\t: options.noTools\n\t\t\t? []\n\t\t\t: defaultActiveToolNames;\n\n\tlet agent: Agent;\n\n\t// Create convertToLlm wrapper that filters images if blockImages is enabled (defense-in-depth)\n\tconst convertToLlmWithBlockImages = (messages: AgentMessage[]): Message[] => {\n\t\tconst converted = convertToLlm(messages);\n\t\t// Check setting dynamically so mid-session changes take effect\n\t\tif (!settingsManager.getBlockImages()) {\n\t\t\treturn converted;\n\t\t}\n\t\t// Filter out ImageContent from all messages, replacing with text placeholder\n\t\treturn converted.map((msg) => {\n\t\t\tif (msg.role === \"user\" || msg.role === \"toolResult\") {\n\t\t\t\tconst content = msg.content;\n\t\t\t\tif (Array.isArray(content)) {\n\t\t\t\t\tconst hasImages = content.some((c) => c.type === \"image\");\n\t\t\t\t\tif (hasImages) {\n\t\t\t\t\t\tconst filteredContent = content\n\t\t\t\t\t\t\t.map((c) =>\n\t\t\t\t\t\t\t\tc.type === \"image\" ? { type: \"text\" as const, text: \"Image reading is disabled.\" } : c,\n\t\t\t\t\t\t\t)\n\t\t\t\t\t\t\t.filter(\n\t\t\t\t\t\t\t\t(c, i, arr) =>\n\t\t\t\t\t\t\t\t\t// Dedupe consecutive \"Image reading is disabled.\" texts\n\t\t\t\t\t\t\t\t\t!(\n\t\t\t\t\t\t\t\t\t\tc.type === \"text\" &&\n\t\t\t\t\t\t\t\t\t\tc.text === \"Image reading is disabled.\" &&\n\t\t\t\t\t\t\t\t\t\ti > 0 &&\n\t\t\t\t\t\t\t\t\t\tarr[i - 1].type === \"text\" &&\n\t\t\t\t\t\t\t\t\t\t(arr[i - 1] as { type: \"text\"; text: string }).text === \"Image reading is disabled.\"\n\t\t\t\t\t\t\t\t\t),\n\t\t\t\t\t\t\t);\n\t\t\t\t\t\treturn { ...msg, content: filteredContent };\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn msg;\n\t\t});\n\t};\n\n\tconst extensionRunnerRef: { current?: ExtensionRunner } = {};\n\n\tagent = new Agent({\n\t\tinitialState: {\n\t\t\tsystemPrompt: \"\",\n\t\t\tmodel,\n\t\t\tthinkingLevel,\n\t\t\ttools: [],\n\t\t},\n\t\tconvertToLlm: convertToLlmWithBlockImages,\n\t\tcreateBackgroundResultMessage: createBackgroundTaskMessage,\n\t\tcreateBackgroundPlaceholder: (toolCall) => createBackgroundPlaceholderText(toolCall),\n\t\t// Report in-process background tool load (e.g. background MCP tools) to the\n\t\t// subagent lifeguard so it widens its heartbeat/timeout tolerance for\n\t\t// concurrently-monitored subagents. Peek (don't create) the pool: background\n\t\t// tools can run before any subagent is ever dispatched.\n\t\tonBackgroundTaskCountChange: (count) => peekSubagentPool()?.setExternalLoad(count),\n\t\tstreamFn: async (model, context, options) => {\n\t\t\tconst auth = await modelRegistry.getApiKeyAndHeaders(model);\n\t\t\tif (!auth.ok) {\n\t\t\t\tthrow new Error(auth.error);\n\t\t\t}\n\t\t\tconst providerRetrySettings = settingsManager.getProviderRetrySettings();\n\t\t\tconst attributionHeaders = getAttributionHeaders(model, settingsManager);\n\t\t\treturn streamSimple(model, context, {\n\t\t\t\t...options,\n\t\t\t\tapiKey: auth.apiKey,\n\t\t\t\ttimeoutMs: options?.timeoutMs ?? providerRetrySettings.timeoutMs,\n\t\t\t\tmaxRetries: options?.maxRetries ?? providerRetrySettings.maxRetries,\n\t\t\t\tmaxRetryDelayMs: options?.maxRetryDelayMs ?? providerRetrySettings.maxRetryDelayMs,\n\t\t\t\theaders:\n\t\t\t\t\tattributionHeaders || auth.headers || options?.headers\n\t\t\t\t\t\t? { ...attributionHeaders, ...auth.headers, ...options?.headers }\n\t\t\t\t\t\t: undefined,\n\t\t\t});\n\t\t},\n\t\tonPayload: async (payload, _model) => {\n\t\t\tconst runner = extensionRunnerRef.current;\n\t\t\tif (!runner?.hasHandlers(\"before_provider_request\")) {\n\t\t\t\treturn payload;\n\t\t\t}\n\t\t\treturn runner.emitBeforeProviderRequest(payload);\n\t\t},\n\t\tonResponse: async (response, _model) => {\n\t\t\tconst runner = extensionRunnerRef.current;\n\t\t\tif (!runner?.hasHandlers(\"after_provider_response\")) {\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tawait runner.emit({\n\t\t\t\ttype: \"after_provider_response\",\n\t\t\t\tstatus: response.status,\n\t\t\t\theaders: response.headers,\n\t\t\t});\n\t\t},\n\t\tsessionId: sessionManager.getSessionId(),\n\t\ttransformContext: async (messages) => {\n\t\t\tconst runner = extensionRunnerRef.current;\n\t\t\tif (!runner) return messages;\n\t\t\treturn runner.emitContext(messages);\n\t\t},\n\t\tsteeringMode: settingsManager.getSteeringMode(),\n\t\tfollowUpMode: settingsManager.getFollowUpMode(),\n\t\ttransport: settingsManager.getTransport(),\n\t\tthinkingBudgets: settingsManager.getThinkingBudgets(),\n\t\tthinkingDisplay: settingsManager.getThinkingDisplay(),\n\t\tmaxRetryDelayMs: settingsManager.getProviderRetrySettings().maxRetryDelayMs,\n\t});\n\n\t// Restore messages if session has existing data\n\tif (hasExistingSession) {\n\t\tagent.state.messages = existingSession.messages;\n\t\tif (!hasThinkingEntry) {\n\t\t\tsessionManager.appendThinkingLevelChange(thinkingLevel);\n\t\t}\n\t} else {\n\t\t// Save initial model and thinking level for new sessions so they can be restored on resume\n\t\tif (model) {\n\t\t\tsessionManager.appendModelChange(model.provider, model.id);\n\t\t}\n\t\tsessionManager.appendThinkingLevelChange(thinkingLevel);\n\t}\n\n\tconst session = new AgentSession({\n\t\tagent,\n\t\tsessionManager,\n\t\tsettingsManager,\n\t\tcwd,\n\t\tscopedModels: options.scopedModels,\n\t\tresourceLoader,\n\t\tcustomTools: options.customTools,\n\t\tmodelRegistry,\n\t\tinitialActiveToolNames,\n\t\tallowedToolNames,\n\t\tdisallowedToolNames: options.disallowedTools,\n\t\textensionRunnerRef,\n\t\tsessionStartEvent: options.sessionStartEvent,\n\t});\n\tconst extensionsResult = resourceLoader.getExtensions();\n\n\treturn {\n\t\tsession,\n\t\textensionsResult,\n\t\tmodelFallbackMessage,\n\t};\n}\n"]}
+{"version":3,"file":"sdk.d.ts","sourceRoot":"","sources":["../../src/core/sdk.ts"],"names":[],"mappings":"AACA,OAAO,EAA4B,KAAK,aAAa,EAAE,MAAM,iCAAiC,CAAC;AAC/F,OAAO,EAAoC,KAAK,KAAK,EAAgB,MAAM,yBAAyB,CAAC;AAErG,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAElD,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAEhD,OAAO,KAAK,EAAmB,oBAAoB,EAAE,iBAAiB,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AAEtH,OAAO,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAC;AAEpD,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AAE3D,OAAO,EAAwB,cAAc,EAAE,MAAM,sBAAsB,CAAC;AAC5E,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AAIxD,OAAO,EACN,cAAc,EACd,iBAAiB,EACjB,cAAc,EACd,cAAc,EACd,cAAc,EACd,YAAY,EACZ,mBAAmB,EACnB,cAAc,EACd,eAAe,EAEf,qBAAqB,EACrB,MAAM,kBAAkB,CAAC;AAE1B,MAAM,WAAW,yBAAyB;IACzC,4EAA4E;IAC5E,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,yDAAyD;IACzD,QAAQ,CAAC,EAAE,MAAM,CAAC;IAElB,oFAAoF;IACpF,WAAW,CAAC,EAAE,WAAW,CAAC;IAC1B,uFAAuF;IACvF,aAAa,CAAC,EAAE,aAAa,CAAC;IAE9B,iEAAiE;IACjE,KAAK,CAAC,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;IACnB,4FAA4F;IAC5F,aAAa,CAAC,EAAE,aAAa,CAAC;IAC9B,gEAAgE;IAChE,YAAY,CAAC,EAAE,KAAK,CAAC;QAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;QAAC,aAAa,CAAC,EAAE,aAAa,CAAA;KAAE,CAAC,CAAC;IAE3E;;;;;;OAMG;IACH,OAAO,CAAC,EAAE,KAAK,GAAG,SAAS,CAAC;IAC5B;;;;;;OAMG;IACH,KAAK,CAAC,EAAE,MAAM,EAAE,CAAC;IACjB;;;OAGG;IACH,eAAe,CAAC,EAAE,MAAM,EAAE,CAAC;IAC3B,gEAAgE;IAChE,WAAW,CAAC,EAAE,cAAc,EAAE,CAAC;IAE/B,oEAAoE;IACpE,cAAc,CAAC,EAAE,cAAc,CAAC;IAEhC,2DAA2D;IAC3D,cAAc,CAAC,EAAE,cAAc,CAAC;IAEhC,uEAAuE;IACvE,eAAe,CAAC,EAAE,eAAe,CAAC;IAClC,kEAAkE;IAClE,iBAAiB,CAAC,EAAE,iBAAiB,CAAC;IACtC;;;;OAIG;IACH,oBAAoB,CAAC,EAAE,OAAO,CAAC;CAC/B;AAED,qCAAqC;AACrC,MAAM,WAAW,wBAAwB;IACxC,0BAA0B;IAC1B,OAAO,EAAE,YAAY,CAAC;IACtB,mEAAmE;IACnE,gBAAgB,EAAE,oBAAoB,CAAC;IACvC,wEAAwE;IACxE,oBAAoB,CAAC,EAAE,MAAM,CAAC;CAC9B;AAID,YAAY,EAAE,eAAe,EAAE,WAAW,EAAE,MAAM,wBAAwB,CAAC;AAC3E,OAAO,EAAE,aAAa,EAAE,qBAAqB,EAAE,iBAAiB,EAAE,MAAM,qBAAqB,CAAC;AAC9F,cAAc,4BAA4B,CAAC;AAC3C,YAAY,EACX,YAAY,EACZ,uBAAuB,EACvB,gBAAgB,EAChB,gBAAgB,EAChB,gBAAgB,EAChB,kBAAkB,EAClB,cAAc,GACd,MAAM,uBAAuB,CAAC;AAC/B,YAAY,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AAC5D,YAAY,EAAE,KAAK,EAAE,MAAM,aAAa,CAAC;AACzC,YAAY,EAAE,IAAI,EAAE,MAAM,kBAAkB,CAAC;AAE7C,OAAO,EACN,qBAAqB,EAErB,iBAAiB,EACjB,mBAAmB,EACnB,cAAc,EACd,cAAc,EACd,cAAc,EACd,eAAe,EACf,cAAc,EACd,cAAc,EACd,YAAY,GACZ,CAAC;AAsCF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,wBAAsB,kBAAkB,CAAC,OAAO,GAAE,yBAA8B,GAAG,OAAO,CAAC,wBAAwB,CAAC,CAsOnH","sourcesContent":["import { join } from \"node:path\";\nimport { Agent, type AgentMessage, type ThinkingLevel } from \"@kolisachint/hoocode-agent-core\";\nimport { clampThinkingLevel, type Message, type Model, streamSimple } from \"@kolisachint/hoocode-ai\";\nimport { getAgentDir } from \"../config.js\";\nimport { AgentSession } from \"./agent-session.js\";\nimport { formatNoModelsAvailableMessage } from \"./auth-guidance.js\";\nimport { AuthStorage } from \"./auth-storage.js\";\nimport { DEFAULT_THINKING_LEVEL } from \"./defaults.js\";\nimport type { ExtensionRunner, LoadExtensionsResult, SessionStartEvent, ToolDefinition } from \"./extensions/index.js\";\nimport { convertToLlm, createBackgroundPlaceholderText, createBackgroundTaskMessage } from \"./messages.js\";\nimport { ModelRegistry } from \"./model-registry.js\";\nimport { findInitialModel } from \"./model-resolver.js\";\nimport type { ResourceLoader } from \"./resource-loader.js\";\nimport { DefaultResourceLoader } from \"./resource-loader.js\";\nimport { getDefaultSessionDir, SessionManager } from \"./session-manager.js\";\nimport { SettingsManager } from \"./settings-manager.js\";\nimport { peekSubagentPool } from \"./subagent-pool-instance.js\";\nimport { isInstallTelemetryEnabled } from \"./telemetry.js\";\nimport { time } from \"./timings.js\";\nimport {\n\tcreateBashTool,\n\tcreateCodingTools,\n\tcreateEditTool,\n\tcreateFindTool,\n\tcreateGrepTool,\n\tcreateLsTool,\n\tcreateReadOnlyTools,\n\tcreateReadTool,\n\tcreateWriteTool,\n\ttype ToolName,\n\twithFileMutationQueue,\n} from \"./tools/index.js\";\n\nexport interface CreateAgentSessionOptions {\n\t/** Working directory for project-local discovery. Default: process.cwd() */\n\tcwd?: string;\n\t/** Global config directory. Default: ~/.hoocode/agent */\n\tagentDir?: string;\n\n\t/** Auth storage for credentials. Default: AuthStorage.create(agentDir/auth.json) */\n\tauthStorage?: AuthStorage;\n\t/** Model registry. Default: ModelRegistry.create(authStorage, agentDir/models.json) */\n\tmodelRegistry?: ModelRegistry;\n\n\t/** Model to use. Default: from settings, else first available */\n\tmodel?: Model<any>;\n\t/** Thinking level. Default: from settings, else 'medium' (clamped to model capabilities) */\n\tthinkingLevel?: ThinkingLevel;\n\t/** Models available for cycling (Ctrl+P in interactive mode) */\n\tscopedModels?: Array<{ model: Model<any>; thinkingLevel?: ThinkingLevel }>;\n\n\t/**\n\t * Optional default tool suppression mode when no explicit allowlist is provided.\n\t *\n\t * - \"all\": start with no tools enabled\n\t * - \"builtin\": disable the default built-in tools (read, bash, edit, write)\n\t *   but keep extension/custom tools enabled\n\t */\n\tnoTools?: \"all\" | \"builtin\";\n\t/**\n\t * Optional allowlist of tool names.\n\t *\n\t * When omitted, hoocode enables the default built-in tools (read, bash, edit, write)\n\t * and leaves extension/custom tools enabled unless `noTools` changes that default.\n\t * When provided, only the listed tool names are enabled.\n\t */\n\ttools?: string[];\n\t/**\n\t * Optional denylist of tool names, subtracted from whatever set is otherwise\n\t * enabled (allowlist or default). Applied to built-in, extension, and custom tools.\n\t */\n\tdisallowedTools?: string[];\n\t/** Custom tools to register (in addition to built-in tools). */\n\tcustomTools?: ToolDefinition[];\n\n\t/** Resource loader. When omitted, DefaultResourceLoader is used. */\n\tresourceLoader?: ResourceLoader;\n\n\t/** Session manager. Default: SessionManager.create(cwd) */\n\tsessionManager?: SessionManager;\n\n\t/** Settings manager. Default: SettingsManager.create(cwd, agentDir) */\n\tsettingsManager?: SettingsManager;\n\t/** Session start event metadata for extension runtime startup. */\n\tsessionStartEvent?: SessionStartEvent;\n\t/**\n\t * Master gate for local-inference routing (compaction + tool-result\n\t * compression via a local executor model). Off by default. Routing also\n\t * requires a `routing` block in models.json. See docs/local-executor-routing.md.\n\t */\n\tenableLocalInference?: boolean;\n}\n\n/** Result from createAgentSession */\nexport interface CreateAgentSessionResult {\n\t/** The created session */\n\tsession: AgentSession;\n\t/** Extensions result (for UI context setup in interactive mode) */\n\textensionsResult: LoadExtensionsResult;\n\t/** Warning if session was restored with a different model than saved */\n\tmodelFallbackMessage?: string;\n}\n\n// Re-exports\n\nexport type { AgentDefinition, AgentSource } from \"./agent-frontmatter.js\";\nexport { AgentRegistry, formatAgentsForPrompt, loadAgentRegistry } from \"./agent-registry.js\";\nexport * from \"./agent-session-runtime.js\";\nexport type {\n\tExtensionAPI,\n\tExtensionCommandContext,\n\tExtensionContext,\n\tExtensionFactory,\n\tSlashCommandInfo,\n\tSlashCommandSource,\n\tToolDefinition,\n} from \"./extensions/index.js\";\nexport type { PromptTemplate } from \"./prompt-templates.js\";\nexport type { Skill } from \"./skills.js\";\nexport type { Tool } from \"./tools/index.js\";\n\nexport {\n\twithFileMutationQueue,\n\t// Tool factories (for custom cwd)\n\tcreateCodingTools,\n\tcreateReadOnlyTools,\n\tcreateReadTool,\n\tcreateBashTool,\n\tcreateEditTool,\n\tcreateWriteTool,\n\tcreateGrepTool,\n\tcreateFindTool,\n\tcreateLsTool,\n};\n\n// Helper Functions\n\nfunction getDefaultAgentDir(): string {\n\treturn getAgentDir();\n}\n\nfunction getAttributionHeaders(\n\tmodel: Model<any>,\n\tsettingsManager: SettingsManager,\n): Record<string, string> | undefined {\n\tif (!isInstallTelemetryEnabled(settingsManager)) {\n\t\treturn undefined;\n\t}\n\n\tif (model.provider === \"openrouter\" || model.baseUrl.includes(\"openrouter.ai\")) {\n\t\treturn {\n\t\t\t\"HTTP-Referer\": \"https://github.com/kolisachint/hoocode\",\n\t\t\t\"X-OpenRouter-Title\": \"hoocode\",\n\t\t\t\"X-OpenRouter-Categories\": \"cli-agent\",\n\t\t};\n\t}\n\n\tif (\n\t\tmodel.provider === \"cloudflare-workers-ai\" ||\n\t\tmodel.provider === \"cloudflare-ai-gateway\" ||\n\t\tmodel.baseUrl.includes(\"api.cloudflare.com\") ||\n\t\tmodel.baseUrl.includes(\"gateway.ai.cloudflare.com\")\n\t) {\n\t\treturn {\n\t\t\t\"User-Agent\": \"hoocode\",\n\t\t};\n\t}\n\n\treturn undefined;\n}\n\n/**\n * Create an AgentSession with the specified options.\n *\n * @example\n * ```typescript\n * // Minimal - uses defaults\n * const { session } = await createAgentSession();\n *\n * // With explicit model\n * import { getModel } from '@kolisachint/hoocode-ai';\n * const { session } = await createAgentSession({\n *   model: getModel('anthropic', 'claude-opus-4-5'),\n *   thinkingLevel: 'high',\n * });\n *\n * // Continue previous session\n * const { session, modelFallbackMessage } = await createAgentSession({\n *   continueSession: true,\n * });\n *\n * // Full control\n * const loader = new DefaultResourceLoader({\n *   cwd: process.cwd(),\n *   agentDir: getAgentDir(),\n *   settingsManager: SettingsManager.create(),\n * });\n * await loader.reload();\n * const { session } = await createAgentSession({\n *   model: myModel,\n *   tools: [readTool, bashTool],\n *   resourceLoader: loader,\n *   sessionManager: SessionManager.inMemory(),\n * });\n * ```\n */\nexport async function createAgentSession(options: CreateAgentSessionOptions = {}): Promise<CreateAgentSessionResult> {\n\tconst cwd = options.cwd ?? options.sessionManager?.getCwd() ?? process.cwd();\n\tconst agentDir = options.agentDir ?? getDefaultAgentDir();\n\tlet resourceLoader = options.resourceLoader;\n\n\t// Use provided or create AuthStorage and ModelRegistry\n\tconst authPath = options.agentDir ? join(agentDir, \"auth.json\") : undefined;\n\tconst modelsPath = options.agentDir ? join(agentDir, \"models.json\") : undefined;\n\tconst authStorage = options.authStorage ?? AuthStorage.create(authPath);\n\tconst modelRegistry = options.modelRegistry ?? ModelRegistry.create(authStorage, modelsPath);\n\n\tconst settingsManager = options.settingsManager ?? SettingsManager.create(cwd, agentDir);\n\tconst sessionManager = options.sessionManager ?? SessionManager.create(cwd, getDefaultSessionDir(cwd, agentDir));\n\n\tif (!resourceLoader) {\n\t\tresourceLoader = new DefaultResourceLoader({ cwd, agentDir, settingsManager });\n\t\tawait resourceLoader.reload();\n\t\ttime(\"resourceLoader.reload\");\n\t}\n\n\t// Check if session has existing data to restore\n\tconst existingSession = sessionManager.buildSessionContext();\n\tconst hasExistingSession = existingSession.messages.length > 0;\n\tconst hasThinkingEntry = sessionManager.getBranch().some((entry) => entry.type === \"thinking_level_change\");\n\n\tlet model = options.model;\n\tlet modelFallbackMessage: string | undefined;\n\n\t// If session has data, try to restore model from it\n\tif (!model && hasExistingSession && existingSession.model) {\n\t\tconst restoredModel = modelRegistry.find(existingSession.model.provider, existingSession.model.modelId);\n\t\tif (restoredModel && modelRegistry.hasConfiguredAuth(restoredModel)) {\n\t\t\tmodel = restoredModel;\n\t\t}\n\t\tif (!model) {\n\t\t\tmodelFallbackMessage = `Could not restore model ${existingSession.model.provider}/${existingSession.model.modelId}`;\n\t\t}\n\t}\n\n\t// If still no model, use findInitialModel (checks settings default, then provider defaults)\n\tif (!model) {\n\t\tconst result = await findInitialModel({\n\t\t\tscopedModels: [],\n\t\t\tisContinuing: hasExistingSession,\n\t\t\tdefaultProvider: settingsManager.getDefaultProvider(),\n\t\t\tdefaultModelId: settingsManager.getDefaultModel(),\n\t\t\tdefaultThinkingLevel: settingsManager.getDefaultThinkingLevel(),\n\t\t\tmodelRegistry,\n\t\t});\n\t\tmodel = result.model;\n\t\tif (!model) {\n\t\t\tmodelFallbackMessage = formatNoModelsAvailableMessage();\n\t\t} else if (modelFallbackMessage) {\n\t\t\tmodelFallbackMessage += `. Using ${model.provider}/${model.id}`;\n\t\t}\n\t}\n\n\tlet thinkingLevel = options.thinkingLevel;\n\n\t// If session has data, restore thinking level from it\n\tif (thinkingLevel === undefined && hasExistingSession) {\n\t\tthinkingLevel = hasThinkingEntry\n\t\t\t? (existingSession.thinkingLevel as ThinkingLevel)\n\t\t\t: (settingsManager.getDefaultThinkingLevel() ?? DEFAULT_THINKING_LEVEL);\n\t}\n\n\t// Fall back to settings default\n\tif (thinkingLevel === undefined) {\n\t\tthinkingLevel = settingsManager.getDefaultThinkingLevel() ?? DEFAULT_THINKING_LEVEL;\n\t}\n\n\t// Clamp to model capabilities\n\tif (!model) {\n\t\tthinkingLevel = \"off\";\n\t} else {\n\t\tthinkingLevel = clampThinkingLevel(model, thinkingLevel) as ThinkingLevel;\n\t}\n\n\tconst defaultActiveToolNames: ToolName[] = [\"read\", \"bash\", \"edit\", \"write\", \"grep\", \"find\", \"ls\"];\n\tconst allowedToolNames = options.tools ?? (options.noTools === \"all\" ? [] : undefined);\n\tconst initialActiveToolNames: string[] = options.tools\n\t\t? [...options.tools]\n\t\t: options.noTools\n\t\t\t? []\n\t\t\t: defaultActiveToolNames;\n\n\tlet agent: Agent;\n\n\t// Create convertToLlm wrapper that filters images if blockImages is enabled (defense-in-depth)\n\tconst convertToLlmWithBlockImages = (messages: AgentMessage[]): Message[] => {\n\t\tconst converted = convertToLlm(messages);\n\t\t// Check setting dynamically so mid-session changes take effect\n\t\tif (!settingsManager.getBlockImages()) {\n\t\t\treturn converted;\n\t\t}\n\t\t// Filter out ImageContent from all messages, replacing with text placeholder\n\t\treturn converted.map((msg) => {\n\t\t\tif (msg.role === \"user\" || msg.role === \"toolResult\") {\n\t\t\t\tconst content = msg.content;\n\t\t\t\tif (Array.isArray(content)) {\n\t\t\t\t\tconst hasImages = content.some((c) => c.type === \"image\");\n\t\t\t\t\tif (hasImages) {\n\t\t\t\t\t\tconst filteredContent = content\n\t\t\t\t\t\t\t.map((c) =>\n\t\t\t\t\t\t\t\tc.type === \"image\" ? { type: \"text\" as const, text: \"Image reading is disabled.\" } : c,\n\t\t\t\t\t\t\t)\n\t\t\t\t\t\t\t.filter(\n\t\t\t\t\t\t\t\t(c, i, arr) =>\n\t\t\t\t\t\t\t\t\t// Dedupe consecutive \"Image reading is disabled.\" texts\n\t\t\t\t\t\t\t\t\t!(\n\t\t\t\t\t\t\t\t\t\tc.type === \"text\" &&\n\t\t\t\t\t\t\t\t\t\tc.text === \"Image reading is disabled.\" &&\n\t\t\t\t\t\t\t\t\t\ti > 0 &&\n\t\t\t\t\t\t\t\t\t\tarr[i - 1].type === \"text\" &&\n\t\t\t\t\t\t\t\t\t\t(arr[i - 1] as { type: \"text\"; text: string }).text === \"Image reading is disabled.\"\n\t\t\t\t\t\t\t\t\t),\n\t\t\t\t\t\t\t);\n\t\t\t\t\t\treturn { ...msg, content: filteredContent };\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn msg;\n\t\t});\n\t};\n\n\tconst extensionRunnerRef: { current?: ExtensionRunner } = {};\n\n\tagent = new Agent({\n\t\tinitialState: {\n\t\t\tsystemPrompt: \"\",\n\t\t\tmodel,\n\t\t\tthinkingLevel,\n\t\t\ttools: [],\n\t\t},\n\t\tconvertToLlm: convertToLlmWithBlockImages,\n\t\tcreateBackgroundResultMessage: createBackgroundTaskMessage,\n\t\tcreateBackgroundPlaceholder: (toolCall) => createBackgroundPlaceholderText(toolCall),\n\t\t// Report in-process background tool load (e.g. background MCP tools) to the\n\t\t// subagent lifeguard so it widens its heartbeat/timeout tolerance for\n\t\t// concurrently-monitored subagents. Peek (don't create) the pool: background\n\t\t// tools can run before any subagent is ever dispatched.\n\t\tonBackgroundTaskCountChange: (count) => peekSubagentPool()?.setExternalLoad(count),\n\t\tstreamFn: async (model, context, options) => {\n\t\t\tconst auth = await modelRegistry.getApiKeyAndHeaders(model);\n\t\t\tif (!auth.ok) {\n\t\t\t\tthrow new Error(auth.error);\n\t\t\t}\n\t\t\tconst providerRetrySettings = settingsManager.getProviderRetrySettings();\n\t\t\tconst attributionHeaders = getAttributionHeaders(model, settingsManager);\n\t\t\treturn streamSimple(model, context, {\n\t\t\t\t...options,\n\t\t\t\tapiKey: auth.apiKey,\n\t\t\t\ttimeoutMs: options?.timeoutMs ?? providerRetrySettings.timeoutMs,\n\t\t\t\tmaxRetries: options?.maxRetries ?? providerRetrySettings.maxRetries,\n\t\t\t\tmaxRetryDelayMs: options?.maxRetryDelayMs ?? providerRetrySettings.maxRetryDelayMs,\n\t\t\t\theaders:\n\t\t\t\t\tattributionHeaders || auth.headers || options?.headers\n\t\t\t\t\t\t? { ...attributionHeaders, ...auth.headers, ...options?.headers }\n\t\t\t\t\t\t: undefined,\n\t\t\t});\n\t\t},\n\t\tonPayload: async (payload, _model) => {\n\t\t\tconst runner = extensionRunnerRef.current;\n\t\t\tif (!runner?.hasHandlers(\"before_provider_request\")) {\n\t\t\t\treturn payload;\n\t\t\t}\n\t\t\treturn runner.emitBeforeProviderRequest(payload);\n\t\t},\n\t\tonResponse: async (response, _model) => {\n\t\t\tconst runner = extensionRunnerRef.current;\n\t\t\tif (!runner?.hasHandlers(\"after_provider_response\")) {\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tawait runner.emit({\n\t\t\t\ttype: \"after_provider_response\",\n\t\t\t\tstatus: response.status,\n\t\t\t\theaders: response.headers,\n\t\t\t});\n\t\t},\n\t\tsessionId: sessionManager.getSessionId(),\n\t\ttransformContext: async (messages) => {\n\t\t\tconst runner = extensionRunnerRef.current;\n\t\t\tif (!runner) return messages;\n\t\t\treturn runner.emitContext(messages);\n\t\t},\n\t\tsteeringMode: settingsManager.getSteeringMode(),\n\t\tfollowUpMode: settingsManager.getFollowUpMode(),\n\t\ttransport: settingsManager.getTransport(),\n\t\tthinkingBudgets: settingsManager.getThinkingBudgets(),\n\t\tthinkingDisplay: settingsManager.getThinkingDisplay(),\n\t\tmaxRetryDelayMs: settingsManager.getProviderRetrySettings().maxRetryDelayMs,\n\t});\n\n\t// Restore messages if session has existing data\n\tif (hasExistingSession) {\n\t\tagent.state.messages = existingSession.messages;\n\t\tif (!hasThinkingEntry) {\n\t\t\tsessionManager.appendThinkingLevelChange(thinkingLevel);\n\t\t}\n\t} else {\n\t\t// Save initial model and thinking level for new sessions so they can be restored on resume\n\t\tif (model) {\n\t\t\tsessionManager.appendModelChange(model.provider, model.id);\n\t\t}\n\t\tsessionManager.appendThinkingLevelChange(thinkingLevel);\n\t}\n\n\tconst session = new AgentSession({\n\t\tagent,\n\t\tsessionManager,\n\t\tsettingsManager,\n\t\tcwd,\n\t\tscopedModels: options.scopedModels,\n\t\tresourceLoader,\n\t\tcustomTools: options.customTools,\n\t\tmodelRegistry,\n\t\tinitialActiveToolNames,\n\t\tallowedToolNames,\n\t\tdisallowedToolNames: options.disallowedTools,\n\t\textensionRunnerRef,\n\t\tsessionStartEvent: options.sessionStartEvent,\n\t\tenableLocalInference: options.enableLocalInference,\n\t});\n\tconst extensionsResult = resourceLoader.getExtensions();\n\n\treturn {\n\t\tsession,\n\t\textensionsResult,\n\t\tmodelFallbackMessage,\n\t};\n}\n"]}

package/dist/core/sdk.js

@@ -281,6 +281,7 @@ export async function createAgentSession(options = {}) {
         disallowedToolNames: options.disallowedTools,
         extensionRunnerRef,
         sessionStartEvent: options.sessionStartEvent,
+        enableLocalInference: options.enableLocalInference,
     });
     const extensionsResult = resourceLoader.getExtensions();
     return {