@kolisachint/hoocode-agent 0.4.72 → 0.4.74

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

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

@@ -19,7 +19,6 @@ export const ROUTING_MODES = [
     "primary-only",
     "executor-for-summarization",
     "executor-for-tool-results",
-    "shadow-executor",
 ];
 /**
  * Tool names whose output is worth compressing (validated). Others pass
@@ -29,15 +28,6 @@ export const ROUTING_MODES = [
  * (grep/find/ls) were never compressible (every line is a distinct fact).
  */
 export const COMPRESSIBLE_TOOLS = new Set(["bash"]);
-/**
- * Global size band (bytes) for local-inference routing. Applies to BOTH
- * tool-result compression and compaction summarization. Inputs below the
- * minimum are not worth offloading; inputs above the maximum are slow and risk
- * GPU OOM on small machines, so they fall back to the primary model. Tunable
- * per-machine via `minBytes`/`maxBytes` in the executor config block.
- */
-export const DEFAULT_MIN_BYTES = 2048;
-export const DEFAULT_MAX_BYTES = 8192;
 function isRoutingMode(value) {
     return typeof value === "string" && ROUTING_MODES.includes(value);
 }
@@ -76,8 +66,11 @@ export class LocalInferenceRouter {
         this.mode = mode;
         this.executorConfig = executorConfig;
         this.executor = executor;
-        this.minBytes = executorConfig?.minBytes ?? DEFAULT_MIN_BYTES;
-        this.maxBytes = executorConfig?.maxBytes ?? DEFAULT_MAX_BYTES;
+        // 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;
@@ -100,8 +93,7 @@ export class LocalInferenceRouter {
     /**
      * 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. `shadow-executor` always returns the primary for the live path (the
-     * executor is exercised separately for measurement).
+     * model.
      */
     selectModel(turnKind, primary) {
         if (!this.isExecutorAvailable() || !this.executor)
@@ -111,7 +103,6 @@ export class LocalInferenceRouter {
                 return turnKind === "summarization" ? this.executor : primary;
             case "executor-for-tool-results":
                 return turnKind === "tool-result" ? this.executor : primary;
-            case "shadow-executor":
             case "primary-only":
                 return primary;
         }

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

@@ -1 +1 @@
-{"version":3,"file":"local-inference.js","sourceRoot":"","sources":["../../../src/core/routing/local-inference.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAcH,MAAM,CAAC,MAAM,aAAa,GAA2B;IACpD,cAAc;IACd,4BAA4B;IAC5B,2BAA2B;IAC3B,iBAAiB;CACR,CAAC;AAEX;;;;;;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"]}
+{"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/subagent-pool-instance.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"subagent-pool-instance.d.ts","sourceRoot":"","sources":["../../src/core/subagent-pool-instance.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAQlD,mFAAmF;AACnF,wBAAgB,eAAe,CAAC,GAAG,EAAE,MAAM,GAAG,YAAY,CAqBzD;AAED;;;;GAIG;AACH,wBAAgB,gBAAgB,IAAI,YAAY,GAAG,SAAS,CAE3D;AAED;;;;;GAKG;AACH,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,IAAI,CAG9D;AAED,mFAAmF;AACnF,wBAAgB,mBAAmB,IAAI,IAAI,CAI1C;AAED;;;GAGG;AACH,wBAAgB,yBAAyB,CAAC,QAAQ,EAAE,YAAY,GAAG,SAAS,GAAG,IAAI,CAElF","sourcesContent":["/**\n * Process-wide SubagentPool singleton.\n *\n * The subagent tool and the `/subagent` command both delegate through one pool\n * so concurrency limits, lifeguard monitoring, and token budgets are shared\n * across every delegation in the session. Created lazily on first use and torn\n * down on process exit.\n */\n\nimport { getSubagentSpawnCommand } from \"../config.js\";\nimport { poolConcurrencyForDepth } from \"./subagent-depth.js\";\nimport { SubagentPool } from \"./subagent-pool.js\";\n\nlet pool: SubagentPool | undefined;\nlet override: SubagentPool | undefined;\nlet exitHandlerRegistered = false;\n/** Latest non-default skill paths to forward to subagents, kept in sync with the resource loader. */\nlet latestSkillPaths: string[] = [];\n\n/** Get the shared pool for a given working directory, creating it on first use. */\nexport function getSubagentPool(cwd: string): SubagentPool {\n\tif (override) return override;\n\tif (!pool) {\n\t\tconst { executable, prefixArgs } = getSubagentSpawnCommand();\n\t\t// Pools created inside a nested subagent (depth >= 1) run with a reduced\n\t\t// concurrency cap so deep delegation trees stay bounded; the root keeps the\n\t\t// SubagentPool default.\n\t\tpool = new SubagentPool({\n\t\t\texecutable,\n\t\t\tprefixArgs,\n\t\t\tcwd,\n\t\t\tskillPaths: latestSkillPaths,\n\t\t\tmaxConcurrency: poolConcurrencyForDepth(),\n\t\t});\n\n\t\tif (!exitHandlerRegistered) {\n\t\t\texitHandlerRegistered = true;\n\t\t\tprocess.once(\"exit\", () => pool?.dispose());\n\t\t}\n\t}\n\treturn pool;\n}\n\n/**\n * Return the shared pool if one already exists, without creating it. Use this for\n * best-effort signaling (e.g. reporting external load) that must not spin up a pool\n * and its lifeguard just because the signal fired before any subagent was dispatched.\n */\nexport function peekSubagentPool(): SubagentPool | undefined {\n\treturn override ?? pool;\n}\n\n/**\n * Update the skill paths forwarded to every subagent.\n * Call this after the resource loader reloads or extends its skill set.\n * If the pool has already been created, updates it immediately.\n * If not, the paths will be passed in when the pool is first created.\n */\nexport function updateSubagentSkillPaths(paths: string[]): void {\n\tlatestSkillPaths = paths;\n\tpool?.updateSkillPaths(paths);\n}\n\n/** Dispose and clear the shared pool. Intended for test isolation and shutdown. */\nexport function disposeSubagentPool(): void {\n\tpool?.dispose();\n\tpool = undefined;\n\tlatestSkillPaths = [];\n}\n\n/**\n * Inject a pool instance for tests, bypassing real child-process spawning.\n * Pass `undefined` to clear the override.\n */\nexport function setSubagentPoolForTesting(testPool: SubagentPool | undefined): void {\n\toverride = testPool;\n}\n"]}
+{"version":3,"file":"subagent-pool-instance.d.ts","sourceRoot":"","sources":["../../src/core/subagent-pool-instance.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AASlD,mFAAmF;AACnF,wBAAgB,eAAe,CAAC,GAAG,EAAE,MAAM,GAAG,YAAY,CAuBzD;AA0BD;;;;GAIG;AACH,wBAAgB,gBAAgB,IAAI,YAAY,GAAG,SAAS,CAE3D;AAED;;;;;GAKG;AACH,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,IAAI,CAG9D;AAED,mFAAmF;AACnF,wBAAgB,mBAAmB,IAAI,IAAI,CAI1C;AAED;;;GAGG;AACH,wBAAgB,yBAAyB,CAAC,QAAQ,EAAE,YAAY,GAAG,SAAS,GAAG,IAAI,CAElF","sourcesContent":["/**\n * Process-wide SubagentPool singleton.\n *\n * The subagent tool and the `/subagent` command both delegate through one pool\n * so concurrency limits, lifeguard monitoring, and token budgets are shared\n * across every delegation in the session. Created lazily on first use and torn\n * down on process exit.\n */\n\nimport { getSubagentSpawnCommand } from \"../config.js\";\nimport { poolConcurrencyForDepth } from \"./subagent-depth.js\";\nimport { SubagentPool } from \"./subagent-pool.js\";\nimport { taskStore } from \"./task-store.js\";\n\nlet pool: SubagentPool | undefined;\nlet override: SubagentPool | undefined;\nlet exitHandlerRegistered = false;\n/** Latest non-default skill paths to forward to subagents, kept in sync with the resource loader. */\nlet latestSkillPaths: string[] = [];\n\n/** Get the shared pool for a given working directory, creating it on first use. */\nexport function getSubagentPool(cwd: string): SubagentPool {\n\tif (override) return override;\n\tif (!pool) {\n\t\tconst { executable, prefixArgs } = getSubagentSpawnCommand();\n\t\t// Pools created inside a nested subagent (depth >= 1) run with a reduced\n\t\t// concurrency cap so deep delegation trees stay bounded; the root keeps the\n\t\t// SubagentPool default.\n\t\tpool = new SubagentPool({\n\t\t\texecutable,\n\t\t\tprefixArgs,\n\t\t\tcwd,\n\t\t\tskillPaths: latestSkillPaths,\n\t\t\tmaxConcurrency: poolConcurrencyForDepth(),\n\t\t});\n\n\t\twireProgressToTaskStore(pool);\n\n\t\tif (!exitHandlerRegistered) {\n\t\t\texitHandlerRegistered = true;\n\t\t\tprocess.once(\"exit\", () => pool?.dispose());\n\t\t}\n\t}\n\treturn pool;\n}\n\n/**\n * Surface live subagent progress on the task panel's agent roster row. The pool\n * forwards only coarse lifecycle events; we map the currently-executing tool onto\n * the agent's `activity` and clear it between tools and on completion. This touches\n * only the roster row (keyed by agent type, a no-op if no such row exists), never\n * task nodes — so it cannot collide with the end-of-run task-tree merge. Render\n * coalescing is handled by the TUI's `requestRender`, so per-event patches are fine.\n */\nfunction wireProgressToTaskStore(p: SubagentPool): void {\n\tp.on(\"task_progress\", (data: { agent_type: string; event: { type?: string; toolName?: string } }) => {\n\t\tconst { agent_type, event } = data;\n\t\tif (event.type === \"tool_execution_start\") {\n\t\t\ttaskStore.patchAgent(agent_type, { activity: typeof event.toolName === \"string\" ? event.toolName : \"\" });\n\t\t} else if (event.type === \"tool_execution_end\" || event.type === \"turn_end\") {\n\t\t\ttaskStore.patchAgent(agent_type, { activity: \"\" });\n\t\t}\n\t});\n\tfor (const terminal of [\"task_done\", \"task_failed\", \"task_stalled\", \"task_timeout\"] as const) {\n\t\tp.on(terminal, (data: { agent_type?: string }) => {\n\t\t\tif (data.agent_type) taskStore.patchAgent(data.agent_type, { activity: \"\" });\n\t\t});\n\t}\n}\n\n/**\n * Return the shared pool if one already exists, without creating it. Use this for\n * best-effort signaling (e.g. reporting external load) that must not spin up a pool\n * and its lifeguard just because the signal fired before any subagent was dispatched.\n */\nexport function peekSubagentPool(): SubagentPool | undefined {\n\treturn override ?? pool;\n}\n\n/**\n * Update the skill paths forwarded to every subagent.\n * Call this after the resource loader reloads or extends its skill set.\n * If the pool has already been created, updates it immediately.\n * If not, the paths will be passed in when the pool is first created.\n */\nexport function updateSubagentSkillPaths(paths: string[]): void {\n\tlatestSkillPaths = paths;\n\tpool?.updateSkillPaths(paths);\n}\n\n/** Dispose and clear the shared pool. Intended for test isolation and shutdown. */\nexport function disposeSubagentPool(): void {\n\tpool?.dispose();\n\tpool = undefined;\n\tlatestSkillPaths = [];\n}\n\n/**\n * Inject a pool instance for tests, bypassing real child-process spawning.\n * Pass `undefined` to clear the override.\n */\nexport function setSubagentPoolForTesting(testPool: SubagentPool | undefined): void {\n\toverride = testPool;\n}\n"]}

package/dist/core/subagent-pool-instance.js

@@ -9,6 +9,7 @@
 import { getSubagentSpawnCommand } from "../config.js";
 import { poolConcurrencyForDepth } from "./subagent-depth.js";
 import { SubagentPool } from "./subagent-pool.js";
+import { taskStore } from "./task-store.js";
 let pool;
 let override;
 let exitHandlerRegistered = false;
@@ -30,6 +31,7 @@ export function getSubagentPool(cwd) {
             skillPaths: latestSkillPaths,
             maxConcurrency: poolConcurrencyForDepth(),
         });
+        wireProgressToTaskStore(pool);
         if (!exitHandlerRegistered) {
             exitHandlerRegistered = true;
             process.once("exit", () => pool?.dispose());
@@ -37,6 +39,31 @@ export function getSubagentPool(cwd) {
     }
     return pool;
 }
+/**
+ * Surface live subagent progress on the task panel's agent roster row. The pool
+ * forwards only coarse lifecycle events; we map the currently-executing tool onto
+ * the agent's `activity` and clear it between tools and on completion. This touches
+ * only the roster row (keyed by agent type, a no-op if no such row exists), never
+ * task nodes — so it cannot collide with the end-of-run task-tree merge. Render
+ * coalescing is handled by the TUI's `requestRender`, so per-event patches are fine.
+ */
+function wireProgressToTaskStore(p) {
+    p.on("task_progress", (data) => {
+        const { agent_type, event } = data;
+        if (event.type === "tool_execution_start") {
+            taskStore.patchAgent(agent_type, { activity: typeof event.toolName === "string" ? event.toolName : "" });
+        }
+        else if (event.type === "tool_execution_end" || event.type === "turn_end") {
+            taskStore.patchAgent(agent_type, { activity: "" });
+        }
+    });
+    for (const terminal of ["task_done", "task_failed", "task_stalled", "task_timeout"]) {
+        p.on(terminal, (data) => {
+            if (data.agent_type)
+                taskStore.patchAgent(data.agent_type, { activity: "" });
+        });
+    }
+}
 /**
  * Return the shared pool if one already exists, without creating it. Use this for
  * best-effort signaling (e.g. reporting external load) that must not spin up a pool

package/dist/core/subagent-pool-instance.js.map

@@ -1 +1 @@
-{"version":3,"file":"subagent-pool-instance.js","sourceRoot":"","sources":["../../src/core/subagent-pool-instance.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,EAAE,uBAAuB,EAAE,MAAM,cAAc,CAAC;AACvD,OAAO,EAAE,uBAAuB,EAAE,MAAM,qBAAqB,CAAC;AAC9D,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAElD,IAAI,IAA8B,CAAC;AACnC,IAAI,QAAkC,CAAC;AACvC,IAAI,qBAAqB,GAAG,KAAK,CAAC;AAClC,qGAAqG;AACrG,IAAI,gBAAgB,GAAa,EAAE,CAAC;AAEpC,mFAAmF;AACnF,MAAM,UAAU,eAAe,CAAC,GAAW,EAAgB;IAC1D,IAAI,QAAQ;QAAE,OAAO,QAAQ,CAAC;IAC9B,IAAI,CAAC,IAAI,EAAE,CAAC;QACX,MAAM,EAAE,UAAU,EAAE,UAAU,EAAE,GAAG,uBAAuB,EAAE,CAAC;QAC7D,yEAAyE;QACzE,4EAA4E;QAC5E,wBAAwB;QACxB,IAAI,GAAG,IAAI,YAAY,CAAC;YACvB,UAAU;YACV,UAAU;YACV,GAAG;YACH,UAAU,EAAE,gBAAgB;YAC5B,cAAc,EAAE,uBAAuB,EAAE;SACzC,CAAC,CAAC;QAEH,IAAI,CAAC,qBAAqB,EAAE,CAAC;YAC5B,qBAAqB,GAAG,IAAI,CAAC;YAC7B,OAAO,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,EAAE,CAAC,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC;QAC7C,CAAC;IACF,CAAC;IACD,OAAO,IAAI,CAAC;AAAA,CACZ;AAED;;;;GAIG;AACH,MAAM,UAAU,gBAAgB,GAA6B;IAC5D,OAAO,QAAQ,IAAI,IAAI,CAAC;AAAA,CACxB;AAED;;;;;GAKG;AACH,MAAM,UAAU,wBAAwB,CAAC,KAAe,EAAQ;IAC/D,gBAAgB,GAAG,KAAK,CAAC;IACzB,IAAI,EAAE,gBAAgB,CAAC,KAAK,CAAC,CAAC;AAAA,CAC9B;AAED,mFAAmF;AACnF,MAAM,UAAU,mBAAmB,GAAS;IAC3C,IAAI,EAAE,OAAO,EAAE,CAAC;IAChB,IAAI,GAAG,SAAS,CAAC;IACjB,gBAAgB,GAAG,EAAE,CAAC;AAAA,CACtB;AAED;;;GAGG;AACH,MAAM,UAAU,yBAAyB,CAAC,QAAkC,EAAQ;IACnF,QAAQ,GAAG,QAAQ,CAAC;AAAA,CACpB","sourcesContent":["/**\n * Process-wide SubagentPool singleton.\n *\n * The subagent tool and the `/subagent` command both delegate through one pool\n * so concurrency limits, lifeguard monitoring, and token budgets are shared\n * across every delegation in the session. Created lazily on first use and torn\n * down on process exit.\n */\n\nimport { getSubagentSpawnCommand } from \"../config.js\";\nimport { poolConcurrencyForDepth } from \"./subagent-depth.js\";\nimport { SubagentPool } from \"./subagent-pool.js\";\n\nlet pool: SubagentPool | undefined;\nlet override: SubagentPool | undefined;\nlet exitHandlerRegistered = false;\n/** Latest non-default skill paths to forward to subagents, kept in sync with the resource loader. */\nlet latestSkillPaths: string[] = [];\n\n/** Get the shared pool for a given working directory, creating it on first use. */\nexport function getSubagentPool(cwd: string): SubagentPool {\n\tif (override) return override;\n\tif (!pool) {\n\t\tconst { executable, prefixArgs } = getSubagentSpawnCommand();\n\t\t// Pools created inside a nested subagent (depth >= 1) run with a reduced\n\t\t// concurrency cap so deep delegation trees stay bounded; the root keeps the\n\t\t// SubagentPool default.\n\t\tpool = new SubagentPool({\n\t\t\texecutable,\n\t\t\tprefixArgs,\n\t\t\tcwd,\n\t\t\tskillPaths: latestSkillPaths,\n\t\t\tmaxConcurrency: poolConcurrencyForDepth(),\n\t\t});\n\n\t\tif (!exitHandlerRegistered) {\n\t\t\texitHandlerRegistered = true;\n\t\t\tprocess.once(\"exit\", () => pool?.dispose());\n\t\t}\n\t}\n\treturn pool;\n}\n\n/**\n * Return the shared pool if one already exists, without creating it. Use this for\n * best-effort signaling (e.g. reporting external load) that must not spin up a pool\n * and its lifeguard just because the signal fired before any subagent was dispatched.\n */\nexport function peekSubagentPool(): SubagentPool | undefined {\n\treturn override ?? pool;\n}\n\n/**\n * Update the skill paths forwarded to every subagent.\n * Call this after the resource loader reloads or extends its skill set.\n * If the pool has already been created, updates it immediately.\n * If not, the paths will be passed in when the pool is first created.\n */\nexport function updateSubagentSkillPaths(paths: string[]): void {\n\tlatestSkillPaths = paths;\n\tpool?.updateSkillPaths(paths);\n}\n\n/** Dispose and clear the shared pool. Intended for test isolation and shutdown. */\nexport function disposeSubagentPool(): void {\n\tpool?.dispose();\n\tpool = undefined;\n\tlatestSkillPaths = [];\n}\n\n/**\n * Inject a pool instance for tests, bypassing real child-process spawning.\n * Pass `undefined` to clear the override.\n */\nexport function setSubagentPoolForTesting(testPool: SubagentPool | undefined): void {\n\toverride = testPool;\n}\n"]}
+{"version":3,"file":"subagent-pool-instance.js","sourceRoot":"","sources":["../../src/core/subagent-pool-instance.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,EAAE,uBAAuB,EAAE,MAAM,cAAc,CAAC;AACvD,OAAO,EAAE,uBAAuB,EAAE,MAAM,qBAAqB,CAAC;AAC9D,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAClD,OAAO,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AAE5C,IAAI,IAA8B,CAAC;AACnC,IAAI,QAAkC,CAAC;AACvC,IAAI,qBAAqB,GAAG,KAAK,CAAC;AAClC,qGAAqG;AACrG,IAAI,gBAAgB,GAAa,EAAE,CAAC;AAEpC,mFAAmF;AACnF,MAAM,UAAU,eAAe,CAAC,GAAW,EAAgB;IAC1D,IAAI,QAAQ;QAAE,OAAO,QAAQ,CAAC;IAC9B,IAAI,CAAC,IAAI,EAAE,CAAC;QACX,MAAM,EAAE,UAAU,EAAE,UAAU,EAAE,GAAG,uBAAuB,EAAE,CAAC;QAC7D,yEAAyE;QACzE,4EAA4E;QAC5E,wBAAwB;QACxB,IAAI,GAAG,IAAI,YAAY,CAAC;YACvB,UAAU;YACV,UAAU;YACV,GAAG;YACH,UAAU,EAAE,gBAAgB;YAC5B,cAAc,EAAE,uBAAuB,EAAE;SACzC,CAAC,CAAC;QAEH,uBAAuB,CAAC,IAAI,CAAC,CAAC;QAE9B,IAAI,CAAC,qBAAqB,EAAE,CAAC;YAC5B,qBAAqB,GAAG,IAAI,CAAC;YAC7B,OAAO,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,EAAE,CAAC,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC;QAC7C,CAAC;IACF,CAAC;IACD,OAAO,IAAI,CAAC;AAAA,CACZ;AAED;;;;;;;GAOG;AACH,SAAS,uBAAuB,CAAC,CAAe,EAAQ;IACvD,CAAC,CAAC,EAAE,CAAC,eAAe,EAAE,CAAC,IAAyE,EAAE,EAAE,CAAC;QACpG,MAAM,EAAE,UAAU,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC;QACnC,IAAI,KAAK,CAAC,IAAI,KAAK,sBAAsB,EAAE,CAAC;YAC3C,SAAS,CAAC,UAAU,CAAC,UAAU,EAAE,EAAE,QAAQ,EAAE,OAAO,KAAK,CAAC,QAAQ,KAAK,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;QAC1G,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,oBAAoB,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YAC7E,SAAS,CAAC,UAAU,CAAC,UAAU,EAAE,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,CAAC;QACpD,CAAC;IAAA,CACD,CAAC,CAAC;IACH,KAAK,MAAM,QAAQ,IAAI,CAAC,WAAW,EAAE,aAAa,EAAE,cAAc,EAAE,cAAc,CAAU,EAAE,CAAC;QAC9F,CAAC,CAAC,EAAE,CAAC,QAAQ,EAAE,CAAC,IAA6B,EAAE,EAAE,CAAC;YACjD,IAAI,IAAI,CAAC,UAAU;gBAAE,SAAS,CAAC,UAAU,CAAC,IAAI,CAAC,UAAU,EAAE,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,CAAC;QAAA,CAC7E,CAAC,CAAC;IACJ,CAAC;AAAA,CACD;AAED;;;;GAIG;AACH,MAAM,UAAU,gBAAgB,GAA6B;IAC5D,OAAO,QAAQ,IAAI,IAAI,CAAC;AAAA,CACxB;AAED;;;;;GAKG;AACH,MAAM,UAAU,wBAAwB,CAAC,KAAe,EAAQ;IAC/D,gBAAgB,GAAG,KAAK,CAAC;IACzB,IAAI,EAAE,gBAAgB,CAAC,KAAK,CAAC,CAAC;AAAA,CAC9B;AAED,mFAAmF;AACnF,MAAM,UAAU,mBAAmB,GAAS;IAC3C,IAAI,EAAE,OAAO,EAAE,CAAC;IAChB,IAAI,GAAG,SAAS,CAAC;IACjB,gBAAgB,GAAG,EAAE,CAAC;AAAA,CACtB;AAED;;;GAGG;AACH,MAAM,UAAU,yBAAyB,CAAC,QAAkC,EAAQ;IACnF,QAAQ,GAAG,QAAQ,CAAC;AAAA,CACpB","sourcesContent":["/**\n * Process-wide SubagentPool singleton.\n *\n * The subagent tool and the `/subagent` command both delegate through one pool\n * so concurrency limits, lifeguard monitoring, and token budgets are shared\n * across every delegation in the session. Created lazily on first use and torn\n * down on process exit.\n */\n\nimport { getSubagentSpawnCommand } from \"../config.js\";\nimport { poolConcurrencyForDepth } from \"./subagent-depth.js\";\nimport { SubagentPool } from \"./subagent-pool.js\";\nimport { taskStore } from \"./task-store.js\";\n\nlet pool: SubagentPool | undefined;\nlet override: SubagentPool | undefined;\nlet exitHandlerRegistered = false;\n/** Latest non-default skill paths to forward to subagents, kept in sync with the resource loader. */\nlet latestSkillPaths: string[] = [];\n\n/** Get the shared pool for a given working directory, creating it on first use. */\nexport function getSubagentPool(cwd: string): SubagentPool {\n\tif (override) return override;\n\tif (!pool) {\n\t\tconst { executable, prefixArgs } = getSubagentSpawnCommand();\n\t\t// Pools created inside a nested subagent (depth >= 1) run with a reduced\n\t\t// concurrency cap so deep delegation trees stay bounded; the root keeps the\n\t\t// SubagentPool default.\n\t\tpool = new SubagentPool({\n\t\t\texecutable,\n\t\t\tprefixArgs,\n\t\t\tcwd,\n\t\t\tskillPaths: latestSkillPaths,\n\t\t\tmaxConcurrency: poolConcurrencyForDepth(),\n\t\t});\n\n\t\twireProgressToTaskStore(pool);\n\n\t\tif (!exitHandlerRegistered) {\n\t\t\texitHandlerRegistered = true;\n\t\t\tprocess.once(\"exit\", () => pool?.dispose());\n\t\t}\n\t}\n\treturn pool;\n}\n\n/**\n * Surface live subagent progress on the task panel's agent roster row. The pool\n * forwards only coarse lifecycle events; we map the currently-executing tool onto\n * the agent's `activity` and clear it between tools and on completion. This touches\n * only the roster row (keyed by agent type, a no-op if no such row exists), never\n * task nodes — so it cannot collide with the end-of-run task-tree merge. Render\n * coalescing is handled by the TUI's `requestRender`, so per-event patches are fine.\n */\nfunction wireProgressToTaskStore(p: SubagentPool): void {\n\tp.on(\"task_progress\", (data: { agent_type: string; event: { type?: string; toolName?: string } }) => {\n\t\tconst { agent_type, event } = data;\n\t\tif (event.type === \"tool_execution_start\") {\n\t\t\ttaskStore.patchAgent(agent_type, { activity: typeof event.toolName === \"string\" ? event.toolName : \"\" });\n\t\t} else if (event.type === \"tool_execution_end\" || event.type === \"turn_end\") {\n\t\t\ttaskStore.patchAgent(agent_type, { activity: \"\" });\n\t\t}\n\t});\n\tfor (const terminal of [\"task_done\", \"task_failed\", \"task_stalled\", \"task_timeout\"] as const) {\n\t\tp.on(terminal, (data: { agent_type?: string }) => {\n\t\t\tif (data.agent_type) taskStore.patchAgent(data.agent_type, { activity: \"\" });\n\t\t});\n\t}\n}\n\n/**\n * Return the shared pool if one already exists, without creating it. Use this for\n * best-effort signaling (e.g. reporting external load) that must not spin up a pool\n * and its lifeguard just because the signal fired before any subagent was dispatched.\n */\nexport function peekSubagentPool(): SubagentPool | undefined {\n\treturn override ?? pool;\n}\n\n/**\n * Update the skill paths forwarded to every subagent.\n * Call this after the resource loader reloads or extends its skill set.\n * If the pool has already been created, updates it immediately.\n * If not, the paths will be passed in when the pool is first created.\n */\nexport function updateSubagentSkillPaths(paths: string[]): void {\n\tlatestSkillPaths = paths;\n\tpool?.updateSkillPaths(paths);\n}\n\n/** Dispose and clear the shared pool. Intended for test isolation and shutdown. */\nexport function disposeSubagentPool(): void {\n\tpool?.dispose();\n\tpool = undefined;\n\tlatestSkillPaths = [];\n}\n\n/**\n * Inject a pool instance for tests, bypassing real child-process spawning.\n * Pass `undefined` to clear the override.\n */\nexport function setSubagentPoolForTesting(testPool: SubagentPool | undefined): void {\n\toverride = testPool;\n}\n"]}

package/dist/core/subagent-pool.d.ts

@@ -93,6 +93,30 @@ export interface SubagentPoolOptions {
  * kills), so this turn cap is the guaranteed hard stop for every subagent.
  */
 export declare const DEFAULT_SUBAGENT_MAX_TURNS = 50;
+/**
+ * AgentSession event `type`s forwarded from a subagent's json event stream as
+ * `task_progress` events. Deliberately coarse: the child also emits per-delta
+ * `message_update` / `tool_execution_update` events (a high-volume firehose) and
+ * large `message_*` bodies, which are dropped here to keep the parent's event loop
+ * and the task panel from thrashing under concurrent subagents.
+ */
+export declare const FORWARDED_SUBAGENT_EVENTS: ReadonlySet<string>;
+/** The action the pool should take for one JSONL line from a subagent's stdout. */
+export type SubagentStdoutLine = {
+    kind: "heartbeat";
+} | {
+    kind: "progress";
+    event: Record<string, unknown>;
+} | {
+    kind: "ignore";
+};
+/**
+ * Classify one JSONL line from a subagent's stdout into the action to take.
+ * Pure (no side effects) so the ping/forward/drop policy is unit-testable without
+ * spawning a child. Line framing — UTF-8-safe reassembly of chunks split mid-line
+ * — is handled upstream by attachJsonlLineReader; this only sees complete lines.
+ */
+export declare function classifySubagentLine(line: string): SubagentStdoutLine;
 /**
  * Pool for running hoocode subagents as child processes with bounded concurrency,
  * FIFO queuing with priority support, and automatic slot refill.
@@ -105,6 +129,8 @@ export declare const DEFAULT_SUBAGENT_MAX_TURNS = 50;
  * - "task_timeout" – hard timeout exceeded, process was SIGKILLed
  * - "budget_warning" – token usage crossed 80% threshold (advisory)
  * - "budget_exceeded" – token usage crossed 100% threshold (advisory; never kills)
+ * - "task_progress" – coarse lifecycle event (turn_end, tool start/end) parsed
+ *                    from the child's json event stream, for live UI updates
  */
 export declare class SubagentPool extends EventEmitter {
     private readonly maxConcurrency;

package/dist/core/subagent-pool.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"subagent-pool.d.ts","sourceRoot":"","sources":["../../src/core/subagent-pool.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,KAAK,EAAE,MAAM,oBAAoB,CAAC;AAC3C,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAa3C,MAAM,WAAW,gBAAgB;IAChC,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB;;;;OAIG;IACH,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,8FAA8F;IAC9F,yBAAyB,CAAC,EAAE,OAAO,CAAC;CACpC;AAED,MAAM,WAAW,YAAY;IAC5B,GAAG,EAAE,MAAM,CAAC;IACZ,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,YAAY,EAAE,MAAM,CAAC;IACrB,OAAO,EAAE,UAAU,CAAC,OAAO,KAAK,CAAC,CAAC;CAClC;AAED,MAAM,WAAW,cAAc;IAC9B,OAAO,EAAE,MAAM,CAAC;IAChB,EAAE,EAAE,OAAO,CAAC;IACZ,MAAM,EAAE,MAAM,CAAC;IACf,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,yEAAyE;IACzE,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B,0DAA0D;IAC1D,MAAM,CAAC,EAAE,UAAU,GAAG,SAAS,GAAG,QAAQ,GAAG,SAAS,GAAG,SAAS,CAAC;IACnE,8EAA8E;IAC9E,WAAW,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IACtC,2FAA2F;IAC3F,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACrC;AAED,MAAM,WAAW,UAAU;IAC1B,qFAAqF;IACrF,cAAc,EAAE,OAAO,CAAC;IACxB,2CAA2C;IAC3C,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,sCAAsC;IACtC,MAAM,CAAC,EAAE,cAAc,CAAC;IACxB,+CAA+C;IAC/C,QAAQ,CAAC,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,eAAe;IAC/B;gFAC4E;IAC5E,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,wEAAwE;IACxE,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,8EAA8E;IAC9E,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,iCAAiC;IACjC,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,kEAAkE;IAClE,WAAW,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,mBAAmB;IACnC,0FAA0F;IAC1F,UAAU,EAAE,MAAM,CAAC;IACnB,+EAA+E;IAC/E,UAAU,CAAC,EAAE,MAAM,EAAE,CAAC;IACtB,yDAAyD;IACzD,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,0EAA0E;IAC1E,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,sDAAsD;IACtD,GAAG,CAAC,EAAE,MAAM,CAAC,UAAU,CAAC;IACxB,oDAAoD;IACpD,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B;;;;OAIG;IACH,UAAU,CAAC,EAAE,MAAM,EAAE,CAAC;CACtB;AAED;;;;GAIG;AACH,eAAO,MAAM,0BAA0B,KAAK,CAAC;AAE7C;;;;;;;;;;;;GAYG;AACH,qBAAa,YAAa,SAAQ,YAAY;IAC7C,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAS;IACxC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAS;IACpC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAW;IACtC,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAS;IAC7B,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAoB;IACxC,OAAO,CAAC,QAAQ,CAAC,kBAAkB,CAAS;IAC5C,+EAA+E;IAC/E,OAAO,CAAC,UAAU,CAAW;IAE7B,OAAO,CAAC,KAAK,CAAmC;IAChD,OAAO,CAAC,KAAK,CAA0B;IACvC,OAAO,CAAC,SAAS,CAAqC;IACtD,OAAO,CAAC,OAAO,CAAkG;IACjH,OAAO,CAAC,OAAO,CAAkC;IACjD,OAAO,CAAC,QAAQ,CAAwB;IACxC,OAAO,CAAC,SAAS,CAAoB;IACrC,OAAO,CAAC,QAAQ,CAAS;IACzB,kFAAkF;IAClF,OAAO,CAAC,QAAQ,CAAC,CAAgB;IACjC,kFAAkF;IAClF,OAAO,CAAC,WAAW,CAA4C;IAC/D,qEAAqE;IACrE,OAAO,CAAC,UAAU,CAAgE;IAElF,YAAY,OAAO,EAAE,mBAAmB,EAmBvC;IAED,qEAAqE;IACrE,gBAAgB,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,IAAI,CAEtC;IAED;;;;;OAKG;IACH,eAAe,CAAC,KAAK,EAAE,MAAM,GAAG,IAAI,CAEnC;IAED,0DAA0D;IAC1D,OAAO,CAAC,WAAW;IAOnB,gDAAgD;IAChD,OAAO,CAAC,UAAU;IAMlB,qDAAqD;IACrD,KAAK,CAAC,IAAI,EAAE,gBAAgB,GAAG,IAAI,CAoBlC;IAED,gCAAgC;IAChC,UAAU,CAAC,OAAO,EAAE,MAAM,GAAG,SAAS,GAAG,QAAQ,GAAG,MAAM,GAAG,QAAQ,GAAG,SAAS,GAAG,SAAS,GAAG,SAAS,CAaxG;IAED,yDAAyD;IACzD,QAAQ,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,cAAc,CAAC,CAcjD;IAED,6CAA6C;IAC7C,aAAa,IAAI,MAAM,CAEtB;IAED,4CAA4C;IAC5C,YAAY,IAAI,MAAM,CAErB;IAED;;;;;;;;OAQG;IACG,QAAQ,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,GAAE,eAAoB,GAAG,OAAO,CAAC,UAAU,CAAC,CAiB/E;IAED;;;OAGG;IACH,gBAAgB,CACf,IAAI,EAAE,MAAM,EACZ,OAAO,GAAE,eAAoB,GAC3B;QAAE,cAAc,EAAE,OAAO,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAA;KAAE,CASrF;IAED;;;OAGG;IACH,OAAO,CAAC,aAAa;IA4CrB;;;;OAIG;IACH,OAAO,CAAC,OAAO,EAAE,MAAM,GAAG,cAAc,GAAG,SAAS,CAEnD;IAED,8DAA8D;IAC9D,cAAc,CAAC,OAAO,EAAE,MAAM,EAAE,GAAG,GAAE,MAAiB,GAAG,MAAM,CAE9D;IAED;;;;OAIG;IACG,MAAM,CACX,OAAO,EAAE,MAAM,EACf,MAAM,EAAE,MAAM,EACd,OAAO,GAAE,IAAI,CAAC,eAAe,EAAE,YAAY,GAAG,aAAa,CAAM,GAC/D,OAAO,CAAC,UAAU,CAAC,CAUrB;IAED,gFAAgF;IAChF,OAAO,CAAC,qBAAqB;IAW7B,OAAO,CAAC,gBAAgB;IA0BxB,OAAO,CAAC,eAAe;IAqBvB;;;OAGG;IACH,OAAO,CAAC,kBAAkB;IAQ1B,+EAA+E;IAC/E,OAAO,IAAI,IAAI,CAyBd;IAED,2DAA2D;IAC3D,OAAO,CAAC,IAAI;IAOZ,sCAAsC;IACtC,OAAO,CAAC,SAAS;IA8FjB,kEAAkE;IAClE,OAAO,CAAC,SAAS;IA2QjB,kFAAkF;IAClF,OAAO,CAAC,6BAA6B;IAiBrC,oFAAoF;IACpF,OAAO,CAAC,6BAA6B;IAUrC,yEAAyE;IACzE,OAAO,CAAC,qBAAqB;IAa7B;;;;OAIG;IACH,OAAO,CAAC,mBAAmB;IAiB3B,OAAO,CAAC,iBAAiB;IAWzB,OAAO,CAAC,aAAa;CAerB","sourcesContent":["import { spawn } from \"node:child_process\";\nimport { EventEmitter } from \"node:events\";\nimport { existsSync, mkdirSync, readFileSync, rmSync, writeFileSync } from \"node:fs\";\nimport { dirname, join } from \"node:path\";\nimport { getDispatchTaskDir } from \"../config.js\";\nimport { waitForChildProcess } from \"../utils/child-process.js\";\nimport { MODEL_INHERIT } from \"./agent-frontmatter.js\";\nimport { type AgentRegistry, loadAgentRegistry } from \"./agent-registry.js\";\nimport { DispatchEvaluator } from \"./dispatch-evaluator.js\";\nimport { SubagentLifeguard } from \"./lifeguard.js\";\nimport { OutputVerifier } from \"./output-verifier.js\";\nimport { currentSubagentDepth, resolveMaxSubagentDepth, SUBAGENT_DEPTH_ENV } from \"./subagent-depth.js\";\nimport { TokenBudget } from \"./token-budget.js\";\n\nexport interface SubagentPoolTask {\n\ttask_id: string;\n\tagent_type: string;\n\ttask: string;\n\tcontext?: string;\n\ttoken_budget?: number;\n\tcwd?: string;\n\tmodel?: string;\n\tprovider?: string;\n\t/**\n\t * Explicit session file for the child to persist/continue. When omitted the\n\t * child uses its own dispatch dir (`<dispatch>/<task_id>/session.jsonl`).\n\t * Resume reuses the original task's session file to continue the transcript.\n\t */\n\tsessionFile?: string;\n\t/** Internal: retry using the caller's model when a built-in agent's preferred model fails. */\n\tuseInheritedModelFallback?: boolean;\n}\n\nexport interface SubagentSlot {\n\tpid: number;\n\tagent_type: string;\n\ttask_id: string;\n\tspawned_at: number;\n\ttoken_budget: number;\n\tprocess: ReturnType<typeof spawn>;\n}\n\nexport interface SubagentResult {\n\ttask_id: string;\n\tok: boolean;\n\tstdout: string;\n\tstderr: string;\n\texit_code: number | null;\n\terror?: string;\n\t/** True when the task exceeded its token budget and was hard-stopped. */\n\tbudget_exceeded?: boolean;\n\t/** Terminal status derived from how the task finished. */\n\tstatus?: \"complete\" | \"partial\" | \"failed\" | \"stalled\" | \"timeout\";\n\t/** Parsed result.json content when available (e.g. on partial completion). */\n\tresult_data?: Record<string, unknown>;\n\t/** True when this run used the inherited-model fallback (preferred model failed first). */\n\tusedInheritedModelFallback?: boolean;\n}\n\nexport interface TaskResult {\n\t/** True when the evaluator decided the task is simple enough for inline handling. */\n\thandled_inline: boolean;\n\t/** Present when the task was delegated. */\n\ttask_id?: string;\n\tagent_type?: string;\n\treason?: string;\n\t/** Subagent result when delegated. */\n\tresult?: SubagentResult;\n\t/** Duration in milliseconds when delegated. */\n\tduration?: number;\n}\n\nexport interface DispatchOptions {\n\t/** Skip evaluation and force this agent type (user/explicit override).\n\t *  Accepts any registry-defined agent name, not just the built-in modes. */\n\tforceAgent?: string;\n\t/** Context distilled from the calling agent, passed to the subagent. */\n\tcontext?: string;\n\t/** Model id for the subagent (defaults to the child's configured default). */\n\tmodel?: string;\n\t/** Provider for the subagent. */\n\tprovider?: string;\n\t/** Explicit session file to persist/continue (used by resume). */\n\tsessionFile?: string;\n}\n\nexport interface SubagentPoolOptions {\n\t/** Path to the hoocode executable (or the runtime, e.g. node, when prefixArgs is set). */\n\texecutable: string;\n\t/** Args inserted before task args (e.g. the CLI entry script for node/tsx). */\n\tprefixArgs?: string[];\n\t/** Maximum concurrent child processes. Defaults to 5. */\n\tmaxConcurrency?: number;\n\t/** Working directory for spawned processes. Defaults to process.cwd(). */\n\tcwd?: string;\n\t/** Environment variables. Defaults to process.env. */\n\tenv?: NodeJS.ProcessEnv;\n\t/** Default token budget per task. Defaults to 0. */\n\tdefaultTokenBudget?: number;\n\t/**\n\t * Non-default skill paths to forward to every spawned subagent via --skill.\n\t * Subagents auto-discover skills from standard locations; only paths that\n\t * won't be found by default discovery need to be forwarded here.\n\t */\n\tskillPaths?: string[];\n}\n\n/**\n * Default hard cap on assistant turns for a spawned subagent when its definition\n * does not set `maxTurns`. The token budget is advisory (it warns but never\n * kills), so this turn cap is the guaranteed hard stop for every subagent.\n */\nexport const DEFAULT_SUBAGENT_MAX_TURNS = 50;\n\n/**\n * Pool for running hoocode subagents as child processes with bounded concurrency,\n * FIFO queuing with priority support, and automatic slot refill.\n *\n * Events:\n * - \"task_done\"    – task completed successfully and output was verified\n * - \"task_failed\"  – task failed (spawn error, bad exit code, verification failure)\n * - \"task_stalled\" – heartbeat missed past the load-scaled threshold (60s base,\n *                    widened under concurrency/event-loop lag), process SIGKILLed\n * - \"task_timeout\" – hard timeout exceeded, process was SIGKILLed\n * - \"budget_warning\" – token usage crossed 80% threshold (advisory)\n * - \"budget_exceeded\" – token usage crossed 100% threshold (advisory; never kills)\n */\nexport class SubagentPool extends EventEmitter {\n\tprivate readonly maxConcurrency: number;\n\tprivate readonly executable: string;\n\tprivate readonly prefixArgs: string[];\n\tprivate readonly cwd: string;\n\tprivate readonly env: NodeJS.ProcessEnv;\n\tprivate readonly defaultTokenBudget: number;\n\t/** Non-default skill paths forwarded to every spawned subagent via --skill. */\n\tprivate skillPaths: string[];\n\n\tprivate slots = new Map<string, SubagentSlot>();\n\tprivate queue: SubagentPoolTask[] = [];\n\tprivate completed = new Map<string, SubagentResult>();\n\tprivate waiters = new Map<string, { resolve: (result: SubagentResult) => void; reject: (err: Error) => void }>();\n\tprivate budgets = new Map<string, TokenBudget>();\n\tprivate verifier = new OutputVerifier();\n\tprivate lifeguard: SubagentLifeguard;\n\tprivate disposed = false;\n\t/** Lazily-loaded agent registry (frontmatter definitions) for this pool's cwd. */\n\tprivate registry?: AgentRegistry;\n\t/** Tracks why a task was killed (stalled / timeout) before exit handler fires. */\n\tprivate killReasons = new Map<string, \"stalled\" | \"timeout\">();\n\t/** Persistent terminal status map, survives wait_for consumption. */\n\tprivate taskStatus = new Map<string, \"done\" | \"failed\" | \"stalled\" | \"timeout\">();\n\n\tconstructor(options: SubagentPoolOptions) {\n\t\tsuper();\n\t\tthis.maxConcurrency = options.maxConcurrency ?? 5;\n\t\tthis.executable = options.executable;\n\t\tthis.prefixArgs = options.prefixArgs ?? [];\n\t\tthis.cwd = options.cwd ?? process.cwd();\n\t\tthis.env = options.env ?? process.env;\n\t\tthis.defaultTokenBudget = options.defaultTokenBudget ?? 0;\n\t\tthis.skillPaths = options.skillPaths ? [...options.skillPaths] : [];\n\t\tthis.verifier = new OutputVerifier(this.cwd);\n\t\tthis.lifeguard = new SubagentLifeguard(this.cwd);\n\t\tthis.lifeguard.on(\"stalled\", (data: { task_id: string; pid: number }) => {\n\t\t\tthis.killReasons.set(data.task_id, \"stalled\");\n\t\t\tthis.emit(\"task_stalled\", data);\n\t\t});\n\t\tthis.lifeguard.on(\"timeout\", (data: { task_id: string; pid: number }) => {\n\t\t\tthis.killReasons.set(data.task_id, \"timeout\");\n\t\t\tthis.emit(\"task_timeout\", data);\n\t\t});\n\t}\n\n\t/** Update the non-default skill paths forwarded to new subagents. */\n\tupdateSkillPaths(paths: string[]): void {\n\t\tthis.skillPaths = [...paths];\n\t}\n\n\t/**\n\t * Report external in-process load (e.g. the number of background MCP tools\n\t * currently executing in the parent) to the lifeguard. This widens its\n\t * heartbeat/timeout tolerance so monitored subagents aren't false-positive\n\t * reaped when the parent's event loop is busy with concurrent background work.\n\t */\n\tsetExternalLoad(count: number): void {\n\t\tthis.lifeguard.setExternalLoad(count);\n\t}\n\n\t/** Lazily load the agent registry for this pool's cwd. */\n\tprivate getRegistry(): AgentRegistry {\n\t\tif (!this.registry) {\n\t\t\tthis.registry = loadAgentRegistry({ cwd: this.cwd });\n\t\t}\n\t\treturn this.registry;\n\t}\n\n\t/** Priority value: higher numbers run first. */\n\tprivate priorityOf(agent_type: string): number {\n\t\t// Read-only investigation (explore/plan) often unblocks downstream work, so\n\t\t// it runs ahead of other agents.\n\t\treturn agent_type === \"explore\" || agent_type === \"plan\" ? 2 : 1;\n\t}\n\n\t/** Queue a task. It will run when a slot is free. */\n\tspawn(task: SubagentPoolTask): void {\n\t\tif (this.disposed) {\n\t\t\tthrow new Error(\"SubagentPool has been disposed\");\n\t\t}\n\t\tif (\n\t\t\tthis.slots.has(task.task_id) ||\n\t\t\tthis.queue.some((t) => t.task_id === task.task_id) ||\n\t\t\tthis.completed.has(task.task_id)\n\t\t) {\n\t\t\tthrow new Error(`Duplicate task_id: ${task.task_id}`);\n\t\t}\n\n\t\tconst p = this.priorityOf(task.agent_type);\n\t\tconst idx = this.queue.findIndex((t) => this.priorityOf(t.agent_type) < p);\n\t\tif (idx === -1) {\n\t\t\tthis.queue.push(task);\n\t\t} else {\n\t\t\tthis.queue.splice(idx, 0, task);\n\t\t}\n\t\tthis.pull();\n\t}\n\n\t/** Current status of a task. */\n\tget_status(task_id: string): \"running\" | \"queued\" | \"done\" | \"failed\" | \"stalled\" | \"timeout\" | \"unknown\" {\n\t\tif (this.slots.has(task_id)) return \"running\";\n\t\tif (this.queue.some((t) => t.task_id === task_id)) return \"queued\";\n\t\tconst persisted = this.taskStatus.get(task_id);\n\t\tif (persisted) return persisted;\n\t\tconst result = this.completed.get(task_id);\n\t\tif (result) {\n\t\t\tif (result.status === \"stalled\") return \"stalled\";\n\t\t\tif (result.status === \"timeout\") return \"timeout\";\n\t\t\tif (result.ok) return \"done\";\n\t\t\treturn \"failed\";\n\t\t}\n\t\treturn \"unknown\";\n\t}\n\n\t/** Wait for a task to complete and return its result. */\n\twait_for(task_id: string): Promise<SubagentResult> {\n\t\tif (this.disposed) {\n\t\t\treturn Promise.reject(new Error(\"SubagentPool has been disposed\"));\n\t\t}\n\n\t\tconst existing = this.completed.get(task_id);\n\t\tif (existing) {\n\t\t\tthis.completed.delete(task_id);\n\t\t\treturn Promise.resolve(existing);\n\t\t}\n\n\t\treturn new Promise((resolve, reject) => {\n\t\t\tthis.waiters.set(task_id, { resolve, reject });\n\t\t});\n\t}\n\n\t/** Number of currently running subagents. */\n\trunning_count(): number {\n\t\treturn this.slots.size;\n\t}\n\n\t/** Number of tasks waiting in the queue. */\n\tqueued_count(): number {\n\t\treturn this.queue.length;\n\t}\n\n\t/**\n\t * Dispatch a task through the evaluator.\n\t *\n\t * - If `options.forceAgent` is provided, skip evaluation and spawn directly.\n\t * - Otherwise evaluate the task. If it should be handled inline, return\n\t *   `{ handled_inline: true }` immediately.\n\t * - If delegating, spawn the subagent, wait for completion, write\n\t *   `output.json`, and return the result.\n\t */\n\tasync dispatch(task: string, options: DispatchOptions = {}): Promise<TaskResult> {\n\t\tif (this.disposed) {\n\t\t\treturn Promise.reject(new Error(\"SubagentPool has been disposed\"));\n\t\t}\n\t\tconst begin = this.beginDispatch(task, options);\n\t\tif (begin.handled_inline) {\n\t\t\treturn { handled_inline: true, reason: begin.reason };\n\t\t}\n\t\tconst result = await this.wait_for(begin.task_id);\n\t\treturn {\n\t\t\thandled_inline: false,\n\t\t\ttask_id: begin.task_id,\n\t\t\tagent_type: begin.agent_type,\n\t\t\treason: begin.reason,\n\t\t\tresult,\n\t\t\tduration: Date.now() - begin.startTime,\n\t\t};\n\t}\n\n\t/**\n\t * Fire-and-forget dispatch for background agents. Spawns the subagent and\n\t * returns its handle immediately; the caller polls get_status()/collect().\n\t */\n\tdispatchDetached(\n\t\ttask: string,\n\t\toptions: DispatchOptions = {},\n\t): { handled_inline: boolean; task_id?: string; agent_type?: string; reason?: string } {\n\t\tif (this.disposed) {\n\t\t\tthrow new Error(\"SubagentPool has been disposed\");\n\t\t}\n\t\tconst begin = this.beginDispatch(task, options);\n\t\tif (begin.handled_inline) {\n\t\t\treturn { handled_inline: true, reason: begin.reason };\n\t\t}\n\t\treturn { handled_inline: false, task_id: begin.task_id, agent_type: begin.agent_type, reason: begin.reason };\n\t}\n\n\t/**\n\t * Evaluate, log, and spawn a task without waiting. Shared by dispatch()\n\t * (blocking) and dispatchDetached() (background).\n\t */\n\tprivate beginDispatch(\n\t\ttask: string,\n\t\toptions: DispatchOptions,\n\t):\n\t\t| { handled_inline: true; reason?: string }\n\t\t| { handled_inline: false; task_id: string; agent_type: string; reason?: string; startTime: number } {\n\t\tconst { forceAgent, context, model, provider, sessionFile } = options;\n\t\tconst evaluator = new DispatchEvaluator();\n\t\tconst analysis = evaluator.evaluate(task);\n\n\t\tif (!forceAgent && !analysis.should_delegate) {\n\t\t\treturn { handled_inline: true, reason: analysis.reason };\n\t\t}\n\n\t\tconst agent_type = forceAgent ?? \"general-purpose\";\n\t\tconst task_id = `dispatch-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;\n\t\tconst reason = forceAgent ? \"user_override\" : analysis.reason;\n\t\tconst complexity = analysis.estimated_complexity;\n\t\t// Depth of the child about to be spawned (this process's depth + 1). Surfaced\n\t\t// so a delegation tree's nesting is visible in logs without extra tooling.\n\t\tconst childDepth = currentSubagentDepth(this.env) + 1;\n\n\t\t// Pre-dispatch logging. Use stderr: stdout is reserved for the JSON event\n\t\t// stream / TUI render and must not be polluted.\n\t\tconsole.error(\n\t\t\t`[DISPATCH] agent=${agent_type} depth=${childDepth} reason=${reason} complexity=${complexity} task_id=${task_id}`,\n\t\t);\n\t\tthis.writeDispatchLog(task_id, agent_type, reason, complexity, task, childDepth);\n\n\t\tconst poolTask: SubagentPoolTask = {\n\t\t\ttask_id,\n\t\t\tagent_type,\n\t\t\ttask,\n\t\t\tcontext,\n\t\t\tmodel,\n\t\t\tprovider,\n\t\t\tsessionFile,\n\t\t\tcwd: this.cwd,\n\t\t};\n\t\tconst startTime = Date.now();\n\t\tthis.spawn(poolTask);\n\t\treturn { handled_inline: false, task_id, agent_type, reason, startTime };\n\t}\n\n\t/**\n\t * Non-destructively read a completed task's result (for background polling).\n\t * Returns undefined while the task is still running/queued, or if its result\n\t * was already consumed via wait_for().\n\t */\n\tcollect(task_id: string): SubagentResult | undefined {\n\t\treturn this.completed.get(task_id);\n\t}\n\n\t/** Absolute path of the persisted session file for a task. */\n\tgetSessionFile(task_id: string, cwd: string = this.cwd): string {\n\t\treturn join(getDispatchTaskDir(cwd, task_id), \"session.jsonl\");\n\t}\n\n\t/**\n\t * Resume a previously dispatched subagent, continuing its persisted session\n\t * with a follow-up prompt. Recovers the original agent type from its dispatch\n\t * log. Rejects if no resumable session exists for the task.\n\t */\n\tasync resume(\n\t\ttask_id: string,\n\t\tprompt: string,\n\t\toptions: Omit<DispatchOptions, \"forceAgent\" | \"sessionFile\"> = {},\n\t): Promise<TaskResult> {\n\t\tif (this.disposed) {\n\t\t\treturn Promise.reject(new Error(\"SubagentPool has been disposed\"));\n\t\t}\n\t\tconst sessionFile = this.getSessionFile(task_id);\n\t\tif (!existsSync(sessionFile)) {\n\t\t\treturn Promise.reject(new Error(`No resumable session for task \"${task_id}\" (expected ${sessionFile}).`));\n\t\t}\n\t\tconst agent_type = this.readDispatchAgentType(task_id) ?? \"general-purpose\";\n\t\treturn this.dispatch(prompt, { ...options, forceAgent: agent_type, sessionFile });\n\t}\n\n\t/** Recover the agent type a task was dispatched with, from its dispatch log. */\n\tprivate readDispatchAgentType(task_id: string): string | undefined {\n\t\tconst path = join(getDispatchTaskDir(this.cwd, task_id), \"dispatch-log.json\");\n\t\tif (!existsSync(path)) return undefined;\n\t\ttry {\n\t\t\tconst parsed = JSON.parse(readFileSync(path, \"utf-8\")) as { agent_type?: string };\n\t\t\treturn typeof parsed.agent_type === \"string\" ? parsed.agent_type : undefined;\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate writeDispatchLog(\n\t\ttask_id: string,\n\t\tagent_type: string,\n\t\treason: string,\n\t\tcomplexity: string,\n\t\ttask: string,\n\t\tdepth: number,\n\t): void {\n\t\tconst log = {\n\t\t\ttimestamp: new Date().toISOString(),\n\t\t\ttask_id,\n\t\t\tagent_type,\n\t\t\tdepth,\n\t\t\treason,\n\t\t\tcomplexity,\n\t\t\ttask,\n\t\t};\n\t\tconst path = join(getDispatchTaskDir(this.cwd, task_id), \"dispatch-log.json\");\n\t\ttry {\n\t\t\tmkdirSync(dirname(path), { recursive: true });\n\t\t\twriteFileSync(path, JSON.stringify(log, null, 2));\n\t\t} catch {\n\t\t\t// Best-effort persistence\n\t\t}\n\t}\n\n\tprivate writeOutputJson(task_id: string, result: SubagentResult): void {\n\t\tconst output = {\n\t\t\ttask_id: result.task_id,\n\t\t\tok: result.ok,\n\t\t\texit_code: result.exit_code,\n\t\t\tstatus: result.status,\n\t\t\tstdout: result.stdout,\n\t\t\tstderr: result.stderr,\n\t\t\terror: result.error,\n\t\t\tbudget_exceeded: result.budget_exceeded,\n\t\t\tresult_data: result.result_data,\n\t\t};\n\t\tconst path = join(getDispatchTaskDir(this.cwd, task_id), \"output.json\");\n\t\ttry {\n\t\t\tmkdirSync(dirname(path), { recursive: true });\n\t\t\twriteFileSync(path, JSON.stringify(output, null, 2));\n\t\t} catch {\n\t\t\t// Best-effort persistence\n\t\t}\n\t}\n\n\t/**\n\t * Remove a task's dispatch dir after a clean, verified success. Best-effort:\n\t * a cleanup failure must never fail an otherwise successful task.\n\t */\n\tprivate cleanupDispatchDir(task_id: string, cwd: string): void {\n\t\ttry {\n\t\t\trmSync(getDispatchTaskDir(cwd, task_id), { recursive: true, force: true });\n\t\t} catch {\n\t\t\t// Best-effort cleanup\n\t\t}\n\t}\n\n\t/** Kill all running processes, clear the queue, and reject pending waiters. */\n\tdispose(): void {\n\t\tif (this.disposed) return;\n\t\tthis.disposed = true;\n\n\t\tfor (const slot of this.slots.values()) {\n\t\t\tif (!slot.process.killed) {\n\t\t\t\tslot.process.kill(\"SIGTERM\");\n\t\t\t}\n\t\t}\n\t\tthis.slots.clear();\n\t\tthis.queue = [];\n\n\t\tfor (const [task_id, waiter] of this.waiters) {\n\t\t\twaiter.reject(new Error(\"SubagentPool disposed\"));\n\t\t\tthis.waiters.delete(task_id);\n\t\t}\n\t\tthis.completed.clear();\n\t\tfor (const budget of this.budgets.values()) {\n\t\t\tbudget.removeAllListeners();\n\t\t}\n\t\tthis.budgets.clear();\n\t\tthis.killReasons.clear();\n\t\tthis.taskStatus.clear();\n\t\tthis.lifeguard.dispose();\n\t\tthis.removeAllListeners();\n\t}\n\n\t/** Pull tasks from the queue while slots are available. */\n\tprivate pull(): void {\n\t\twhile (this.slots.size < this.maxConcurrency && this.queue.length > 0) {\n\t\t\tconst task = this.queue.shift()!;\n\t\t\tthis.startTask(task, false);\n\t\t}\n\t}\n\n\t/** Build CLI arguments for a task. */\n\tprivate buildArgs(task: SubagentPoolTask): string[] {\n\t\t// Persist the child's session so a finished/interrupted subagent can be\n\t\t// resumed later (see resume()). SessionManager.open() creates the file on\n\t\t// first run and continues it on subsequent runs.\n\t\tconst sessionFile = task.sessionFile ?? this.getSessionFile(task.task_id, task.cwd ?? this.cwd);\n\t\tconst args: string[] = [\n\t\t\t...this.prefixArgs,\n\t\t\t\"--mode\",\n\t\t\t\"json\",\n\t\t\t\"--session\",\n\t\t\tsessionFile,\n\t\t\t\"--task-id\",\n\t\t\ttask.task_id,\n\t\t];\n\n\t\t// Prefer the data-driven agent definition from the registry; fall back to the\n\t\t// built-in mode prompt/allowlist for legacy modes not present in the registry.\n\t\tconst def = task.agent_type ? this.getRegistry().get(task.agent_type) : undefined;\n\n\t\tif (task.agent_type) {\n\t\t\tconst systemPrompt = def?.prompt;\n\t\t\tif (systemPrompt) {\n\t\t\t\targs.push(\"--system-prompt\", systemPrompt);\n\t\t\t}\n\n\t\t\t// A `delegate: true` agent may itself dispatch via the Task tool, but only\n\t\t\t// while the child it becomes can still nest (childDepth < cap) — so the\n\t\t\t// deepest permitted level cannot delegate further. Gating here keeps the\n\t\t\t// authorization explicit and bounded by the same cap as the depth guard.\n\t\t\tconst childDepth = currentSubagentDepth(this.env) + 1;\n\t\t\tconst canChildDelegate = def?.delegate === true && childDepth < resolveMaxSubagentDepth(undefined, this.env);\n\n\t\t\t// Tool allowlist comes from the agent definition's frontmatter `tools`\n\t\t\t// field (read-only built-ins declare their own sandbox). When omitted, no\n\t\t\t// --tools is passed and the subagent inherits all parent tools (so the Task\n\t\t\t// tool already survives). A delegating agent with an explicit allowlist must\n\t\t\t// have Task/TaskOutput added, or the child would filter them out.\n\t\t\tconst tools = def?.tools ? [...def.tools] : undefined;\n\t\t\tif (canChildDelegate && tools) {\n\t\t\t\tfor (const t of [\"Task\", \"TaskOutput\"]) {\n\t\t\t\t\tif (!tools.includes(t)) tools.push(t);\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (tools && tools.length > 0) {\n\t\t\t\targs.push(\"--tools\", tools.join(\",\"));\n\t\t\t}\n\t\t\tif (def?.disallowedTools && def.disallowedTools.length > 0) {\n\t\t\t\targs.push(\"--disallowed-tools\", def.disallowedTools.join(\",\"));\n\t\t\t}\n\n\t\t\t// Propagate subagent enablement so the child registers the Task tool; without\n\t\t\t// this the flag-based enablement would not reach a spawned child.\n\t\t\tif (canChildDelegate) {\n\t\t\t\targs.push(\"--enable-subagents\");\n\t\t\t\t// Scoped delegation: restrict which agent types this child may spawn.\n\t\t\t\tif (def?.delegateTo && def.delegateTo.length > 0) {\n\t\t\t\t\targs.push(\"--delegate-allow\", def.delegateTo.join(\",\"));\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\n\t\t// Model precedence: a definition's explicit model wins (unless it is the\n\t\t// `inherit` sentinel), otherwise use the caller-provided model. Built-in\n\t\t// agents can retry with the inherited model when their preferred model is\n\t\t// unavailable or quota-limited.\n\t\tconst explicitModel =\n\t\t\t!task.useInheritedModelFallback && def?.model && def.model !== MODEL_INHERIT ? def.model : undefined;\n\t\tconst modelToUse = explicitModel ?? task.model;\n\t\tif (modelToUse) {\n\t\t\targs.push(\"--model\", modelToUse);\n\t\t}\n\t\tif (task.provider) {\n\t\t\targs.push(\"--provider\", task.provider);\n\t\t}\n\n\t\t// Always give subagents a hard turn cap. With the token budget now advisory\n\t\t// (warn-only), this is the guaranteed hard stop for a runaway subagent.\n\t\tconst maxTurns = def?.maxTurns && def.maxTurns > 0 ? def.maxTurns : DEFAULT_SUBAGENT_MAX_TURNS;\n\t\targs.push(\"--max-turns\", String(maxTurns));\n\n\t\t// Forward non-default skill paths so the subagent has access to all parent skills.\n\t\t// Standard discovery locations (~/.hoocode/, .hoocode/, .claude/) are found automatically.\n\t\tfor (const skillPath of this.skillPaths) {\n\t\t\targs.push(\"--skill\", skillPath);\n\t\t}\n\n\t\tconst prompt = task.context?.trim()\n\t\t\t? `Context from the calling agent:\\n\\n${task.context.trim()}\\n\\nTask: ${task.task.trim()}`\n\t\t\t: `Task: ${task.task.trim()}`;\n\t\targs.push(prompt);\n\n\t\treturn args;\n\t}\n\n\t/** Start a task in a child process, with one retry on failure. */\n\tprivate startTask(task: SubagentPoolTask, isRetry: boolean): void {\n\t\t// Get or create a TokenBudget tracker. On retry, reuse the existing one\n\t\t// so cumulative usage persists across retries.\n\t\tlet budget = this.budgets.get(task.task_id);\n\t\tif (!budget) {\n\t\t\tbudget = new TokenBudget(task.task_id, task.agent_type, {\n\t\t\t\tlimit: task.token_budget,\n\t\t\t\tcwd: task.cwd ?? this.cwd,\n\t\t\t});\n\t\t\tbudget.on(\"budget_warning\", (data: { task_id: string; message: string; used: number; limit: number }) => {\n\t\t\t\tthis.emit(\"budget_warning\", data);\n\t\t\t});\n\t\t\t// The token budget is advisory: surface telemetry but never kill. The\n\t\t\t// guaranteed hard stop is the per-subagent turn cap (--max-turns); see\n\t\t\t// DEFAULT_SUBAGENT_MAX_TURNS.\n\t\t\tbudget.on(\"budget_exceeded\", (data: { task_id: string; used: number; limit: number }) => {\n\t\t\t\tthis.emit(\"budget_exceeded\", data);\n\t\t\t});\n\t\t\tthis.budgets.set(task.task_id, budget);\n\t\t}\n\n\t\tlet proc: ReturnType<typeof spawn>;\n\t\ttry {\n\t\t\tproc = spawn(this.executable, this.buildArgs(task), {\n\t\t\t\tcwd: task.cwd ?? this.cwd,\n\t\t\t\t// Stamp the child's depth (parent depth + 1) so its own guard knows where\n\t\t\t\t// it sits in the tree. The tree-wide cap (HOOCODE_SUBAGENT_MAX_DEPTH) is\n\t\t\t\t// inherited via the spread; at the default cap of 1 the child lands at\n\t\t\t\t// depth 1 and cannot spawn further subagents.\n\t\t\t\tenv: {\n\t\t\t\t\t...this.env,\n\t\t\t\t\t[SUBAGENT_DEPTH_ENV]: String(currentSubagentDepth(this.env) + 1),\n\t\t\t\t},\n\t\t\t\tshell: false,\n\t\t\t\tstdio: [\"ignore\", \"pipe\", \"pipe\"],\n\t\t\t});\n\t\t} catch {\n\t\t\tif (!isRetry) {\n\t\t\t\tthis.startTask(task, true);\n\t\t\t} else {\n\t\t\t\tthis.emit(\"task_failed\", {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\terror: \"Spawn failed synchronously\",\n\t\t\t\t});\n\t\t\t\tthis.resolveWaiter(task.task_id, {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\tok: false,\n\t\t\t\t\tstdout: \"\",\n\t\t\t\t\tstderr: \"\",\n\t\t\t\t\texit_code: null,\n\t\t\t\t\terror: \"Spawn failed synchronously\",\n\t\t\t\t\tstatus: \"failed\",\n\t\t\t\t});\n\t\t\t\tthis.pull();\n\t\t\t}\n\t\t\treturn;\n\t\t}\n\n\t\tconst slot: SubagentSlot = {\n\t\t\tpid: proc.pid ?? 0,\n\t\t\tagent_type: task.agent_type,\n\t\t\ttask_id: task.task_id,\n\t\t\tspawned_at: Date.now(),\n\t\t\ttoken_budget: task.token_budget ?? this.defaultTokenBudget,\n\t\t\tprocess: proc,\n\t\t};\n\n\t\tthis.slots.set(task.task_id, slot);\n\t\tthis.lifeguard.monitor(task.task_id, task.agent_type, proc);\n\n\t\tlet stdout = \"\";\n\t\tlet stderr = \"\";\n\n\t\tproc.stdout?.on(\"data\", (data: Buffer) => {\n\t\t\tconst chunk = data.toString();\n\t\t\tstdout += chunk;\n\t\t\tbudget.processStdout(chunk);\n\n\t\t\t// Any output proves the child is alive and working, so treat it as a\n\t\t\t// heartbeat. The dedicated {\"ping\":true} line below still matters for\n\t\t\t// quiet phases (e.g. a long single model turn that emits nothing), but\n\t\t\t// relying on it alone falsely reaps subagents that are busily streaming\n\t\t\t// events while the parent's event loop is starved by concurrent load.\n\t\t\tthis.lifeguard.recordHeartbeat(task.task_id);\n\n\t\t\t// Heartbeat detection: look for {\"ping\":true} JSON lines\n\t\t\tfor (const raw of chunk.split(\"\\n\")) {\n\t\t\t\tconst line = raw.trim();\n\t\t\t\tif (!line.startsWith(\"{\")) continue;\n\t\t\t\ttry {\n\t\t\t\t\tconst parsed = JSON.parse(line) as Record<string, unknown>;\n\t\t\t\t\tif (parsed.ping === true) {\n\t\t\t\t\t\tthis.lifeguard.recordHeartbeat(task.task_id);\n\t\t\t\t\t}\n\t\t\t\t} catch {\n\t\t\t\t\t// Not a ping line, ignore\n\t\t\t\t}\n\t\t\t}\n\t\t});\n\t\tproc.stderr?.on(\"data\", (data: Buffer) => {\n\t\t\tstderr += data.toString();\n\t\t});\n\n\t\twaitForChildProcess(proc)\n\t\t\t.then((code) => {\n\t\t\t\tthis.slots.delete(task.task_id);\n\t\t\t\tbudget.flush();\n\n\t\t\t\tconst killReason = this.killReasons.get(task.task_id);\n\t\t\t\tthis.killReasons.delete(task.task_id);\n\n\t\t\t\tconst duration = Date.now() - slot.spawned_at;\n\t\t\t\tconst tokens_used = budget.getUsed();\n\t\t\t\tconst budgetExceeded = budget.isExceeded();\n\n\t\t\t\t// A subagent's success is defined by a valid, verified result.json, not by\n\t\t\t\t// its exit code. A child that finished its work and wrote a valid result can\n\t\t\t\t// still be SIGKILLed by the lifeguard before it exits on its own (lingering\n\t\t\t\t// open handles delay a natural exit past the heartbeat threshold), which forces\n\t\t\t\t// exit_code === null. Keying completion off the verified result, not code === 0,\n\t\t\t\t// honors that genuine success instead of discarding it as a false stall.\n\t\t\t\tconst verification = this.verifier.verify(task.task_id, task.cwd ?? this.cwd);\n\t\t\t\t// A well-formed result.json counts as clean completion unless its own\n\t\t\t\t// status field declares failure (e.g. \"failed\" from a provider quota\n\t\t\t\t// error). Without this check the pool would treat a child that wrote a\n\t\t\t\t// valid-but-failed result.json and exited non-zero as a success.\n\t\t\t\tlet cleanlyCompleted = code === 0 || verification.valid;\n\t\t\t\tif (cleanlyCompleted && verification.valid) {\n\t\t\t\t\tconst rd = this.tryReadResultJson(task.task_id, task.cwd ?? this.cwd);\n\t\t\t\t\tif (rd && (rd as Record<string, unknown>).status === \"failed\") {\n\t\t\t\t\t\tcleanlyCompleted = false;\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\t// If killed by lifeguard before producing a valid result, honor the kill.\n\t\t\t\tif ((killReason === \"stalled\" || killReason === \"timeout\") && !verification.valid) {\n\t\t\t\t\tconst result: SubagentResult = {\n\t\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\t\tok: false,\n\t\t\t\t\t\tstdout,\n\t\t\t\t\t\tstderr,\n\t\t\t\t\t\texit_code: code,\n\t\t\t\t\t\tstatus: killReason,\n\t\t\t\t\t};\n\t\t\t\t\tthis.writeOutputJson(task.task_id, result);\n\t\t\t\t\tthis.emit(`task_${killReason}`, {\n\t\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\t\tagent_type: task.agent_type,\n\t\t\t\t\t\tduration,\n\t\t\t\t\t\ttokens_used,\n\t\t\t\t\t});\n\t\t\t\t\tthis.resolveWaiter(task.task_id, result);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\n\t\t\t\tconst result: SubagentResult = {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\tok: cleanlyCompleted,\n\t\t\t\t\tstdout,\n\t\t\t\t\tstderr,\n\t\t\t\t\texit_code: code,\n\t\t\t\t\t// Advisory telemetry only: exceeding the budget never fails the task.\n\t\t\t\t\tbudget_exceeded: budgetExceeded,\n\t\t\t\t\tstatus: cleanlyCompleted ? \"complete\" : \"failed\",\n\t\t\t\t\tusedInheritedModelFallback: task.useInheritedModelFallback === true,\n\t\t\t\t};\n\n\t\t\t\tif (result.ok) {\n\t\t\t\t\tif (!verification.valid) {\n\t\t\t\t\t\tresult.ok = false;\n\t\t\t\t\t\tresult.error = verification.reason;\n\t\t\t\t\t\tresult.status = \"failed\";\n\t\t\t\t\t\tthis.writeOutputJson(task.task_id, result);\n\t\t\t\t\t\tthis.emit(\"task_failed\", {\n\t\t\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\t\t\tagent_type: task.agent_type,\n\t\t\t\t\t\t\tduration,\n\t\t\t\t\t\t\ttokens_used,\n\t\t\t\t\t\t\terror: verification.reason,\n\t\t\t\t\t\t});\n\t\t\t\t\t\tthis.resolveWaiter(task.task_id, result);\n\t\t\t\t\t\treturn;\n\t\t\t\t\t}\n\t\t\t\t\t// Attach the verified result.json so callers can read the summary\n\t\t\t\t\t// without parsing the raw event stream.\n\t\t\t\t\tresult.result_data = this.tryReadResultJson(task.task_id, task.cwd ?? this.cwd);\n\n\t\t\t\t\t// Clean success: discard the per-task dispatch dir entirely\n\t\t\t\t\t// (session.jsonl, result.json, dispatch-log.json, budget.json). The\n\t\t\t\t\t// in-memory result already carries result_data, so callers lose\n\t\t\t\t\t// nothing. Trade-off: resume() only works for non-successful tasks.\n\t\t\t\t\tthis.cleanupDispatchDir(task.task_id, task.cwd ?? this.cwd);\n\n\t\t\t\t\tthis.emit(\"task_done\", {\n\t\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\t\tagent_type: task.agent_type,\n\t\t\t\t\t\tduration,\n\t\t\t\t\t\ttokens_used,\n\t\t\t\t\t\tstatus: \"complete\",\n\t\t\t\t\t});\n\t\t\t\t\tthis.resolveWaiter(task.task_id, result);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\n\t\t\t\t// Failure path: keep the dispatch dir for debugging and persist output.\n\t\t\t\t// Attach the child's result.json (if any) and derive a concrete failure\n\t\t\t\t// reason so callers see the real cause (e.g. a provider usage/quota\n\t\t\t\t// error) instead of a generic \"subagent failed\".\n\t\t\t\tresult.result_data = this.tryReadResultJson(task.task_id, task.cwd ?? this.cwd);\n\t\t\t\tif (!result.error) {\n\t\t\t\t\tresult.error = this.deriveFailureReason(result);\n\t\t\t\t}\n\t\t\t\tif (this.shouldRetryWithInheritedModel(task, result)) {\n\t\t\t\t\tconsole.error(\n\t\t\t\t\t\t`[DISPATCH] agent=${task.agent_type} task_id=${task.task_id} preferred model failed; retrying with inherited model`,\n\t\t\t\t\t);\n\t\t\t\t\tthis.cleanupRetryArtifacts(task);\n\t\t\t\t\tthis.queue.unshift({ ...task, useInheritedModelFallback: true });\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tthis.writeOutputJson(task.task_id, result);\n\t\t\t\tthis.emit(\"task_failed\", {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\tagent_type: task.agent_type,\n\t\t\t\t\tduration,\n\t\t\t\t\ttokens_used,\n\t\t\t\t\terror: result.error ?? `Exited with code ${code}`,\n\t\t\t\t});\n\t\t\t\tthis.resolveWaiter(task.task_id, result);\n\t\t\t})\n\t\t\t.catch((err) => {\n\t\t\t\tthis.slots.delete(task.task_id);\n\t\t\t\tbudget.flush();\n\t\t\t\tconst duration = Date.now() - slot.spawned_at;\n\t\t\t\tconst tokens_used = budget.getUsed();\n\t\t\t\tif (!isRetry) {\n\t\t\t\t\tthis.startTask(task, true);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tconst error = err instanceof Error ? err.message : String(err);\n\t\t\t\tconst result: SubagentResult = {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\tok: false,\n\t\t\t\t\tstdout,\n\t\t\t\t\tstderr,\n\t\t\t\t\texit_code: null,\n\t\t\t\t\terror,\n\t\t\t\t\tstatus: \"failed\",\n\t\t\t\t\tusedInheritedModelFallback: task.useInheritedModelFallback === true,\n\t\t\t\t};\n\t\t\t\tthis.writeOutputJson(task.task_id, result);\n\t\t\t\tthis.emit(\"task_failed\", {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\tagent_type: task.agent_type,\n\t\t\t\t\tduration,\n\t\t\t\t\ttokens_used,\n\t\t\t\t\terror,\n\t\t\t\t});\n\t\t\t\tthis.resolveWaiter(task.task_id, result);\n\t\t\t})\n\t\t\t.finally(() => {\n\t\t\t\tbudget.removeAllListeners();\n\t\t\t\tthis.budgets.delete(task.task_id);\n\t\t\t\tthis.pull();\n\t\t\t});\n\t}\n\n\t/** Whether a failed built-in subagent should be retried with `model: inherit`. */\n\tprivate shouldRetryWithInheritedModel(task: SubagentPoolTask, result: SubagentResult): boolean {\n\t\tif (task.useInheritedModelFallback) return false;\n\t\tif (task.sessionFile) return false;\n\t\t// Only the parent model is required: the provider may be unset when the\n\t\t// harness routes through a gateway. The retry inherits the parent model and\n\t\t// lets the child resolve the provider from its own default when none was threaded through.\n\t\tif (!task.model) return false;\n\n\t\tconst def = task.agent_type ? this.getRegistry().get(task.agent_type) : undefined;\n\t\t// Built-in agents always inherit; project agents may pin an explicit model in\n\t\t// frontmatter, so let them fall back too when that model is rejected.\n\t\tif (def?.source !== \"builtin\" && def?.source !== \"project\") return false;\n\t\tif (!def.model || def.model === MODEL_INHERIT) return false;\n\n\t\treturn this.isInheritedModelFallbackError(result);\n\t}\n\n\t/** Detect provider/model failures where inheriting the parent model can recover. */\n\tprivate isInheritedModelFallbackError(result: SubagentResult): boolean {\n\t\tconst text = [result.error, result.stderr, JSON.stringify(result.result_data ?? {})]\n\t\t\t.filter((part): part is string => typeof part === \"string\" && part.length > 0)\n\t\t\t.join(\"\\n\");\n\n\t\treturn /usage[_\\s-]?limit|subscription|quota|rate.?limit|too many requests|429|insufficient|out of credit|credit balance|billing|payment required|402|model[^\\n]*(not found|unavailable|not available|not supported|does not exist|invalid|unsupported)|no api key|no auth configured|authentication|unauthorized|forbidden|permission/i.test(\n\t\t\ttext,\n\t\t);\n\t}\n\n\t/** Remove failed attempt artifacts before rerunning the same task id. */\n\tprivate cleanupRetryArtifacts(task: SubagentPoolTask): void {\n\t\tconst cwd = task.cwd ?? this.cwd;\n\t\tconst taskDir = getDispatchTaskDir(cwd, task.task_id);\n\t\tconst sessionFile = task.sessionFile ?? this.getSessionFile(task.task_id, cwd);\n\t\ttry {\n\t\t\trmSync(sessionFile, { force: true });\n\t\t\trmSync(join(taskDir, \"result.json\"), { force: true });\n\t\t\trmSync(join(taskDir, \"output.json\"), { force: true });\n\t\t} catch {\n\t\t\t// Best-effort cleanup; retry can still proceed with existing artifacts.\n\t\t}\n\t}\n\n\t/**\n\t * Best-effort concrete failure reason for a non-zero-exit subagent. Prefers\n\t * the child's result.json summary (which carries the provider/model error\n\t * message on failure), then the tail of stderr, then the exit code.\n\t */\n\tprivate deriveFailureReason(result: SubagentResult): string {\n\t\tconst summary = (result.result_data as { summary?: string } | undefined)?.summary?.trim();\n\t\tif (summary) {\n\t\t\treturn summary;\n\t\t}\n\t\tconst stderrTail = result.stderr\n\t\t\t.split(\"\\n\")\n\t\t\t.map((line) => line.trim())\n\t\t\t.filter((line) => line.length > 0)\n\t\t\t.slice(-5)\n\t\t\t.join(\"\\n\");\n\t\tif (stderrTail) {\n\t\t\treturn stderrTail;\n\t\t}\n\t\treturn `Exited with code ${result.exit_code}`;\n\t}\n\n\tprivate tryReadResultJson(task_id: string, cwd: string): Record<string, unknown> | undefined {\n\t\tconst path = join(getDispatchTaskDir(cwd, task_id), \"result.json\");\n\t\tif (!existsSync(path)) return undefined;\n\t\ttry {\n\t\t\tconst raw = readFileSync(path, \"utf-8\");\n\t\t\treturn JSON.parse(raw) as Record<string, unknown>;\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate resolveWaiter(task_id: string, result: SubagentResult): void {\n\t\t// Persist terminal status for get_status() even after wait_for consumes the result\n\t\tif (result.status === \"stalled\") this.taskStatus.set(task_id, \"stalled\");\n\t\telse if (result.status === \"timeout\") this.taskStatus.set(task_id, \"timeout\");\n\t\telse if (result.ok) this.taskStatus.set(task_id, \"done\");\n\t\telse this.taskStatus.set(task_id, \"failed\");\n\n\t\tconst waiter = this.waiters.get(task_id);\n\t\tif (waiter) {\n\t\t\twaiter.resolve(result);\n\t\t\tthis.waiters.delete(task_id);\n\t\t\treturn;\n\t\t}\n\t\tthis.completed.set(task_id, result);\n\t}\n}\n"]}
+{"version":3,"file":"subagent-pool.d.ts","sourceRoot":"","sources":["../../src/core/subagent-pool.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,KAAK,EAAE,MAAM,oBAAoB,CAAC;AAC3C,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAc3C,MAAM,WAAW,gBAAgB;IAChC,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB;;;;OAIG;IACH,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,8FAA8F;IAC9F,yBAAyB,CAAC,EAAE,OAAO,CAAC;CACpC;AAED,MAAM,WAAW,YAAY;IAC5B,GAAG,EAAE,MAAM,CAAC;IACZ,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,YAAY,EAAE,MAAM,CAAC;IACrB,OAAO,EAAE,UAAU,CAAC,OAAO,KAAK,CAAC,CAAC;CAClC;AAED,MAAM,WAAW,cAAc;IAC9B,OAAO,EAAE,MAAM,CAAC;IAChB,EAAE,EAAE,OAAO,CAAC;IACZ,MAAM,EAAE,MAAM,CAAC;IACf,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,yEAAyE;IACzE,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B,0DAA0D;IAC1D,MAAM,CAAC,EAAE,UAAU,GAAG,SAAS,GAAG,QAAQ,GAAG,SAAS,GAAG,SAAS,CAAC;IACnE,8EAA8E;IAC9E,WAAW,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IACtC,2FAA2F;IAC3F,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACrC;AAED,MAAM,WAAW,UAAU;IAC1B,qFAAqF;IACrF,cAAc,EAAE,OAAO,CAAC;IACxB,2CAA2C;IAC3C,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,sCAAsC;IACtC,MAAM,CAAC,EAAE,cAAc,CAAC;IACxB,+CAA+C;IAC/C,QAAQ,CAAC,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,eAAe;IAC/B;gFAC4E;IAC5E,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,wEAAwE;IACxE,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,8EAA8E;IAC9E,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,iCAAiC;IACjC,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,kEAAkE;IAClE,WAAW,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,mBAAmB;IACnC,0FAA0F;IAC1F,UAAU,EAAE,MAAM,CAAC;IACnB,+EAA+E;IAC/E,UAAU,CAAC,EAAE,MAAM,EAAE,CAAC;IACtB,yDAAyD;IACzD,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,0EAA0E;IAC1E,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,sDAAsD;IACtD,GAAG,CAAC,EAAE,MAAM,CAAC,UAAU,CAAC;IACxB,oDAAoD;IACpD,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B;;;;OAIG;IACH,UAAU,CAAC,EAAE,MAAM,EAAE,CAAC;CACtB;AAED;;;;GAIG;AACH,eAAO,MAAM,0BAA0B,KAAK,CAAC;AAE7C;;;;;;GAMG;AACH,eAAO,MAAM,yBAAyB,EAAE,WAAW,CAAC,MAAM,CAIxD,CAAC;AAEH,mFAAmF;AACnF,MAAM,MAAM,kBAAkB,GAC3B;IAAE,IAAI,EAAE,WAAW,CAAA;CAAE,GACrB;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CAAE,GACpD;IAAE,IAAI,EAAE,QAAQ,CAAA;CAAE,CAAC;AAEtB;;;;;GAKG;AACH,wBAAgB,oBAAoB,CAAC,IAAI,EAAE,MAAM,GAAG,kBAAkB,CAcrE;AAED;;;;;;;;;;;;;;GAcG;AACH,qBAAa,YAAa,SAAQ,YAAY;IAC7C,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAS;IACxC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAS;IACpC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAW;IACtC,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAS;IAC7B,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAoB;IACxC,OAAO,CAAC,QAAQ,CAAC,kBAAkB,CAAS;IAC5C,+EAA+E;IAC/E,OAAO,CAAC,UAAU,CAAW;IAE7B,OAAO,CAAC,KAAK,CAAmC;IAChD,OAAO,CAAC,KAAK,CAA0B;IACvC,OAAO,CAAC,SAAS,CAAqC;IACtD,OAAO,CAAC,OAAO,CAAkG;IACjH,OAAO,CAAC,OAAO,CAAkC;IACjD,OAAO,CAAC,QAAQ,CAAwB;IACxC,OAAO,CAAC,SAAS,CAAoB;IACrC,OAAO,CAAC,QAAQ,CAAS;IACzB,kFAAkF;IAClF,OAAO,CAAC,QAAQ,CAAC,CAAgB;IACjC,kFAAkF;IAClF,OAAO,CAAC,WAAW,CAA4C;IAC/D,qEAAqE;IACrE,OAAO,CAAC,UAAU,CAAgE;IAElF,YAAY,OAAO,EAAE,mBAAmB,EAmBvC;IAED,qEAAqE;IACrE,gBAAgB,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,IAAI,CAEtC;IAED;;;;;OAKG;IACH,eAAe,CAAC,KAAK,EAAE,MAAM,GAAG,IAAI,CAEnC;IAED,0DAA0D;IAC1D,OAAO,CAAC,WAAW;IAOnB,gDAAgD;IAChD,OAAO,CAAC,UAAU;IAMlB,qDAAqD;IACrD,KAAK,CAAC,IAAI,EAAE,gBAAgB,GAAG,IAAI,CAoBlC;IAED,gCAAgC;IAChC,UAAU,CAAC,OAAO,EAAE,MAAM,GAAG,SAAS,GAAG,QAAQ,GAAG,MAAM,GAAG,QAAQ,GAAG,SAAS,GAAG,SAAS,GAAG,SAAS,CAaxG;IAED,yDAAyD;IACzD,QAAQ,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,cAAc,CAAC,CAcjD;IAED,6CAA6C;IAC7C,aAAa,IAAI,MAAM,CAEtB;IAED,4CAA4C;IAC5C,YAAY,IAAI,MAAM,CAErB;IAED;;;;;;;;OAQG;IACG,QAAQ,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,GAAE,eAAoB,GAAG,OAAO,CAAC,UAAU,CAAC,CAiB/E;IAED;;;OAGG;IACH,gBAAgB,CACf,IAAI,EAAE,MAAM,EACZ,OAAO,GAAE,eAAoB,GAC3B;QAAE,cAAc,EAAE,OAAO,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAA;KAAE,CASrF;IAED;;;OAGG;IACH,OAAO,CAAC,aAAa;IA4CrB;;;;OAIG;IACH,OAAO,CAAC,OAAO,EAAE,MAAM,GAAG,cAAc,GAAG,SAAS,CAEnD;IAED,8DAA8D;IAC9D,cAAc,CAAC,OAAO,EAAE,MAAM,EAAE,GAAG,GAAE,MAAiB,GAAG,MAAM,CAE9D;IAED;;;;OAIG;IACG,MAAM,CACX,OAAO,EAAE,MAAM,EACf,MAAM,EAAE,MAAM,EACd,OAAO,GAAE,IAAI,CAAC,eAAe,EAAE,YAAY,GAAG,aAAa,CAAM,GAC/D,OAAO,CAAC,UAAU,CAAC,CAUrB;IAED,gFAAgF;IAChF,OAAO,CAAC,qBAAqB;IAW7B,OAAO,CAAC,gBAAgB;IA0BxB,OAAO,CAAC,eAAe;IAqBvB;;;OAGG;IACH,OAAO,CAAC,kBAAkB;IAQ1B,+EAA+E;IAC/E,OAAO,IAAI,IAAI,CAyBd;IAED,2DAA2D;IAC3D,OAAO,CAAC,IAAI;IAOZ,sCAAsC;IACtC,OAAO,CAAC,SAAS;IA8FjB,kEAAkE;IAClE,OAAO,CAAC,SAAS;IAoRjB,kFAAkF;IAClF,OAAO,CAAC,6BAA6B;IAiBrC,oFAAoF;IACpF,OAAO,CAAC,6BAA6B;IAUrC,yEAAyE;IACzE,OAAO,CAAC,qBAAqB;IAa7B;;;;OAIG;IACH,OAAO,CAAC,mBAAmB;IAiB3B,OAAO,CAAC,iBAAiB;IAWzB,OAAO,CAAC,aAAa;CAerB","sourcesContent":["import { spawn } from \"node:child_process\";\nimport { EventEmitter } from \"node:events\";\nimport { existsSync, mkdirSync, readFileSync, rmSync, writeFileSync } from \"node:fs\";\nimport { dirname, join } from \"node:path\";\nimport { getDispatchTaskDir } from \"../config.js\";\nimport { attachJsonlLineReader } from \"../modes/rpc/jsonl.js\";\nimport { waitForChildProcess } from \"../utils/child-process.js\";\nimport { MODEL_INHERIT } from \"./agent-frontmatter.js\";\nimport { type AgentRegistry, loadAgentRegistry } from \"./agent-registry.js\";\nimport { DispatchEvaluator } from \"./dispatch-evaluator.js\";\nimport { SubagentLifeguard } from \"./lifeguard.js\";\nimport { OutputVerifier } from \"./output-verifier.js\";\nimport { currentSubagentDepth, resolveMaxSubagentDepth, SUBAGENT_DEPTH_ENV } from \"./subagent-depth.js\";\nimport { TokenBudget } from \"./token-budget.js\";\n\nexport interface SubagentPoolTask {\n\ttask_id: string;\n\tagent_type: string;\n\ttask: string;\n\tcontext?: string;\n\ttoken_budget?: number;\n\tcwd?: string;\n\tmodel?: string;\n\tprovider?: string;\n\t/**\n\t * Explicit session file for the child to persist/continue. When omitted the\n\t * child uses its own dispatch dir (`<dispatch>/<task_id>/session.jsonl`).\n\t * Resume reuses the original task's session file to continue the transcript.\n\t */\n\tsessionFile?: string;\n\t/** Internal: retry using the caller's model when a built-in agent's preferred model fails. */\n\tuseInheritedModelFallback?: boolean;\n}\n\nexport interface SubagentSlot {\n\tpid: number;\n\tagent_type: string;\n\ttask_id: string;\n\tspawned_at: number;\n\ttoken_budget: number;\n\tprocess: ReturnType<typeof spawn>;\n}\n\nexport interface SubagentResult {\n\ttask_id: string;\n\tok: boolean;\n\tstdout: string;\n\tstderr: string;\n\texit_code: number | null;\n\terror?: string;\n\t/** True when the task exceeded its token budget and was hard-stopped. */\n\tbudget_exceeded?: boolean;\n\t/** Terminal status derived from how the task finished. */\n\tstatus?: \"complete\" | \"partial\" | \"failed\" | \"stalled\" | \"timeout\";\n\t/** Parsed result.json content when available (e.g. on partial completion). */\n\tresult_data?: Record<string, unknown>;\n\t/** True when this run used the inherited-model fallback (preferred model failed first). */\n\tusedInheritedModelFallback?: boolean;\n}\n\nexport interface TaskResult {\n\t/** True when the evaluator decided the task is simple enough for inline handling. */\n\thandled_inline: boolean;\n\t/** Present when the task was delegated. */\n\ttask_id?: string;\n\tagent_type?: string;\n\treason?: string;\n\t/** Subagent result when delegated. */\n\tresult?: SubagentResult;\n\t/** Duration in milliseconds when delegated. */\n\tduration?: number;\n}\n\nexport interface DispatchOptions {\n\t/** Skip evaluation and force this agent type (user/explicit override).\n\t *  Accepts any registry-defined agent name, not just the built-in modes. */\n\tforceAgent?: string;\n\t/** Context distilled from the calling agent, passed to the subagent. */\n\tcontext?: string;\n\t/** Model id for the subagent (defaults to the child's configured default). */\n\tmodel?: string;\n\t/** Provider for the subagent. */\n\tprovider?: string;\n\t/** Explicit session file to persist/continue (used by resume). */\n\tsessionFile?: string;\n}\n\nexport interface SubagentPoolOptions {\n\t/** Path to the hoocode executable (or the runtime, e.g. node, when prefixArgs is set). */\n\texecutable: string;\n\t/** Args inserted before task args (e.g. the CLI entry script for node/tsx). */\n\tprefixArgs?: string[];\n\t/** Maximum concurrent child processes. Defaults to 5. */\n\tmaxConcurrency?: number;\n\t/** Working directory for spawned processes. Defaults to process.cwd(). */\n\tcwd?: string;\n\t/** Environment variables. Defaults to process.env. */\n\tenv?: NodeJS.ProcessEnv;\n\t/** Default token budget per task. Defaults to 0. */\n\tdefaultTokenBudget?: number;\n\t/**\n\t * Non-default skill paths to forward to every spawned subagent via --skill.\n\t * Subagents auto-discover skills from standard locations; only paths that\n\t * won't be found by default discovery need to be forwarded here.\n\t */\n\tskillPaths?: string[];\n}\n\n/**\n * Default hard cap on assistant turns for a spawned subagent when its definition\n * does not set `maxTurns`. The token budget is advisory (it warns but never\n * kills), so this turn cap is the guaranteed hard stop for every subagent.\n */\nexport const DEFAULT_SUBAGENT_MAX_TURNS = 50;\n\n/**\n * AgentSession event `type`s forwarded from a subagent's json event stream as\n * `task_progress` events. Deliberately coarse: the child also emits per-delta\n * `message_update` / `tool_execution_update` events (a high-volume firehose) and\n * large `message_*` bodies, which are dropped here to keep the parent's event loop\n * and the task panel from thrashing under concurrent subagents.\n */\nexport const FORWARDED_SUBAGENT_EVENTS: ReadonlySet<string> = new Set([\n\t\"turn_end\",\n\t\"tool_execution_start\",\n\t\"tool_execution_end\",\n]);\n\n/** The action the pool should take for one JSONL line from a subagent's stdout. */\nexport type SubagentStdoutLine =\n\t| { kind: \"heartbeat\" }\n\t| { kind: \"progress\"; event: Record<string, unknown> }\n\t| { kind: \"ignore\" };\n\n/**\n * Classify one JSONL line from a subagent's stdout into the action to take.\n * Pure (no side effects) so the ping/forward/drop policy is unit-testable without\n * spawning a child. Line framing — UTF-8-safe reassembly of chunks split mid-line\n * — is handled upstream by attachJsonlLineReader; this only sees complete lines.\n */\nexport function classifySubagentLine(line: string): SubagentStdoutLine {\n\tconst trimmed = line.trim();\n\tif (!trimmed.startsWith(\"{\")) return { kind: \"ignore\" };\n\tlet parsed: Record<string, unknown>;\n\ttry {\n\t\tparsed = JSON.parse(trimmed) as Record<string, unknown>;\n\t} catch {\n\t\treturn { kind: \"ignore\" };\n\t}\n\tif (parsed.ping === true) return { kind: \"heartbeat\" };\n\tif (typeof parsed.type === \"string\" && FORWARDED_SUBAGENT_EVENTS.has(parsed.type)) {\n\t\treturn { kind: \"progress\", event: parsed };\n\t}\n\treturn { kind: \"ignore\" };\n}\n\n/**\n * Pool for running hoocode subagents as child processes with bounded concurrency,\n * FIFO queuing with priority support, and automatic slot refill.\n *\n * Events:\n * - \"task_done\"    – task completed successfully and output was verified\n * - \"task_failed\"  – task failed (spawn error, bad exit code, verification failure)\n * - \"task_stalled\" – heartbeat missed past the load-scaled threshold (60s base,\n *                    widened under concurrency/event-loop lag), process SIGKILLed\n * - \"task_timeout\" – hard timeout exceeded, process was SIGKILLed\n * - \"budget_warning\" – token usage crossed 80% threshold (advisory)\n * - \"budget_exceeded\" – token usage crossed 100% threshold (advisory; never kills)\n * - \"task_progress\" – coarse lifecycle event (turn_end, tool start/end) parsed\n *                    from the child's json event stream, for live UI updates\n */\nexport class SubagentPool extends EventEmitter {\n\tprivate readonly maxConcurrency: number;\n\tprivate readonly executable: string;\n\tprivate readonly prefixArgs: string[];\n\tprivate readonly cwd: string;\n\tprivate readonly env: NodeJS.ProcessEnv;\n\tprivate readonly defaultTokenBudget: number;\n\t/** Non-default skill paths forwarded to every spawned subagent via --skill. */\n\tprivate skillPaths: string[];\n\n\tprivate slots = new Map<string, SubagentSlot>();\n\tprivate queue: SubagentPoolTask[] = [];\n\tprivate completed = new Map<string, SubagentResult>();\n\tprivate waiters = new Map<string, { resolve: (result: SubagentResult) => void; reject: (err: Error) => void }>();\n\tprivate budgets = new Map<string, TokenBudget>();\n\tprivate verifier = new OutputVerifier();\n\tprivate lifeguard: SubagentLifeguard;\n\tprivate disposed = false;\n\t/** Lazily-loaded agent registry (frontmatter definitions) for this pool's cwd. */\n\tprivate registry?: AgentRegistry;\n\t/** Tracks why a task was killed (stalled / timeout) before exit handler fires. */\n\tprivate killReasons = new Map<string, \"stalled\" | \"timeout\">();\n\t/** Persistent terminal status map, survives wait_for consumption. */\n\tprivate taskStatus = new Map<string, \"done\" | \"failed\" | \"stalled\" | \"timeout\">();\n\n\tconstructor(options: SubagentPoolOptions) {\n\t\tsuper();\n\t\tthis.maxConcurrency = options.maxConcurrency ?? 5;\n\t\tthis.executable = options.executable;\n\t\tthis.prefixArgs = options.prefixArgs ?? [];\n\t\tthis.cwd = options.cwd ?? process.cwd();\n\t\tthis.env = options.env ?? process.env;\n\t\tthis.defaultTokenBudget = options.defaultTokenBudget ?? 0;\n\t\tthis.skillPaths = options.skillPaths ? [...options.skillPaths] : [];\n\t\tthis.verifier = new OutputVerifier(this.cwd);\n\t\tthis.lifeguard = new SubagentLifeguard(this.cwd);\n\t\tthis.lifeguard.on(\"stalled\", (data: { task_id: string; pid: number }) => {\n\t\t\tthis.killReasons.set(data.task_id, \"stalled\");\n\t\t\tthis.emit(\"task_stalled\", data);\n\t\t});\n\t\tthis.lifeguard.on(\"timeout\", (data: { task_id: string; pid: number }) => {\n\t\t\tthis.killReasons.set(data.task_id, \"timeout\");\n\t\t\tthis.emit(\"task_timeout\", data);\n\t\t});\n\t}\n\n\t/** Update the non-default skill paths forwarded to new subagents. */\n\tupdateSkillPaths(paths: string[]): void {\n\t\tthis.skillPaths = [...paths];\n\t}\n\n\t/**\n\t * Report external in-process load (e.g. the number of background MCP tools\n\t * currently executing in the parent) to the lifeguard. This widens its\n\t * heartbeat/timeout tolerance so monitored subagents aren't false-positive\n\t * reaped when the parent's event loop is busy with concurrent background work.\n\t */\n\tsetExternalLoad(count: number): void {\n\t\tthis.lifeguard.setExternalLoad(count);\n\t}\n\n\t/** Lazily load the agent registry for this pool's cwd. */\n\tprivate getRegistry(): AgentRegistry {\n\t\tif (!this.registry) {\n\t\t\tthis.registry = loadAgentRegistry({ cwd: this.cwd });\n\t\t}\n\t\treturn this.registry;\n\t}\n\n\t/** Priority value: higher numbers run first. */\n\tprivate priorityOf(agent_type: string): number {\n\t\t// Read-only investigation (explore/plan) often unblocks downstream work, so\n\t\t// it runs ahead of other agents.\n\t\treturn agent_type === \"explore\" || agent_type === \"plan\" ? 2 : 1;\n\t}\n\n\t/** Queue a task. It will run when a slot is free. */\n\tspawn(task: SubagentPoolTask): void {\n\t\tif (this.disposed) {\n\t\t\tthrow new Error(\"SubagentPool has been disposed\");\n\t\t}\n\t\tif (\n\t\t\tthis.slots.has(task.task_id) ||\n\t\t\tthis.queue.some((t) => t.task_id === task.task_id) ||\n\t\t\tthis.completed.has(task.task_id)\n\t\t) {\n\t\t\tthrow new Error(`Duplicate task_id: ${task.task_id}`);\n\t\t}\n\n\t\tconst p = this.priorityOf(task.agent_type);\n\t\tconst idx = this.queue.findIndex((t) => this.priorityOf(t.agent_type) < p);\n\t\tif (idx === -1) {\n\t\t\tthis.queue.push(task);\n\t\t} else {\n\t\t\tthis.queue.splice(idx, 0, task);\n\t\t}\n\t\tthis.pull();\n\t}\n\n\t/** Current status of a task. */\n\tget_status(task_id: string): \"running\" | \"queued\" | \"done\" | \"failed\" | \"stalled\" | \"timeout\" | \"unknown\" {\n\t\tif (this.slots.has(task_id)) return \"running\";\n\t\tif (this.queue.some((t) => t.task_id === task_id)) return \"queued\";\n\t\tconst persisted = this.taskStatus.get(task_id);\n\t\tif (persisted) return persisted;\n\t\tconst result = this.completed.get(task_id);\n\t\tif (result) {\n\t\t\tif (result.status === \"stalled\") return \"stalled\";\n\t\t\tif (result.status === \"timeout\") return \"timeout\";\n\t\t\tif (result.ok) return \"done\";\n\t\t\treturn \"failed\";\n\t\t}\n\t\treturn \"unknown\";\n\t}\n\n\t/** Wait for a task to complete and return its result. */\n\twait_for(task_id: string): Promise<SubagentResult> {\n\t\tif (this.disposed) {\n\t\t\treturn Promise.reject(new Error(\"SubagentPool has been disposed\"));\n\t\t}\n\n\t\tconst existing = this.completed.get(task_id);\n\t\tif (existing) {\n\t\t\tthis.completed.delete(task_id);\n\t\t\treturn Promise.resolve(existing);\n\t\t}\n\n\t\treturn new Promise((resolve, reject) => {\n\t\t\tthis.waiters.set(task_id, { resolve, reject });\n\t\t});\n\t}\n\n\t/** Number of currently running subagents. */\n\trunning_count(): number {\n\t\treturn this.slots.size;\n\t}\n\n\t/** Number of tasks waiting in the queue. */\n\tqueued_count(): number {\n\t\treturn this.queue.length;\n\t}\n\n\t/**\n\t * Dispatch a task through the evaluator.\n\t *\n\t * - If `options.forceAgent` is provided, skip evaluation and spawn directly.\n\t * - Otherwise evaluate the task. If it should be handled inline, return\n\t *   `{ handled_inline: true }` immediately.\n\t * - If delegating, spawn the subagent, wait for completion, write\n\t *   `output.json`, and return the result.\n\t */\n\tasync dispatch(task: string, options: DispatchOptions = {}): Promise<TaskResult> {\n\t\tif (this.disposed) {\n\t\t\treturn Promise.reject(new Error(\"SubagentPool has been disposed\"));\n\t\t}\n\t\tconst begin = this.beginDispatch(task, options);\n\t\tif (begin.handled_inline) {\n\t\t\treturn { handled_inline: true, reason: begin.reason };\n\t\t}\n\t\tconst result = await this.wait_for(begin.task_id);\n\t\treturn {\n\t\t\thandled_inline: false,\n\t\t\ttask_id: begin.task_id,\n\t\t\tagent_type: begin.agent_type,\n\t\t\treason: begin.reason,\n\t\t\tresult,\n\t\t\tduration: Date.now() - begin.startTime,\n\t\t};\n\t}\n\n\t/**\n\t * Fire-and-forget dispatch for background agents. Spawns the subagent and\n\t * returns its handle immediately; the caller polls get_status()/collect().\n\t */\n\tdispatchDetached(\n\t\ttask: string,\n\t\toptions: DispatchOptions = {},\n\t): { handled_inline: boolean; task_id?: string; agent_type?: string; reason?: string } {\n\t\tif (this.disposed) {\n\t\t\tthrow new Error(\"SubagentPool has been disposed\");\n\t\t}\n\t\tconst begin = this.beginDispatch(task, options);\n\t\tif (begin.handled_inline) {\n\t\t\treturn { handled_inline: true, reason: begin.reason };\n\t\t}\n\t\treturn { handled_inline: false, task_id: begin.task_id, agent_type: begin.agent_type, reason: begin.reason };\n\t}\n\n\t/**\n\t * Evaluate, log, and spawn a task without waiting. Shared by dispatch()\n\t * (blocking) and dispatchDetached() (background).\n\t */\n\tprivate beginDispatch(\n\t\ttask: string,\n\t\toptions: DispatchOptions,\n\t):\n\t\t| { handled_inline: true; reason?: string }\n\t\t| { handled_inline: false; task_id: string; agent_type: string; reason?: string; startTime: number } {\n\t\tconst { forceAgent, context, model, provider, sessionFile } = options;\n\t\tconst evaluator = new DispatchEvaluator();\n\t\tconst analysis = evaluator.evaluate(task);\n\n\t\tif (!forceAgent && !analysis.should_delegate) {\n\t\t\treturn { handled_inline: true, reason: analysis.reason };\n\t\t}\n\n\t\tconst agent_type = forceAgent ?? \"general-purpose\";\n\t\tconst task_id = `dispatch-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;\n\t\tconst reason = forceAgent ? \"user_override\" : analysis.reason;\n\t\tconst complexity = analysis.estimated_complexity;\n\t\t// Depth of the child about to be spawned (this process's depth + 1). Surfaced\n\t\t// so a delegation tree's nesting is visible in logs without extra tooling.\n\t\tconst childDepth = currentSubagentDepth(this.env) + 1;\n\n\t\t// Pre-dispatch logging. Use stderr: stdout is reserved for the JSON event\n\t\t// stream / TUI render and must not be polluted.\n\t\tconsole.error(\n\t\t\t`[DISPATCH] agent=${agent_type} depth=${childDepth} reason=${reason} complexity=${complexity} task_id=${task_id}`,\n\t\t);\n\t\tthis.writeDispatchLog(task_id, agent_type, reason, complexity, task, childDepth);\n\n\t\tconst poolTask: SubagentPoolTask = {\n\t\t\ttask_id,\n\t\t\tagent_type,\n\t\t\ttask,\n\t\t\tcontext,\n\t\t\tmodel,\n\t\t\tprovider,\n\t\t\tsessionFile,\n\t\t\tcwd: this.cwd,\n\t\t};\n\t\tconst startTime = Date.now();\n\t\tthis.spawn(poolTask);\n\t\treturn { handled_inline: false, task_id, agent_type, reason, startTime };\n\t}\n\n\t/**\n\t * Non-destructively read a completed task's result (for background polling).\n\t * Returns undefined while the task is still running/queued, or if its result\n\t * was already consumed via wait_for().\n\t */\n\tcollect(task_id: string): SubagentResult | undefined {\n\t\treturn this.completed.get(task_id);\n\t}\n\n\t/** Absolute path of the persisted session file for a task. */\n\tgetSessionFile(task_id: string, cwd: string = this.cwd): string {\n\t\treturn join(getDispatchTaskDir(cwd, task_id), \"session.jsonl\");\n\t}\n\n\t/**\n\t * Resume a previously dispatched subagent, continuing its persisted session\n\t * with a follow-up prompt. Recovers the original agent type from its dispatch\n\t * log. Rejects if no resumable session exists for the task.\n\t */\n\tasync resume(\n\t\ttask_id: string,\n\t\tprompt: string,\n\t\toptions: Omit<DispatchOptions, \"forceAgent\" | \"sessionFile\"> = {},\n\t): Promise<TaskResult> {\n\t\tif (this.disposed) {\n\t\t\treturn Promise.reject(new Error(\"SubagentPool has been disposed\"));\n\t\t}\n\t\tconst sessionFile = this.getSessionFile(task_id);\n\t\tif (!existsSync(sessionFile)) {\n\t\t\treturn Promise.reject(new Error(`No resumable session for task \"${task_id}\" (expected ${sessionFile}).`));\n\t\t}\n\t\tconst agent_type = this.readDispatchAgentType(task_id) ?? \"general-purpose\";\n\t\treturn this.dispatch(prompt, { ...options, forceAgent: agent_type, sessionFile });\n\t}\n\n\t/** Recover the agent type a task was dispatched with, from its dispatch log. */\n\tprivate readDispatchAgentType(task_id: string): string | undefined {\n\t\tconst path = join(getDispatchTaskDir(this.cwd, task_id), \"dispatch-log.json\");\n\t\tif (!existsSync(path)) return undefined;\n\t\ttry {\n\t\t\tconst parsed = JSON.parse(readFileSync(path, \"utf-8\")) as { agent_type?: string };\n\t\t\treturn typeof parsed.agent_type === \"string\" ? parsed.agent_type : undefined;\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate writeDispatchLog(\n\t\ttask_id: string,\n\t\tagent_type: string,\n\t\treason: string,\n\t\tcomplexity: string,\n\t\ttask: string,\n\t\tdepth: number,\n\t): void {\n\t\tconst log = {\n\t\t\ttimestamp: new Date().toISOString(),\n\t\t\ttask_id,\n\t\t\tagent_type,\n\t\t\tdepth,\n\t\t\treason,\n\t\t\tcomplexity,\n\t\t\ttask,\n\t\t};\n\t\tconst path = join(getDispatchTaskDir(this.cwd, task_id), \"dispatch-log.json\");\n\t\ttry {\n\t\t\tmkdirSync(dirname(path), { recursive: true });\n\t\t\twriteFileSync(path, JSON.stringify(log, null, 2));\n\t\t} catch {\n\t\t\t// Best-effort persistence\n\t\t}\n\t}\n\n\tprivate writeOutputJson(task_id: string, result: SubagentResult): void {\n\t\tconst output = {\n\t\t\ttask_id: result.task_id,\n\t\t\tok: result.ok,\n\t\t\texit_code: result.exit_code,\n\t\t\tstatus: result.status,\n\t\t\tstdout: result.stdout,\n\t\t\tstderr: result.stderr,\n\t\t\terror: result.error,\n\t\t\tbudget_exceeded: result.budget_exceeded,\n\t\t\tresult_data: result.result_data,\n\t\t};\n\t\tconst path = join(getDispatchTaskDir(this.cwd, task_id), \"output.json\");\n\t\ttry {\n\t\t\tmkdirSync(dirname(path), { recursive: true });\n\t\t\twriteFileSync(path, JSON.stringify(output, null, 2));\n\t\t} catch {\n\t\t\t// Best-effort persistence\n\t\t}\n\t}\n\n\t/**\n\t * Remove a task's dispatch dir after a clean, verified success. Best-effort:\n\t * a cleanup failure must never fail an otherwise successful task.\n\t */\n\tprivate cleanupDispatchDir(task_id: string, cwd: string): void {\n\t\ttry {\n\t\t\trmSync(getDispatchTaskDir(cwd, task_id), { recursive: true, force: true });\n\t\t} catch {\n\t\t\t// Best-effort cleanup\n\t\t}\n\t}\n\n\t/** Kill all running processes, clear the queue, and reject pending waiters. */\n\tdispose(): void {\n\t\tif (this.disposed) return;\n\t\tthis.disposed = true;\n\n\t\tfor (const slot of this.slots.values()) {\n\t\t\tif (!slot.process.killed) {\n\t\t\t\tslot.process.kill(\"SIGTERM\");\n\t\t\t}\n\t\t}\n\t\tthis.slots.clear();\n\t\tthis.queue = [];\n\n\t\tfor (const [task_id, waiter] of this.waiters) {\n\t\t\twaiter.reject(new Error(\"SubagentPool disposed\"));\n\t\t\tthis.waiters.delete(task_id);\n\t\t}\n\t\tthis.completed.clear();\n\t\tfor (const budget of this.budgets.values()) {\n\t\t\tbudget.removeAllListeners();\n\t\t}\n\t\tthis.budgets.clear();\n\t\tthis.killReasons.clear();\n\t\tthis.taskStatus.clear();\n\t\tthis.lifeguard.dispose();\n\t\tthis.removeAllListeners();\n\t}\n\n\t/** Pull tasks from the queue while slots are available. */\n\tprivate pull(): void {\n\t\twhile (this.slots.size < this.maxConcurrency && this.queue.length > 0) {\n\t\t\tconst task = this.queue.shift()!;\n\t\t\tthis.startTask(task, false);\n\t\t}\n\t}\n\n\t/** Build CLI arguments for a task. */\n\tprivate buildArgs(task: SubagentPoolTask): string[] {\n\t\t// Persist the child's session so a finished/interrupted subagent can be\n\t\t// resumed later (see resume()). SessionManager.open() creates the file on\n\t\t// first run and continues it on subsequent runs.\n\t\tconst sessionFile = task.sessionFile ?? this.getSessionFile(task.task_id, task.cwd ?? this.cwd);\n\t\tconst args: string[] = [\n\t\t\t...this.prefixArgs,\n\t\t\t\"--mode\",\n\t\t\t\"json\",\n\t\t\t\"--session\",\n\t\t\tsessionFile,\n\t\t\t\"--task-id\",\n\t\t\ttask.task_id,\n\t\t];\n\n\t\t// Prefer the data-driven agent definition from the registry; fall back to the\n\t\t// built-in mode prompt/allowlist for legacy modes not present in the registry.\n\t\tconst def = task.agent_type ? this.getRegistry().get(task.agent_type) : undefined;\n\n\t\tif (task.agent_type) {\n\t\t\tconst systemPrompt = def?.prompt;\n\t\t\tif (systemPrompt) {\n\t\t\t\targs.push(\"--system-prompt\", systemPrompt);\n\t\t\t}\n\n\t\t\t// A `delegate: true` agent may itself dispatch via the Task tool, but only\n\t\t\t// while the child it becomes can still nest (childDepth < cap) — so the\n\t\t\t// deepest permitted level cannot delegate further. Gating here keeps the\n\t\t\t// authorization explicit and bounded by the same cap as the depth guard.\n\t\t\tconst childDepth = currentSubagentDepth(this.env) + 1;\n\t\t\tconst canChildDelegate = def?.delegate === true && childDepth < resolveMaxSubagentDepth(undefined, this.env);\n\n\t\t\t// Tool allowlist comes from the agent definition's frontmatter `tools`\n\t\t\t// field (read-only built-ins declare their own sandbox). When omitted, no\n\t\t\t// --tools is passed and the subagent inherits all parent tools (so the Task\n\t\t\t// tool already survives). A delegating agent with an explicit allowlist must\n\t\t\t// have Task/TaskOutput added, or the child would filter them out.\n\t\t\tconst tools = def?.tools ? [...def.tools] : undefined;\n\t\t\tif (canChildDelegate && tools) {\n\t\t\t\tfor (const t of [\"Task\", \"TaskOutput\"]) {\n\t\t\t\t\tif (!tools.includes(t)) tools.push(t);\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (tools && tools.length > 0) {\n\t\t\t\targs.push(\"--tools\", tools.join(\",\"));\n\t\t\t}\n\t\t\tif (def?.disallowedTools && def.disallowedTools.length > 0) {\n\t\t\t\targs.push(\"--disallowed-tools\", def.disallowedTools.join(\",\"));\n\t\t\t}\n\n\t\t\t// Propagate subagent enablement so the child registers the Task tool; without\n\t\t\t// this the flag-based enablement would not reach a spawned child.\n\t\t\tif (canChildDelegate) {\n\t\t\t\targs.push(\"--enable-subagents\");\n\t\t\t\t// Scoped delegation: restrict which agent types this child may spawn.\n\t\t\t\tif (def?.delegateTo && def.delegateTo.length > 0) {\n\t\t\t\t\targs.push(\"--delegate-allow\", def.delegateTo.join(\",\"));\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\n\t\t// Model precedence: a definition's explicit model wins (unless it is the\n\t\t// `inherit` sentinel), otherwise use the caller-provided model. Built-in\n\t\t// agents can retry with the inherited model when their preferred model is\n\t\t// unavailable or quota-limited.\n\t\tconst explicitModel =\n\t\t\t!task.useInheritedModelFallback && def?.model && def.model !== MODEL_INHERIT ? def.model : undefined;\n\t\tconst modelToUse = explicitModel ?? task.model;\n\t\tif (modelToUse) {\n\t\t\targs.push(\"--model\", modelToUse);\n\t\t}\n\t\tif (task.provider) {\n\t\t\targs.push(\"--provider\", task.provider);\n\t\t}\n\n\t\t// Always give subagents a hard turn cap. With the token budget now advisory\n\t\t// (warn-only), this is the guaranteed hard stop for a runaway subagent.\n\t\tconst maxTurns = def?.maxTurns && def.maxTurns > 0 ? def.maxTurns : DEFAULT_SUBAGENT_MAX_TURNS;\n\t\targs.push(\"--max-turns\", String(maxTurns));\n\n\t\t// Forward non-default skill paths so the subagent has access to all parent skills.\n\t\t// Standard discovery locations (~/.hoocode/, .hoocode/, .claude/) are found automatically.\n\t\tfor (const skillPath of this.skillPaths) {\n\t\t\targs.push(\"--skill\", skillPath);\n\t\t}\n\n\t\tconst prompt = task.context?.trim()\n\t\t\t? `Context from the calling agent:\\n\\n${task.context.trim()}\\n\\nTask: ${task.task.trim()}`\n\t\t\t: `Task: ${task.task.trim()}`;\n\t\targs.push(prompt);\n\n\t\treturn args;\n\t}\n\n\t/** Start a task in a child process, with one retry on failure. */\n\tprivate startTask(task: SubagentPoolTask, isRetry: boolean): void {\n\t\t// Get or create a TokenBudget tracker. On retry, reuse the existing one\n\t\t// so cumulative usage persists across retries.\n\t\tlet budget = this.budgets.get(task.task_id);\n\t\tif (!budget) {\n\t\t\tbudget = new TokenBudget(task.task_id, task.agent_type, {\n\t\t\t\tlimit: task.token_budget,\n\t\t\t\tcwd: task.cwd ?? this.cwd,\n\t\t\t});\n\t\t\tbudget.on(\"budget_warning\", (data: { task_id: string; message: string; used: number; limit: number }) => {\n\t\t\t\tthis.emit(\"budget_warning\", data);\n\t\t\t});\n\t\t\t// The token budget is advisory: surface telemetry but never kill. The\n\t\t\t// guaranteed hard stop is the per-subagent turn cap (--max-turns); see\n\t\t\t// DEFAULT_SUBAGENT_MAX_TURNS.\n\t\t\tbudget.on(\"budget_exceeded\", (data: { task_id: string; used: number; limit: number }) => {\n\t\t\t\tthis.emit(\"budget_exceeded\", data);\n\t\t\t});\n\t\t\tthis.budgets.set(task.task_id, budget);\n\t\t}\n\n\t\tlet proc: ReturnType<typeof spawn>;\n\t\ttry {\n\t\t\tproc = spawn(this.executable, this.buildArgs(task), {\n\t\t\t\tcwd: task.cwd ?? this.cwd,\n\t\t\t\t// Stamp the child's depth (parent depth + 1) so its own guard knows where\n\t\t\t\t// it sits in the tree. The tree-wide cap (HOOCODE_SUBAGENT_MAX_DEPTH) is\n\t\t\t\t// inherited via the spread; at the default cap of 1 the child lands at\n\t\t\t\t// depth 1 and cannot spawn further subagents.\n\t\t\t\tenv: {\n\t\t\t\t\t...this.env,\n\t\t\t\t\t[SUBAGENT_DEPTH_ENV]: String(currentSubagentDepth(this.env) + 1),\n\t\t\t\t},\n\t\t\t\tshell: false,\n\t\t\t\tstdio: [\"ignore\", \"pipe\", \"pipe\"],\n\t\t\t});\n\t\t} catch {\n\t\t\tif (!isRetry) {\n\t\t\t\tthis.startTask(task, true);\n\t\t\t} else {\n\t\t\t\tthis.emit(\"task_failed\", {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\terror: \"Spawn failed synchronously\",\n\t\t\t\t});\n\t\t\t\tthis.resolveWaiter(task.task_id, {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\tok: false,\n\t\t\t\t\tstdout: \"\",\n\t\t\t\t\tstderr: \"\",\n\t\t\t\t\texit_code: null,\n\t\t\t\t\terror: \"Spawn failed synchronously\",\n\t\t\t\t\tstatus: \"failed\",\n\t\t\t\t});\n\t\t\t\tthis.pull();\n\t\t\t}\n\t\t\treturn;\n\t\t}\n\n\t\tconst slot: SubagentSlot = {\n\t\t\tpid: proc.pid ?? 0,\n\t\t\tagent_type: task.agent_type,\n\t\t\ttask_id: task.task_id,\n\t\t\tspawned_at: Date.now(),\n\t\t\ttoken_budget: task.token_budget ?? this.defaultTokenBudget,\n\t\t\tprocess: proc,\n\t\t};\n\n\t\tthis.slots.set(task.task_id, slot);\n\t\tthis.lifeguard.monitor(task.task_id, task.agent_type, proc);\n\n\t\tlet stdout = \"\";\n\t\tlet stderr = \"\";\n\t\tlet detachStdoutReader: (() => void) | undefined;\n\n\t\tproc.stdout?.on(\"data\", (data: Buffer) => {\n\t\t\tconst chunk = data.toString();\n\t\t\tstdout += chunk;\n\t\t\tbudget.processStdout(chunk);\n\n\t\t\t// Any output proves the child is alive and working, so treat it as a\n\t\t\t// heartbeat. The dedicated {\"ping\":true} line (parsed below via the JSONL\n\t\t\t// reader) still matters for quiet phases (e.g. a long single model turn\n\t\t\t// that emits nothing), but relying on it alone falsely reaps subagents\n\t\t\t// that are busily streaming events while the parent's event loop is\n\t\t\t// starved by concurrent load.\n\t\t\tthis.lifeguard.recordHeartbeat(task.task_id);\n\t\t});\n\n\t\t// Parse the child's newline-delimited JSON event stream with UTF-8-safe,\n\t\t// LF-only framing — multi-byte characters and large events split across pipe\n\t\t// chunks are reassembled before parsing, which the raw handler above cannot do.\n\t\t// Pings refresh the heartbeat; coarse lifecycle events are forwarded for live\n\t\t// UI. Detached when the child exits (see cleanup below).\n\t\tif (proc.stdout) {\n\t\t\tdetachStdoutReader = attachJsonlLineReader(proc.stdout, (line) => {\n\t\t\t\tconst action = classifySubagentLine(line);\n\t\t\t\tif (action.kind === \"heartbeat\") {\n\t\t\t\t\tthis.lifeguard.recordHeartbeat(task.task_id);\n\t\t\t\t} else if (action.kind === \"progress\") {\n\t\t\t\t\tthis.emit(\"task_progress\", {\n\t\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\t\tagent_type: task.agent_type,\n\t\t\t\t\t\tevent: action.event,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t});\n\t\t}\n\t\tproc.stderr?.on(\"data\", (data: Buffer) => {\n\t\t\tstderr += data.toString();\n\t\t});\n\n\t\twaitForChildProcess(proc)\n\t\t\t.then((code) => {\n\t\t\t\tthis.slots.delete(task.task_id);\n\t\t\t\tdetachStdoutReader?.();\n\t\t\t\tbudget.flush();\n\n\t\t\t\tconst killReason = this.killReasons.get(task.task_id);\n\t\t\t\tthis.killReasons.delete(task.task_id);\n\n\t\t\t\tconst duration = Date.now() - slot.spawned_at;\n\t\t\t\tconst tokens_used = budget.getUsed();\n\t\t\t\tconst budgetExceeded = budget.isExceeded();\n\n\t\t\t\t// A subagent's success is defined by a valid, verified result.json, not by\n\t\t\t\t// its exit code. A child that finished its work and wrote a valid result can\n\t\t\t\t// still be SIGKILLed by the lifeguard before it exits on its own (lingering\n\t\t\t\t// open handles delay a natural exit past the heartbeat threshold), which forces\n\t\t\t\t// exit_code === null. Keying completion off the verified result, not code === 0,\n\t\t\t\t// honors that genuine success instead of discarding it as a false stall.\n\t\t\t\tconst verification = this.verifier.verify(task.task_id, task.cwd ?? this.cwd);\n\t\t\t\t// A well-formed result.json counts as clean completion unless its own\n\t\t\t\t// status field declares failure (e.g. \"failed\" from a provider quota\n\t\t\t\t// error). Without this check the pool would treat a child that wrote a\n\t\t\t\t// valid-but-failed result.json and exited non-zero as a success.\n\t\t\t\tlet cleanlyCompleted = code === 0 || verification.valid;\n\t\t\t\tif (cleanlyCompleted && verification.valid) {\n\t\t\t\t\tconst rd = this.tryReadResultJson(task.task_id, task.cwd ?? this.cwd);\n\t\t\t\t\tif (rd && (rd as Record<string, unknown>).status === \"failed\") {\n\t\t\t\t\t\tcleanlyCompleted = false;\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\t// If killed by lifeguard before producing a valid result, honor the kill.\n\t\t\t\tif ((killReason === \"stalled\" || killReason === \"timeout\") && !verification.valid) {\n\t\t\t\t\tconst result: SubagentResult = {\n\t\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\t\tok: false,\n\t\t\t\t\t\tstdout,\n\t\t\t\t\t\tstderr,\n\t\t\t\t\t\texit_code: code,\n\t\t\t\t\t\tstatus: killReason,\n\t\t\t\t\t};\n\t\t\t\t\tthis.writeOutputJson(task.task_id, result);\n\t\t\t\t\tthis.emit(`task_${killReason}`, {\n\t\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\t\tagent_type: task.agent_type,\n\t\t\t\t\t\tduration,\n\t\t\t\t\t\ttokens_used,\n\t\t\t\t\t});\n\t\t\t\t\tthis.resolveWaiter(task.task_id, result);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\n\t\t\t\tconst result: SubagentResult = {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\tok: cleanlyCompleted,\n\t\t\t\t\tstdout,\n\t\t\t\t\tstderr,\n\t\t\t\t\texit_code: code,\n\t\t\t\t\t// Advisory telemetry only: exceeding the budget never fails the task.\n\t\t\t\t\tbudget_exceeded: budgetExceeded,\n\t\t\t\t\tstatus: cleanlyCompleted ? \"complete\" : \"failed\",\n\t\t\t\t\tusedInheritedModelFallback: task.useInheritedModelFallback === true,\n\t\t\t\t};\n\n\t\t\t\tif (result.ok) {\n\t\t\t\t\tif (!verification.valid) {\n\t\t\t\t\t\tresult.ok = false;\n\t\t\t\t\t\tresult.error = verification.reason;\n\t\t\t\t\t\tresult.status = \"failed\";\n\t\t\t\t\t\tthis.writeOutputJson(task.task_id, result);\n\t\t\t\t\t\tthis.emit(\"task_failed\", {\n\t\t\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\t\t\tagent_type: task.agent_type,\n\t\t\t\t\t\t\tduration,\n\t\t\t\t\t\t\ttokens_used,\n\t\t\t\t\t\t\terror: verification.reason,\n\t\t\t\t\t\t});\n\t\t\t\t\t\tthis.resolveWaiter(task.task_id, result);\n\t\t\t\t\t\treturn;\n\t\t\t\t\t}\n\t\t\t\t\t// Attach the verified result.json so callers can read the summary\n\t\t\t\t\t// without parsing the raw event stream.\n\t\t\t\t\tresult.result_data = this.tryReadResultJson(task.task_id, task.cwd ?? this.cwd);\n\n\t\t\t\t\t// Clean success: discard the per-task dispatch dir entirely\n\t\t\t\t\t// (session.jsonl, result.json, dispatch-log.json, budget.json). The\n\t\t\t\t\t// in-memory result already carries result_data, so callers lose\n\t\t\t\t\t// nothing. Trade-off: resume() only works for non-successful tasks.\n\t\t\t\t\tthis.cleanupDispatchDir(task.task_id, task.cwd ?? this.cwd);\n\n\t\t\t\t\tthis.emit(\"task_done\", {\n\t\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\t\tagent_type: task.agent_type,\n\t\t\t\t\t\tduration,\n\t\t\t\t\t\ttokens_used,\n\t\t\t\t\t\tstatus: \"complete\",\n\t\t\t\t\t});\n\t\t\t\t\tthis.resolveWaiter(task.task_id, result);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\n\t\t\t\t// Failure path: keep the dispatch dir for debugging and persist output.\n\t\t\t\t// Attach the child's result.json (if any) and derive a concrete failure\n\t\t\t\t// reason so callers see the real cause (e.g. a provider usage/quota\n\t\t\t\t// error) instead of a generic \"subagent failed\".\n\t\t\t\tresult.result_data = this.tryReadResultJson(task.task_id, task.cwd ?? this.cwd);\n\t\t\t\tif (!result.error) {\n\t\t\t\t\tresult.error = this.deriveFailureReason(result);\n\t\t\t\t}\n\t\t\t\tif (this.shouldRetryWithInheritedModel(task, result)) {\n\t\t\t\t\tconsole.error(\n\t\t\t\t\t\t`[DISPATCH] agent=${task.agent_type} task_id=${task.task_id} preferred model failed; retrying with inherited model`,\n\t\t\t\t\t);\n\t\t\t\t\tthis.cleanupRetryArtifacts(task);\n\t\t\t\t\tthis.queue.unshift({ ...task, useInheritedModelFallback: true });\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tthis.writeOutputJson(task.task_id, result);\n\t\t\t\tthis.emit(\"task_failed\", {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\tagent_type: task.agent_type,\n\t\t\t\t\tduration,\n\t\t\t\t\ttokens_used,\n\t\t\t\t\terror: result.error ?? `Exited with code ${code}`,\n\t\t\t\t});\n\t\t\t\tthis.resolveWaiter(task.task_id, result);\n\t\t\t})\n\t\t\t.catch((err) => {\n\t\t\t\tthis.slots.delete(task.task_id);\n\t\t\t\tbudget.flush();\n\t\t\t\tconst duration = Date.now() - slot.spawned_at;\n\t\t\t\tconst tokens_used = budget.getUsed();\n\t\t\t\tif (!isRetry) {\n\t\t\t\t\tthis.startTask(task, true);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tconst error = err instanceof Error ? err.message : String(err);\n\t\t\t\tconst result: SubagentResult = {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\tok: false,\n\t\t\t\t\tstdout,\n\t\t\t\t\tstderr,\n\t\t\t\t\texit_code: null,\n\t\t\t\t\terror,\n\t\t\t\t\tstatus: \"failed\",\n\t\t\t\t\tusedInheritedModelFallback: task.useInheritedModelFallback === true,\n\t\t\t\t};\n\t\t\t\tthis.writeOutputJson(task.task_id, result);\n\t\t\t\tthis.emit(\"task_failed\", {\n\t\t\t\t\ttask_id: task.task_id,\n\t\t\t\t\tagent_type: task.agent_type,\n\t\t\t\t\tduration,\n\t\t\t\t\ttokens_used,\n\t\t\t\t\terror,\n\t\t\t\t});\n\t\t\t\tthis.resolveWaiter(task.task_id, result);\n\t\t\t})\n\t\t\t.finally(() => {\n\t\t\t\tbudget.removeAllListeners();\n\t\t\t\tthis.budgets.delete(task.task_id);\n\t\t\t\tthis.pull();\n\t\t\t});\n\t}\n\n\t/** Whether a failed built-in subagent should be retried with `model: inherit`. */\n\tprivate shouldRetryWithInheritedModel(task: SubagentPoolTask, result: SubagentResult): boolean {\n\t\tif (task.useInheritedModelFallback) return false;\n\t\tif (task.sessionFile) return false;\n\t\t// Only the parent model is required: the provider may be unset when the\n\t\t// harness routes through a gateway. The retry inherits the parent model and\n\t\t// lets the child resolve the provider from its own default when none was threaded through.\n\t\tif (!task.model) return false;\n\n\t\tconst def = task.agent_type ? this.getRegistry().get(task.agent_type) : undefined;\n\t\t// Built-in agents always inherit; project agents may pin an explicit model in\n\t\t// frontmatter, so let them fall back too when that model is rejected.\n\t\tif (def?.source !== \"builtin\" && def?.source !== \"project\") return false;\n\t\tif (!def.model || def.model === MODEL_INHERIT) return false;\n\n\t\treturn this.isInheritedModelFallbackError(result);\n\t}\n\n\t/** Detect provider/model failures where inheriting the parent model can recover. */\n\tprivate isInheritedModelFallbackError(result: SubagentResult): boolean {\n\t\tconst text = [result.error, result.stderr, JSON.stringify(result.result_data ?? {})]\n\t\t\t.filter((part): part is string => typeof part === \"string\" && part.length > 0)\n\t\t\t.join(\"\\n\");\n\n\t\treturn /usage[_\\s-]?limit|subscription|quota|rate.?limit|too many requests|429|insufficient|out of credit|credit balance|billing|payment required|402|model[^\\n]*(not found|unavailable|not available|not supported|does not exist|invalid|unsupported)|no api key|no auth configured|authentication|unauthorized|forbidden|permission/i.test(\n\t\t\ttext,\n\t\t);\n\t}\n\n\t/** Remove failed attempt artifacts before rerunning the same task id. */\n\tprivate cleanupRetryArtifacts(task: SubagentPoolTask): void {\n\t\tconst cwd = task.cwd ?? this.cwd;\n\t\tconst taskDir = getDispatchTaskDir(cwd, task.task_id);\n\t\tconst sessionFile = task.sessionFile ?? this.getSessionFile(task.task_id, cwd);\n\t\ttry {\n\t\t\trmSync(sessionFile, { force: true });\n\t\t\trmSync(join(taskDir, \"result.json\"), { force: true });\n\t\t\trmSync(join(taskDir, \"output.json\"), { force: true });\n\t\t} catch {\n\t\t\t// Best-effort cleanup; retry can still proceed with existing artifacts.\n\t\t}\n\t}\n\n\t/**\n\t * Best-effort concrete failure reason for a non-zero-exit subagent. Prefers\n\t * the child's result.json summary (which carries the provider/model error\n\t * message on failure), then the tail of stderr, then the exit code.\n\t */\n\tprivate deriveFailureReason(result: SubagentResult): string {\n\t\tconst summary = (result.result_data as { summary?: string } | undefined)?.summary?.trim();\n\t\tif (summary) {\n\t\t\treturn summary;\n\t\t}\n\t\tconst stderrTail = result.stderr\n\t\t\t.split(\"\\n\")\n\t\t\t.map((line) => line.trim())\n\t\t\t.filter((line) => line.length > 0)\n\t\t\t.slice(-5)\n\t\t\t.join(\"\\n\");\n\t\tif (stderrTail) {\n\t\t\treturn stderrTail;\n\t\t}\n\t\treturn `Exited with code ${result.exit_code}`;\n\t}\n\n\tprivate tryReadResultJson(task_id: string, cwd: string): Record<string, unknown> | undefined {\n\t\tconst path = join(getDispatchTaskDir(cwd, task_id), \"result.json\");\n\t\tif (!existsSync(path)) return undefined;\n\t\ttry {\n\t\t\tconst raw = readFileSync(path, \"utf-8\");\n\t\t\treturn JSON.parse(raw) as Record<string, unknown>;\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate resolveWaiter(task_id: string, result: SubagentResult): void {\n\t\t// Persist terminal status for get_status() even after wait_for consumes the result\n\t\tif (result.status === \"stalled\") this.taskStatus.set(task_id, \"stalled\");\n\t\telse if (result.status === \"timeout\") this.taskStatus.set(task_id, \"timeout\");\n\t\telse if (result.ok) this.taskStatus.set(task_id, \"done\");\n\t\telse this.taskStatus.set(task_id, \"failed\");\n\n\t\tconst waiter = this.waiters.get(task_id);\n\t\tif (waiter) {\n\t\t\twaiter.resolve(result);\n\t\t\tthis.waiters.delete(task_id);\n\t\t\treturn;\n\t\t}\n\t\tthis.completed.set(task_id, result);\n\t}\n}\n"]}