@caupulican/pi-adaptative 0.81.37 → 0.81.39

package/dist/core/model-router-controller.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"model-router-controller.d.ts","sourceRoot":"","sources":["../../src/core/model-router-controller.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAEH,OAAO,KAAK,EAAE,KAAK,EAAE,YAAY,EAAiB,MAAM,2BAA2B,CAAC;AACpF,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,gCAAgC,CAAC;AACrE,OAAO,KAAK,EAAE,GAAG,EAAW,KAAK,EAAE,KAAK,EAAE,MAAM,mBAAmB,CAAC;AAEpE,OAAO,KAAK,EAAE,iBAAiB,EAAE,yBAAyB,EAAE,wBAAwB,EAAE,MAAM,oBAAoB,CAAC;AACjH,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,yBAAyB,CAAC;AAC7D,OAAO,EAAkC,KAAK,sBAAsB,EAAE,MAAM,gCAAgC,CAAC;AAG7G,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAC;AAkBzD,OAAO,EAIN,KAAK,yBAAyB,EAC9B,KAAK,yBAAyB,EAE9B,MAAM,0BAA0B,CAAC;AAGlC,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AAO7D,wGAAwG;AACxG,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,MAAM,CAEhE;AA+BD,MAAM,WAAW,yBAAyB;IACzC;iFAC6E;IAC7E,QAAQ,IAAI,KAAK,CAAC;IAClB,4FAA4F;IAC5F,QAAQ,IAAI,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CAAC;IACnC,oFAAoF;IACpF,kBAAkB,IAAI,eAAe,CAAC;IACtC,4GAA4G;IAC5G,iBAAiB,IAAI,cAAc,CAAC;IACpC,uFAAuF;IACvF,gBAAgB,IAAI,aAAa,CAAC;IAClC,4DAA4D;IAC5D,gBAAgB,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,OAAO,CAAC;IAC7C,4EAA4E;IAC5E,iBAAiB,IAAI,yBAAyB,CAAC;IAC/C,kGAAkG;IAClG,WAAW,IAAI,MAAM,CAAC;IACtB,0EAA0E;IAC1E,mBAAmB,IAAI,WAAW,CAAC;IACnC,qGAAmG;IACnG,mBAAmB,IAAI,MAAM,CAAC;IAC9B,6GAA6G;IAC7G,cAAc,CAAC,QAAQ,EAAE,YAAY,GAAG,YAAY,EAAE,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACvE,iGAAiG;IACjG,6BAA6B,CAAC,SAAS,EAAE,MAAM,EAAE,GAAG,MAAM,CAAC;IAC3D,0GAA0G;IAC1G,+BAA+B,IAAI,IAAI,CAAC;IACxC,yGAAuG;IACvG,qBAAqB,CAAC,IAAI,EAAE,yBAAyB,GAAG,OAAO,CAAC,wBAAwB,CAAC,CAAC;IAC1F,6DAA6D;IAC7D,eAAe,CACd,KAAK,EAAE,KAAK,EACZ,IAAI,CAAC,EAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,eAAe,CAAC,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,EAAE,MAAM,CAAA;KAAE,GACpE,MAAM,GAAG,SAAS,CAAC;IACtB,oDAAoD;IACpD,IAAI,CAAC,KAAK,EAAE,iBAAiB,GAAG,IAAI,CAAC;IACrC,wFAAwF;IACxF,qBAAqB,CAAC,KAAK,EAAE,sBAAsB,GAAG,IAAI,CAAC;IAC3D,6EAA6E;IAC7E,gBAAgB,CAAC,OAAO,EAAE,MAAM,GAAG,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CAAC;IAC1D,sGAAsG;IACtG,yBAAyB,IAAI,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CAAC;CACpD;AAED;;;GAGG;AACH,qBAAa,qBAAqB;IACjC,gFAAgF;IAChF,OAAO,CAAC,wBAAwB,CAAC,CAAoB;IACrD,OAAO,CAAC,uBAAuB,CAAC,CAAgB;IAChD,OAAO,CAAC,yBAAyB,CAAC,CAA2B;IAC7D,OAAO,CAAC,+BAA+B,CAAS;IAChD,OAAO,CAAC,mBAAmB,CAAS;IACpC,OAAO,CAAC,wBAAwB,CAAC,CAA4B;IAC7D,OAAO,CAAC,0BAA0B,CAAC,CAAS;IAC5C,OAAO,CAAC,sBAAsB,CAAC,CAAoB;IAEnD,OAAO,CAAC,QAAQ,CAAC,IAAI,CAA4B;IAEjD,YAAY,IAAI,EAAE,yBAAyB,EAE1C;IAED,6GAA6G;IAC7G,eAAe,IAAI,OAAO,CAEzB;IAED,qFAAqF;IACrF,eAAe,IAAI,yBAAyB,GAAG,SAAS,CAEvD;IAED;;;;OAIG;IACH,qBAAqB,CAAC,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,OAAO,GAAG;QAAE,KAAK,EAAE,IAAI,CAAC;QAAC,MAAM,EAAE,MAAM,CAAA;KAAE,GAAG,SAAS,CAmBlG;IAED;;;;OAIG;IACH,qBAAqB,CAAC,OAAO,EAAE,YAAY,GAAG,OAAO,CAYpD;IAED,OAAO,CAAC,0BAA0B;IAMlC,OAAO,CAAC,qBAAqB;IAO7B,OAAO,CAAC,qBAAqB;IAI7B,OAAO,CAAC,qBAAqB;IAI7B,OAAO,CAAC,6BAA6B;IAmBrC,OAAO,CAAC,8BAA8B;IAiCtC,OAAO,CAAC,qBAAqB;IA6B7B,0GAA0G;IAC1G,OAAO,CAAC,2BAA2B;YAYrB,2BAA2B;IAkCzC,OAAO,CAAC,4BAA4B;IA6FpC,OAAO,CAAC,iCAAiC;IAUzC,0BAA0B,CAAC,IAAI,EAAE,OAAO,GAAG,QAAQ,GAAG,WAAW,GAAG,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CASzF;IAED;;;;;;;OAOG;IACG,sBAAsB,CAC3B,MAAM,EAAE,MAAM,EACd,OAAO,CAAC,EAAE;QAAE,SAAS,CAAC,EAAE,OAAO,CAAA;KAAE,GAC/B,OAAO,CAAC;QAAE,QAAQ,EAAE,aAAa,CAAC;QAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAA;KAAE,GAAG,SAAS,CAAC,CA2FrE;IAGD,OAAO,CAAC,4BAA4B;IAKpC,SAAS,CAAC,WAAW,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,MAAM,GAAG,MAAM,CA6BzD;IAEK,aAAa,CAClB,QAAQ,EAAE,YAAY,GAAG,YAAY,EAAE,EACvC,WAAW,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,EACnC,aAAa,EAAE,aAAa,GAAG,SAAS,EACxC,eAAe,UAAO,GACpB,OAAO,CAAC,IAAI,CAAC,CAsPf;CACD","sourcesContent":["/**\n * Model-router turn routing: the session's per-turn model-selection subsystem — the regex/executor\n * route resolver, the optional bounded routing judge, the executor lane (Level-0 toolkit direct hit\n * + speculative brain-refined retry), the per-tier thinking/tool-surface swap around a routed turn,\n * the cheap-research-turn session buffer with mutating-tool escalation to an expensive retry, and\n * the router status/diagnostics report.\n *\n * Extracted verbatim from agent-session.ts (god-file decomposition). Owns the transient per-turn\n * route state — the active intent/route, the cheap-turn session buffer, the escalation-requested and\n * retry-in-flight flags — and the sticky last-decision/last-skip-reason/last-intent used by the\n * status report. Everything else it needs — the live agent + its state, the current model, the\n * session/settings managers, the model registry, the agent dir, the reflection abort signal, the\n * base system prompt, the isolated-completion primitive, spawned-usage accounting, the event/telemetry\n * emitters, and the recently-extracted BackgroundLaneController (resolveLaneModel) / ContextPipeline\n * (resolveCurationModelIfFit) collaborators — is reached through narrow deps accessors rather than the\n * whole AgentSession.\n *\n * Drive-path boundary (deliberate): the actual agent.prompt()/continue() loop stays host-side in\n * AgentSession._runAgentPrompt; this controller's parallel routed drive path ({@link runRoutedTurn})\n * owns only the route decision/escalation/tier bookkeeping and delegates every agent turn back through\n * {@link ModelRouterControllerDeps.runAgentPrompt}, so the drive-loop logic is never duplicated. The\n * host keeps a one-line delegation at each call-in: the routing prep + routed-turn entry in\n * _promptUnserialized, the beforeToolCall escalation branch ({@link maybeEscalateToolCall}), the\n * message_end cheap-turn buffering ({@link captureSessionMessage}), the retry-event suppression\n * ({@link isRetryInFlight}), and the public getModelRouterStatus / autonomy-telemetry reads.\n */\n\nimport type { Agent, AgentMessage, ThinkingLevel } from \"@caupulican/pi-agent-core\";\nimport type { SessionManager } from \"@caupulican/pi-agent-core/node\";\nimport type { Api, Message, Model, Usage } from \"@caupulican/pi-ai\";\nimport { clampThinkingLevel, modelsAreEqual } from \"@caupulican/pi-ai\";\nimport type { AgentSessionEvent, IsolatedCompletionOptions, IsolatedCompletionResult } from \"./agent-session.ts\";\nimport type { RouteDecision } from \"./autonomy/contracts.ts\";\nimport { AUTONOMY_TELEMETRY_EVENT_TYPES, type AutonomyTelemetryEvent } from \"./autonomy/telemetry-events.ts\";\nimport { latestUserPromptText } from \"./context-pipeline.ts\";\nimport { deriveModelCapabilityProfile, filterToolNamesForCapability } from \"./model-capability.ts\";\nimport type { ModelRegistry } from \"./model-registry.ts\";\nimport { resolveCliModel } from \"./model-resolver.ts\";\nimport { collectModelRouterConfigDiagnostics } from \"./model-router/config-diagnostics.ts\";\nimport { classifyExecutorTurn } from \"./model-router/executor-route.ts\";\nimport {\n\tevaluateSurfaceFitness,\n\ttype FitnessGatedSurface,\n\ttype FitnessGateVerdict,\n} from \"./model-router/fitness-gate.ts\";\nimport { classifyModelRouterRoute, type ModelRouterIntent } from \"./model-router/intent-classifier.ts\";\nimport { ROUTE_JUDGE_MAX_OUTPUT_TOKENS, runRouteJudge } from \"./model-router/route-judge.ts\";\nimport {\n\tbufferModelRouterSessionCustomMessage,\n\tbufferModelRouterSessionMessage,\n\tcreateModelRouterSessionBuffer,\n\tflushModelRouterSessionBuffer,\n\ttype ModelRouterSessionBuffer,\n} from \"./model-router/session-buffer.ts\";\nimport {\n\tformatModelRouterStatus,\n\tgetRecentModelRouterDecisions,\n\tMODEL_ROUTER_DECISION_CUSTOM_TYPE,\n\ttype ModelRouterDecisionStatus,\n\ttype ModelRouterFailoverStatus,\n\ttype ModelRouterFitnessStatuses,\n} from \"./model-router/status.ts\";\nimport { shouldEscalateModelRouterTool } from \"./model-router/tool-escalation.ts\";\nimport { FitnessStore } from \"./models/fitness-store.ts\";\nimport type { SettingsManager } from \"./settings-manager.ts\";\nimport {\n\tbuildReflexUserPrompt,\n\tparseReflexPlan,\n\tREFLEX_INTERPRETER_SYSTEM_PROMPT,\n} from \"./toolkit/reflex-interpreter.ts\";\n\n/** Canonical `provider/id` label for a routed/resolved model, as it appears in decisions and status. */\nexport function formatModelRouterModel(model: Model<Api>): string {\n\treturn `${model.provider}/${model.id}`;\n}\n\nconst ROUTE_JUDGE_STATIC_FAST_PATH_REASON_CODES = new Set([\n\t\"empty_prompt\",\n\t\"read_only_question\",\n\t\"release_or_publish\",\n\t\"security_or_auth\",\n\t\"destructive_or_git_history\",\n\t\"settings_or_self_modification\",\n\t\"architecture_or_ambiguous\",\n]);\n\nfunction shouldSkipRouteJudgeForStaticDecision(decision: RouteDecision): boolean {\n\treturn ROUTE_JUDGE_STATIC_FAST_PATH_REASON_CODES.has(decision.reasonCode);\n}\n\nfunction withJudgeUnavailableFallback(decision: RouteDecision, reason: string): RouteDecision {\n\treturn {\n\t\t...decision,\n\t\treasonCode: \"judge_unavailable_fallback\",\n\t\treasons: [...decision.reasons, reason],\n\t};\n}\n\nfunction persistModelRouterDecision(\n\tsessionManager: Pick<SessionManager, \"appendCustomEntry\">,\n\tdecision: ModelRouterDecisionStatus,\n): void {\n\tsessionManager.appendCustomEntry(MODEL_ROUTER_DECISION_CUSTOM_TYPE, decision);\n}\n\nexport interface ModelRouterControllerDeps {\n\t/** Live agent — the controller reads/writes agent.state.{model,thinkingLevel,tools,systemPrompt,messages}\n\t * for the per-turn tier swap and aborts it on a mutating-tool escalation. */\n\tgetAgent(): Agent;\n\t/** Current session model, used to decide whether a routed turn actually swaps the model. */\n\tgetModel(): Model<Api> | undefined;\n\t/** Router/executor/judge/thinking settings + capability mode (all opt-in gates). */\n\tgetSettingsManager(): SettingsManager;\n\t/** Session log: routed-turn message buffering/persistence, decision persistence, recent-decision status. */\n\tgetSessionManager(): SessionManager;\n\t/** Resolves configured route/judge/executor model patterns against configured auth. */\n\tgetModelRegistry(): ModelRegistry;\n\t/** Session-scoped provider/model quota exhaustion guard. */\n\tisModelExhausted(model: Model<Api>): boolean;\n\t/** Status snapshot for B5 exhausted models and the last failover notice. */\n\tgetFailoverStatus(): ModelRouterFailoverStatus;\n\t/** Root dir the host-keyed {@link FitnessStore} lives under (executor tool-call fitness gate). */\n\tgetAgentDir(): string;\n\t/** Aborts the judge's bounded completion when the session is disposed. */\n\tgetReflectionSignal(): AbortSignal;\n\t/** Base (extension-free) system prompt — the tier swap only sheds tools when the turn is on it. */\n\tgetBaseSystemPrompt(): string;\n\t/** The host-side drive loop (agent.prompt()/continue()); the routed drive path delegates every turn here. */\n\trunAgentPrompt(messages: AgentMessage | AgentMessage[]): Promise<void>;\n\t/** Rebuilds the system prompt for a filtered tool surface (routed-model capability shedding). */\n\tbuildSystemPromptForToolNames(toolNames: string[]): string;\n\t/** Re-resolves the restored model against the registry after a routed turn (provider override safety). */\n\trefreshCurrentModelFromRegistry(): void;\n\t/** One-shot, tool-less LLM call — the routing judge and the executor reflex-brain warmup ride this. */\n\trunIsolatedCompletion(opts: IsolatedCompletionOptions): Promise<IsolatedCompletionResult>;\n\t/** Rolls judge/brain spend into spawned-usage accounting. */\n\taddSpawnedUsage(\n\t\tusage: Usage,\n\t\topts?: { label?: string; sourceSessionId?: string; reportId?: string },\n\t): string | undefined;\n\t/** Session event stream (executor-miss warning). */\n\temit(event: AgentSessionEvent): void;\n\t/** Autonomy telemetry stream (one route-decision event per user-facing routed turn). */\n\temitAutonomyTelemetry(event: AutonomyTelemetryEvent): void;\n\t/** Resolves the judge model pattern via {@link BackgroundLaneController}. */\n\tresolveLaneModel(pattern: string): Model<Api> | undefined;\n\t/** Fitness-gated reflex-brain model via {@link ContextPipeline} (executor speculative refinement). */\n\tresolveCurationModelIfFit(): Model<Api> | undefined;\n}\n\n/**\n * Owns the model-router turn routing extracted from {@link AgentSession}. See the module header for the\n * drive-path boundary that keeps the agent.prompt()/continue() loop host-side.\n */\nexport class ModelRouterController {\n\t/** Active model-router intent for the current transient routed turn, if any. */\n\tprivate _activeModelRouterIntent?: ModelRouterIntent;\n\tprivate _activeModelRouterRoute?: RouteDecision;\n\tprivate _modelRouterSessionBuffer?: ModelRouterSessionBuffer;\n\tprivate _modelRouterEscalationRequested = false;\n\tprivate _isModelRouterRetry = false;\n\tprivate _lastModelRouterDecision?: ModelRouterDecisionStatus;\n\tprivate _lastModelRouterSkipReason?: string;\n\tprivate _lastModelRouterIntent?: ModelRouterIntent;\n\n\tprivate readonly deps: ModelRouterControllerDeps;\n\n\tconstructor(deps: ModelRouterControllerDeps) {\n\t\tthis.deps = deps;\n\t}\n\n\t/** True while the escalation retry turn is running, so the host can suppress its duplicate prompt events. */\n\tisRetryInFlight(): boolean {\n\t\treturn this._isModelRouterRetry;\n\t}\n\n\t/** Latest completed route decision (sticky), for the autonomy telemetry snapshot. */\n\tgetLastDecision(): ModelRouterDecisionStatus | undefined {\n\t\treturn this._lastModelRouterDecision;\n\t}\n\n\t/**\n\t * beforeToolCall escalation gate: a cheap research turn that reaches for a mutating tool aborts the\n\t * turn and requests a retry on the expensive model. Returns the block result the host hook forwards,\n\t * or undefined when no escalation is required.\n\t */\n\tmaybeEscalateToolCall(toolName: string, args: unknown): { block: true; reason: string } | undefined {\n\t\tif (\n\t\t\tthis._activeModelRouterRoute &&\n\t\t\tshouldEscalateModelRouterTool({\n\t\t\t\ttier: this._activeModelRouterRoute.tier,\n\t\t\t\ttoolName,\n\t\t\t\targs,\n\t\t\t\treasonCode: this._activeModelRouterRoute.reasonCode,\n\t\t\t})\n\t\t) {\n\t\t\tthis._modelRouterEscalationRequested = true;\n\t\t\tthis.deps.getAgent().abort();\n\t\t\treturn {\n\t\t\t\tblock: true,\n\t\t\t\treason:\n\t\t\t\t\t\"Model router escalation required: a cheap research turn attempted a mutating tool. Retry the turn on the configured expensive model.\",\n\t\t\t};\n\t\t}\n\t\treturn undefined;\n\t}\n\n\t/**\n\t * message_end hook: while a cheap routed turn is buffering, capture its messages into the session\n\t * buffer instead of persisting them (they are flushed on success or discarded on escalation).\n\t * Returns true when the message was buffered, so the host skips its own persistence.\n\t */\n\tcaptureSessionMessage(message: AgentMessage): boolean {\n\t\tconst modelRouterBuffer = this._modelRouterSessionBuffer;\n\t\tif (!modelRouterBuffer) return false;\n\t\tif (message.role === \"custom\") {\n\t\t\tbufferModelRouterSessionCustomMessage(modelRouterBuffer, message);\n\t\t\treturn true;\n\t\t}\n\t\tif (message.role === \"user\" || message.role === \"assistant\" || message.role === \"toolResult\") {\n\t\t\tbufferModelRouterSessionMessage(modelRouterBuffer, message as Message);\n\t\t\treturn true;\n\t\t}\n\t\treturn false;\n\t}\n\n\tprivate _isModelAvailableAndAuthed(pattern: string): boolean {\n\t\tconst resolved = resolveCliModel({ cliModel: pattern, modelRegistry: this.deps.getModelRegistry() });\n\t\tif (!resolved.model) return false;\n\t\treturn this.deps.getModelRegistry().hasConfiguredAuth(resolved.model);\n\t}\n\n\tprivate _evaluateModelFitness(surface: FitnessGatedSurface, model: Model<Api>): FitnessGateVerdict {\n\t\tconst fitness = FitnessStore.forAgentDir(this.deps.getAgentDir())\n\t\t\t.getForHost()\n\t\t\t.find((entry) => entry.model === formatModelRouterModel(model));\n\t\treturn evaluateSurfaceFitness(surface, fitness?.report);\n\t}\n\n\tprivate _formatFitnessFailure(verdict: Exclude<FitnessGateVerdict, { fit: true }>): string {\n\t\treturn verdict.reason === \"unprobed\" ? \"unprobed\" : `${verdict.lane} ${verdict.succeeded}/${verdict.total}`;\n\t}\n\n\tprivate _routerSurfaceForTier(tier: \"cheap\" | \"medium\" | \"expensive\"): FitnessGatedSurface {\n\t\treturn tier === \"cheap\" ? \"router_cheap\" : tier === \"medium\" ? \"router_medium\" : \"router_expensive\";\n\t}\n\n\tprivate _getRouterTierFitnessStatuses(): ModelRouterFitnessStatuses {\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst statuses: ModelRouterFitnessStatuses = {};\n\t\tfor (const tier of [\"cheap\", \"medium\", \"expensive\"] as const) {\n\t\t\tconst pattern =\n\t\t\t\ttier === \"cheap\" ? settings.cheapModel : tier === \"medium\" ? settings.mediumModel : settings.expensiveModel;\n\t\t\tif (!pattern) continue;\n\t\t\tconst resolved = resolveCliModel({ cliModel: pattern, modelRegistry: this.deps.getModelRegistry() });\n\t\t\tif (!resolved.model || !this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) continue;\n\t\t\tconst verdict = this._evaluateModelFitness(this._routerSurfaceForTier(tier), resolved.model);\n\t\t\tstatuses[tier] = verdict.fit\n\t\t\t\t? { status: verdict.probed ? \"fit\" : \"unprobed\" }\n\t\t\t\t: verdict.reason === \"unprobed\"\n\t\t\t\t\t? { status: \"unprobed\" }\n\t\t\t\t\t: { status: \"unfit\", lane: verdict.lane, succeeded: verdict.succeeded, total: verdict.total };\n\t\t}\n\t\treturn statuses;\n\t}\n\n\tprivate _resolveExpensiveFallbackRoute(\n\t\tdecision: RouteDecision,\n\t\treasonCode: string,\n\t\treason: string,\n\t): { decision: RouteDecision; model: Model<Api> } | undefined {\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst expensivePattern = settings.expensiveModel;\n\t\tif (!expensivePattern || !this._isModelAvailableAndAuthed(expensivePattern)) return undefined;\n\t\tconst resolvedExpensive = resolveCliModel({\n\t\t\tcliModel: expensivePattern,\n\t\t\tmodelRegistry: this.deps.getModelRegistry(),\n\t\t});\n\t\tif (!resolvedExpensive.model) return undefined;\n\t\tif (this.deps.isModelExhausted(resolvedExpensive.model)) {\n\t\t\tthis._lastModelRouterSkipReason = \"expensive model exhausted: quota\";\n\t\t\treturn undefined;\n\t\t}\n\t\tif (settings.fitnessGate) {\n\t\t\tconst verdict = this._evaluateModelFitness(\"router_expensive\", resolvedExpensive.model);\n\t\t\tif (!verdict.fit) {\n\t\t\t\tthis._lastModelRouterSkipReason = `expensive model unfit: ${this._formatFitnessFailure(verdict)} (fitness gate)`;\n\t\t\t\treturn undefined;\n\t\t\t}\n\t\t}\n\t\tdecision.fallbackFrom = \"medium\";\n\t\tdecision.tier = \"expensive\";\n\t\tdecision.reasonCode = reasonCode;\n\t\tdecision.reasons = [...decision.reasons, reason];\n\t\tdecision.model = formatModelRouterModel(resolvedExpensive.model);\n\t\tthis._lastModelRouterSkipReason = undefined;\n\t\treturn { decision, model: resolvedExpensive.model };\n\t}\n\n\tprivate _resolveExecutorRoute(\n\t\tprompt: string,\n\t\texecutorPattern: string | undefined,\n\t): { decision: RouteDecision; model: Model<Api> } | undefined {\n\t\tif (!executorPattern) return undefined;\n\t\ttry {\n\t\t\tconst verdict = classifyExecutorTurn(prompt, this.deps.getSettingsManager().getToolkitScripts());\n\t\t\tif (!verdict.execute) return undefined;\n\t\t\tconst resolved = resolveCliModel({ cliModel: executorPattern, modelRegistry: this.deps.getModelRegistry() });\n\t\t\tif (!resolved.model || !this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) return undefined;\n\t\t\t// Fitness gate: the executor must have PROVEN tool-calling on this host (same\n\t\t\t// canonical-ref discipline as the curation gate).\n\t\t\tif (!this._evaluateModelFitness(\"executor\", resolved.model).fit) return undefined;\n\t\t\tthis._lastModelRouterIntent = \"research\";\n\t\t\treturn {\n\t\t\t\tdecision: {\n\t\t\t\t\ttier: \"cheap\",\n\t\t\t\t\trisk: \"scoped-write\",\n\t\t\t\t\tconfidence: 1,\n\t\t\t\t\treasonCode: \"executor_direct\",\n\t\t\t\t\treasons: [`Executor lane: Level-0 direct hit on toolkit script \"${verdict.scriptName}\"`],\n\t\t\t\t},\n\t\t\t\tmodel: resolved.model,\n\t\t\t};\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\t/** True if a run_toolkit_script tool result since `fromIndex` actually EXECUTED (not error/ambiguous). */\n\tprivate _executorTurnExecutedScript(fromIndex: number): boolean {\n\t\tfor (const message of this.deps.getAgent().state.messages.slice(fromIndex)) {\n\t\t\tif ((message as { role?: string }).role !== \"toolResult\") continue;\n\t\t\tif ((message as { toolName?: string }).toolName !== \"run_toolkit_script\") continue;\n\t\t\tif ((message as { isError?: boolean }).isError === true) continue;\n\t\t\tconst outcome = (message as { details?: { outcome?: unknown } }).details?.outcome;\n\t\t\tif (outcome === \"executed\") return true;\n\t\t}\n\t\treturn false;\n\t}\n\n\t/** Ask the reflex brain to refine the last user request into an explicit toolkit instruction. */\n\tprivate async _buildExecutorRefinedPrompt(messages: AgentMessage | AgentMessage[]): Promise<string | undefined> {\n\t\ttry {\n\t\t\tconst model = this.deps.resolveCurationModelIfFit();\n\t\t\tif (!model) return undefined;\n\t\t\tconst list = Array.isArray(messages) ? messages : [messages];\n\t\t\tconst request = latestUserPromptText(list.filter((m): m is AgentMessage => true));\n\t\t\tif (!request) return undefined;\n\t\t\tconst scripts = this.deps.getSettingsManager().getToolkitScripts();\n\t\t\tconst completion = await this.deps.runIsolatedCompletion({\n\t\t\t\tsystemPrompt: REFLEX_INTERPRETER_SYSTEM_PROMPT,\n\t\t\t\tmessages: [\n\t\t\t\t\t{\n\t\t\t\t\t\trole: \"user\",\n\t\t\t\t\t\tcontent: [{ type: \"text\", text: buildReflexUserPrompt(request, scripts) }],\n\t\t\t\t\t\ttimestamp: Date.now(),\n\t\t\t\t\t},\n\t\t\t\t],\n\t\t\t\tmodel,\n\t\t\t\tthinkingLevel: \"off\",\n\t\t\t\tmaxTokens: 256,\n\t\t\t\tcacheRetention: \"short\",\n\t\t\t});\n\t\t\tif (completion.usage.cost.total > 0 || completion.usage.totalTokens > 0) {\n\t\t\t\tthis.deps.addSpawnedUsage(completion.usage, { label: \"executor-brain-warmup\" });\n\t\t\t}\n\t\t\tconst plan = parseReflexPlan(completion.text);\n\t\t\tif (!plan || plan.script === \"none\") return undefined;\n\t\t\tconst argHint = plan.args.length > 0 ? ` with args ${JSON.stringify(plan.args)}` : \"\";\n\t\t\treturn `Run the toolkit script \"${plan.script}\"${argHint} using run_toolkit_script, then report its result exactly.`;\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate _resolveModelRouterTurnRoute(prompt: string): { decision: RouteDecision; model: Model<Api> } | undefined {\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tif (!settings.enabled) {\n\t\t\tthis._lastModelRouterSkipReason = \"disabled\";\n\t\t\treturn undefined;\n\t\t}\n\n\t\t// G16 executor lane: a Level-0 DIRECT toolkit hit on a command-shaped prompt routes the\n\t\t// whole turn to the configured local executor (tool-call-fitness-gated) instead of\n\t\t// spending the frontier model on a one-tool reflex. Ambiguity never routes here — it\n\t\t// stays with the big model and the reflex brain. Deterministic, so the judge is skipped.\n\t\tconst executorRoute = this._resolveExecutorRoute(prompt, settings.executorModel);\n\t\tif (executorRoute) return executorRoute;\n\n\t\tconst decision = classifyModelRouterRoute(prompt);\n\t\tthis._lastModelRouterIntent = decision.tier === \"cheap\" ? \"research\" : \"modify\";\n\n\t\t// Learning tier must not be selected for normal user prompts\n\t\tif (decision.tier === \"learning\") {\n\t\t\tthis._lastModelRouterSkipReason = \"learning tier not supported for user prompts\";\n\t\t\treturn undefined;\n\t\t}\n\n\t\tconst modelPattern =\n\t\t\tsettings[\n\t\t\t\tdecision.tier === \"cheap\" ? \"cheapModel\" : decision.tier === \"medium\" ? \"mediumModel\" : \"expensiveModel\"\n\t\t\t];\n\t\tconst label =\n\t\t\tdecision.tier === \"cheap\" ? \"cheap model\" : decision.tier === \"medium\" ? \"medium model\" : \"expensive model\";\n\n\t\tif (decision.tier === \"medium\" && (!modelPattern || !this._isModelAvailableAndAuthed(modelPattern))) {\n\t\t\tconst fallback = this._resolveExpensiveFallbackRoute(\n\t\t\t\tdecision,\n\t\t\t\t\"medium_unavailable_fallback_expensive\",\n\t\t\t\t\"Medium model is unavailable, falling back to expensive model\",\n\t\t\t);\n\t\t\tif (fallback) return fallback;\n\t\t\tthis._lastModelRouterSkipReason ??= \"medium model and expensive fallback are unavailable\";\n\t\t\treturn undefined;\n\t\t}\n\n\t\tif (!modelPattern) {\n\t\t\tthis._lastModelRouterSkipReason = `${label} unset`;\n\t\t\treturn undefined;\n\t\t}\n\n\t\tconst resolved = resolveCliModel({ cliModel: modelPattern, modelRegistry: this.deps.getModelRegistry() });\n\t\tif (!resolved.model) {\n\t\t\tthis._lastModelRouterSkipReason = `${label} unresolved: ${modelPattern}`;\n\t\t\treturn undefined;\n\t\t}\n\n\t\tconst resolvedName = formatModelRouterModel(resolved.model);\n\t\tif (!this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) {\n\t\t\tthis._lastModelRouterSkipReason = `${label} missing auth: ${resolvedName}`;\n\t\t\treturn undefined;\n\t\t}\n\n\t\tif (this.deps.isModelExhausted(resolved.model)) {\n\t\t\tif (decision.tier === \"medium\") {\n\t\t\t\tconst fallback = this._resolveExpensiveFallbackRoute(\n\t\t\t\t\tdecision,\n\t\t\t\t\t\"medium_exhausted_fallback_expensive\",\n\t\t\t\t\t\"Medium model exhausted: quota; falling back to expensive model\",\n\t\t\t\t);\n\t\t\t\tif (fallback) return fallback;\n\t\t\t}\n\t\t\tthis._lastModelRouterSkipReason = `${decision.tier} model exhausted: quota`;\n\t\t\treturn undefined;\n\t\t}\n\n\t\tif (settings.fitnessGate) {\n\t\t\tconst verdict = this._evaluateModelFitness(this._routerSurfaceForTier(decision.tier), resolved.model);\n\t\t\tif (!verdict.fit) {\n\t\t\t\tif (decision.tier === \"medium\") {\n\t\t\t\t\tconst failure = this._formatFitnessFailure(verdict);\n\t\t\t\t\tconst fallback = this._resolveExpensiveFallbackRoute(\n\t\t\t\t\t\tdecision,\n\t\t\t\t\t\t\"medium_unfit_fallback_expensive\",\n\t\t\t\t\t\t`Medium model is unfit (${failure}); falling back to expensive model`,\n\t\t\t\t\t);\n\t\t\t\t\tif (fallback) return fallback;\n\t\t\t\t}\n\t\t\t\tthis._lastModelRouterSkipReason = `${decision.tier} model unfit: ${this._formatFitnessFailure(verdict)} (fitness gate)`;\n\t\t\t\treturn undefined;\n\t\t\t}\n\t\t}\n\n\t\tthis._lastModelRouterSkipReason = undefined;\n\t\tdecision.model = resolvedName;\n\t\treturn { decision, model: resolved.model };\n\t}\n\n\tprivate _resolveModelRouterModelForIntent(intent: ModelRouterIntent): Model<Api> | undefined {\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst modelPattern = intent === \"research\" ? settings.cheapModel : settings.expensiveModel;\n\t\tif (!modelPattern) return undefined;\n\t\tconst resolved = resolveCliModel({ cliModel: modelPattern, modelRegistry: this.deps.getModelRegistry() });\n\t\tif (!resolved.model) return undefined;\n\t\tif (!this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) return undefined;\n\t\treturn this.deps.isModelExhausted(resolved.model) ? undefined : resolved.model;\n\t}\n\n\tresolveConfiguredTierModel(tier: \"cheap\" | \"medium\" | \"expensive\"): Model<Api> | undefined {\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst pattern =\n\t\t\ttier === \"cheap\" ? settings.cheapModel : tier === \"medium\" ? settings.mediumModel : settings.expensiveModel;\n\t\tif (!pattern) return undefined;\n\t\tconst resolved = resolveCliModel({ cliModel: pattern, modelRegistry: this.deps.getModelRegistry() });\n\t\tif (!resolved.model) return undefined;\n\t\tif (!this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) return undefined;\n\t\treturn this.deps.isModelExhausted(resolved.model) ? undefined : resolved.model;\n\t}\n\n\t/**\n\t * Router resolution with the routing judge (auto-on with the router): the regex classifier's\n\t * decision is the baseline; when a judge model resolves (judgeModel, else mediumModel), one\n\t * bounded, tool-less completion may move the tier between cheap/medium/expensive — never to\n\t * learning. Core rule encoded in the judge prompt: planning is never cheap unless genuinely\n\t * trivial. Every fallback stays visible in the decision reasons, and judge spend reports\n\t * through spawned-usage accounting.\n\t */\n\tasync resolveTurnRouteJudged(\n\t\tprompt: string,\n\t\toptions?: { skipJudge?: boolean },\n\t): Promise<{ decision: RouteDecision; model: Model<Api> } | undefined> {\n\t\tconst baseline = this._resolveModelRouterTurnRoute(prompt);\n\t\tif (!baseline) return undefined;\n\t\tif (options?.skipJudge) return baseline;\n\t\t// Deterministic executor routes need no judge (Level-0 already decided).\n\t\tif (baseline.decision.reasonCode === \"executor_direct\") return baseline;\n\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tif (!settings.judgeEnabled) return baseline;\n\t\tif (shouldSkipRouteJudgeForStaticDecision(baseline.decision)) return baseline;\n\t\tconst judgePattern = settings.judgeModel ?? settings.mediumModel;\n\t\tif (!judgePattern) return baseline;\n\t\tconst judgeModel = this.deps.resolveLaneModel(judgePattern);\n\t\tif (!judgeModel) {\n\t\t\treturn {\n\t\t\t\tdecision: withJudgeUnavailableFallback(\n\t\t\t\t\tbaseline.decision,\n\t\t\t\t\t`routing judge unavailable: ${judgePattern} did not resolve; baseline kept`,\n\t\t\t\t),\n\t\t\t\tmodel: baseline.model,\n\t\t\t};\n\t\t}\n\t\tif (settings.fitnessGate) {\n\t\t\tconst verdict = this._evaluateModelFitness(\"router_judge\", judgeModel);\n\t\t\tif (!verdict.fit) {\n\t\t\t\treturn {\n\t\t\t\t\tdecision: {\n\t\t\t\t\t\t...baseline.decision,\n\t\t\t\t\t\treasons: [\n\t\t\t\t\t\t\t...baseline.decision.reasons,\n\t\t\t\t\t\t\t`routing judge skipped: ${formatModelRouterModel(judgeModel)} unfit (${this._formatFitnessFailure(verdict)})`,\n\t\t\t\t\t\t],\n\t\t\t\t\t},\n\t\t\t\t\tmodel: baseline.model,\n\t\t\t\t};\n\t\t\t}\n\t\t}\n\n\t\tlet spentUsage: Usage | undefined;\n\t\tconst judged = await runRouteJudge({\n\t\t\tprompt,\n\t\t\tbaseline: baseline.decision,\n\t\t\tsignal: this.deps.getReflectionSignal(),\n\t\t\tcomplete: async ({ systemPrompt, userPrompt, signal }) => {\n\t\t\t\tconst completion = await this.deps.runIsolatedCompletion({\n\t\t\t\t\tsystemPrompt,\n\t\t\t\t\tmessages: [{ role: \"user\", content: [{ type: \"text\", text: userPrompt }], timestamp: Date.now() }],\n\t\t\t\t\tmodel: judgeModel,\n\t\t\t\t\t// Per-tier thinking (R1): judgeThinking overrides the judge's own completion; unset\n\t\t\t\t\t// keeps today's \"off\" (the judge is a cheap classification call by default).\n\t\t\t\t\tthinkingLevel: settings.judgeThinking ?? \"off\",\n\t\t\t\t\tmaxTokens: ROUTE_JUDGE_MAX_OUTPUT_TOKENS,\n\t\t\t\t\tsignal,\n\t\t\t\t\t// The judge system prompt is static — the provider can cache the prefix.\n\t\t\t\t\tcacheRetention: \"short\",\n\t\t\t\t});\n\t\t\t\tspentUsage = completion.usage;\n\t\t\t\treturn {\n\t\t\t\t\ttext: completion.text,\n\t\t\t\t\tcostUsd: completion.usage.cost.total,\n\t\t\t\t\tstopReason: String(completion.stopReason),\n\t\t\t\t};\n\t\t\t},\n\t\t});\n\t\tif (spentUsage && (spentUsage.cost.total > 0 || spentUsage.totalTokens > 0)) {\n\t\t\tthis.deps.addSpawnedUsage(spentUsage, { label: \"router-judge\" });\n\t\t}\n\n\t\tif (!judged.verdict || judged.decision.tier === baseline.decision.tier) {\n\t\t\t// Same tier (or judge fell back): keep the baseline model, carry the annotated decision.\n\t\t\treturn { decision: judged.decision, model: baseline.model };\n\t\t}\n\n\t\tconst judgedTier = judged.decision.tier;\n\t\tif (judgedTier !== \"cheap\" && judgedTier !== \"medium\" && judgedTier !== \"expensive\") {\n\t\t\treturn { decision: baseline.decision, model: baseline.model };\n\t\t}\n\t\tconst judgedModel = this.resolveConfiguredTierModel(judgedTier);\n\t\tif (!judgedModel) {\n\t\t\treturn {\n\t\t\t\tdecision: {\n\t\t\t\t\t...baseline.decision,\n\t\t\t\t\treasons: [\n\t\t\t\t\t\t...baseline.decision.reasons,\n\t\t\t\t\t\t`Route judge chose ${judgedTier} but no model resolves for that tier; baseline kept`,\n\t\t\t\t\t],\n\t\t\t\t},\n\t\t\t\tmodel: baseline.model,\n\t\t\t};\n\t\t}\n\t\treturn { decision: { ...judged.decision, model: formatModelRouterModel(judgedModel) }, model: judgedModel };\n\t}\n\n\t// biome-ignore lint/correctness/noUnusedPrivateClassMembers: test seam\n\tprivate _resolveModelRouterTurnModel(prompt: string): Model<Api> | undefined {\n\t\tconst resolved = this._resolveModelRouterTurnRoute(prompt);\n\t\treturn resolved?.model;\n\t}\n\n\tgetStatus(formatLabel?: (label: string) => string): string {\n\t\tconst recentDecisions = getRecentModelRouterDecisions(this.deps.getSessionManager().getEntries());\n\t\tconst lastDecision = this._lastModelRouterDecision ?? recentDecisions.at(-1);\n\t\tconst historicalDecisions = this._lastModelRouterDecision ? recentDecisions : recentDecisions.slice(0, -1);\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst lines = [\n\t\t\tformatModelRouterStatus(\n\t\t\t\tsettings,\n\t\t\t\tlastDecision,\n\t\t\t\tformatLabel,\n\t\t\t\thistoricalDecisions,\n\t\t\t\tthis._lastModelRouterSkipReason,\n\t\t\t\tthis._lastModelRouterIntent ?? lastDecision?.intent,\n\t\t\t\tsettings.fitnessGate ? this._getRouterTierFitnessStatuses() : undefined,\n\t\t\t\tthis.deps.getFailoverStatus(),\n\t\t\t),\n\t\t];\n\t\tconst diagnostics = collectModelRouterConfigDiagnostics(\n\t\t\tsettings,\n\t\t\tthis.deps.getModelRegistry(),\n\t\t\tthis.deps.getAgentDir(),\n\t\t);\n\t\tif (diagnostics.length > 0) {\n\t\t\tlines.push(formatLabel ? formatLabel(\"Config diagnostics:\") : \"Config diagnostics:\");\n\t\t\tfor (const diagnostic of diagnostics) {\n\t\t\t\tlines.push(`- ${diagnostic}`);\n\t\t\t}\n\t\t}\n\t\treturn lines.join(\"\\n\");\n\t}\n\n\tasync runRoutedTurn(\n\t\tmessages: AgentMessage | AgentMessage[],\n\t\troutedModel: Model<Api> | undefined,\n\t\trouteDecision: RouteDecision | undefined,\n\t\tpersistDecision = true,\n\t): Promise<void> {\n\t\tif (!routedModel) {\n\t\t\tawait this.deps.runAgentPrompt(messages);\n\t\t\treturn;\n\t\t}\n\n\t\tconst agent = this.deps.getAgent();\n\t\tconst previousModel = agent.state.model;\n\t\tconst previousThinkingLevel = agent.state.thinkingLevel;\n\t\tconst previousTurnTools = agent.state.tools;\n\t\tconst previousSystemPrompt = agent.state.systemPrompt;\n\t\t// G4 swap bookkeeping (Bug G): the exact references the swap below assigns, so the finally can\n\t\t// restore ONLY what IT put there — never assigned when no swap happens (e.g. a full-class\n\t\t// routed profile).\n\t\tlet swappedTools: typeof previousTurnTools | undefined;\n\t\tlet swappedSystemPrompt: typeof previousSystemPrompt | undefined;\n\t\tconst previousActiveModelRouterIntent = this._activeModelRouterIntent;\n\t\tconst previousActiveModelRouterRoute = this._activeModelRouterRoute;\n\t\tconst previousModelRouterSessionBuffer = this._modelRouterSessionBuffer;\n\t\tconst previousModelRouterEscalationRequested = this._modelRouterEscalationRequested;\n\t\tconst bufferRoutedTurn = routeDecision?.tier === \"cheap\";\n\t\tconst originalHistoryLength = agent.state.messages.length;\n\t\tlet retryModel: Model<Api> | undefined;\n\t\tlet completedDecision: ModelRouterDecisionStatus | undefined = routeDecision\n\t\t\t? {\n\t\t\t\t\troute: routeDecision,\n\t\t\t\t\troutedModel: formatModelRouterModel(routedModel),\n\t\t\t\t\toutcome: \"routed\",\n\t\t\t\t\tintent: routeDecision.tier === \"cheap\" ? \"research\" : \"modify\",\n\t\t\t\t}\n\t\t\t: undefined;\n\t\tlet thrownError: unknown;\n\t\tif (routeDecision) {\n\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t}\n\t\tthis._activeModelRouterIntent = routeDecision\n\t\t\t? routeDecision.tier === \"cheap\"\n\t\t\t\t? \"research\"\n\t\t\t\t: \"modify\"\n\t\t\t: undefined;\n\t\tthis._activeModelRouterRoute = routeDecision;\n\t\tif (bufferRoutedTurn) {\n\t\t\tthis._modelRouterSessionBuffer = createModelRouterSessionBuffer();\n\t\t\tthis._modelRouterEscalationRequested = false;\n\t\t}\n\t\tconst routerThinkingSettings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst configuredThinking = !routeDecision\n\t\t\t? undefined\n\t\t\t: routeDecision.reasonCode === \"executor_direct\"\n\t\t\t\t? routerThinkingSettings.executorThinking\n\t\t\t\t: routeDecision.tier === \"cheap\"\n\t\t\t\t\t? routerThinkingSettings.cheapThinking\n\t\t\t\t\t: routeDecision.tier === \"medium\"\n\t\t\t\t\t\t? routerThinkingSettings.mediumThinking\n\t\t\t\t\t\t: routeDecision.tier === \"expensive\"\n\t\t\t\t\t\t\t? routerThinkingSettings.expensiveThinking\n\t\t\t\t\t\t\t: undefined;\n\t\tconst routedThinkingLevel = clampThinkingLevel(\n\t\t\troutedModel,\n\t\t\tconfiguredThinking ?? previousThinkingLevel,\n\t\t) as ThinkingLevel;\n\t\tconst modelChanged = !modelsAreEqual(this.deps.getModel(), routedModel);\n\t\tconst thinkingChanged = routedThinkingLevel !== previousThinkingLevel;\n\t\tif (modelChanged || thinkingChanged) {\n\t\t\tagent.state.model = routedModel;\n\t\t\t// Per-tier thinking (R1): a configured tier/executor thinking level overrides the inherited\n\t\t\t// session thinking for THIS routed turn only; unset falls back to exactly today's\n\t\t\t// inherit-and-clamp behavior. Executor routes carry tier \"cheap\" too, so reasonCode is\n\t\t\t// checked first — otherwise an executor turn would silently pick up cheapThinking instead.\n\t\t\t// The judge's own completion has a separate knob (judgeThinking) applied at its call site.\n\t\t\tagent.state.thinkingLevel = routedThinkingLevel;\n\t\t\t// G4: capability tool-filtering follows the ROUTED model for the turn. Without this a\n\t\t\t// cheap/local routed model inherits the session model's full tool surface — schemas it\n\t\t\t// pays for on every request and may not be able to drive at all.\n\t\t\tif (modelChanged) {\n\t\t\t\tconst routedProfile = deriveModelCapabilityProfile({\n\t\t\t\t\tcontextWindow: routedModel.contextWindow,\n\t\t\t\t\tmode: this.deps.getSettingsManager().getModelCapabilitySettings().mode,\n\t\t\t\t});\n\t\t\t\tif (routedProfile.class !== \"full\") {\n\t\t\t\t\tconst allowed = new Set(\n\t\t\t\t\t\tfilterToolNamesForCapability(\n\t\t\t\t\t\t\tpreviousTurnTools.map((tool) => tool.name),\n\t\t\t\t\t\t\troutedProfile,\n\t\t\t\t\t\t),\n\t\t\t\t\t);\n\t\t\t\t\tswappedTools = previousTurnTools.filter((tool) => allowed.has(tool.name));\n\t\t\t\t\tagent.state.tools = swappedTools;\n\t\t\t\t}\n\t\t\t}\n\t\t\t// The routed prompt follows both the routed tool surface and its effective thinking mode.\n\t\t\t// This removes an Ultra policy during max/Luna routes, including same-model tier overrides.\n\t\t\t// Per-turn only; a live extension override is preserved rather than silently replaced.\n\t\t\tif (agent.state.systemPrompt === this.deps.getBaseSystemPrompt()) {\n\t\t\t\tswappedSystemPrompt = this.deps.buildSystemPromptForToolNames(agent.state.tools.map((tool) => tool.name));\n\t\t\t\tagent.state.systemPrompt = swappedSystemPrompt;\n\t\t\t}\n\t\t}\n\t\ttry {\n\t\t\tawait this.deps.runAgentPrompt(messages);\n\t\t\t// Speculative muscle-retry (G16 refinement): an executor-routed turn is a bet that the\n\t\t\t// small model can run the toolkit command directly. If it ends WITHOUT a successful\n\t\t\t// run_toolkit_script execution, retry ONCE on the same executor with the brain's\n\t\t\t// refined instruction injected — the brain warms while the muscle tries, so the retry\n\t\t\t// pays only when the muscle actually missed.\n\t\t\tif (\n\t\t\t\trouteDecision?.reasonCode === \"executor_direct\" &&\n\t\t\t\t!this._isModelRouterRetry &&\n\t\t\t\t!this._executorTurnExecutedScript(originalHistoryLength)\n\t\t\t) {\n\t\t\t\tconst refined = await this._buildExecutorRefinedPrompt(messages);\n\t\t\t\tif (refined) {\n\t\t\t\t\tagent.state.messages.splice(originalHistoryLength);\n\t\t\t\t\tif (bufferRoutedTurn) this._modelRouterSessionBuffer = createModelRouterSessionBuffer();\n\t\t\t\t\tawait this.deps.runAgentPrompt([\n\t\t\t\t\t\t{ role: \"user\", content: [{ type: \"text\", text: refined }], timestamp: Date.now() },\n\t\t\t\t\t]);\n\t\t\t\t\tcompletedDecision = {\n\t\t\t\t\t\troute: {\n\t\t\t\t\t\t\t...routeDecision,\n\t\t\t\t\t\t\treasonCode: \"executor_speculative_retry\",\n\t\t\t\t\t\t\treasons: [\n\t\t\t\t\t\t\t\t...routeDecision.reasons,\n\t\t\t\t\t\t\t\t\"Executor missed on first try; retried with brain-refined instruction\",\n\t\t\t\t\t\t\t],\n\t\t\t\t\t\t},\n\t\t\t\t\t\troutedModel: formatModelRouterModel(routedModel),\n\t\t\t\t\t\toutcome: \"routed\",\n\t\t\t\t\t\tintent: \"research\",\n\t\t\t\t\t};\n\t\t\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t\t\t} else {\n\t\t\t\t\t// The muscle missed AND the reflex brain could not refine the request into a toolkit\n\t\t\t\t\t// instruction (no fit brain model, or no confident plan). There is deliberately NO\n\t\t\t\t\t// frontier fallback here, so surface the miss instead of letting it stand silently —\n\t\t\t\t\t// otherwise the routed turn just ends with an unrun command and no explanation.\n\t\t\t\t\tthis.deps.emit({\n\t\t\t\t\t\ttype: \"warning\",\n\t\t\t\t\t\tmessage:\n\t\t\t\t\t\t\t\"Executor lane: the toolkit command did not run and the reflex brain could not refine it into an explicit instruction; leaving the turn as-is (no automatic escalation).\",\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (bufferRoutedTurn && this._modelRouterEscalationRequested) {\n\t\t\t\tagent.state.messages.splice(originalHistoryLength);\n\t\t\t\tretryModel = this._resolveModelRouterModelForIntent(\"modify\") ?? previousModel;\n\t\t\t\tcompletedDecision = {\n\t\t\t\t\troute: routeDecision!,\n\t\t\t\t\troutedModel: formatModelRouterModel(routedModel),\n\t\t\t\t\toutcome: \"escalated\",\n\t\t\t\t\tretryModel: formatModelRouterModel(retryModel),\n\t\t\t\t\tintent: routeDecision!.tier === \"cheap\" ? \"research\" : \"modify\",\n\t\t\t\t};\n\t\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t\t} else if (bufferRoutedTurn && this._modelRouterSessionBuffer) {\n\t\t\t\tflushModelRouterSessionBuffer(\n\t\t\t\t\tthis._modelRouterSessionBuffer,\n\t\t\t\t\t(message) => {\n\t\t\t\t\t\tthis.deps.getSessionManager().appendMessage(message);\n\t\t\t\t\t},\n\t\t\t\t\t(customType, content, display, details) => {\n\t\t\t\t\t\tthis.deps.getSessionManager().appendCustomMessageEntry(customType, content, display, details);\n\t\t\t\t\t},\n\t\t\t\t);\n\t\t\t}\n\t\t} catch (error) {\n\t\t\tthrownError = error;\n\t\t\tif (completedDecision) {\n\t\t\t\tcompletedDecision = { ...completedDecision, outcome: \"failed\" };\n\t\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t\t}\n\t\t} finally {\n\t\t\t// Restore the pre-route model ONLY if the routed model is still in place: a command\n\t\t\t// handler may have legitimately changed the session model mid-turn (setModel or a\n\t\t\t// provider re-registration), and clobbering that would silently undo the change.\n\t\t\tif (modelsAreEqual(agent.state.model, routedModel)) {\n\t\t\t\tagent.state.model = previousModel;\n\t\t\t\tagent.state.thinkingLevel = previousThinkingLevel;\n\t\t\t\t// Symmetric restore (Bug G): undo tools/systemPrompt only if each is STILL the exact\n\t\t\t\t// reference/string the G4 swap above assigned (never assigned at all when the routed\n\t\t\t\t// profile was full-class — then there is nothing to restore either). An extension calling\n\t\t\t\t// setActiveToolsByName mid-turn reassigns both to its own values without touching the\n\t\t\t\t// model — the model guard above still passes, but that live change is legitimate and must\n\t\t\t\t// survive rather than being silently reverted to the stale pre-turn snapshot.\n\t\t\t\tif (swappedTools !== undefined && agent.state.tools === swappedTools) {\n\t\t\t\t\tagent.state.tools = previousTurnTools;\n\t\t\t\t}\n\t\t\t\tif (swappedSystemPrompt !== undefined && agent.state.systemPrompt === swappedSystemPrompt) {\n\t\t\t\t\tagent.state.systemPrompt = previousSystemPrompt;\n\t\t\t\t}\n\t\t\t\t// The registry may have changed mid-turn (command-time registerProvider): re-resolve\n\t\t\t\t// the restored model so a provider override is not dropped with the routed model.\n\t\t\t\tthis.deps.refreshCurrentModelFromRegistry();\n\t\t\t}\n\t\t\tthis._activeModelRouterIntent = previousActiveModelRouterIntent;\n\t\t\tthis._activeModelRouterRoute = previousActiveModelRouterRoute;\n\t\t\tthis._modelRouterSessionBuffer = previousModelRouterSessionBuffer;\n\t\t\tthis._modelRouterEscalationRequested = previousModelRouterEscalationRequested;\n\t\t}\n\n\t\tif (retryModel && !thrownError) {\n\t\t\tconst previousIsModelRouterRetry = this._isModelRouterRetry;\n\t\t\ttry {\n\t\t\t\tthis._isModelRouterRetry = true;\n\t\t\t\tconst retryDecision: RouteDecision = {\n\t\t\t\t\ttier: \"expensive\",\n\t\t\t\t\trisk: \"high-impact\",\n\t\t\t\t\tconfidence: 1.0,\n\t\t\t\t\treasonCode: \"cheap_mutating_tool_escalation\",\n\t\t\t\t\treasons: [\"Cheap research turn attempted a mutating tool and escalated\"],\n\t\t\t\t\tfallbackFrom: \"cheap\",\n\t\t\t\t\tmodel: formatModelRouterModel(retryModel),\n\t\t\t\t};\n\t\t\t\tawait this.runRoutedTurn(messages, retryModel, retryDecision, false);\n\t\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t\t} catch (error) {\n\t\t\t\tthrownError = error;\n\t\t\t\tif (completedDecision) {\n\t\t\t\t\tcompletedDecision = { ...completedDecision, outcome: \"failed\" };\n\t\t\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t\t\t}\n\t\t\t} finally {\n\t\t\t\tthis._isModelRouterRetry = previousIsModelRouterRetry;\n\t\t\t}\n\t\t}\n\n\t\tif (persistDecision && completedDecision) {\n\t\t\tpersistModelRouterDecision(this.deps.getSessionManager(), completedDecision);\n\t\t\t// G3: one route event per user-facing routed turn (the escalation retry runs with\n\t\t\t// persistDecision=false, so it does not double-emit). Codes/numbers only — no prompt text.\n\t\t\tthis.deps.emitAutonomyTelemetry({\n\t\t\t\ttype: AUTONOMY_TELEMETRY_EVENT_TYPES.routeDecision,\n\t\t\t\ttimestamp: new Date().toISOString(),\n\t\t\t\tpayload: {\n\t\t\t\t\ttier: completedDecision.route.tier,\n\t\t\t\t\trisk: completedDecision.route.risk,\n\t\t\t\t\treasonCode: completedDecision.route.reasonCode,\n\t\t\t\t\tconfidence: completedDecision.route.confidence,\n\t\t\t\t\toutcome: completedDecision.outcome,\n\t\t\t\t},\n\t\t\t});\n\t\t}\n\n\t\tif (thrownError) {\n\t\t\tthrow thrownError;\n\t\t}\n\t}\n}\n"]}
+{"version":3,"file":"model-router-controller.d.ts","sourceRoot":"","sources":["../../src/core/model-router-controller.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;AAGH,OAAO,KAAK,EAAE,KAAK,EAAE,YAAY,EAAiB,MAAM,2BAA2B,CAAC;AACpF,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,gCAAgC,CAAC;AACrE,OAAO,KAAK,EAAE,GAAG,EAAW,KAAK,EAAE,KAAK,EAAE,MAAM,mBAAmB,CAAC;AAEpE,OAAO,KAAK,EAAE,iBAAiB,EAAE,yBAAyB,EAAE,wBAAwB,EAAE,MAAM,oBAAoB,CAAC;AACjH,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,yBAAyB,CAAC;AAC7D,OAAO,EAAkC,KAAK,sBAAsB,EAAE,MAAM,gCAAgC,CAAC;AAG7G,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAC;AAkBzD,OAAO,EAIN,KAAK,yBAAyB,EAC9B,KAAK,yBAAyB,EAE9B,MAAM,0BAA0B,CAAC;AAElC,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,8BAA8B,CAAC;AAE1E,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AAO7D,wGAAwG;AACxG,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,MAAM,CAEhE;AA0CD,MAAM,WAAW,yBAAyB;IACzC;iFAC6E;IAC7E,QAAQ,IAAI,KAAK,CAAC;IAClB,4FAA4F;IAC5F,QAAQ,IAAI,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CAAC;IACnC,oFAAoF;IACpF,kBAAkB,IAAI,eAAe,CAAC;IACtC,4GAA4G;IAC5G,iBAAiB,IAAI,cAAc,CAAC;IACpC,uFAAuF;IACvF,gBAAgB,IAAI,aAAa,CAAC;IAClC,4DAA4D;IAC5D,gBAAgB,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,OAAO,CAAC;IAC7C,yEAAyE;IACzE,iBAAiB,IAAI,yBAAyB,CAAC;IAC/C,kGAAkG;IAClG,WAAW,IAAI,MAAM,CAAC;IACtB,0EAA0E;IAC1E,mBAAmB,IAAI,WAAW,CAAC;IACnC,qGAAmG;IACnG,mBAAmB,IAAI,MAAM,CAAC;IAC9B,6GAA6G;IAC7G,cAAc,CAAC,QAAQ,EAAE,YAAY,GAAG,YAAY,EAAE,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACvE,iGAAiG;IACjG,6BAA6B,CAAC,SAAS,EAAE,MAAM,EAAE,GAAG,MAAM,CAAC;IAC3D,0GAA0G;IAC1G,+BAA+B,IAAI,IAAI,CAAC;IACxC,yGAAuG;IACvG,qBAAqB,CAAC,IAAI,EAAE,yBAAyB,GAAG,OAAO,CAAC,wBAAwB,CAAC,CAAC;IAC1F;kGAC8F;IAC9F,eAAe,CACd,KAAK,EAAE,KAAK,EACZ,IAAI,EAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,eAAe,CAAC,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,GAClE,MAAM,GAAG,SAAS,CAAC;IACtB,oDAAoD;IACpD,IAAI,CAAC,KAAK,EAAE,iBAAiB,GAAG,IAAI,CAAC;IACrC,wFAAwF;IACxF,qBAAqB,CAAC,KAAK,EAAE,sBAAsB,GAAG,IAAI,CAAC;IAC3D,6EAA6E;IAC7E,gBAAgB,CAAC,OAAO,EAAE,MAAM,GAAG,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CAAC;IAC1D,sGAAsG;IACtG,yBAAyB,IAAI,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CAAC;IACpD;;4EAEwE;IACxE,mBAAmB,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,qBAAqB,GAAG,SAAS,CAAC;CAC1E;AAED;;;GAGG;AACH,qBAAa,qBAAqB;IACjC,gFAAgF;IAChF,OAAO,CAAC,wBAAwB,CAAC,CAAoB;IACrD,OAAO,CAAC,uBAAuB,CAAC,CAAgB;IAChD,OAAO,CAAC,yBAAyB,CAAC,CAA2B;IAC7D,OAAO,CAAC,+BAA+B,CAAS;IAChD,OAAO,CAAC,mBAAmB,CAAS;IACpC,OAAO,CAAC,wBAAwB,CAAC,CAA4B;IAC7D,OAAO,CAAC,0BAA0B,CAAC,CAAS;IAC5C,OAAO,CAAC,sBAAsB,CAAC,CAAoB;IAEnD,OAAO,CAAC,QAAQ,CAAC,IAAI,CAA4B;IAEjD,YAAY,IAAI,EAAE,yBAAyB,EAE1C;IAED,6GAA6G;IAC7G,eAAe,IAAI,OAAO,CAEzB;IAED,qFAAqF;IACrF,eAAe,IAAI,yBAAyB,GAAG,SAAS,CAEvD;IAED;;;;OAIG;IACH,qBAAqB,CAAC,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,OAAO,GAAG;QAAE,KAAK,EAAE,IAAI,CAAC;QAAC,MAAM,EAAE,MAAM,CAAA;KAAE,GAAG,SAAS,CAmBlG;IAED;;;;;;;;;;;;;OAaG;IACH,kCAAkC,IAAI,IAAI,CAIzC;IAED;;;;OAIG;IACH,qBAAqB,CAAC,OAAO,EAAE,YAAY,GAAG,OAAO,CAYpD;IAED,OAAO,CAAC,0BAA0B;IAMlC,OAAO,CAAC,qBAAqB;IAO7B,OAAO,CAAC,qBAAqB;IAI7B,OAAO,CAAC,qBAAqB;IAI7B,OAAO,CAAC,6BAA6B;IAmBrC,OAAO,CAAC,8BAA8B;IAiCtC,OAAO,CAAC,qBAAqB;IA6B7B,0GAA0G;IAC1G,OAAO,CAAC,2BAA2B;YAYrB,2BAA2B;IA2CzC,OAAO,CAAC,4BAA4B;IAiHpC,OAAO,CAAC,iCAAiC;IAUzC,0BAA0B,CAAC,IAAI,EAAE,OAAO,GAAG,QAAQ,GAAG,WAAW,GAAG,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CAezF;IAED;;;;;;;OAOG;IACG,sBAAsB,CAC3B,MAAM,EAAE,MAAM,EACd,OAAO,CAAC,EAAE;QAAE,SAAS,CAAC,EAAE,OAAO,CAAA;KAAE,GAC/B,OAAO,CAAC;QAAE,QAAQ,EAAE,aAAa,CAAC;QAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAA;KAAE,GAAG,SAAS,CAAC,CAqGrE;IAGD,OAAO,CAAC,4BAA4B;IAKpC,SAAS,CAAC,WAAW,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,MAAM,GAAG,MAAM,CA6BzD;IAEK,aAAa,CAClB,QAAQ,EAAE,YAAY,GAAG,YAAY,EAAE,EACvC,WAAW,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,EACnC,aAAa,EAAE,aAAa,GAAG,SAAS,EACxC,eAAe,UAAO,GACpB,OAAO,CAAC,IAAI,CAAC,CA8Pf;CACD","sourcesContent":["/**\n * Model-router turn routing: the session's per-turn model-selection subsystem — the regex/executor\n * route resolver, the optional bounded routing judge, the executor lane (Level-0 toolkit direct hit\n * + speculative brain-refined retry), the per-tier thinking/tool-surface swap around a routed turn,\n * the cheap-research-turn session buffer with mutating-tool escalation to an expensive retry, and\n * the router status/diagnostics report.\n *\n * Extracted verbatim from agent-session.ts (god-file decomposition). Owns the transient per-turn\n * route state — the active intent/route, the cheap-turn session buffer, the escalation-requested and\n * retry-in-flight flags — and the sticky last-decision/last-skip-reason/last-intent used by the\n * status report. Everything else it needs — the live agent + its state, the current model, the\n * session/settings managers, the model registry, the agent dir, the reflection abort signal, the\n * base system prompt, the isolated-completion primitive, spawned-usage accounting, the event/telemetry\n * emitters, and the recently-extracted BackgroundLaneController (resolveLaneModel) / ContextPipeline\n * (resolveCurationModelIfFit) collaborators — is reached through narrow deps accessors rather than the\n * whole AgentSession.\n *\n * Drive-path boundary (deliberate): the actual agent.prompt()/continue() loop stays host-side in\n * AgentSession._runAgentPrompt; this controller's parallel routed drive path ({@link runRoutedTurn})\n * owns only the route decision/escalation/tier bookkeeping and delegates every agent turn back through\n * {@link ModelRouterControllerDeps.runAgentPrompt}, so the drive-loop logic is never duplicated. The\n * host keeps a one-line delegation at each call-in: the routing prep + routed-turn entry in\n * _promptUnserialized, the beforeToolCall MUTATION escalation branch ({@link\n * maybeEscalateToolCall}), the tool-name-agnostic VALIDATION-FAILURE escalation branch for cloud\n * models ({@link requestValidationFailureEscalation} — de-conflated from the mutation gate; see the\n * capability-gate spine doctrine, which routes local/managed models to an evidence-gated\n * native→phone auto-probe on AgentSession instead), the message_end cheap-turn buffering ({@link\n * captureSessionMessage}), the retry-event suppression ({@link isRetryInFlight}), the public\n * getModelRouterStatus / autonomy-telemetry reads, and the tier resolution's consultation of the\n * persisted `/toolprobe` verdict for local/managed tier models ({@link\n * ModelRouterControllerDeps.getToolProbeVerdict}).\n */\n\nimport { createHash } from \"node:crypto\";\nimport type { Agent, AgentMessage, ThinkingLevel } from \"@caupulican/pi-agent-core\";\nimport type { SessionManager } from \"@caupulican/pi-agent-core/node\";\nimport type { Api, Message, Model, Usage } from \"@caupulican/pi-ai\";\nimport { clampThinkingLevel, modelsAreEqual } from \"@caupulican/pi-ai\";\nimport type { AgentSessionEvent, IsolatedCompletionOptions, IsolatedCompletionResult } from \"./agent-session.ts\";\nimport type { RouteDecision } from \"./autonomy/contracts.ts\";\nimport { AUTONOMY_TELEMETRY_EVENT_TYPES, type AutonomyTelemetryEvent } from \"./autonomy/telemetry-events.ts\";\nimport { latestUserPromptText } from \"./context-pipeline.ts\";\nimport { deriveModelCapabilityProfile, filterToolNamesForCapability } from \"./model-capability.ts\";\nimport type { ModelRegistry } from \"./model-registry.ts\";\nimport { resolveCliModel } from \"./model-resolver.ts\";\nimport { collectModelRouterConfigDiagnostics } from \"./model-router/config-diagnostics.ts\";\nimport { classifyExecutorTurn } from \"./model-router/executor-route.ts\";\nimport {\n\tevaluateSurfaceFitness,\n\ttype FitnessGatedSurface,\n\ttype FitnessGateVerdict,\n} from \"./model-router/fitness-gate.ts\";\nimport { classifyModelRouterRoute, type ModelRouterIntent } from \"./model-router/intent-classifier.ts\";\nimport { ROUTE_JUDGE_MAX_OUTPUT_TOKENS, runRouteJudge } from \"./model-router/route-judge.ts\";\nimport {\n\tbufferModelRouterSessionCustomMessage,\n\tbufferModelRouterSessionMessage,\n\tcreateModelRouterSessionBuffer,\n\tflushModelRouterSessionBuffer,\n\ttype ModelRouterSessionBuffer,\n} from \"./model-router/session-buffer.ts\";\nimport {\n\tformatModelRouterStatus,\n\tgetRecentModelRouterDecisions,\n\tMODEL_ROUTER_DECISION_CUSTOM_TYPE,\n\ttype ModelRouterDecisionStatus,\n\ttype ModelRouterFailoverStatus,\n\ttype ModelRouterFitnessStatuses,\n} from \"./model-router/status.ts\";\nimport { isLocalOrManagedRouterModel, shouldEscalateModelRouterTool } from \"./model-router/tool-escalation.ts\";\nimport type { ModelToolProbeVerdict } from \"./models/adaptation-store.ts\";\nimport { FitnessStore } from \"./models/fitness-store.ts\";\nimport type { SettingsManager } from \"./settings-manager.ts\";\nimport {\n\tbuildReflexUserPrompt,\n\tparseReflexPlan,\n\tREFLEX_INTERPRETER_SYSTEM_PROMPT,\n} from \"./toolkit/reflex-interpreter.ts\";\n\n/** Canonical `provider/id` label for a routed/resolved model, as it appears in decisions and status. */\nexport function formatModelRouterModel(model: Model<Api>): string {\n\treturn `${model.provider}/${model.id}`;\n}\n\nconst ROUTE_JUDGE_STATIC_FAST_PATH_REASON_CODES = new Set([\n\t\"empty_prompt\",\n\t\"read_only_question\",\n\t\"release_or_publish\",\n\t\"security_or_auth\",\n\t\"destructive_or_git_history\",\n\t\"settings_or_self_modification\",\n\t\"architecture_or_ambiguous\",\n]);\n\nfunction shouldSkipRouteJudgeForStaticDecision(decision: RouteDecision): boolean {\n\treturn ROUTE_JUDGE_STATIC_FAST_PATH_REASON_CODES.has(decision.reasonCode);\n}\n\nfunction withJudgeUnavailableFallback(decision: RouteDecision, reason: string): RouteDecision {\n\treturn {\n\t\t...decision,\n\t\treasonCode: \"judge_unavailable_fallback\",\n\t\treasons: [...decision.reasons, reason],\n\t};\n}\n\nfunction persistModelRouterDecision(\n\tsessionManager: Pick<SessionManager, \"appendCustomEntry\">,\n\tdecision: ModelRouterDecisionStatus,\n): void {\n\tsessionManager.appendCustomEntry(MODEL_ROUTER_DECISION_CUSTOM_TYPE, decision);\n}\n\n/**\n * Deterministic `addSpawnedUsage` reportId: `kind` + session id + a content hash of `identity`\n * (the text driving the call — never `Date.now`/random). Same identity on a retry of the same\n * logical work unit yields the same id, so the ledger's `seenSubagentReportIds` dedupe catches a\n * duplicate report instead of double-counting spend.\n */\nfunction deriveSpawnedUsageReportId(kind: string, sessionId: string, identity: string): string {\n\tconst digest = createHash(\"sha256\").update(identity).digest(\"hex\").slice(0, 16);\n\treturn `${kind}:${sessionId}:${digest}`;\n}\n\nexport interface ModelRouterControllerDeps {\n\t/** Live agent — the controller reads/writes agent.state.{model,thinkingLevel,tools,systemPrompt,messages}\n\t * for the per-turn tier swap and aborts it on a mutating-tool escalation. */\n\tgetAgent(): Agent;\n\t/** Current session model, used to decide whether a routed turn actually swaps the model. */\n\tgetModel(): Model<Api> | undefined;\n\t/** Router/executor/judge/thinking settings + capability mode (all opt-in gates). */\n\tgetSettingsManager(): SettingsManager;\n\t/** Session log: routed-turn message buffering/persistence, decision persistence, recent-decision status. */\n\tgetSessionManager(): SessionManager;\n\t/** Resolves configured route/judge/executor model patterns against configured auth. */\n\tgetModelRegistry(): ModelRegistry;\n\t/** Session-scoped provider/model quota exhaustion guard. */\n\tisModelExhausted(model: Model<Api>): boolean;\n\t/** Status snapshot for exhausted models and the last failover notice. */\n\tgetFailoverStatus(): ModelRouterFailoverStatus;\n\t/** Root dir the host-keyed {@link FitnessStore} lives under (executor tool-call fitness gate). */\n\tgetAgentDir(): string;\n\t/** Aborts the judge's bounded completion when the session is disposed. */\n\tgetReflectionSignal(): AbortSignal;\n\t/** Base (extension-free) system prompt — the tier swap only sheds tools when the turn is on it. */\n\tgetBaseSystemPrompt(): string;\n\t/** The host-side drive loop (agent.prompt()/continue()); the routed drive path delegates every turn here. */\n\trunAgentPrompt(messages: AgentMessage | AgentMessage[]): Promise<void>;\n\t/** Rebuilds the system prompt for a filtered tool surface (routed-model capability shedding). */\n\tbuildSystemPromptForToolNames(toolNames: string[]): string;\n\t/** Re-resolves the restored model against the registry after a routed turn (provider override safety). */\n\trefreshCurrentModelFromRegistry(): void;\n\t/** One-shot, tool-less LLM call — the routing judge and the executor reflex-brain warmup ride this. */\n\trunIsolatedCompletion(opts: IsolatedCompletionOptions): Promise<IsolatedCompletionResult>;\n\t/** Rolls judge/brain spend into spawned-usage accounting. `reportId` is REQUIRED: every\n\t * caller derives a stable id from the work unit's identity so a retry cannot double-count. */\n\taddSpawnedUsage(\n\t\tusage: Usage,\n\t\topts: { label?: string; sourceSessionId?: string; reportId: string },\n\t): string | undefined;\n\t/** Session event stream (executor-miss warning). */\n\temit(event: AgentSessionEvent): void;\n\t/** Autonomy telemetry stream (one route-decision event per user-facing routed turn). */\n\temitAutonomyTelemetry(event: AutonomyTelemetryEvent): void;\n\t/** Resolves the judge model pattern via {@link BackgroundLaneController}. */\n\tresolveLaneModel(pattern: string): Model<Api> | undefined;\n\t/** Fitness-gated reflex-brain model via {@link ContextPipeline} (executor speculative refinement). */\n\tresolveCurationModelIfFit(): Model<Api> | undefined;\n\t/** Persisted `/toolprobe` verdict for this model (native / text-protocol / none), or undefined\n\t * when never probed. Tier-resolution's consultation reads this ONLY for local/managed models\n\t * ({@link isLocalOrManagedRouterModel}); cloud models never call it. */\n\tgetToolProbeVerdict(model: Model<Api>): ModelToolProbeVerdict | undefined;\n}\n\n/**\n * Owns the model-router turn routing extracted from {@link AgentSession}. See the module header for the\n * drive-path boundary that keeps the agent.prompt()/continue() loop host-side.\n */\nexport class ModelRouterController {\n\t/** Active model-router intent for the current transient routed turn, if any. */\n\tprivate _activeModelRouterIntent?: ModelRouterIntent;\n\tprivate _activeModelRouterRoute?: RouteDecision;\n\tprivate _modelRouterSessionBuffer?: ModelRouterSessionBuffer;\n\tprivate _modelRouterEscalationRequested = false;\n\tprivate _isModelRouterRetry = false;\n\tprivate _lastModelRouterDecision?: ModelRouterDecisionStatus;\n\tprivate _lastModelRouterSkipReason?: string;\n\tprivate _lastModelRouterIntent?: ModelRouterIntent;\n\n\tprivate readonly deps: ModelRouterControllerDeps;\n\n\tconstructor(deps: ModelRouterControllerDeps) {\n\t\tthis.deps = deps;\n\t}\n\n\t/** True while the escalation retry turn is running, so the host can suppress its duplicate prompt events. */\n\tisRetryInFlight(): boolean {\n\t\treturn this._isModelRouterRetry;\n\t}\n\n\t/** Latest completed route decision (sticky), for the autonomy telemetry snapshot. */\n\tgetLastDecision(): ModelRouterDecisionStatus | undefined {\n\t\treturn this._lastModelRouterDecision;\n\t}\n\n\t/**\n\t * beforeToolCall escalation gate: a cheap research turn that reaches for a mutating tool aborts the\n\t * turn and requests a retry on the expensive model. Returns the block result the host hook forwards,\n\t * or undefined when no escalation is required.\n\t */\n\tmaybeEscalateToolCall(toolName: string, args: unknown): { block: true; reason: string } | undefined {\n\t\tif (\n\t\t\tthis._activeModelRouterRoute &&\n\t\t\tshouldEscalateModelRouterTool({\n\t\t\t\ttier: this._activeModelRouterRoute.tier,\n\t\t\t\ttoolName,\n\t\t\t\targs,\n\t\t\t\treasonCode: this._activeModelRouterRoute.reasonCode,\n\t\t\t})\n\t\t) {\n\t\t\tthis._modelRouterEscalationRequested = true;\n\t\t\tthis.deps.getAgent().abort();\n\t\t\treturn {\n\t\t\t\tblock: true,\n\t\t\t\treason:\n\t\t\t\t\t\"Model router escalation required: a cheap research turn attempted a mutating tool. Retry the turn on the configured expensive model.\",\n\t\t\t};\n\t\t}\n\t\treturn undefined;\n\t}\n\n\t/**\n\t * Tool-name-agnostic validation-failure escalation gate, called from\n\t * AgentSession's onToolValidationEscalation handler for a CLOUD model only — see the\n\t * capability-gate-spine doctrine (local/managed models never reach this method; they trigger\n\t * the evidence-gated native→phone auto-probe instead). Unlike {@link maybeEscalateToolCall} (the\n\t * beforeToolCall MUTATION gate, still governed by `shouldEscalateModelRouterTool`/\n\t * READ_ONLY_TOOL_NAMES — a legitimate, unrelated mutation-blast-radius policy), \"the model\n\t * repeatedly cannot construct valid arguments for this tool\" is evidence about the MODEL's\n\t * capability, not about the tool's mutation status, so a read-only tool's repeated validation\n\t * failure now escalates a cheap routed turn exactly like a mutating tool's would — takes no\n\t * tool-name/args input at all (unlike maybeEscalateToolCall), because every repeated validation\n\t * failure escalates regardless of which tool or model triggered it. No-op outside an active\n\t * cheap-tier routed turn, same scoping as maybeEscalateToolCall.\n\t */\n\trequestValidationFailureEscalation(): void {\n\t\tif (this._activeModelRouterRoute?.tier !== \"cheap\") return;\n\t\tthis._modelRouterEscalationRequested = true;\n\t\tthis.deps.getAgent().abort();\n\t}\n\n\t/**\n\t * message_end hook: while a cheap routed turn is buffering, capture its messages into the session\n\t * buffer instead of persisting them (they are flushed on success or discarded on escalation).\n\t * Returns true when the message was buffered, so the host skips its own persistence.\n\t */\n\tcaptureSessionMessage(message: AgentMessage): boolean {\n\t\tconst modelRouterBuffer = this._modelRouterSessionBuffer;\n\t\tif (!modelRouterBuffer) return false;\n\t\tif (message.role === \"custom\") {\n\t\t\tbufferModelRouterSessionCustomMessage(modelRouterBuffer, message);\n\t\t\treturn true;\n\t\t}\n\t\tif (message.role === \"user\" || message.role === \"assistant\" || message.role === \"toolResult\") {\n\t\t\tbufferModelRouterSessionMessage(modelRouterBuffer, message as Message);\n\t\t\treturn true;\n\t\t}\n\t\treturn false;\n\t}\n\n\tprivate _isModelAvailableAndAuthed(pattern: string): boolean {\n\t\tconst resolved = resolveCliModel({ cliModel: pattern, modelRegistry: this.deps.getModelRegistry() });\n\t\tif (!resolved.model) return false;\n\t\treturn this.deps.getModelRegistry().hasConfiguredAuth(resolved.model);\n\t}\n\n\tprivate _evaluateModelFitness(surface: FitnessGatedSurface, model: Model<Api>): FitnessGateVerdict {\n\t\tconst fitness = FitnessStore.forAgentDir(this.deps.getAgentDir())\n\t\t\t.getForHost()\n\t\t\t.find((entry) => entry.model === formatModelRouterModel(model));\n\t\treturn evaluateSurfaceFitness(surface, fitness?.report);\n\t}\n\n\tprivate _formatFitnessFailure(verdict: Exclude<FitnessGateVerdict, { fit: true }>): string {\n\t\treturn verdict.reason === \"unprobed\" ? \"unprobed\" : `${verdict.lane} ${verdict.succeeded}/${verdict.total}`;\n\t}\n\n\tprivate _routerSurfaceForTier(tier: \"cheap\" | \"medium\" | \"expensive\"): FitnessGatedSurface {\n\t\treturn tier === \"cheap\" ? \"router_cheap\" : tier === \"medium\" ? \"router_medium\" : \"router_expensive\";\n\t}\n\n\tprivate _getRouterTierFitnessStatuses(): ModelRouterFitnessStatuses {\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst statuses: ModelRouterFitnessStatuses = {};\n\t\tfor (const tier of [\"cheap\", \"medium\", \"expensive\"] as const) {\n\t\t\tconst pattern =\n\t\t\t\ttier === \"cheap\" ? settings.cheapModel : tier === \"medium\" ? settings.mediumModel : settings.expensiveModel;\n\t\t\tif (!pattern) continue;\n\t\t\tconst resolved = resolveCliModel({ cliModel: pattern, modelRegistry: this.deps.getModelRegistry() });\n\t\t\tif (!resolved.model || !this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) continue;\n\t\t\tconst verdict = this._evaluateModelFitness(this._routerSurfaceForTier(tier), resolved.model);\n\t\t\tstatuses[tier] = verdict.fit\n\t\t\t\t? { status: verdict.probed ? \"fit\" : \"unprobed\" }\n\t\t\t\t: verdict.reason === \"unprobed\"\n\t\t\t\t\t? { status: \"unprobed\" }\n\t\t\t\t\t: { status: \"unfit\", lane: verdict.lane, succeeded: verdict.succeeded, total: verdict.total };\n\t\t}\n\t\treturn statuses;\n\t}\n\n\tprivate _resolveExpensiveFallbackRoute(\n\t\tdecision: RouteDecision,\n\t\treasonCode: string,\n\t\treason: string,\n\t): { decision: RouteDecision; model: Model<Api> } | undefined {\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst expensivePattern = settings.expensiveModel;\n\t\tif (!expensivePattern || !this._isModelAvailableAndAuthed(expensivePattern)) return undefined;\n\t\tconst resolvedExpensive = resolveCliModel({\n\t\t\tcliModel: expensivePattern,\n\t\t\tmodelRegistry: this.deps.getModelRegistry(),\n\t\t});\n\t\tif (!resolvedExpensive.model) return undefined;\n\t\tif (this.deps.isModelExhausted(resolvedExpensive.model)) {\n\t\t\tthis._lastModelRouterSkipReason = \"expensive model exhausted: quota\";\n\t\t\treturn undefined;\n\t\t}\n\t\tif (settings.fitnessGate) {\n\t\t\tconst verdict = this._evaluateModelFitness(\"router_expensive\", resolvedExpensive.model);\n\t\t\tif (!verdict.fit) {\n\t\t\t\tthis._lastModelRouterSkipReason = `expensive model unfit: ${this._formatFitnessFailure(verdict)} (fitness gate)`;\n\t\t\t\treturn undefined;\n\t\t\t}\n\t\t}\n\t\tdecision.fallbackFrom = \"medium\";\n\t\tdecision.tier = \"expensive\";\n\t\tdecision.reasonCode = reasonCode;\n\t\tdecision.reasons = [...decision.reasons, reason];\n\t\tdecision.model = formatModelRouterModel(resolvedExpensive.model);\n\t\tthis._lastModelRouterSkipReason = undefined;\n\t\treturn { decision, model: resolvedExpensive.model };\n\t}\n\n\tprivate _resolveExecutorRoute(\n\t\tprompt: string,\n\t\texecutorPattern: string | undefined,\n\t): { decision: RouteDecision; model: Model<Api> } | undefined {\n\t\tif (!executorPattern) return undefined;\n\t\ttry {\n\t\t\tconst verdict = classifyExecutorTurn(prompt, this.deps.getSettingsManager().getToolkitScripts());\n\t\t\tif (!verdict.execute) return undefined;\n\t\t\tconst resolved = resolveCliModel({ cliModel: executorPattern, modelRegistry: this.deps.getModelRegistry() });\n\t\t\tif (!resolved.model || !this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) return undefined;\n\t\t\t// Fitness gate: the executor must have PROVEN tool-calling on this host (same\n\t\t\t// canonical-ref discipline as the curation gate).\n\t\t\tif (!this._evaluateModelFitness(\"executor\", resolved.model).fit) return undefined;\n\t\t\tthis._lastModelRouterIntent = \"research\";\n\t\t\treturn {\n\t\t\t\tdecision: {\n\t\t\t\t\ttier: \"cheap\",\n\t\t\t\t\trisk: \"scoped-write\",\n\t\t\t\t\tconfidence: 1,\n\t\t\t\t\treasonCode: \"executor_direct\",\n\t\t\t\t\treasons: [`Executor lane: Level-0 direct hit on toolkit script \"${verdict.scriptName}\"`],\n\t\t\t\t},\n\t\t\t\tmodel: resolved.model,\n\t\t\t};\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\t/** True if a run_toolkit_script tool result since `fromIndex` actually EXECUTED (not error/ambiguous). */\n\tprivate _executorTurnExecutedScript(fromIndex: number): boolean {\n\t\tfor (const message of this.deps.getAgent().state.messages.slice(fromIndex)) {\n\t\t\tif ((message as { role?: string }).role !== \"toolResult\") continue;\n\t\t\tif ((message as { toolName?: string }).toolName !== \"run_toolkit_script\") continue;\n\t\t\tif ((message as { isError?: boolean }).isError === true) continue;\n\t\t\tconst outcome = (message as { details?: { outcome?: unknown } }).details?.outcome;\n\t\t\tif (outcome === \"executed\") return true;\n\t\t}\n\t\treturn false;\n\t}\n\n\t/** Ask the reflex brain to refine the last user request into an explicit toolkit instruction. */\n\tprivate async _buildExecutorRefinedPrompt(messages: AgentMessage | AgentMessage[]): Promise<string | undefined> {\n\t\ttry {\n\t\t\tconst model = this.deps.resolveCurationModelIfFit();\n\t\t\tif (!model) return undefined;\n\t\t\tconst list = Array.isArray(messages) ? messages : [messages];\n\t\t\tconst request = latestUserPromptText(list.filter((m): m is AgentMessage => true));\n\t\t\tif (!request) return undefined;\n\t\t\tconst scripts = this.deps.getSettingsManager().getToolkitScripts();\n\t\t\tconst completion = await this.deps.runIsolatedCompletion({\n\t\t\t\tsystemPrompt: REFLEX_INTERPRETER_SYSTEM_PROMPT,\n\t\t\t\tmessages: [\n\t\t\t\t\t{\n\t\t\t\t\t\trole: \"user\",\n\t\t\t\t\t\tcontent: [{ type: \"text\", text: buildReflexUserPrompt(request, scripts) }],\n\t\t\t\t\t\ttimestamp: Date.now(),\n\t\t\t\t\t},\n\t\t\t\t],\n\t\t\t\tmodel,\n\t\t\t\tthinkingLevel: \"off\",\n\t\t\t\tmaxTokens: 256,\n\t\t\t\tcacheRetention: \"short\",\n\t\t\t\t// Stable per-lane synthetic affinity key for repeat executor-brain warmups.\n\t\t\t\tlaneKind: \"executor\",\n\t\t\t});\n\t\t\tif (completion.usage.cost.total > 0 || completion.usage.totalTokens > 0) {\n\t\t\t\t// `reportId` keyed on the request text driving THIS refinement — stable across a\n\t\t\t\t// retry of the same routed turn, distinct across genuinely different requests.\n\t\t\t\tconst reportId = deriveSpawnedUsageReportId(\n\t\t\t\t\t\"executor-brain\",\n\t\t\t\t\tthis.deps.getSessionManager().getSessionId(),\n\t\t\t\t\trequest,\n\t\t\t\t);\n\t\t\t\tthis.deps.addSpawnedUsage(completion.usage, { label: \"executor-brain-warmup\", reportId });\n\t\t\t}\n\t\t\tconst plan = parseReflexPlan(completion.text);\n\t\t\tif (!plan || plan.script === \"none\") return undefined;\n\t\t\tconst argHint = plan.args.length > 0 ? ` with args ${JSON.stringify(plan.args)}` : \"\";\n\t\t\treturn `Run the toolkit script \"${plan.script}\"${argHint} using run_toolkit_script, then report its result exactly.`;\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate _resolveModelRouterTurnRoute(prompt: string): { decision: RouteDecision; model: Model<Api> } | undefined {\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tif (!settings.enabled) {\n\t\t\tthis._lastModelRouterSkipReason = \"disabled\";\n\t\t\treturn undefined;\n\t\t}\n\n\t\t// Executor lane: a Level-0 DIRECT toolkit hit on a command-shaped prompt routes the\n\t\t// whole turn to the configured local executor (tool-call-fitness-gated) instead of\n\t\t// spending the frontier model on a one-tool reflex. Ambiguity never routes here — it\n\t\t// stays with the big model and the reflex brain. Deterministic, so the judge is skipped.\n\t\tconst executorRoute = this._resolveExecutorRoute(prompt, settings.executorModel);\n\t\tif (executorRoute) return executorRoute;\n\n\t\tconst decision = classifyModelRouterRoute(prompt);\n\t\tthis._lastModelRouterIntent = decision.tier === \"cheap\" ? \"research\" : \"modify\";\n\n\t\t// Learning tier must not be selected for normal user prompts\n\t\tif (decision.tier === \"learning\") {\n\t\t\tthis._lastModelRouterSkipReason = \"learning tier not supported for user prompts\";\n\t\t\treturn undefined;\n\t\t}\n\n\t\tconst modelPattern =\n\t\t\tsettings[\n\t\t\t\tdecision.tier === \"cheap\" ? \"cheapModel\" : decision.tier === \"medium\" ? \"mediumModel\" : \"expensiveModel\"\n\t\t\t];\n\t\tconst label =\n\t\t\tdecision.tier === \"cheap\" ? \"cheap model\" : decision.tier === \"medium\" ? \"medium model\" : \"expensive model\";\n\n\t\tif (decision.tier === \"medium\" && (!modelPattern || !this._isModelAvailableAndAuthed(modelPattern))) {\n\t\t\tconst fallback = this._resolveExpensiveFallbackRoute(\n\t\t\t\tdecision,\n\t\t\t\t\"medium_unavailable_fallback_expensive\",\n\t\t\t\t\"Medium model is unavailable, falling back to expensive model\",\n\t\t\t);\n\t\t\tif (fallback) return fallback;\n\t\t\tthis._lastModelRouterSkipReason ??= \"medium model and expensive fallback are unavailable\";\n\t\t\treturn undefined;\n\t\t}\n\n\t\tif (!modelPattern) {\n\t\t\tthis._lastModelRouterSkipReason = `${label} unset`;\n\t\t\treturn undefined;\n\t\t}\n\n\t\tconst resolved = resolveCliModel({ cliModel: modelPattern, modelRegistry: this.deps.getModelRegistry() });\n\t\tif (!resolved.model) {\n\t\t\tthis._lastModelRouterSkipReason = `${label} unresolved: ${modelPattern}`;\n\t\t\treturn undefined;\n\t\t}\n\n\t\tconst resolvedName = formatModelRouterModel(resolved.model);\n\t\tif (!this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) {\n\t\t\tthis._lastModelRouterSkipReason = `${label} missing auth: ${resolvedName}`;\n\t\t\treturn undefined;\n\t\t}\n\n\t\tif (this.deps.isModelExhausted(resolved.model)) {\n\t\t\tif (decision.tier === \"medium\") {\n\t\t\t\tconst fallback = this._resolveExpensiveFallbackRoute(\n\t\t\t\t\tdecision,\n\t\t\t\t\t\"medium_exhausted_fallback_expensive\",\n\t\t\t\t\t\"Medium model exhausted: quota; falling back to expensive model\",\n\t\t\t\t);\n\t\t\t\tif (fallback) return fallback;\n\t\t\t}\n\t\t\tthis._lastModelRouterSkipReason = `${decision.tier} model exhausted: quota`;\n\t\t\treturn undefined;\n\t\t}\n\n\t\t// For a LOCAL/MANAGED tier model (never cloud — isLocalOrManagedRouterModel), honor a\n\t\t// persisted \"no working tool-call path\" probe verdict. This is an ALWAYS-ON doctrine gate,\n\t\t// deliberately not behind the opt-in `fitnessGate` setting below (cloud fitness gating stays\n\t\t// opt-in; this local/managed check is unconditional, matching the auto-probe that writes\n\t\t// the verdict). \"native\"/\"text-protocol\"/unprobed all fall through unchanged: native wins\n\t\t// when it works, the phone lane engages downstream via _textProtocolFlag, and an unprobed\n\t\t// model routes native-first (the evidence loop, not a speculative pre-block).\n\t\tif (isLocalOrManagedRouterModel(resolved.model) && this.deps.getToolProbeVerdict(resolved.model) === \"none\") {\n\t\t\tif (decision.tier === \"medium\") {\n\t\t\t\tconst fallback = this._resolveExpensiveFallbackRoute(\n\t\t\t\t\tdecision,\n\t\t\t\t\t\"medium_no_tool_path_fallback_expensive\",\n\t\t\t\t\t\"Medium model has no working tool-call path (native and text-protocol probe both failed); falling back to expensive model\",\n\t\t\t\t);\n\t\t\t\tif (fallback) return fallback;\n\t\t\t}\n\t\t\tthis._lastModelRouterSkipReason = `${decision.tier} model has no working tool-call path (native and text-protocol probe both failed)`;\n\t\t\treturn undefined;\n\t\t}\n\n\t\tif (settings.fitnessGate) {\n\t\t\tconst verdict = this._evaluateModelFitness(this._routerSurfaceForTier(decision.tier), resolved.model);\n\t\t\tif (!verdict.fit) {\n\t\t\t\tif (decision.tier === \"medium\") {\n\t\t\t\t\tconst failure = this._formatFitnessFailure(verdict);\n\t\t\t\t\tconst fallback = this._resolveExpensiveFallbackRoute(\n\t\t\t\t\t\tdecision,\n\t\t\t\t\t\t\"medium_unfit_fallback_expensive\",\n\t\t\t\t\t\t`Medium model is unfit (${failure}); falling back to expensive model`,\n\t\t\t\t\t);\n\t\t\t\t\tif (fallback) return fallback;\n\t\t\t\t}\n\t\t\t\tthis._lastModelRouterSkipReason = `${decision.tier} model unfit: ${this._formatFitnessFailure(verdict)} (fitness gate)`;\n\t\t\t\treturn undefined;\n\t\t\t}\n\t\t}\n\n\t\tthis._lastModelRouterSkipReason = undefined;\n\t\tdecision.model = resolvedName;\n\t\treturn { decision, model: resolved.model };\n\t}\n\n\tprivate _resolveModelRouterModelForIntent(intent: ModelRouterIntent): Model<Api> | undefined {\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst modelPattern = intent === \"research\" ? settings.cheapModel : settings.expensiveModel;\n\t\tif (!modelPattern) return undefined;\n\t\tconst resolved = resolveCliModel({ cliModel: modelPattern, modelRegistry: this.deps.getModelRegistry() });\n\t\tif (!resolved.model) return undefined;\n\t\tif (!this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) return undefined;\n\t\treturn this.deps.isModelExhausted(resolved.model) ? undefined : resolved.model;\n\t}\n\n\tresolveConfiguredTierModel(tier: \"cheap\" | \"medium\" | \"expensive\"): Model<Api> | undefined {\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst pattern =\n\t\t\ttier === \"cheap\" ? settings.cheapModel : tier === \"medium\" ? settings.mediumModel : settings.expensiveModel;\n\t\tif (!pattern) return undefined;\n\t\tconst resolved = resolveCliModel({ cliModel: pattern, modelRegistry: this.deps.getModelRegistry() });\n\t\tif (!resolved.model) return undefined;\n\t\tif (!this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) return undefined;\n\t\tif (this.deps.isModelExhausted(resolved.model)) return undefined;\n\t\t// (Same doctrine as _resolveModelRouterTurnRoute above): never resolve a local/managed\n\t\t// model the probe has already graded as having no working tool-call path.\n\t\tif (isLocalOrManagedRouterModel(resolved.model) && this.deps.getToolProbeVerdict(resolved.model) === \"none\") {\n\t\t\treturn undefined;\n\t\t}\n\t\treturn resolved.model;\n\t}\n\n\t/**\n\t * Router resolution with the routing judge (auto-on with the router): the regex classifier's\n\t * decision is the baseline; when a judge model resolves (judgeModel, else mediumModel), one\n\t * bounded, tool-less completion may move the tier between cheap/medium/expensive — never to\n\t * learning. Core rule encoded in the judge prompt: planning is never cheap unless genuinely\n\t * trivial. Every fallback stays visible in the decision reasons, and judge spend reports\n\t * through spawned-usage accounting.\n\t */\n\tasync resolveTurnRouteJudged(\n\t\tprompt: string,\n\t\toptions?: { skipJudge?: boolean },\n\t): Promise<{ decision: RouteDecision; model: Model<Api> } | undefined> {\n\t\tconst baseline = this._resolveModelRouterTurnRoute(prompt);\n\t\tif (!baseline) return undefined;\n\t\tif (options?.skipJudge) return baseline;\n\t\t// Deterministic executor routes need no judge (Level-0 already decided).\n\t\tif (baseline.decision.reasonCode === \"executor_direct\") return baseline;\n\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tif (!settings.judgeEnabled) return baseline;\n\t\tif (shouldSkipRouteJudgeForStaticDecision(baseline.decision)) return baseline;\n\t\tconst judgePattern = settings.judgeModel ?? settings.mediumModel;\n\t\tif (!judgePattern) return baseline;\n\t\tconst judgeModel = this.deps.resolveLaneModel(judgePattern);\n\t\tif (!judgeModel) {\n\t\t\treturn {\n\t\t\t\tdecision: withJudgeUnavailableFallback(\n\t\t\t\t\tbaseline.decision,\n\t\t\t\t\t`routing judge unavailable: ${judgePattern} did not resolve; baseline kept`,\n\t\t\t\t),\n\t\t\t\tmodel: baseline.model,\n\t\t\t};\n\t\t}\n\t\tif (settings.fitnessGate) {\n\t\t\tconst verdict = this._evaluateModelFitness(\"router_judge\", judgeModel);\n\t\t\tif (!verdict.fit) {\n\t\t\t\treturn {\n\t\t\t\t\tdecision: {\n\t\t\t\t\t\t...baseline.decision,\n\t\t\t\t\t\treasons: [\n\t\t\t\t\t\t\t...baseline.decision.reasons,\n\t\t\t\t\t\t\t`routing judge skipped: ${formatModelRouterModel(judgeModel)} unfit (${this._formatFitnessFailure(verdict)})`,\n\t\t\t\t\t\t],\n\t\t\t\t\t},\n\t\t\t\t\tmodel: baseline.model,\n\t\t\t\t};\n\t\t\t}\n\t\t}\n\n\t\tlet spentUsage: Usage | undefined;\n\t\tconst judged = await runRouteJudge({\n\t\t\tprompt,\n\t\t\tbaseline: baseline.decision,\n\t\t\tsignal: this.deps.getReflectionSignal(),\n\t\t\tcomplete: async ({ systemPrompt, userPrompt, signal }) => {\n\t\t\t\tconst completion = await this.deps.runIsolatedCompletion({\n\t\t\t\t\tsystemPrompt,\n\t\t\t\t\tmessages: [{ role: \"user\", content: [{ type: \"text\", text: userPrompt }], timestamp: Date.now() }],\n\t\t\t\t\tmodel: judgeModel,\n\t\t\t\t\t// Per-tier thinking: judgeThinking overrides the judge's own completion; unset\n\t\t\t\t\t// keeps today's \"off\" (the judge is a cheap classification call by default).\n\t\t\t\t\tthinkingLevel: settings.judgeThinking ?? \"off\",\n\t\t\t\t\tmaxTokens: ROUTE_JUDGE_MAX_OUTPUT_TOKENS,\n\t\t\t\t\tsignal,\n\t\t\t\t\t// The judge system prompt is static — the provider can cache the prefix.\n\t\t\t\t\tcacheRetention: \"short\",\n\t\t\t\t\t// Stable per-lane synthetic affinity key so repeat judge calls hit the same\n\t\t\t\t\t// cache-warm backend.\n\t\t\t\t\tlaneKind: \"route-judge\",\n\t\t\t\t});\n\t\t\t\tspentUsage = completion.usage;\n\t\t\t\treturn {\n\t\t\t\t\ttext: completion.text,\n\t\t\t\t\tcostUsd: completion.usage.cost.total,\n\t\t\t\t\tstopReason: String(completion.stopReason),\n\t\t\t\t};\n\t\t\t},\n\t\t});\n\t\tif (spentUsage && (spentUsage.cost.total > 0 || spentUsage.totalTokens > 0)) {\n\t\t\t// `reportId` keyed on the routed prompt text — stable across a retry of the same\n\t\t\t// judged turn, distinct across genuinely different prompts.\n\t\t\tconst reportId = deriveSpawnedUsageReportId(\n\t\t\t\t\"route-judge\",\n\t\t\t\tthis.deps.getSessionManager().getSessionId(),\n\t\t\t\tprompt,\n\t\t\t);\n\t\t\tthis.deps.addSpawnedUsage(spentUsage, { label: \"router-judge\", reportId });\n\t\t}\n\n\t\tif (!judged.verdict || judged.decision.tier === baseline.decision.tier) {\n\t\t\t// Same tier (or judge fell back): keep the baseline model, carry the annotated decision.\n\t\t\treturn { decision: judged.decision, model: baseline.model };\n\t\t}\n\n\t\tconst judgedTier = judged.decision.tier;\n\t\tif (judgedTier !== \"cheap\" && judgedTier !== \"medium\" && judgedTier !== \"expensive\") {\n\t\t\treturn { decision: baseline.decision, model: baseline.model };\n\t\t}\n\t\tconst judgedModel = this.resolveConfiguredTierModel(judgedTier);\n\t\tif (!judgedModel) {\n\t\t\treturn {\n\t\t\t\tdecision: {\n\t\t\t\t\t...baseline.decision,\n\t\t\t\t\treasons: [\n\t\t\t\t\t\t...baseline.decision.reasons,\n\t\t\t\t\t\t`Route judge chose ${judgedTier} but no model resolves for that tier; baseline kept`,\n\t\t\t\t\t],\n\t\t\t\t},\n\t\t\t\tmodel: baseline.model,\n\t\t\t};\n\t\t}\n\t\treturn { decision: { ...judged.decision, model: formatModelRouterModel(judgedModel) }, model: judgedModel };\n\t}\n\n\t// biome-ignore lint/correctness/noUnusedPrivateClassMembers: test seam\n\tprivate _resolveModelRouterTurnModel(prompt: string): Model<Api> | undefined {\n\t\tconst resolved = this._resolveModelRouterTurnRoute(prompt);\n\t\treturn resolved?.model;\n\t}\n\n\tgetStatus(formatLabel?: (label: string) => string): string {\n\t\tconst recentDecisions = getRecentModelRouterDecisions(this.deps.getSessionManager().getEntries());\n\t\tconst lastDecision = this._lastModelRouterDecision ?? recentDecisions.at(-1);\n\t\tconst historicalDecisions = this._lastModelRouterDecision ? recentDecisions : recentDecisions.slice(0, -1);\n\t\tconst settings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst lines = [\n\t\t\tformatModelRouterStatus(\n\t\t\t\tsettings,\n\t\t\t\tlastDecision,\n\t\t\t\tformatLabel,\n\t\t\t\thistoricalDecisions,\n\t\t\t\tthis._lastModelRouterSkipReason,\n\t\t\t\tthis._lastModelRouterIntent ?? lastDecision?.intent,\n\t\t\t\tsettings.fitnessGate ? this._getRouterTierFitnessStatuses() : undefined,\n\t\t\t\tthis.deps.getFailoverStatus(),\n\t\t\t),\n\t\t];\n\t\tconst diagnostics = collectModelRouterConfigDiagnostics(\n\t\t\tsettings,\n\t\t\tthis.deps.getModelRegistry(),\n\t\t\tthis.deps.getAgentDir(),\n\t\t);\n\t\tif (diagnostics.length > 0) {\n\t\t\tlines.push(formatLabel ? formatLabel(\"Config diagnostics:\") : \"Config diagnostics:\");\n\t\t\tfor (const diagnostic of diagnostics) {\n\t\t\t\tlines.push(`- ${diagnostic}`);\n\t\t\t}\n\t\t}\n\t\treturn lines.join(\"\\n\");\n\t}\n\n\tasync runRoutedTurn(\n\t\tmessages: AgentMessage | AgentMessage[],\n\t\troutedModel: Model<Api> | undefined,\n\t\trouteDecision: RouteDecision | undefined,\n\t\tpersistDecision = true,\n\t): Promise<void> {\n\t\tif (!routedModel) {\n\t\t\tawait this.deps.runAgentPrompt(messages);\n\t\t\treturn;\n\t\t}\n\n\t\tconst agent = this.deps.getAgent();\n\t\tconst previousModel = agent.state.model;\n\t\tconst previousThinkingLevel = agent.state.thinkingLevel;\n\t\tconst previousTurnTools = agent.state.tools;\n\t\tconst previousSystemPrompt = agent.state.systemPrompt;\n\t\t// Swap bookkeeping: the exact references the swap below assigns, so the finally can\n\t\t// restore ONLY what IT put there — never assigned when no swap happens (e.g. a full-class\n\t\t// routed profile).\n\t\tlet swappedTools: typeof previousTurnTools | undefined;\n\t\tlet swappedSystemPrompt: typeof previousSystemPrompt | undefined;\n\t\tconst previousActiveModelRouterIntent = this._activeModelRouterIntent;\n\t\tconst previousActiveModelRouterRoute = this._activeModelRouterRoute;\n\t\tconst previousModelRouterSessionBuffer = this._modelRouterSessionBuffer;\n\t\tconst previousModelRouterEscalationRequested = this._modelRouterEscalationRequested;\n\t\tconst bufferRoutedTurn = routeDecision?.tier === \"cheap\";\n\t\tconst originalHistoryLength = agent.state.messages.length;\n\t\tlet retryModel: Model<Api> | undefined;\n\t\tlet completedDecision: ModelRouterDecisionStatus | undefined = routeDecision\n\t\t\t? {\n\t\t\t\t\troute: routeDecision,\n\t\t\t\t\troutedModel: formatModelRouterModel(routedModel),\n\t\t\t\t\toutcome: \"routed\",\n\t\t\t\t\tintent: routeDecision.tier === \"cheap\" ? \"research\" : \"modify\",\n\t\t\t\t}\n\t\t\t: undefined;\n\t\tlet thrownError: unknown;\n\t\tif (routeDecision) {\n\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t}\n\t\tthis._activeModelRouterIntent = routeDecision\n\t\t\t? routeDecision.tier === \"cheap\"\n\t\t\t\t? \"research\"\n\t\t\t\t: \"modify\"\n\t\t\t: undefined;\n\t\tthis._activeModelRouterRoute = routeDecision;\n\t\tif (bufferRoutedTurn) {\n\t\t\tthis._modelRouterSessionBuffer = createModelRouterSessionBuffer();\n\t\t\tthis._modelRouterEscalationRequested = false;\n\t\t}\n\t\tconst routerThinkingSettings = this.deps.getSettingsManager().getModelRouterSettings();\n\t\tconst configuredThinking = !routeDecision\n\t\t\t? undefined\n\t\t\t: routeDecision.reasonCode === \"executor_direct\"\n\t\t\t\t? routerThinkingSettings.executorThinking\n\t\t\t\t: routeDecision.tier === \"cheap\"\n\t\t\t\t\t? routerThinkingSettings.cheapThinking\n\t\t\t\t\t: routeDecision.tier === \"medium\"\n\t\t\t\t\t\t? routerThinkingSettings.mediumThinking\n\t\t\t\t\t\t: routeDecision.tier === \"expensive\"\n\t\t\t\t\t\t\t? routerThinkingSettings.expensiveThinking\n\t\t\t\t\t\t\t: undefined;\n\t\tconst routedThinkingLevel = clampThinkingLevel(\n\t\t\troutedModel,\n\t\t\tconfiguredThinking ?? previousThinkingLevel,\n\t\t) as ThinkingLevel;\n\t\tconst modelChanged = !modelsAreEqual(this.deps.getModel(), routedModel);\n\t\tconst thinkingChanged = routedThinkingLevel !== previousThinkingLevel;\n\t\tif (modelChanged || thinkingChanged) {\n\t\t\tagent.state.model = routedModel;\n\t\t\t// Per-tier thinking: a configured tier/executor thinking level overrides the inherited\n\t\t\t// session thinking for THIS routed turn only; unset falls back to exactly today's\n\t\t\t// inherit-and-clamp behavior. Executor routes carry tier \"cheap\" too, so reasonCode is\n\t\t\t// checked first — otherwise an executor turn would silently pick up cheapThinking instead.\n\t\t\t// The judge's own completion has a separate knob (judgeThinking) applied at its call site.\n\t\t\tagent.state.thinkingLevel = routedThinkingLevel;\n\t\t\t// Capability tool-filtering follows the ROUTED model for the turn. Without this a\n\t\t\t// cheap/local routed model inherits the session model's full tool surface — schemas it\n\t\t\t// pays for on every request and may not be able to drive at all.\n\t\t\tif (modelChanged) {\n\t\t\t\tconst routedProfile = deriveModelCapabilityProfile({\n\t\t\t\t\tcontextWindow: routedModel.contextWindow,\n\t\t\t\t\tmode: this.deps.getSettingsManager().getModelCapabilitySettings().mode,\n\t\t\t\t});\n\t\t\t\tif (routedProfile.class !== \"full\") {\n\t\t\t\t\tconst allowed = new Set(\n\t\t\t\t\t\tfilterToolNamesForCapability(\n\t\t\t\t\t\t\tpreviousTurnTools.map((tool) => tool.name),\n\t\t\t\t\t\t\troutedProfile,\n\t\t\t\t\t\t),\n\t\t\t\t\t);\n\t\t\t\t\tswappedTools = previousTurnTools.filter((tool) => allowed.has(tool.name));\n\t\t\t\t\tagent.state.tools = swappedTools;\n\t\t\t\t}\n\t\t\t}\n\t\t\t// The routed prompt follows both the routed tool surface and its effective thinking mode.\n\t\t\t// This removes an Ultra policy during max/Luna routes, including same-model tier overrides.\n\t\t\t// Per-turn only; a live extension override is preserved rather than silently replaced.\n\t\t\tif (agent.state.systemPrompt === this.deps.getBaseSystemPrompt()) {\n\t\t\t\tswappedSystemPrompt = this.deps.buildSystemPromptForToolNames(agent.state.tools.map((tool) => tool.name));\n\t\t\t\tagent.state.systemPrompt = swappedSystemPrompt;\n\t\t\t}\n\t\t}\n\t\ttry {\n\t\t\tawait this.deps.runAgentPrompt(messages);\n\t\t\t// Speculative muscle-retry: an executor-routed turn is a bet that the\n\t\t\t// small model can run the toolkit command directly. If it ends WITHOUT a successful\n\t\t\t// run_toolkit_script execution, retry ONCE on the same executor with the brain's\n\t\t\t// refined instruction injected — the brain warms while the muscle tries, so the retry\n\t\t\t// pays only when the muscle actually missed.\n\t\t\tif (\n\t\t\t\trouteDecision?.reasonCode === \"executor_direct\" &&\n\t\t\t\t!this._isModelRouterRetry &&\n\t\t\t\t!this._executorTurnExecutedScript(originalHistoryLength)\n\t\t\t) {\n\t\t\t\tconst refined = await this._buildExecutorRefinedPrompt(messages);\n\t\t\t\tif (refined) {\n\t\t\t\t\tagent.state.messages.splice(originalHistoryLength);\n\t\t\t\t\tif (bufferRoutedTurn) this._modelRouterSessionBuffer = createModelRouterSessionBuffer();\n\t\t\t\t\tawait this.deps.runAgentPrompt([\n\t\t\t\t\t\t{ role: \"user\", content: [{ type: \"text\", text: refined }], timestamp: Date.now() },\n\t\t\t\t\t]);\n\t\t\t\t\tcompletedDecision = {\n\t\t\t\t\t\troute: {\n\t\t\t\t\t\t\t...routeDecision,\n\t\t\t\t\t\t\treasonCode: \"executor_speculative_retry\",\n\t\t\t\t\t\t\treasons: [\n\t\t\t\t\t\t\t\t...routeDecision.reasons,\n\t\t\t\t\t\t\t\t\"Executor missed on first try; retried with brain-refined instruction\",\n\t\t\t\t\t\t\t],\n\t\t\t\t\t\t},\n\t\t\t\t\t\troutedModel: formatModelRouterModel(routedModel),\n\t\t\t\t\t\toutcome: \"routed\",\n\t\t\t\t\t\tintent: \"research\",\n\t\t\t\t\t};\n\t\t\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t\t\t} else {\n\t\t\t\t\t// The muscle missed AND the reflex brain could not refine the request into a toolkit\n\t\t\t\t\t// instruction (no fit brain model, or no confident plan). There is deliberately NO\n\t\t\t\t\t// frontier fallback here, so surface the miss instead of letting it stand silently —\n\t\t\t\t\t// otherwise the routed turn just ends with an unrun command and no explanation.\n\t\t\t\t\tthis.deps.emit({\n\t\t\t\t\t\ttype: \"warning\",\n\t\t\t\t\t\tmessage:\n\t\t\t\t\t\t\t\"Executor lane: the toolkit command did not run and the reflex brain could not refine it into an explicit instruction; leaving the turn as-is (no automatic escalation).\",\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (bufferRoutedTurn && this._modelRouterEscalationRequested) {\n\t\t\t\tagent.state.messages.splice(originalHistoryLength);\n\t\t\t\tretryModel = this._resolveModelRouterModelForIntent(\"modify\") ?? previousModel;\n\t\t\t\tcompletedDecision = {\n\t\t\t\t\troute: routeDecision!,\n\t\t\t\t\troutedModel: formatModelRouterModel(routedModel),\n\t\t\t\t\toutcome: \"escalated\",\n\t\t\t\t\tretryModel: formatModelRouterModel(retryModel),\n\t\t\t\t\tintent: routeDecision!.tier === \"cheap\" ? \"research\" : \"modify\",\n\t\t\t\t};\n\t\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t\t} else if (bufferRoutedTurn && this._modelRouterSessionBuffer) {\n\t\t\t\tflushModelRouterSessionBuffer(\n\t\t\t\t\tthis._modelRouterSessionBuffer,\n\t\t\t\t\t(message) => {\n\t\t\t\t\t\tthis.deps.getSessionManager().appendMessage(message);\n\t\t\t\t\t},\n\t\t\t\t\t(customType, content, display, details) => {\n\t\t\t\t\t\tthis.deps.getSessionManager().appendCustomMessageEntry(customType, content, display, details);\n\t\t\t\t\t},\n\t\t\t\t);\n\t\t\t}\n\t\t} catch (error) {\n\t\t\tthrownError = error;\n\t\t\t// Mirror the escalation splice above (~:812): a buffered cheap-tier turn never flushes its\n\t\t\t// live messages to the session on a genuine throw, so agent.state.messages must be rolled\n\t\t\t// back to the pre-turn length here too — otherwise the never-persisted buffered messages\n\t\t\t// permanently diverge from the persisted session (same shape as the W1.3 ghost-turn bug,\n\t\t\t// but on the error path instead of the success/escalation paths).\n\t\t\tif (bufferRoutedTurn) {\n\t\t\t\tagent.state.messages.splice(originalHistoryLength);\n\t\t\t}\n\t\t\tif (completedDecision) {\n\t\t\t\tcompletedDecision = { ...completedDecision, outcome: \"failed\" };\n\t\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t\t}\n\t\t} finally {\n\t\t\t// Restore the pre-route model ONLY if the routed model is still in place: a command\n\t\t\t// handler may have legitimately changed the session model mid-turn (setModel or a\n\t\t\t// provider re-registration), and clobbering that would silently undo the change.\n\t\t\tif (modelsAreEqual(agent.state.model, routedModel)) {\n\t\t\t\tagent.state.model = previousModel;\n\t\t\t\tagent.state.thinkingLevel = previousThinkingLevel;\n\t\t\t\t// Symmetric restore: undo tools/systemPrompt only if each is STILL the exact\n\t\t\t\t// reference/string the swap above assigned (never assigned at all when the routed\n\t\t\t\t// profile was full-class — then there is nothing to restore either). An extension calling\n\t\t\t\t// setActiveToolsByName mid-turn reassigns both to its own values without touching the\n\t\t\t\t// model — the model guard above still passes, but that live change is legitimate and must\n\t\t\t\t// survive rather than being silently reverted to the stale pre-turn snapshot.\n\t\t\t\tif (swappedTools !== undefined && agent.state.tools === swappedTools) {\n\t\t\t\t\tagent.state.tools = previousTurnTools;\n\t\t\t\t}\n\t\t\t\tif (swappedSystemPrompt !== undefined && agent.state.systemPrompt === swappedSystemPrompt) {\n\t\t\t\t\tagent.state.systemPrompt = previousSystemPrompt;\n\t\t\t\t}\n\t\t\t\t// The registry may have changed mid-turn (command-time registerProvider): re-resolve\n\t\t\t\t// the restored model so a provider override is not dropped with the routed model.\n\t\t\t\tthis.deps.refreshCurrentModelFromRegistry();\n\t\t\t}\n\t\t\tthis._activeModelRouterIntent = previousActiveModelRouterIntent;\n\t\t\tthis._activeModelRouterRoute = previousActiveModelRouterRoute;\n\t\t\tthis._modelRouterSessionBuffer = previousModelRouterSessionBuffer;\n\t\t\tthis._modelRouterEscalationRequested = previousModelRouterEscalationRequested;\n\t\t}\n\n\t\tif (retryModel && !thrownError) {\n\t\t\tconst previousIsModelRouterRetry = this._isModelRouterRetry;\n\t\t\ttry {\n\t\t\t\tthis._isModelRouterRetry = true;\n\t\t\t\tconst retryDecision: RouteDecision = {\n\t\t\t\t\ttier: \"expensive\",\n\t\t\t\t\trisk: \"high-impact\",\n\t\t\t\t\tconfidence: 1.0,\n\t\t\t\t\treasonCode: \"cheap_mutating_tool_escalation\",\n\t\t\t\t\treasons: [\"Cheap research turn attempted a mutating tool and escalated\"],\n\t\t\t\t\tfallbackFrom: \"cheap\",\n\t\t\t\t\tmodel: formatModelRouterModel(retryModel),\n\t\t\t\t};\n\t\t\t\tawait this.runRoutedTurn(messages, retryModel, retryDecision, false);\n\t\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t\t} catch (error) {\n\t\t\t\tthrownError = error;\n\t\t\t\tif (completedDecision) {\n\t\t\t\t\tcompletedDecision = { ...completedDecision, outcome: \"failed\" };\n\t\t\t\t\tthis._lastModelRouterDecision = completedDecision;\n\t\t\t\t}\n\t\t\t} finally {\n\t\t\t\tthis._isModelRouterRetry = previousIsModelRouterRetry;\n\t\t\t}\n\t\t}\n\n\t\tif (persistDecision && completedDecision) {\n\t\t\tpersistModelRouterDecision(this.deps.getSessionManager(), completedDecision);\n\t\t\t// One route event per user-facing routed turn (the escalation retry runs with\n\t\t\t// persistDecision=false, so it does not double-emit). Codes/numbers only — no prompt text.\n\t\t\tthis.deps.emitAutonomyTelemetry({\n\t\t\t\ttype: AUTONOMY_TELEMETRY_EVENT_TYPES.routeDecision,\n\t\t\t\ttimestamp: new Date().toISOString(),\n\t\t\t\tpayload: {\n\t\t\t\t\ttier: completedDecision.route.tier,\n\t\t\t\t\trisk: completedDecision.route.risk,\n\t\t\t\t\treasonCode: completedDecision.route.reasonCode,\n\t\t\t\t\tconfidence: completedDecision.route.confidence,\n\t\t\t\t\toutcome: completedDecision.outcome,\n\t\t\t\t},\n\t\t\t});\n\t\t}\n\n\t\tif (thrownError) {\n\t\t\tthrow thrownError;\n\t\t}\n\t}\n}\n"]}

package/dist/core/model-router-controller.js

@@ -20,10 +20,17 @@
  * owns only the route decision/escalation/tier bookkeeping and delegates every agent turn back through
  * {@link ModelRouterControllerDeps.runAgentPrompt}, so the drive-loop logic is never duplicated. The
  * host keeps a one-line delegation at each call-in: the routing prep + routed-turn entry in
- * _promptUnserialized, the beforeToolCall escalation branch ({@link maybeEscalateToolCall}), the
- * message_end cheap-turn buffering ({@link captureSessionMessage}), the retry-event suppression
- * ({@link isRetryInFlight}), and the public getModelRouterStatus / autonomy-telemetry reads.
+ * _promptUnserialized, the beforeToolCall MUTATION escalation branch ({@link
+ * maybeEscalateToolCall}), the tool-name-agnostic VALIDATION-FAILURE escalation branch for cloud
+ * models ({@link requestValidationFailureEscalation} — de-conflated from the mutation gate; see the
+ * capability-gate spine doctrine, which routes local/managed models to an evidence-gated
+ * native→phone auto-probe on AgentSession instead), the message_end cheap-turn buffering ({@link
+ * captureSessionMessage}), the retry-event suppression ({@link isRetryInFlight}), the public
+ * getModelRouterStatus / autonomy-telemetry reads, and the tier resolution's consultation of the
+ * persisted `/toolprobe` verdict for local/managed tier models ({@link
+ * ModelRouterControllerDeps.getToolProbeVerdict}).
  */
+import { createHash } from "node:crypto";
 import { clampThinkingLevel, modelsAreEqual } from "@caupulican/pi-ai";
 import { AUTONOMY_TELEMETRY_EVENT_TYPES } from "./autonomy/telemetry-events.js";
 import { latestUserPromptText } from "./context-pipeline.js";
@@ -36,7 +43,7 @@ import { classifyModelRouterRoute } from "./model-router/intent-classifier.js";
 import { ROUTE_JUDGE_MAX_OUTPUT_TOKENS, runRouteJudge } from "./model-router/route-judge.js";
 import { bufferModelRouterSessionCustomMessage, bufferModelRouterSessionMessage, createModelRouterSessionBuffer, flushModelRouterSessionBuffer, } from "./model-router/session-buffer.js";
 import { formatModelRouterStatus, getRecentModelRouterDecisions, MODEL_ROUTER_DECISION_CUSTOM_TYPE, } from "./model-router/status.js";
-import { shouldEscalateModelRouterTool } from "./model-router/tool-escalation.js";
+import { isLocalOrManagedRouterModel, shouldEscalateModelRouterTool } from "./model-router/tool-escalation.js";
 import { FitnessStore } from "./models/fitness-store.js";
 import { buildReflexUserPrompt, parseReflexPlan, REFLEX_INTERPRETER_SYSTEM_PROMPT, } from "./toolkit/reflex-interpreter.js";
 /** Canonical `provider/id` label for a routed/resolved model, as it appears in decisions and status. */
@@ -65,6 +72,16 @@ function withJudgeUnavailableFallback(decision, reason) {
 function persistModelRouterDecision(sessionManager, decision) {
     sessionManager.appendCustomEntry(MODEL_ROUTER_DECISION_CUSTOM_TYPE, decision);
 }
+/**
+ * Deterministic `addSpawnedUsage` reportId: `kind` + session id + a content hash of `identity`
+ * (the text driving the call — never `Date.now`/random). Same identity on a retry of the same
+ * logical work unit yields the same id, so the ledger's `seenSubagentReportIds` dedupe catches a
+ * duplicate report instead of double-counting spend.
+ */
+function deriveSpawnedUsageReportId(kind, sessionId, identity) {
+    const digest = createHash("sha256").update(identity).digest("hex").slice(0, 16);
+    return `${kind}:${sessionId}:${digest}`;
+}
 /**
  * Owns the model-router turn routing extracted from {@link AgentSession}. See the module header for the
  * drive-path boundary that keeps the agent.prompt()/continue() loop host-side.
@@ -113,6 +130,26 @@ export class ModelRouterController {
         }
         return undefined;
     }
+    /**
+     * Tool-name-agnostic validation-failure escalation gate, called from
+     * AgentSession's onToolValidationEscalation handler for a CLOUD model only — see the
+     * capability-gate-spine doctrine (local/managed models never reach this method; they trigger
+     * the evidence-gated native→phone auto-probe instead). Unlike {@link maybeEscalateToolCall} (the
+     * beforeToolCall MUTATION gate, still governed by `shouldEscalateModelRouterTool`/
+     * READ_ONLY_TOOL_NAMES — a legitimate, unrelated mutation-blast-radius policy), "the model
+     * repeatedly cannot construct valid arguments for this tool" is evidence about the MODEL's
+     * capability, not about the tool's mutation status, so a read-only tool's repeated validation
+     * failure now escalates a cheap routed turn exactly like a mutating tool's would — takes no
+     * tool-name/args input at all (unlike maybeEscalateToolCall), because every repeated validation
+     * failure escalates regardless of which tool or model triggered it. No-op outside an active
+     * cheap-tier routed turn, same scoping as maybeEscalateToolCall.
+     */
+    requestValidationFailureEscalation() {
+        if (this._activeModelRouterRoute?.tier !== "cheap")
+            return;
+        this._modelRouterEscalationRequested = true;
+        this.deps.getAgent().abort();
+    }
     /**
      * message_end hook: while a cheap routed turn is buffering, capture its messages into the session
      * buffer instead of persisting them (they are flushed on success or discarded on escalation).
@@ -268,9 +305,14 @@ export class ModelRouterController {
                 thinkingLevel: "off",
                 maxTokens: 256,
                 cacheRetention: "short",
+                // Stable per-lane synthetic affinity key for repeat executor-brain warmups.
+                laneKind: "executor",
             });
             if (completion.usage.cost.total > 0 || completion.usage.totalTokens > 0) {
-                this.deps.addSpawnedUsage(completion.usage, { label: "executor-brain-warmup" });
+                // `reportId` keyed on the request text driving THIS refinement — stable across a
+                // retry of the same routed turn, distinct across genuinely different requests.
+                const reportId = deriveSpawnedUsageReportId("executor-brain", this.deps.getSessionManager().getSessionId(), request);
+                this.deps.addSpawnedUsage(completion.usage, { label: "executor-brain-warmup", reportId });
             }
             const plan = parseReflexPlan(completion.text);
             if (!plan || plan.script === "none")
@@ -288,7 +330,7 @@ export class ModelRouterController {
             this._lastModelRouterSkipReason = "disabled";
             return undefined;
         }
-        // G16 executor lane: a Level-0 DIRECT toolkit hit on a command-shaped prompt routes the
+        // Executor lane: a Level-0 DIRECT toolkit hit on a command-shaped prompt routes the
         // whole turn to the configured local executor (tool-call-fitness-gated) instead of
         // spending the frontier model on a one-tool reflex. Ambiguity never routes here — it
         // stays with the big model and the reflex brain. Deterministic, so the judge is skipped.
@@ -334,6 +376,22 @@ export class ModelRouterController {
             this._lastModelRouterSkipReason = `${decision.tier} model exhausted: quota`;
             return undefined;
         }
+        // For a LOCAL/MANAGED tier model (never cloud — isLocalOrManagedRouterModel), honor a
+        // persisted "no working tool-call path" probe verdict. This is an ALWAYS-ON doctrine gate,
+        // deliberately not behind the opt-in `fitnessGate` setting below (cloud fitness gating stays
+        // opt-in; this local/managed check is unconditional, matching the auto-probe that writes
+        // the verdict). "native"/"text-protocol"/unprobed all fall through unchanged: native wins
+        // when it works, the phone lane engages downstream via _textProtocolFlag, and an unprobed
+        // model routes native-first (the evidence loop, not a speculative pre-block).
+        if (isLocalOrManagedRouterModel(resolved.model) && this.deps.getToolProbeVerdict(resolved.model) === "none") {
+            if (decision.tier === "medium") {
+                const fallback = this._resolveExpensiveFallbackRoute(decision, "medium_no_tool_path_fallback_expensive", "Medium model has no working tool-call path (native and text-protocol probe both failed); falling back to expensive model");
+                if (fallback)
+                    return fallback;
+            }
+            this._lastModelRouterSkipReason = `${decision.tier} model has no working tool-call path (native and text-protocol probe both failed)`;
+            return undefined;
+        }
         if (settings.fitnessGate) {
             const verdict = this._evaluateModelFitness(this._routerSurfaceForTier(decision.tier), resolved.model);
             if (!verdict.fit) {
@@ -373,7 +431,14 @@ export class ModelRouterController {
             return undefined;
         if (!this.deps.getModelRegistry().hasConfiguredAuth(resolved.model))
             return undefined;
-        return this.deps.isModelExhausted(resolved.model) ? undefined : resolved.model;
+        if (this.deps.isModelExhausted(resolved.model))
+            return undefined;
+        // (Same doctrine as _resolveModelRouterTurnRoute above): never resolve a local/managed
+        // model the probe has already graded as having no working tool-call path.
+        if (isLocalOrManagedRouterModel(resolved.model) && this.deps.getToolProbeVerdict(resolved.model) === "none") {
+            return undefined;
+        }
+        return resolved.model;
     }
     /**
      * Router resolution with the routing judge (auto-on with the router): the regex classifier's
@@ -432,13 +497,16 @@ export class ModelRouterController {
                     systemPrompt,
                     messages: [{ role: "user", content: [{ type: "text", text: userPrompt }], timestamp: Date.now() }],
                     model: judgeModel,
-                    // Per-tier thinking (R1): judgeThinking overrides the judge's own completion; unset
+                    // Per-tier thinking: judgeThinking overrides the judge's own completion; unset
                     // keeps today's "off" (the judge is a cheap classification call by default).
                     thinkingLevel: settings.judgeThinking ?? "off",
                     maxTokens: ROUTE_JUDGE_MAX_OUTPUT_TOKENS,
                     signal,
                     // The judge system prompt is static — the provider can cache the prefix.
                     cacheRetention: "short",
+                    // Stable per-lane synthetic affinity key so repeat judge calls hit the same
+                    // cache-warm backend.
+                    laneKind: "route-judge",
                 });
                 spentUsage = completion.usage;
                 return {
@@ -449,7 +517,10 @@ export class ModelRouterController {
             },
         });
         if (spentUsage && (spentUsage.cost.total > 0 || spentUsage.totalTokens > 0)) {
-            this.deps.addSpawnedUsage(spentUsage, { label: "router-judge" });
+            // `reportId` keyed on the routed prompt text — stable across a retry of the same
+            // judged turn, distinct across genuinely different prompts.
+            const reportId = deriveSpawnedUsageReportId("route-judge", this.deps.getSessionManager().getSessionId(), prompt);
+            this.deps.addSpawnedUsage(spentUsage, { label: "router-judge", reportId });
         }
         if (!judged.verdict || judged.decision.tier === baseline.decision.tier) {
             // Same tier (or judge fell back): keep the baseline model, carry the annotated decision.
@@ -506,7 +577,7 @@ export class ModelRouterController {
         const previousThinkingLevel = agent.state.thinkingLevel;
         const previousTurnTools = agent.state.tools;
         const previousSystemPrompt = agent.state.systemPrompt;
-        // G4 swap bookkeeping (Bug G): the exact references the swap below assigns, so the finally can
+        // Swap bookkeeping: the exact references the swap below assigns, so the finally can
         // restore ONLY what IT put there — never assigned when no swap happens (e.g. a full-class
         // routed profile).
         let swappedTools;
@@ -557,13 +628,13 @@ export class ModelRouterController {
         const thinkingChanged = routedThinkingLevel !== previousThinkingLevel;
         if (modelChanged || thinkingChanged) {
             agent.state.model = routedModel;
-            // Per-tier thinking (R1): a configured tier/executor thinking level overrides the inherited
+            // Per-tier thinking: a configured tier/executor thinking level overrides the inherited
             // session thinking for THIS routed turn only; unset falls back to exactly today's
             // inherit-and-clamp behavior. Executor routes carry tier "cheap" too, so reasonCode is
             // checked first — otherwise an executor turn would silently pick up cheapThinking instead.
             // The judge's own completion has a separate knob (judgeThinking) applied at its call site.
             agent.state.thinkingLevel = routedThinkingLevel;
-            // G4: capability tool-filtering follows the ROUTED model for the turn. Without this a
+            // Capability tool-filtering follows the ROUTED model for the turn. Without this a
             // cheap/local routed model inherits the session model's full tool surface — schemas it
             // pays for on every request and may not be able to drive at all.
             if (modelChanged) {
@@ -587,7 +658,7 @@ export class ModelRouterController {
         }
         try {
             await this.deps.runAgentPrompt(messages);
-            // Speculative muscle-retry (G16 refinement): an executor-routed turn is a bet that the
+            // Speculative muscle-retry: an executor-routed turn is a bet that the
             // small model can run the toolkit command directly. If it ends WITHOUT a successful
             // run_toolkit_script execution, retry ONCE on the same executor with the brain's
             // refined instruction injected — the brain warms while the muscle tries, so the retry
@@ -651,6 +722,14 @@ export class ModelRouterController {
         }
         catch (error) {
             thrownError = error;
+            // Mirror the escalation splice above (~:812): a buffered cheap-tier turn never flushes its
+            // live messages to the session on a genuine throw, so agent.state.messages must be rolled
+            // back to the pre-turn length here too — otherwise the never-persisted buffered messages
+            // permanently diverge from the persisted session (same shape as the W1.3 ghost-turn bug,
+            // but on the error path instead of the success/escalation paths).
+            if (bufferRoutedTurn) {
+                agent.state.messages.splice(originalHistoryLength);
+            }
             if (completedDecision) {
                 completedDecision = { ...completedDecision, outcome: "failed" };
                 this._lastModelRouterDecision = completedDecision;
@@ -663,8 +742,8 @@ export class ModelRouterController {
             if (modelsAreEqual(agent.state.model, routedModel)) {
                 agent.state.model = previousModel;
                 agent.state.thinkingLevel = previousThinkingLevel;
-                // Symmetric restore (Bug G): undo tools/systemPrompt only if each is STILL the exact
-                // reference/string the G4 swap above assigned (never assigned at all when the routed
+                // Symmetric restore: undo tools/systemPrompt only if each is STILL the exact
+                // reference/string the swap above assigned (never assigned at all when the routed
                 // profile was full-class — then there is nothing to restore either). An extension calling
                 // setActiveToolsByName mid-turn reassigns both to its own values without touching the
                 // model — the model guard above still passes, but that live change is legitimate and must
@@ -713,7 +792,7 @@ export class ModelRouterController {
         }
         if (persistDecision && completedDecision) {
             persistModelRouterDecision(this.deps.getSessionManager(), completedDecision);
-            // G3: one route event per user-facing routed turn (the escalation retry runs with
+            // One route event per user-facing routed turn (the escalation retry runs with
             // persistDecision=false, so it does not double-emit). Codes/numbers only — no prompt text.
             this.deps.emitAutonomyTelemetry({
                 type: AUTONOMY_TELEMETRY_EVENT_TYPES.routeDecision,