@caupulican/pi-adaptative 0.81.2 → 0.81.3

package/dist/core/billing-failover-controller.js

@@ -0,0 +1,72 @@
+import { classifyFailure } from "@caupulican/pi-agent-core";
+import { decideBillingFailover } from "./billing-failover.js";
+const DEFAULT_MODEL_PER_PROVIDER = {
+    "openai-codex": "gpt-5.5",
+};
+export class ExhaustedProviderRegistry {
+    exhausted = new Map();
+    markExhausted(ref, until) {
+        this.exhausted.set(ref, until);
+    }
+    isExhausted(ref) {
+        const until = this.exhausted.get(ref);
+        if (until === undefined)
+            return this.exhausted.has(ref);
+        if (Date.now() < until)
+            return true;
+        this.exhausted.delete(ref);
+        return false;
+    }
+    snapshot() {
+        return [...this.exhausted.keys()].filter((ref) => this.isExhausted(ref));
+    }
+}
+export class BillingFailoverController {
+    deps;
+    lastNotice;
+    constructor(deps) {
+        this.deps = deps;
+    }
+    isExhausted(ref) {
+        return this.deps.exhausted.isExhausted(ref);
+    }
+    snapshotExhausted() {
+        return this.deps.exhausted.snapshot();
+    }
+    getLastNotice() {
+        return this.lastNotice;
+    }
+    getStatus() {
+        return { exhausted: this.snapshotExhausted(), lastNotice: this.lastNotice };
+    }
+    async handleAssistantError(message) {
+        if (message.stopReason !== "error")
+            return false;
+        const classified = classifyFailure({ message: message.errorMessage ?? "", provider: message.provider });
+        if (classified.reason !== "billing_or_quota")
+            return false;
+        const failedModel = this.deps.modelRegistry.find(message.provider, message.model) ?? this.deps.agent.state.model;
+        const failedRef = `${failedModel.provider}/${failedModel.id}`;
+        this.deps.exhausted.markExhausted(failedRef, expiryFromRetryAfter(classified.retryAfterMs));
+        const defaultModelId = DEFAULT_MODEL_PER_PROVIDER[failedModel.provider];
+        const hop = defaultModelId ? this.deps.modelRegistry.find(failedModel.provider, defaultModelId) : undefined;
+        const action = decideBillingFailover({
+            failedModel: { provider: failedModel.provider, id: failedModel.id },
+            billingClass: this.deps.modelRegistry.isUsingOAuth(failedModel) ? "subscription" : "metered",
+            providerDefaultModelId: defaultModelId,
+            hopResolvesWithAuth: Boolean(hop && this.deps.modelRegistry.hasConfiguredAuth(hop)),
+            hopExhausted: hop ? this.deps.exhausted.isExhausted(`${hop.provider}/${hop.id}`) : false,
+            subscriptionHop: this.deps.subscriptionHop,
+        });
+        if (action.action === "failover") {
+            this.deps.agent.state.model = hop;
+        }
+        this.lastNotice = action.notice;
+        this.deps.emit({ type: "warning", message: action.notice });
+        return true;
+    }
+}
+function expiryFromRetryAfter(retryAfterMs) {
+    return retryAfterMs === undefined ? undefined : Date.now() + retryAfterMs;
+}
+//# sourceMappingURL=billing-failover-controller.js.map

package/dist/core/billing-failover-controller.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"billing-failover-controller.js","sourceRoot":"","sources":["../../src/core/billing-failover-controller.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,eAAe,EAAE,MAAM,2BAA2B,CAAC;AAE5D,OAAO,EAAE,qBAAqB,EAAE,MAAM,uBAAuB,CAAC;AAI9D,MAAM,0BAA0B,GAA2B;IAC1D,cAAc,EAAE,SAAS;CACzB,CAAC;AAEF,MAAM,OAAO,yBAAyB;IACpB,SAAS,GAAG,IAAI,GAAG,EAA8B,CAAC;IAEnE,aAAa,CAAC,GAAW,EAAE,KAAc,EAAQ;QAChD,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;IAAA,CAC/B;IAED,WAAW,CAAC,GAAW,EAAW;QACjC,MAAM,KAAK,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QACtC,IAAI,KAAK,KAAK,SAAS;YAAE,OAAO,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QACxD,IAAI,IAAI,CAAC,GAAG,EAAE,GAAG,KAAK;YAAE,OAAO,IAAI,CAAC;QACpC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QAC3B,OAAO,KAAK,CAAC;IAAA,CACb;IAED,QAAQ,GAAa;QACpB,OAAO,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC;IAAA,CACzE;CACD;AAUD,MAAM,OAAO,yBAAyB;IACpB,IAAI,CAAgC;IAC7C,UAAU,CAAqB;IAEvC,YAAY,IAAmC,EAAE;QAChD,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;IAAA,CACjB;IAED,WAAW,CAAC,GAAW,EAAW;QACjC,OAAO,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;IAAA,CAC5C;IAED,iBAAiB,GAAa;QAC7B,OAAO,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,QAAQ,EAAE,CAAC;IAAA,CACtC;IAED,aAAa,GAAuB;QACnC,OAAO,IAAI,CAAC,UAAU,CAAC;IAAA,CACvB;IAED,SAAS,GAA8B;QACtC,OAAO,EAAE,SAAS,EAAE,IAAI,CAAC,iBAAiB,EAAE,EAAE,UAAU,EAAE,IAAI,CAAC,UAAU,EAAE,CAAC;IAAA,CAC5E;IAED,KAAK,CAAC,oBAAoB,CAAC,OAAyB,EAAoB;QACvE,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO;YAAE,OAAO,KAAK,CAAC;QACjD,MAAM,UAAU,GAAG,eAAe,CAAC,EAAE,OAAO,EAAE,OAAO,CAAC,YAAY,IAAI,EAAE,EAAE,QAAQ,EAAE,OAAO,CAAC,QAAQ,EAAE,CAAC,CAAC;QACxG,IAAI,UAAU,CAAC,MAAM,KAAK,kBAAkB;YAAE,OAAO,KAAK,CAAC;QAC3D,MAAM,WAAW,GAAG,IAAI,CAAC,IAAI,CAAC,aAAa,CAAC,IAAI,CAAC,OAAO,CAAC,QAAQ,EAAE,OAAO,CAAC,KAAK,CAAC,IAAI,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,KAAK,CAAC;QACjH,MAAM,SAAS,GAAG,GAAG,WAAW,CAAC,QAAQ,IAAI,WAAW,CAAC,EAAE,EAAE,CAAC;QAC9D,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,aAAa,CAAC,SAAS,EAAE,oBAAoB,CAAC,UAAU,CAAC,YAAY,CAAC,CAAC,CAAC;QAE5F,MAAM,cAAc,GAAG,0BAA0B,CAAC,WAAW,CAAC,QAAQ,CAAC,CAAC;QACxE,MAAM,GAAG,GAAG,cAAc,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,aAAa,CAAC,IAAI,CAAC,WAAW,CAAC,QAAQ,EAAE,cAAc,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAC5G,MAAM,MAAM,GAAG,qBAAqB,CAAC;YACpC,WAAW,EAAE,EAAE,QAAQ,EAAE,WAAW,CAAC,QAAQ,EAAE,EAAE,EAAE,WAAW,CAAC,EAAE,EAAE;YACnE,YAAY,EAAE,IAAI,CAAC,IAAI,CAAC,aAAa,CAAC,YAAY,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,cAAc,CAAC,CAAC,CAAC,SAAS;YAC5F,sBAAsB,EAAE,cAAc;YACtC,mBAAmB,EAAE,OAAO,CAAC,GAAG,IAAI,IAAI,CAAC,IAAI,CAAC,aAAa,CAAC,iBAAiB,CAAC,GAAG,CAAC,CAAC;YACnF,YAAY,EAAE,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,WAAW,CAAC,GAAG,GAAG,CAAC,QAAQ,IAAI,GAAG,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,KAAK;YACxF,eAAe,EAAE,IAAI,CAAC,IAAI,CAAC,eAAe;SAC1C,CAAC,CAAC;QACH,IAAI,MAAM,CAAC,MAAM,KAAK,UAAU,EAAE,CAAC;YAClC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,KAAK,GAAG,GAAiB,CAAC;QACjD,CAAC;QACD,IAAI,CAAC,UAAU,GAAG,MAAM,CAAC,MAAM,CAAC;QAChC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,SAAS,EAAE,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,CAAC,CAAC;QAC5D,OAAO,IAAI,CAAC;IAAA,CACZ;CACD;AAED,SAAS,oBAAoB,CAAC,YAAgC,EAAsB;IACnF,OAAO,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,EAAE,GAAG,YAAY,CAAC;AAAA,CAC1E","sourcesContent":["import type { Agent } from \"@caupulican/pi-agent-core\";\nimport { classifyFailure } from \"@caupulican/pi-agent-core\";\nimport type { Api, AssistantMessage, Model } from \"@caupulican/pi-ai\";\nimport { decideBillingFailover } from \"./billing-failover.ts\";\nimport type { ModelRegistry } from \"./model-registry.ts\";\nimport type { ModelRouterFailoverStatus } from \"./model-router/status.ts\";\n\nconst DEFAULT_MODEL_PER_PROVIDER: Record<string, string> = {\n\t\"openai-codex\": \"gpt-5.5\",\n};\n\nexport class ExhaustedProviderRegistry {\n\tprivate readonly exhausted = new Map<string, number | undefined>();\n\n\tmarkExhausted(ref: string, until?: number): void {\n\t\tthis.exhausted.set(ref, until);\n\t}\n\n\tisExhausted(ref: string): boolean {\n\t\tconst until = this.exhausted.get(ref);\n\t\tif (until === undefined) return this.exhausted.has(ref);\n\t\tif (Date.now() < until) return true;\n\t\tthis.exhausted.delete(ref);\n\t\treturn false;\n\t}\n\n\tsnapshot(): string[] {\n\t\treturn [...this.exhausted.keys()].filter((ref) => this.isExhausted(ref));\n\t}\n}\n\nexport interface BillingFailoverControllerDeps {\n\tagent: Agent;\n\tmodelRegistry: ModelRegistry;\n\temit(event: { type: \"warning\"; message: string }): void;\n\texhausted: ExhaustedProviderRegistry;\n\tsubscriptionHop?: boolean;\n}\n\nexport class BillingFailoverController {\n\tprivate readonly deps: BillingFailoverControllerDeps;\n\tprivate lastNotice: string | undefined;\n\n\tconstructor(deps: BillingFailoverControllerDeps) {\n\t\tthis.deps = deps;\n\t}\n\n\tisExhausted(ref: string): boolean {\n\t\treturn this.deps.exhausted.isExhausted(ref);\n\t}\n\n\tsnapshotExhausted(): string[] {\n\t\treturn this.deps.exhausted.snapshot();\n\t}\n\n\tgetLastNotice(): string | undefined {\n\t\treturn this.lastNotice;\n\t}\n\n\tgetStatus(): ModelRouterFailoverStatus {\n\t\treturn { exhausted: this.snapshotExhausted(), lastNotice: this.lastNotice };\n\t}\n\n\tasync handleAssistantError(message: AssistantMessage): Promise<boolean> {\n\t\tif (message.stopReason !== \"error\") return false;\n\t\tconst classified = classifyFailure({ message: message.errorMessage ?? \"\", provider: message.provider });\n\t\tif (classified.reason !== \"billing_or_quota\") return false;\n\t\tconst failedModel = this.deps.modelRegistry.find(message.provider, message.model) ?? this.deps.agent.state.model;\n\t\tconst failedRef = `${failedModel.provider}/${failedModel.id}`;\n\t\tthis.deps.exhausted.markExhausted(failedRef, expiryFromRetryAfter(classified.retryAfterMs));\n\n\t\tconst defaultModelId = DEFAULT_MODEL_PER_PROVIDER[failedModel.provider];\n\t\tconst hop = defaultModelId ? this.deps.modelRegistry.find(failedModel.provider, defaultModelId) : undefined;\n\t\tconst action = decideBillingFailover({\n\t\t\tfailedModel: { provider: failedModel.provider, id: failedModel.id },\n\t\t\tbillingClass: this.deps.modelRegistry.isUsingOAuth(failedModel) ? \"subscription\" : \"metered\",\n\t\t\tproviderDefaultModelId: defaultModelId,\n\t\t\thopResolvesWithAuth: Boolean(hop && this.deps.modelRegistry.hasConfiguredAuth(hop)),\n\t\t\thopExhausted: hop ? this.deps.exhausted.isExhausted(`${hop.provider}/${hop.id}`) : false,\n\t\t\tsubscriptionHop: this.deps.subscriptionHop,\n\t\t});\n\t\tif (action.action === \"failover\") {\n\t\t\tthis.deps.agent.state.model = hop as Model<Api>;\n\t\t}\n\t\tthis.lastNotice = action.notice;\n\t\tthis.deps.emit({ type: \"warning\", message: action.notice });\n\t\treturn true;\n\t}\n}\n\nfunction expiryFromRetryAfter(retryAfterMs: number | undefined): number | undefined {\n\treturn retryAfterMs === undefined ? undefined : Date.now() + retryAfterMs;\n}\n"]}

package/dist/core/billing-failover.d.ts

@@ -0,0 +1,25 @@
+export type BillingClass = "subscription" | "metered";
+export type BillingFailoverAction = {
+    action: "failover";
+    to: {
+        provider: string;
+        modelId: string;
+    };
+    notice: string;
+} | {
+    action: "halt_ask";
+    notice: string;
+};
+export interface BillingFailoverInput {
+    failedModel: {
+        provider: string;
+        id: string;
+    };
+    billingClass: BillingClass;
+    providerDefaultModelId: string | undefined;
+    hopResolvesWithAuth: boolean;
+    hopExhausted: boolean;
+    subscriptionHop?: boolean;
+}
+export declare function decideBillingFailover(input: BillingFailoverInput): BillingFailoverAction;
+//# sourceMappingURL=billing-failover.d.ts.map

package/dist/core/billing-failover.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"billing-failover.d.ts","sourceRoot":"","sources":["../../src/core/billing-failover.ts"],"names":[],"mappings":"AAAA,MAAM,MAAM,YAAY,GAAG,cAAc,GAAG,SAAS,CAAC;AAEtD,MAAM,MAAM,qBAAqB,GAC9B;IAAE,MAAM,EAAE,UAAU,CAAC;IAAC,EAAE,EAAE;QAAE,QAAQ,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,GACjF;IAAE,MAAM,EAAE,UAAU,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,CAAC;AAE1C,MAAM,WAAW,oBAAoB;IACpC,WAAW,EAAE;QAAE,QAAQ,EAAE,MAAM,CAAC;QAAC,EAAE,EAAE,MAAM,CAAA;KAAE,CAAC;IAC9C,YAAY,EAAE,YAAY,CAAC;IAC3B,sBAAsB,EAAE,MAAM,GAAG,SAAS,CAAC;IAC3C,mBAAmB,EAAE,OAAO,CAAC;IAC7B,YAAY,EAAE,OAAO,CAAC;IACtB,eAAe,CAAC,EAAE,OAAO,CAAC;CAC1B;AAMD,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,oBAAoB,GAAG,qBAAqB,CAkBxF","sourcesContent":["export type BillingClass = \"subscription\" | \"metered\";\n\nexport type BillingFailoverAction =\n\t| { action: \"failover\"; to: { provider: string; modelId: string }; notice: string }\n\t| { action: \"halt_ask\"; notice: string };\n\nexport interface BillingFailoverInput {\n\tfailedModel: { provider: string; id: string };\n\tbillingClass: BillingClass;\n\tproviderDefaultModelId: string | undefined;\n\thopResolvesWithAuth: boolean;\n\thopExhausted: boolean;\n\tsubscriptionHop?: boolean;\n}\n\nfunction haltNotice(provider: string, modelId: string): string {\n\treturn `Provider quota/limit reached for ${provider}/${modelId}: switch models (/model), wait for the limit window, or re-send to retry`;\n}\n\nexport function decideBillingFailover(input: BillingFailoverInput): BillingFailoverAction {\n\tconst { failedModel, providerDefaultModelId } = input;\n\tconst subscriptionHop = input.subscriptionHop ?? true;\n\tif (\n\t\tinput.billingClass === \"subscription\" &&\n\t\tsubscriptionHop &&\n\t\tproviderDefaultModelId &&\n\t\tproviderDefaultModelId !== failedModel.id &&\n\t\tinput.hopResolvesWithAuth &&\n\t\t!input.hopExhausted\n\t) {\n\t\treturn {\n\t\t\taction: \"failover\",\n\t\t\tto: { provider: failedModel.provider, modelId: providerDefaultModelId },\n\t\t\tnotice: `${failedModel.id} quota reached — switched to ${failedModel.provider}/${providerDefaultModelId}`,\n\t\t};\n\t}\n\treturn { action: \"halt_ask\", notice: haltNotice(failedModel.provider, failedModel.id) };\n}\n"]}

package/dist/core/billing-failover.js

@@ -0,0 +1,21 @@
+function haltNotice(provider, modelId) {
+    return `Provider quota/limit reached for ${provider}/${modelId}: switch models (/model), wait for the limit window, or re-send to retry`;
+}
+export function decideBillingFailover(input) {
+    const { failedModel, providerDefaultModelId } = input;
+    const subscriptionHop = input.subscriptionHop ?? true;
+    if (input.billingClass === "subscription" &&
+        subscriptionHop &&
+        providerDefaultModelId &&
+        providerDefaultModelId !== failedModel.id &&
+        input.hopResolvesWithAuth &&
+        !input.hopExhausted) {
+        return {
+            action: "failover",
+            to: { provider: failedModel.provider, modelId: providerDefaultModelId },
+            notice: `${failedModel.id} quota reached — switched to ${failedModel.provider}/${providerDefaultModelId}`,
+        };
+    }
+    return { action: "halt_ask", notice: haltNotice(failedModel.provider, failedModel.id) };
+}
+//# sourceMappingURL=billing-failover.js.map

package/dist/core/billing-failover.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"billing-failover.js","sourceRoot":"","sources":["../../src/core/billing-failover.ts"],"names":[],"mappings":"AAeA,SAAS,UAAU,CAAC,QAAgB,EAAE,OAAe,EAAU;IAC9D,OAAO,oCAAoC,QAAQ,IAAI,OAAO,0EAA0E,CAAC;AAAA,CACzI;AAED,MAAM,UAAU,qBAAqB,CAAC,KAA2B,EAAyB;IACzF,MAAM,EAAE,WAAW,EAAE,sBAAsB,EAAE,GAAG,KAAK,CAAC;IACtD,MAAM,eAAe,GAAG,KAAK,CAAC,eAAe,IAAI,IAAI,CAAC;IACtD,IACC,KAAK,CAAC,YAAY,KAAK,cAAc;QACrC,eAAe;QACf,sBAAsB;QACtB,sBAAsB,KAAK,WAAW,CAAC,EAAE;QACzC,KAAK,CAAC,mBAAmB;QACzB,CAAC,KAAK,CAAC,YAAY,EAClB,CAAC;QACF,OAAO;YACN,MAAM,EAAE,UAAU;YAClB,EAAE,EAAE,EAAE,QAAQ,EAAE,WAAW,CAAC,QAAQ,EAAE,OAAO,EAAE,sBAAsB,EAAE;YACvE,MAAM,EAAE,GAAG,WAAW,CAAC,EAAE,kCAAgC,WAAW,CAAC,QAAQ,IAAI,sBAAsB,EAAE;SACzG,CAAC;IACH,CAAC;IACD,OAAO,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,EAAE,UAAU,CAAC,WAAW,CAAC,QAAQ,EAAE,WAAW,CAAC,EAAE,CAAC,EAAE,CAAC;AAAA,CACxF","sourcesContent":["export type BillingClass = \"subscription\" | \"metered\";\n\nexport type BillingFailoverAction =\n\t| { action: \"failover\"; to: { provider: string; modelId: string }; notice: string }\n\t| { action: \"halt_ask\"; notice: string };\n\nexport interface BillingFailoverInput {\n\tfailedModel: { provider: string; id: string };\n\tbillingClass: BillingClass;\n\tproviderDefaultModelId: string | undefined;\n\thopResolvesWithAuth: boolean;\n\thopExhausted: boolean;\n\tsubscriptionHop?: boolean;\n}\n\nfunction haltNotice(provider: string, modelId: string): string {\n\treturn `Provider quota/limit reached for ${provider}/${modelId}: switch models (/model), wait for the limit window, or re-send to retry`;\n}\n\nexport function decideBillingFailover(input: BillingFailoverInput): BillingFailoverAction {\n\tconst { failedModel, providerDefaultModelId } = input;\n\tconst subscriptionHop = input.subscriptionHop ?? true;\n\tif (\n\t\tinput.billingClass === \"subscription\" &&\n\t\tsubscriptionHop &&\n\t\tproviderDefaultModelId &&\n\t\tproviderDefaultModelId !== failedModel.id &&\n\t\tinput.hopResolvesWithAuth &&\n\t\t!input.hopExhausted\n\t) {\n\t\treturn {\n\t\t\taction: \"failover\",\n\t\t\tto: { provider: failedModel.provider, modelId: providerDefaultModelId },\n\t\t\tnotice: `${failedModel.id} quota reached — switched to ${failedModel.provider}/${providerDefaultModelId}`,\n\t\t};\n\t}\n\treturn { action: \"halt_ask\", notice: haltNotice(failedModel.provider, failedModel.id) };\n}\n"]}

package/dist/core/context-pipeline.d.ts

@@ -77,6 +77,7 @@ export declare class ContextPipeline {
     private _latestPromptPolicyReport;
     private _latestPromptPolicyGcCorrelation;
     private _latestPromptEnforcementReport;
+    private readonly _tokenBudget;
     private readonly deps;
     constructor(deps: ContextPipelineDeps);
     private _contextGcStorageDir;

package/dist/core/context-pipeline.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"context-pipeline.d.ts","sourceRoot":"","sources":["../../src/core/context-pipeline.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AAGH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,2BAA2B,CAAC;AAC9D,OAAO,EAAmD,KAAK,cAAc,EAAE,MAAM,gCAAgC,CAAC;AACtH,OAAO,KAAK,EAAE,GAAG,EAAoB,KAAK,EAAE,KAAK,EAAE,MAAM,mBAAmB,CAAC;AAC7E,OAAO,KAAK,EAAE,yBAAyB,EAAE,wBAAwB,EAAE,MAAM,oBAAoB,CAAC;AAC9F,OAAO,EAAgB,KAAK,yBAAyB,EAA6B,MAAM,4BAA4B,CAAC;AACrH,OAAO,EAAE,KAAK,aAAa,EAA2B,MAAM,gCAAgC,CAAC;AAC7F,OAAO,EAAE,KAAK,kBAAkB,EAAmB,MAAM,4BAA4B,CAAC;AACtF,OAAO,EAAuB,KAAK,uBAAuB,EAAE,MAAM,yCAAyC,CAAC;AAC5G,OAAO,EAEN,KAAK,+BAA+B,EACpC,KAAK,wBAAwB,EAE7B,MAAM,oCAAoC,CAAC;AAC5C,OAAO,EAAkB,KAAK,eAAe,EAAE,MAAM,iBAAiB,CAAC;AACvE,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,4BAA4B,CAAC;AAChE,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAC;AAGzD,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AAE7D,qGAAqG;AACrG,wBAAgB,oBAAoB,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,MAAM,CAYrE;AAWD,MAAM,WAAW,mBAAmB;IACnC,yEAAyE;IACzE,YAAY,IAAI,MAAM,CAAC;IACvB,+GAA+G;IAC/G,iBAAiB,IAAI,cAAc,CAAC;IACpC,8EAA8E;IAC9E,kBAAkB,IAAI,eAAe,CAAC;IACtC,4EAA4E;IAC5E,gBAAgB,IAAI,aAAa,CAAC;IAClC,mGAAmG;IACnG,WAAW,IAAI,MAAM,CAAC;IACtB,qDAAqD;IACrD,MAAM,IAAI,MAAM,CAAC;IACjB,yGAAuG;IACvG,kBAAkB,IAAI,MAAM,EAAE,CAAC;IAC/B,yEAAyE;IACzE,UAAU,IAAI,OAAO,CAAC;IACtB,gGAA8F;IAC9F,gBAAgB,IAAI,aAAa,CAAC;IAClC,+FAA+F;IAC/F,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,8GAA4G;IAC5G,qBAAqB,CAAC,IAAI,EAAE,yBAAyB,GAAG,OAAO,CAAC,wBAAwB,CAAC,CAAC;CAC1F;AAED,qBAAa,eAAe;IAC3B,OAAO,CAAC,sBAAsB,CAA0C;IACxE;iGAC6F;IAC7F,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAsB;IACpD,OAAO,CAAC,uBAAuB,CAAiC;IAChE,OAAO,CAAC,wBAAwB,CAAiC;IACjE,OAAO,CAAC,kBAAkB,CAAwC;IAClE,OAAO,CAAC,yBAAyB,CAA6C;IAC9E,OAAO,CAAC,yBAAyB,CAAmD;IACpF,OAAO,CAAC,gCAAgC,CAA0D;IAClG,OAAO,CAAC,8BAA8B,CAAkD;IAExF,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAsB;IAE3C,YAAY,IAAI,EAAE,mBAAmB,EAEpC;IAED,OAAO,CAAC,oBAAoB;IAI5B,OAAO,CAAC,iBAAiB;IAIzB;;;;;;;;;;;;;;;OAeG;IACH,oBAAoB,IAAI,aAAa,CAGpC;IAED;;;;OAIG;IACH,iCAAiC,IAAI,IAAI,CAExC;IAED;;;;;;OAMG;IACH,OAAO,CAAC,0BAA0B;IAUlC;;;;;;;;OAQG;IACH,eAAe,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,kBAAkB,CAc5D;IAED;;;;OAIG;IACH,qBAAqB,CAAC,QAAQ,CAAC,EAAE,YAAY,EAAE,GAAG,kBAAkB,CAGnE;IAED;;;;;OAKG;IACH,uBAAuB,CAAC,WAAW,EAAE,kBAAkB,GAAG,wBAAwB,CAUjF;IAED;;;;OAIG;IACH,qBAAqB,CAAC,QAAQ,CAAC,EAAE,YAAY,EAAE,GAAG,wBAAwB,CAGzE;IAED;;;;;OAKG;IACH,kCAAkC,CAAC,QAAQ,EAAE,eAAe,GAAG,IAAI,CAOlE;IAED,iGAAiG;IACjG,4BAA4B,IAAI,+BAA+B,CAE9D;IAED;;;;;;;OAOG;IACH,oBAAoB,CACnB,QAAQ,EAAE,YAAY,EAAE,EACxB,YAAY,EAAE,wBAAwB,GACpC;QAAE,QAAQ,EAAE,YAAY,EAAE,CAAC;QAAC,MAAM,EAAE,uBAAuB,CAAA;KAAE,CAqB/D;IAED;;;;OAIG;IACH,wBAAwB,CAAC,QAAQ,EAAE,YAAY,EAAE,EAAE,YAAY,EAAE,wBAAwB,GAAG,IAAI,CAwB/F;IAED;;;;OAIG;IACH;;;;;OAKG;IACH,yBAAyB,IAAI,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CA+BlD;IAED,uBAAuB,IAAI,IAAI,CAU9B;IAED;;;;;OAKG;IACH,wBAAwB,IAAI,CAAC,CAAC,IAAI,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW,KAAK,OAAO,CAAC,MAAM,CAAC,CAAC,GAAG,SAAS,CA8ChG;YAEa,mBAAmB;IA8DjC,gGAAgG;IAChG,wBAAwB,IAAI;QAC3B,OAAO,EAAE,OAAO,CAAC;QACjB,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,SAAS,EAAE,yBAAyB,CAAC;QACrC,cAAc,CAAC,EAAE,MAAM,CAAC;QACxB,uBAAuB,CAAC,EAAE,MAAM,CAAC;KACjC,CASA;IAED,yFAAyF;IACzF,0BAA0B,IAAI,uBAAuB,CAEpD;IAED,cAAc,CACb,QAAQ,EAAE,YAAY,EAAE,EACxB,aAAa,EAAE,OAAO,GACpB;QAAE,QAAQ,EAAE,YAAY,EAAE,CAAC;QAAC,MAAM,EAAE,eAAe,CAAA;KAAE,CA6DvD;IAED;;;;;;;;;;;;OAYG;IACH,OAAO,CAAC,kCAAkC;IAiB1C,kBAAkB,CAAC,QAAQ,CAAC,EAAE,YAAY,EAAE,GAAG,eAAe,CAY7D;IAED,4BAA4B,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,MAAM,CAgB7D;CACD","sourcesContent":["/**\n * Context pipeline: the session's per-turn context-shaping subsystem — the observe-only context\n * audit, the shadow prompt-policy plan and its correlation with legacy context-gc, the enforcement\n * pilot, the relevance-curation queue + the fitness-gated curation model resolver / brain-curation\n * drain + compaction pre-digest, the legacy context-gc pass and its packed-artifact reference\n * release, the tool-output artifact store, and the current-context token estimate.\n *\n * Extracted verbatim from agent-session.ts (god-file decomposition). Owns the latest\n * audit/policy/correlation/enforcement/gc reports, the {@link BrainCurator} sidecar and its last\n * skip reasons, and the lazily-built tool-artifact store. Everything else it needs — the turn\n * index, the session/settings managers, the model registry, agent/workspace dirs, the active tool\n * names, the disposed flag, the isolated-completion primitive, spawned-usage accounting, and the\n * live {@link MemoryManager} — is reached through narrow deps accessors rather than the whole\n * AgentSession.\n *\n * Context-transform boundary (deliberate): the per-turn stages ({@link estimateCurrentContextTokens},\n * {@link runContextAudit}, {@link runPromptPolicyPlanning}, {@link applyContextGc},\n * {@link correlatePromptPolicyWithContextGc}, {@link runPromptEnforcement},\n * {@link enqueueRelevanceCuration}, {@link maybeDrainBrainCuration}) are invoked from the session's\n * context transform as one-line delegations, so the transform stays the single owner of the pass\n * ordering. This controller reaches {@link MemoryController} functionality only through\n * {@link ContextPipelineDeps.getMemoryManager} (never imports it), and MemoryController never imports\n * the pipeline — keeping the transform the one place the two subsystems meet.\n */\n\nimport { join } from \"node:path\";\nimport type { AgentMessage } from \"@caupulican/pi-agent-core\";\nimport { estimateContextTokens, getLatestCompactionEntry, type SessionManager } from \"@caupulican/pi-agent-core/node\";\nimport type { Api, AssistantMessage, Model, Usage } from \"@caupulican/pi-ai\";\nimport type { IsolatedCompletionOptions, IsolatedCompletionResult } from \"./agent-session.ts\";\nimport { BrainCurator, type CurationTelemetrySnapshot, preDigestConversationText } from \"./context/brain-curator.ts\";\nimport { type ArtifactStore, createFileArtifactStore } from \"./context/context-artifacts.ts\";\nimport { type ContextAuditReport, runContextAudit } from \"./context/context-audit.ts\";\nimport { enforcePromptPolicy, type PromptEnforcementReport } from \"./context/context-prompt-enforcement.ts\";\nimport {\n\tcorrelateWithContextGc,\n\ttype PromptPolicyGcCorrelationReport,\n\ttype PromptPolicyShadowReport,\n\tplanPromptPolicy,\n} from \"./context/context-prompt-policy.ts\";\nimport { applyContextGc, type ContextGcReport } from \"./context-gc.ts\";\nimport type { MemoryManager } from \"./memory/memory-manager.ts\";\nimport type { ModelRegistry } from \"./model-registry.ts\";\nimport { resolveCliModel } from \"./model-resolver.ts\";\nimport { FitnessStore } from \"./models/fitness-store.ts\";\nimport type { SettingsManager } from \"./settings-manager.ts\";\n\n/** Latest user prompt text in the provider-visible array (curation goal line; bounded by caller). */\nexport function latestUserPromptText(messages: AgentMessage[]): string {\n\tfor (let index = messages.length - 1; index >= 0; index--) {\n\t\tconst message = messages[index];\n\t\tif (!message || message.role !== \"user\") continue;\n\t\tif (typeof message.content === \"string\") return message.content;\n\t\tconst text = message.content\n\t\t\t.filter((part): part is { type: \"text\"; text: string } => (part as { type?: string }).type === \"text\")\n\t\t\t.map((part) => part.text)\n\t\t\t.join(\"\\n\");\n\t\tif (text.length > 0) return text;\n\t}\n\treturn \"\";\n}\n\n/** Read a packed grep/find tool result's `details.artifactId`, if present, without `any`. */\nfunction extractArtifactId(message: AgentMessage | undefined): string | undefined {\n\tif (!message || message.role !== \"toolResult\") return undefined;\n\tconst details = (message as { details?: unknown }).details;\n\tif (typeof details !== \"object\" || details === null) return undefined;\n\tconst artifactId = (details as { artifactId?: unknown }).artifactId;\n\treturn typeof artifactId === \"string\" ? artifactId : undefined;\n}\n\nexport interface ContextPipelineDeps {\n\t/** Current turn index, stamped into audit/policy/enforcement reports. */\n\tgetTurnIndex(): number;\n\t/** Session log: audit lookup, gc/artifact storage dirs, curation entries, token-estimate compaction anchor. */\n\tgetSessionManager(): SessionManager;\n\t/** Context-gc / prompt-enforcement / curation settings (all opt-in gates). */\n\tgetSettingsManager(): SettingsManager;\n\t/** Resolves a configured curation model pattern against configured auth. */\n\tgetModelRegistry(): ModelRegistry;\n\t/** Root dir the host-keyed {@link FitnessStore} and per-session gc/artifact storage live under. */\n\tgetAgentDir(): string;\n\t/** Workspace root, passed to the context-gc pass. */\n\tgetCwd(): string;\n\t/** Currently-active tool names — enforcement checks whether artifact_retrieve is a live affordance. */\n\tgetActiveToolNames(): string[];\n\t/** A disposed session must never persist a curation/pre-digest entry. */\n\tisDisposed(): boolean;\n\t/** The live memory manager — the active providers' page markers feed the semantic-gc scan. */\n\tgetMemoryManager(): MemoryManager;\n\t/** Roll a curation drain's spawned usage into session accounting (idempotent per reportId). */\n\taddSpawnedUsage(\n\t\tusage: Usage,\n\t\topts?: { label?: string; sourceSessionId?: string; reportId?: string },\n\t): string | undefined;\n\t/** One-shot LLM call fully isolated from the main session — the curation/pre-digest execution primitive. */\n\trunIsolatedCompletion(opts: IsolatedCompletionOptions): Promise<IsolatedCompletionResult>;\n}\n\nexport class ContextPipeline {\n\tprivate _latestContextGcReport: ContextGcReport | undefined = undefined;\n\t/** Brain-curation sidecar (design: brain-context-curation-design.md). Inert unless the\n\t * contextPolicy.curation setting is enabled AND the model passes the digest fitness gate. */\n\tprivate readonly _brainCurator = new BrainCurator();\n\tprivate _lastCurationSkipReason: string | undefined = undefined;\n\tprivate _lastPreDigestSkipReason: string | undefined = undefined;\n\tprivate _toolArtifactStore: ArtifactStore | undefined = undefined;\n\tprivate _latestContextAuditReport: ContextAuditReport | undefined = undefined;\n\tprivate _latestPromptPolicyReport: PromptPolicyShadowReport | undefined = undefined;\n\tprivate _latestPromptPolicyGcCorrelation: PromptPolicyGcCorrelationReport | undefined = undefined;\n\tprivate _latestPromptEnforcementReport: PromptEnforcementReport | undefined = undefined;\n\n\tprivate readonly deps: ContextPipelineDeps;\n\n\tconstructor(deps: ContextPipelineDeps) {\n\t\tthis.deps = deps;\n\t}\n\n\tprivate _contextGcStorageDir(): string {\n\t\treturn join(this.deps.getAgentDir(), \"context-gc\", this.deps.getSessionManager().getSessionId());\n\t}\n\n\tprivate _toolArtifactsDir(): string {\n\t\treturn join(this.deps.getAgentDir(), \"context-artifacts\", this.deps.getSessionManager().getSessionId());\n\t}\n\n\t/**\n\t * Session-scoped, filesystem-backed artifact store for first-capture-then-bound tool\n\t * output (grep/find only, for now -- see tool-output-artifacts.md). Lazily created and\n\t * cached so every tool construction in this session shares one store instance.\n\t *\n\t * `packToolOutput()` registers a reference (the packing tool call's id) at pack time\n\t * and fails closed, so packed artifacts are never prematurely collected.\n\t * `_releaseGcPackedArtifactReferences()` (called from `applyContextGc()`) releases\n\t * that reference once context-gc packs the result out of live context, and\n\t * opportunistically reclaims now-unreferenced artifacts via `cleanup()`.\n\t * Remaining carry-forward gap: cleanup() now also runs at dispose(), but only reclaims\n\t * already-released (zero-reference) artifacts. A session that ends before context-gc\n\t * ever evicts a result never releases that reference, so its artifact stays on disk by\n\t * design (resolvable on resume). Reclaiming those requires an explicit cross-session\n\t * expiry/liveness policy, not just a sweep.\n\t */\n\tgetToolArtifactStore(): ArtifactStore {\n\t\tthis._toolArtifactStore ??= createFileArtifactStore({ baseDir: this._toolArtifactsDir() });\n\t\treturn this._toolArtifactStore;\n\t}\n\n\t/**\n\t * Best-effort final sweep of any already-released (zero-reference) tool-output artifact at\n\t * session dispose. Reads the field (not the getter) so a session that never packed anything\n\t * doesn't force-create a store/dir just to sweep it.\n\t */\n\tcleanupToolArtifactStoreOnDispose(): void {\n\t\tthis._toolArtifactStore?.cleanup();\n\t}\n\n\t/**\n\t * One pass over the current branch, mapping each toolResult's toolCallId to its\n\t * persisted session-entry id. Rebuilt every audit pass (O(branch) per turn), so this is\n\t * O(n^2) over a long session. Fine at current scale; after the artifact-read fix this is\n\t * the next per-turn audit cost to optimize if it ever matters (e.g. cache/incrementally\n\t * update instead of a full rebuild).\n\t */\n\tprivate _buildSessionEntryIdLookup(): (toolCallId: string) => string | undefined {\n\t\tconst map = new Map<string, string>();\n\t\tfor (const entry of this.deps.getSessionManager().getBranch()) {\n\t\t\tif (entry.type === \"message\" && entry.message.role === \"toolResult\") {\n\t\t\t\tmap.set(entry.message.toolCallId, entry.id);\n\t\t\t}\n\t\t}\n\t\treturn (toolCallId: string) => map.get(toolCallId);\n\t}\n\n\t/**\n\t * Phase 1 observe-only audit pass (see context/context-audit.ts): converts live\n\t * toolResult messages into ContextItems and runs the existing retention/hard-constraint\n\t * evaluators over them, storing the latest deterministic report for tests/debugging.\n\t * Read-only with respect to messages, the transcript, and artifact references -- uses\n\t * `_toolArtifactStore` (the field), not `getToolArtifactStore()` (the getter), so a\n\t * session that never packed anything doesn't force-create a store/dir just to audit.\n\t * Never throws into a live turn: any failure degrades to an empty report.\n\t */\n\trunContextAudit(messages: AgentMessage[]): ContextAuditReport {\n\t\ttry {\n\t\t\tconst report = runContextAudit(messages, {\n\t\t\t\tturnIndex: this.deps.getTurnIndex(),\n\t\t\t\tartifactStore: this._toolArtifactStore,\n\t\t\t\tsessionEntryIdForToolCallId: this._buildSessionEntryIdLookup(),\n\t\t\t});\n\t\t\tthis._latestContextAuditReport = report;\n\t\t\treturn report;\n\t\t} catch {\n\t\t\tconst report: ContextAuditReport = { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t\t\tthis._latestContextAuditReport = report;\n\t\t\treturn report;\n\t\t}\n\t}\n\n\t/**\n\t * Read-only inspection of the context audit. With `messages`, recomputes fresh against\n\t * the given array (still no mutation of messages/transcript/artifact refs); without,\n\t * returns the last report computed during a real transform pass.\n\t */\n\tgetContextAuditReport(messages?: AgentMessage[]): ContextAuditReport {\n\t\tif (messages) return this.runContextAudit(messages);\n\t\treturn this._latestContextAuditReport ?? { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t}\n\n\t/**\n\t * Observe-first shadow/planning pass (see context/context-prompt-policy.ts): re-shapes\n\t * the audit report into a per-item policy plan whose `appliedAction` is always\n\t * \"keep_raw\" -- this never enforces anything, it only records what the policy engine\n\t * would say. Never throws into a live turn: any failure degrades to an empty report.\n\t */\n\trunPromptPolicyPlanning(auditReport: ContextAuditReport): PromptPolicyShadowReport {\n\t\ttry {\n\t\t\tconst report = planPromptPolicy(auditReport);\n\t\t\tthis._latestPromptPolicyReport = report;\n\t\t\treturn report;\n\t\t} catch {\n\t\t\tconst report: PromptPolicyShadowReport = { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t\t\tthis._latestPromptPolicyReport = report;\n\t\t\treturn report;\n\t\t}\n\t}\n\n\t/**\n\t * Read-only inspection of the shadow policy plan. With `messages`, recomputes fresh\n\t * (audit + plan) against the given array; without, returns the last plan computed\n\t * during a real transform pass. Never mutates messages/transcript/artifact refs.\n\t */\n\tgetPromptPolicyReport(messages?: AgentMessage[]): PromptPolicyShadowReport {\n\t\tif (messages) return this.runPromptPolicyPlanning(this.runContextAudit(messages));\n\t\treturn this._latestPromptPolicyReport ?? { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t}\n\n\t/**\n\t * Report-only correlation between the shadow plan just computed this turn and what the\n\t * legacy context-gc pass actually packed. Runs after `applyContextGc()` has already\n\t * produced its report; never influences context-gc itself. Never throws into a live\n\t * turn: any failure degrades to an empty correlation.\n\t */\n\tcorrelatePromptPolicyWithContextGc(gcReport: ContextGcReport): void {\n\t\tconst shadowReport = this._latestPromptPolicyReport ?? { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t\ttry {\n\t\t\tthis._latestPromptPolicyGcCorrelation = correlateWithContextGc(shadowReport, gcReport);\n\t\t} catch {\n\t\t\tthis._latestPromptPolicyGcCorrelation = { turnIndex: this.deps.getTurnIndex(), entries: [] };\n\t\t}\n\t}\n\n\t/** Read-only inspection of the latest shadow-plan/legacy-gc correlation, for tests/debugging. */\n\tgetPromptPolicyGcCorrelation(): PromptPolicyGcCorrelationReport {\n\t\treturn this._latestPromptPolicyGcCorrelation ?? { turnIndex: this.deps.getTurnIndex(), entries: [] };\n\t}\n\n\t/**\n\t * First enforcement pilot (see context/context-prompt-enforcement.ts): opt-in,\n\t * default-disabled stub-in-place of stale artifact-backed tool_output results in the\n\t * provider-visible message array only. Runs on `messages` AFTER context-gc has already\n\t * produced its own result, so legacy context-gc's own packing/reporting is completely\n\t * unaffected by this pass -- it only ever acts on messages gc left untouched this turn.\n\t * Never throws into a live turn: any failure degrades to returning `messages` unchanged.\n\t */\n\trunPromptEnforcement(\n\t\tmessages: AgentMessage[],\n\t\tshadowReport: PromptPolicyShadowReport,\n\t): { messages: AgentMessage[]; report: PromptEnforcementReport } {\n\t\ttry {\n\t\t\tconst persistedSettings = this.deps.getSettingsManager().getContextPromptEnforcementSettings();\n\t\t\tconst curationEnabled = this.deps.getSettingsManager().getContextCurationSettings().enabled;\n\t\t\tconst settings = {\n\t\t\t\t...persistedSettings,\n\t\t\t\t// Runtime fact, never assumed: artifact_retrieve is a companion affordance\n\t\t\t\t// (auto-activated alongside grep/find), not a default/global tool, so active\n\t\t\t\t// tools can differ turn to turn -- see context-prompt-enforcement.ts's doc\n\t\t\t\t// comment on why this is checked separately from hasAvailableRetrievalPath.\n\t\t\t\tretrievalToolAvailable: this.deps.getActiveToolNames().includes(\"artifact_retrieve\"),\n\t\t\t\tbrainRelevance: curationEnabled ? (itemId: string) => this._brainCurator.getRelevance(itemId) : undefined,\n\t\t\t};\n\t\t\tconst result = enforcePromptPolicy(messages, shadowReport, settings);\n\t\t\tthis._latestPromptEnforcementReport = result.report;\n\t\t\treturn result;\n\t\t} catch {\n\t\t\tconst report: PromptEnforcementReport = { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t\t\tthis._latestPromptEnforcementReport = report;\n\t\t\treturn { messages, report };\n\t\t}\n\t}\n\n\t/**\n\t * Enqueue relevance-scoring jobs for stale, artifact-backed tool outputs the enforcement\n\t * pilot could act on. Pure queueing — the verdicts only ever take effect through the\n\t * asymmetric advisory lever inside enforcePromptPolicy. Never throws into a turn.\n\t */\n\tenqueueRelevanceCuration(messages: AgentMessage[], shadowReport: PromptPolicyShadowReport): void {\n\t\ttry {\n\t\t\tconst settings = this.deps.getSettingsManager().getContextCurationSettings();\n\t\t\tif (!settings.enabled) return;\n\t\t\tconst goal = latestUserPromptText(messages).slice(0, 400);\n\t\t\tfor (const item of shadowReport.items) {\n\t\t\t\tif (!item.hasAvailableRetrievalPath) continue;\n\t\t\t\tconst message = messages[item.messageIndex];\n\t\t\t\tif (!message || message.role !== \"toolResult\" || message.toolCallId !== item.toolCallId) continue;\n\t\t\t\tif (message.isError) continue;\n\t\t\t\tconst details = message.details as\n\t\t\t\t\t| { contextGc?: { packed?: unknown }; promptPolicy?: { enforced?: unknown } }\n\t\t\t\t\t| undefined;\n\t\t\t\tif (details?.contextGc?.packed === true || details?.promptPolicy?.enforced === true) continue;\n\t\t\t\tconst text = message.content\n\t\t\t\t\t.filter((part): part is { type: \"text\"; text: string } => part.type === \"text\")\n\t\t\t\t\t.map((part) => part.text)\n\t\t\t\t\t.join(\"\\n\");\n\t\t\t\tif (text.length === 0) continue;\n\t\t\t\tthis._brainCurator.enqueue({ kind: \"relevance\", key: item.itemId, content: text.slice(0, 4000), goal });\n\t\t\t}\n\t\t} catch {\n\t\t\t// curation is a sidecar; it must never disrupt a turn\n\t\t}\n\t}\n\n\t/**\n\t * Drain gate: settings on, model configured+authed, and the model has PASSED the digest\n\t * fitness probe on THIS host (design: unfit or unprobed models are refused with a visible\n\t * reason, never silently degraded). Fire-and-forget; never throws into a turn.\n\t */\n\t/**\n\t * Resolve the curation model IFF every gate passes: setting enabled, model configured,\n\t * resolvable+authed, and digest-fitness-proven on THIS host (canonical \"provider/id\" ref —\n\t * runModelFitness stores reports under it, while settings.model may be a bare id or pattern).\n\t * Sets _lastCurationSkipReason on refusal; never throws.\n\t */\n\tresolveCurationModelIfFit(): Model<Api> | undefined {\n\t\tconst settings = this.deps.getSettingsManager().getContextCurationSettings();\n\t\tif (!settings.enabled) {\n\t\t\t// Never surface a stale refusal reason for a feature the user has since disabled.\n\t\t\tthis._lastCurationSkipReason = undefined;\n\t\t\treturn undefined;\n\t\t}\n\t\tif (!settings.model) {\n\t\t\tthis._lastCurationSkipReason = \"curation_model_unset\";\n\t\t\treturn undefined;\n\t\t}\n\t\tconst resolved = resolveCliModel({ cliModel: settings.model, modelRegistry: this.deps.getModelRegistry() });\n\t\tif (!resolved.model || !this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) {\n\t\t\tthis._lastCurationSkipReason = \"curation_model_unresolved\";\n\t\t\treturn undefined;\n\t\t}\n\t\tconst canonicalRef = `${resolved.model.provider}/${resolved.model.id}`;\n\t\tconst fitness = FitnessStore.forAgentDir(this.deps.getAgentDir())\n\t\t\t.getForHost()\n\t\t\t.find((entry) => entry.model === canonicalRef);\n\t\tconst digestScore = fitness?.report.digest;\n\t\tif (!digestScore) {\n\t\t\tthis._lastCurationSkipReason = \"curation_model_unprobed\";\n\t\t\treturn undefined;\n\t\t}\n\t\tif (digestScore.succeeded < Math.ceil(digestScore.total * (2 / 3))) {\n\t\t\tthis._lastCurationSkipReason = \"curation_model_digest_unfit\";\n\t\t\treturn undefined;\n\t\t}\n\t\tthis._lastCurationSkipReason = undefined;\n\t\treturn resolved.model;\n\t}\n\n\tmaybeDrainBrainCuration(): void {\n\t\ttry {\n\t\t\tif (!this._brainCurator.hasWork() || this._brainCurator.isDraining) return;\n\t\t\tconst model = this.resolveCurationModelIfFit();\n\t\t\tif (!model) return;\n\t\t\tconst settings = this.deps.getSettingsManager().getContextCurationSettings();\n\t\t\tvoid this._drainBrainCuration(model, settings.maxJobsPerTurn);\n\t\t} catch {\n\t\t\t// curation is a sidecar; it must never disrupt a turn\n\t\t}\n\t}\n\n\t/**\n\t * Compaction pre-digest gate (design surface 3): everything the drain gate requires PLUS a\n\t * RUNTIME reliability proof — the curator must have run >=5 jobs on this session with a parse\n\t * failure rate <=5% before it is trusted to pre-digest compaction input. Returns undefined\n\t * (verbatim compaction, byte-for-byte today's behavior) whenever any gate refuses.\n\t */\n\tbuildCompactionPreDigest(): ((text: string, signal?: AbortSignal) => Promise<string>) | undefined {\n\t\ttry {\n\t\t\tconst model = this.resolveCurationModelIfFit();\n\t\t\tif (!model) return undefined;\n\t\t\tconst telemetry = this._brainCurator.telemetry();\n\t\t\tif (telemetry.jobsRun < 5 || telemetry.parseFailures / telemetry.jobsRun > 0.05) {\n\t\t\t\tthis._lastPreDigestSkipReason = \"curation_predigest_reliability_unproven\";\n\t\t\t\treturn undefined;\n\t\t\t}\n\t\t\tthis._lastPreDigestSkipReason = undefined;\n\t\t\treturn async (text, signal) => {\n\t\t\t\tconst result = await preDigestConversationText({\n\t\t\t\t\ttext,\n\t\t\t\t\tsignal,\n\t\t\t\t\tcomplete: async ({ systemPrompt, userPrompt, signal: chunkSignal }) => {\n\t\t\t\t\t\tconst completion = await this.deps.runIsolatedCompletion({\n\t\t\t\t\t\t\tsystemPrompt,\n\t\t\t\t\t\t\tmessages: [{ role: \"user\", content: [{ type: \"text\", text: userPrompt }], timestamp: Date.now() }],\n\t\t\t\t\t\t\tmodel,\n\t\t\t\t\t\t\tthinkingLevel: \"off\",\n\t\t\t\t\t\t\tmaxTokens: 512,\n\t\t\t\t\t\t\tsignal: chunkSignal,\n\t\t\t\t\t\t\tcacheRetention: \"short\",\n\t\t\t\t\t\t});\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttext: completion.text,\n\t\t\t\t\t\t\tcostUsd: completion.usage.cost.total,\n\t\t\t\t\t\t\tstopReason: String(completion.stopReason),\n\t\t\t\t\t\t};\n\t\t\t\t\t},\n\t\t\t\t});\n\t\t\t\tif (!this.deps.isDisposed() && result.totalChunks > 0) {\n\t\t\t\t\tthis.deps.getSessionManager().appendCustomEntry(\"brain-curation-predigest\", {\n\t\t\t\t\t\tversion: 1,\n\t\t\t\t\t\ttotalChunks: result.totalChunks,\n\t\t\t\t\t\tdigested: result.digested,\n\t\t\t\t\t\tfailed: result.failed,\n\t\t\t\t\t\tcharsBefore: text.length,\n\t\t\t\t\t\tcharsAfter: result.text.length,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t\treturn result.text;\n\t\t\t};\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate async _drainBrainCuration(model: Model<Api>, maxJobs: number): Promise<void> {\n\t\ttry {\n\t\t\t// ACCUMULATE across all drained jobs (the drain runs the completer once PER job) —\n\t\t\t// keeping only the last job's usage would under-report every multi-job drain.\n\t\t\tlet spentUsage: AssistantMessage[\"usage\"] | undefined;\n\t\t\tconst results = await this._brainCurator.drain({\n\t\t\t\tmaxJobs,\n\t\t\t\tcomplete: async ({ systemPrompt, userPrompt, signal }) => {\n\t\t\t\t\tconst completion = await this.deps.runIsolatedCompletion({\n\t\t\t\t\t\tsystemPrompt,\n\t\t\t\t\t\tmessages: [{ role: \"user\", content: [{ type: \"text\", text: userPrompt }], timestamp: Date.now() }],\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t\tthinkingLevel: \"off\",\n\t\t\t\t\t\tmaxTokens: 256,\n\t\t\t\t\t\tsignal,\n\t\t\t\t\t\t// Both curation system prompts are static — the provider can cache the prefix.\n\t\t\t\t\t\tcacheRetention: \"short\",\n\t\t\t\t\t});\n\t\t\t\t\tconst usage = completion.usage;\n\t\t\t\t\tif (!spentUsage) {\n\t\t\t\t\t\tspentUsage = structuredClone(usage);\n\t\t\t\t\t} else {\n\t\t\t\t\t\tspentUsage.input += usage.input;\n\t\t\t\t\t\tspentUsage.output += usage.output;\n\t\t\t\t\t\tspentUsage.cacheRead += usage.cacheRead;\n\t\t\t\t\t\tspentUsage.cacheWrite += usage.cacheWrite;\n\t\t\t\t\t\tspentUsage.totalTokens += usage.totalTokens;\n\t\t\t\t\t\tspentUsage.cost.input += usage.cost.input;\n\t\t\t\t\t\tspentUsage.cost.output += usage.cost.output;\n\t\t\t\t\t\tspentUsage.cost.cacheRead += usage.cost.cacheRead;\n\t\t\t\t\t\tspentUsage.cost.cacheWrite += usage.cost.cacheWrite;\n\t\t\t\t\t\tspentUsage.cost.total += usage.cost.total;\n\t\t\t\t\t}\n\t\t\t\t\treturn {\n\t\t\t\t\t\ttext: completion.text,\n\t\t\t\t\t\tcostUsd: completion.usage.cost.total,\n\t\t\t\t\t\tstopReason: String(completion.stopReason),\n\t\t\t\t\t};\n\t\t\t\t},\n\t\t\t});\n\t\t\t// Honest accounting even for free local models: token visibility is the contract.\n\t\t\tif (spentUsage && (spentUsage.cost.total > 0 || spentUsage.totalTokens > 0)) {\n\t\t\t\tthis.deps.addSpawnedUsage(spentUsage, { label: \"context-curator\" });\n\t\t\t}\n\t\t\tif (this.deps.isDisposed() || results.length === 0) return;\n\t\t\tthis.deps.getSessionManager().appendCustomEntry(\"brain-curation\", {\n\t\t\t\tversion: 1,\n\t\t\t\tresults: results.map((result) => ({\n\t\t\t\t\tkey: result.key,\n\t\t\t\t\tkind: result.kind,\n\t\t\t\t\tok: result.ok,\n\t\t\t\t\tms: result.ms,\n\t\t\t\t\t...(result.digest !== undefined ? { digest: result.digest } : {}),\n\t\t\t\t\t...(result.relevant !== undefined ? { relevant: result.relevant, confidence: result.confidence } : {}),\n\t\t\t\t})),\n\t\t\t\ttelemetry: this._brainCurator.telemetry(),\n\t\t\t});\n\t\t} catch {\n\t\t\t// curation is a sidecar; it must never disrupt a turn\n\t\t}\n\t}\n\n\t/** Curation status for diagnostics/dashboard: settings, live telemetry, last refusal reason. */\n\tgetContextCurationStatus(): {\n\t\tenabled: boolean;\n\t\tmodel?: string;\n\t\ttelemetry: CurationTelemetrySnapshot;\n\t\tlastSkipReason?: string;\n\t\tlastPreDigestSkipReason?: string;\n\t} {\n\t\tconst settings = this.deps.getSettingsManager().getContextCurationSettings();\n\t\treturn {\n\t\t\tenabled: settings.enabled,\n\t\t\tmodel: settings.model,\n\t\t\ttelemetry: this._brainCurator.telemetry(),\n\t\t\tlastSkipReason: this._lastCurationSkipReason,\n\t\t\tlastPreDigestSkipReason: this._lastPreDigestSkipReason,\n\t\t};\n\t}\n\n\t/** Read-only inspection of the latest prompt-enforcement report, for tests/debugging. */\n\tgetPromptEnforcementReport(): PromptEnforcementReport {\n\t\treturn this._latestPromptEnforcementReport ?? { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t}\n\n\tapplyContextGc(\n\t\tmessages: AgentMessage[],\n\t\twritePayloads: boolean,\n\t): { messages: AgentMessage[]; report: ContextGcReport } {\n\t\ttry {\n\t\t\tconst settings = this.deps.getSettingsManager().getContextGcSettings();\n\t\t\t// Merge the ACTIVE memory providers' own page markers (e.g. transcript-recall's\n\t\t\t// \"<memory_context\") into the semantic-memory marker list. The settings default is\n\t\t\t// provider-agnostic and non-empty, so without this merge the recall pages the bundled\n\t\t\t// default provider actually emits are never recognized as semantic-memory pages and\n\t\t\t// accumulate raw for the life of the session — the exact growth Bug #7 GC exists to stop.\n\t\t\tconst providerMarkers = this.deps.getMemoryManager().getContextMarkers();\n\t\t\tconst curationSettings = this.deps.getSettingsManager().getContextCurationSettings();\n\t\t\tconst result = applyContextGc(messages, {\n\t\t\t\t...settings,\n\t\t\t\tsemanticMemory: {\n\t\t\t\t\t...settings.semanticMemory,\n\t\t\t\t\tmarkers: [...new Set([...settings.semanticMemory.markers, ...providerMarkers])],\n\t\t\t\t},\n\t\t\t\tcwd: this.deps.getCwd(),\n\t\t\t\tstorageDir: this._contextGcStorageDir(),\n\t\t\t\twritePayloads,\n\t\t\t\tcuration: curationSettings.enabled\n\t\t\t\t\t? {\n\t\t\t\t\t\t\tresolveDigest: (digestKey) => {\n\t\t\t\t\t\t\t\tconst digest = this._brainCurator.getDigest(digestKey);\n\t\t\t\t\t\t\t\t// Count serves on the REAL per-turn pass only, never the report path.\n\t\t\t\t\t\t\t\tif (digest !== undefined && writePayloads) this._brainCurator.noteDigestServed();\n\t\t\t\t\t\t\t\treturn digest;\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t// Only the real per-turn pass enqueues work; the read-only report path\n\t\t\t\t\t\t\t// (writePayloads=false) stays side-effect free.\n\t\t\t\t\t\t\tonPacked: writePayloads\n\t\t\t\t\t\t\t\t? (record, originalText) => {\n\t\t\t\t\t\t\t\t\t\tthis._brainCurator.enqueue({\n\t\t\t\t\t\t\t\t\t\t\tkind: \"stub_digest\",\n\t\t\t\t\t\t\t\t\t\t\tkey: record.key ?? record.toolCallId,\n\t\t\t\t\t\t\t\t\t\t\tcontent: originalText,\n\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t: undefined,\n\t\t\t\t\t\t}\n\t\t\t\t\t: undefined,\n\t\t\t});\n\t\t\tthis._latestContextGcReport = result.report;\n\t\t\t// Only release/reclaim on the real per-turn pass (writePayloads=true), never on\n\t\t\t// the read-only status-report path (getContextGcReport with writePayloads=false),\n\t\t\t// so merely inspecting the report can't have side effects.\n\t\t\tif (writePayloads && result.report.packedCount > 0) {\n\t\t\t\tthis._releaseGcPackedArtifactReferences(messages, result.report);\n\t\t\t}\n\t\t\treturn result;\n\t\t} catch {\n\t\t\tconst report: ContextGcReport = {\n\t\t\t\tenabled: false,\n\t\t\t\tpackedCount: 0,\n\t\t\t\toriginalTokens: 0,\n\t\t\t\tpackedTokens: 0,\n\t\t\t\tsavedTokens: 0,\n\t\t\t\trecords: [],\n\t\t\t};\n\t\t\tthis._latestContextGcReport = report;\n\t\t\treturn { messages, report };\n\t\t}\n\t}\n\n\t/**\n\t * Reference-release + cleanup lifecycle: once context-gc has packed a grep/find tool\n\t * result out of the live prompt (the message is no longer current/active working\n\t * context -- see contracts-and-retention.md's \"ephemeral\"/\"expired\" retention\n\t * classes), release the pack-time reference `packToolOutput()` registered for it, and\n\t * opportunistically reclaim now-unreferenced artifacts. This is the other half of the\n\t * D2b-1 gate: artifacts were being registered but never released, so they accumulated\n\t * for the life of the session.\n\t *\n\t * `record.toolCallId` (from context-gc's packed record) is exactly the holder id\n\t * `packToolOutput()` used when it called `addReference()` -- both trace back to the\n\t * same tool call's id -- so no separate bookkeeping is needed to find it.\n\t */\n\tprivate _releaseGcPackedArtifactReferences(messages: AgentMessage[], report: ContextGcReport): void {\n\t\tconst store = this._toolArtifactStore;\n\t\tif (!store) return; // no store was ever constructed, so nothing could have been packed to one\n\n\t\tlet releasedAny = false;\n\t\tfor (const record of report.records) {\n\t\t\tif (record.toolName !== \"grep\" && record.toolName !== \"find\") continue;\n\t\t\tconst artifactId = extractArtifactId(messages[record.messageIndex]);\n\t\t\tif (!artifactId) continue;\n\t\t\tif (store.removeReference(artifactId, record.toolCallId)) releasedAny = true;\n\t\t}\n\t\t// Cleanup only runs immediately after a release actually happened in this pass, so\n\t\t// a long session doesn't re-scan the artifact directory on every turn once nothing\n\t\t// new became eligible for release.\n\t\tif (releasedAny) store.cleanup();\n\t}\n\n\tgetContextGcReport(messages?: AgentMessage[]): ContextGcReport {\n\t\tif (messages) return this.applyContextGc(messages, false).report;\n\t\treturn (\n\t\t\tthis._latestContextGcReport ?? {\n\t\t\t\tenabled: this.deps.getSettingsManager().getContextGcSettings().enabled,\n\t\t\t\tpackedCount: 0,\n\t\t\t\toriginalTokens: 0,\n\t\t\t\tpackedTokens: 0,\n\t\t\t\tsavedTokens: 0,\n\t\t\t\trecords: [],\n\t\t\t}\n\t\t);\n\t}\n\n\testimateCurrentContextTokens(messages: AgentMessage[]): number {\n\t\tconst estimate = estimateContextTokens(messages);\n\t\tconst compactionEntry = getLatestCompactionEntry(this.deps.getSessionManager().getBranch());\n\t\tif (estimate.lastUsageIndex === null || !compactionEntry) {\n\t\t\treturn estimate.tokens;\n\t\t}\n\t\tconst usageMessage = messages[estimate.lastUsageIndex];\n\t\tif (usageMessage?.role !== \"assistant\") {\n\t\t\treturn estimate.tokens;\n\t\t}\n\t\tconst usageTimestamp = (usageMessage as AssistantMessage).timestamp;\n\t\tconst compactionTimestamp = new Date(compactionEntry.timestamp).getTime();\n\t\tif (usageTimestamp <= compactionTimestamp) {\n\t\t\treturn estimate.trailingTokens;\n\t\t}\n\t\treturn estimate.tokens;\n\t}\n}\n"]}
+{"version":3,"file":"context-pipeline.d.ts","sourceRoot":"","sources":["../../src/core/context-pipeline.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AAGH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,2BAA2B,CAAC;AAC9D,OAAO,EAGN,KAAK,cAAc,EAEnB,MAAM,gCAAgC,CAAC;AACxC,OAAO,KAAK,EAAE,GAAG,EAAoB,KAAK,EAAE,KAAK,EAAE,MAAM,mBAAmB,CAAC;AAC7E,OAAO,KAAK,EAAE,yBAAyB,EAAE,wBAAwB,EAAE,MAAM,oBAAoB,CAAC;AAC9F,OAAO,EAAgB,KAAK,yBAAyB,EAA6B,MAAM,4BAA4B,CAAC;AACrH,OAAO,EAAE,KAAK,aAAa,EAA2B,MAAM,gCAAgC,CAAC;AAC7F,OAAO,EAAE,KAAK,kBAAkB,EAAmB,MAAM,4BAA4B,CAAC;AACtF,OAAO,EAAuB,KAAK,uBAAuB,EAAE,MAAM,yCAAyC,CAAC;AAC5G,OAAO,EAEN,KAAK,+BAA+B,EACpC,KAAK,wBAAwB,EAE7B,MAAM,oCAAoC,CAAC;AAC5C,OAAO,EAAkB,KAAK,eAAe,EAAE,MAAM,iBAAiB,CAAC;AACvE,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,4BAA4B,CAAC;AAChE,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAC;AAIzD,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AAE7D,qGAAqG;AACrG,wBAAgB,oBAAoB,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,MAAM,CAYrE;AAWD,MAAM,WAAW,mBAAmB;IACnC,yEAAyE;IACzE,YAAY,IAAI,MAAM,CAAC;IACvB,+GAA+G;IAC/G,iBAAiB,IAAI,cAAc,CAAC;IACpC,8EAA8E;IAC9E,kBAAkB,IAAI,eAAe,CAAC;IACtC,4EAA4E;IAC5E,gBAAgB,IAAI,aAAa,CAAC;IAClC,mGAAmG;IACnG,WAAW,IAAI,MAAM,CAAC;IACtB,qDAAqD;IACrD,MAAM,IAAI,MAAM,CAAC;IACjB,yGAAuG;IACvG,kBAAkB,IAAI,MAAM,EAAE,CAAC;IAC/B,yEAAyE;IACzE,UAAU,IAAI,OAAO,CAAC;IACtB,gGAA8F;IAC9F,gBAAgB,IAAI,aAAa,CAAC;IAClC,+FAA+F;IAC/F,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,8GAA4G;IAC5G,qBAAqB,CAAC,IAAI,EAAE,yBAAyB,GAAG,OAAO,CAAC,wBAAwB,CAAC,CAAC;CAC1F;AAED,qBAAa,eAAe;IAC3B,OAAO,CAAC,sBAAsB,CAA0C;IACxE;iGAC6F;IAC7F,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAsB;IACpD,OAAO,CAAC,uBAAuB,CAAiC;IAChE,OAAO,CAAC,wBAAwB,CAAiC;IACjE,OAAO,CAAC,kBAAkB,CAAwC;IAClE,OAAO,CAAC,yBAAyB,CAA6C;IAC9E,OAAO,CAAC,yBAAyB,CAAmD;IACpF,OAAO,CAAC,gCAAgC,CAA0D;IAClG,OAAO,CAAC,8BAA8B,CAAkD;IACxF,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAqB;IAElD,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAsB;IAE3C,YAAY,IAAI,EAAE,mBAAmB,EAEpC;IAED,OAAO,CAAC,oBAAoB;IAI5B,OAAO,CAAC,iBAAiB;IAIzB;;;;;;;;;;;;;;;OAeG;IACH,oBAAoB,IAAI,aAAa,CAGpC;IAED;;;;OAIG;IACH,iCAAiC,IAAI,IAAI,CAExC;IAED;;;;;;OAMG;IACH,OAAO,CAAC,0BAA0B;IAUlC;;;;;;;;OAQG;IACH,eAAe,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,kBAAkB,CAc5D;IAED;;;;OAIG;IACH,qBAAqB,CAAC,QAAQ,CAAC,EAAE,YAAY,EAAE,GAAG,kBAAkB,CAGnE;IAED;;;;;OAKG;IACH,uBAAuB,CAAC,WAAW,EAAE,kBAAkB,GAAG,wBAAwB,CAUjF;IAED;;;;OAIG;IACH,qBAAqB,CAAC,QAAQ,CAAC,EAAE,YAAY,EAAE,GAAG,wBAAwB,CAGzE;IAED;;;;;OAKG;IACH,kCAAkC,CAAC,QAAQ,EAAE,eAAe,GAAG,IAAI,CAOlE;IAED,iGAAiG;IACjG,4BAA4B,IAAI,+BAA+B,CAE9D;IAED;;;;;;;OAOG;IACH,oBAAoB,CACnB,QAAQ,EAAE,YAAY,EAAE,EACxB,YAAY,EAAE,wBAAwB,GACpC;QAAE,QAAQ,EAAE,YAAY,EAAE,CAAC;QAAC,MAAM,EAAE,uBAAuB,CAAA;KAAE,CAqB/D;IAED;;;;OAIG;IACH,wBAAwB,CAAC,QAAQ,EAAE,YAAY,EAAE,EAAE,YAAY,EAAE,wBAAwB,GAAG,IAAI,CAwB/F;IAED;;;;OAIG;IACH;;;;;OAKG;IACH,yBAAyB,IAAI,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CA4BlD;IAED,uBAAuB,IAAI,IAAI,CAU9B;IAED;;;;;OAKG;IACH,wBAAwB,IAAI,CAAC,CAAC,IAAI,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW,KAAK,OAAO,CAAC,MAAM,CAAC,CAAC,GAAG,SAAS,CA8ChG;YAEa,mBAAmB;IA8DjC,gGAAgG;IAChG,wBAAwB,IAAI;QAC3B,OAAO,EAAE,OAAO,CAAC;QACjB,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,SAAS,EAAE,yBAAyB,CAAC;QACrC,cAAc,CAAC,EAAE,MAAM,CAAC;QACxB,uBAAuB,CAAC,EAAE,MAAM,CAAC;KACjC,CASA;IAED,yFAAyF;IACzF,0BAA0B,IAAI,uBAAuB,CAEpD;IAED,cAAc,CACb,QAAQ,EAAE,YAAY,EAAE,EACxB,aAAa,EAAE,OAAO,GACpB;QAAE,QAAQ,EAAE,YAAY,EAAE,CAAC;QAAC,MAAM,EAAE,eAAe,CAAA;KAAE,CA6DvD;IAED;;;;;;;;;;;;OAYG;IACH,OAAO,CAAC,kCAAkC;IAiB1C,kBAAkB,CAAC,QAAQ,CAAC,EAAE,YAAY,EAAE,GAAG,eAAe,CAY7D;IAED,4BAA4B,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,MAAM,CAqB7D;CACD","sourcesContent":["/**\n * Context pipeline: the session's per-turn context-shaping subsystem — the observe-only context\n * audit, the shadow prompt-policy plan and its correlation with legacy context-gc, the enforcement\n * pilot, the relevance-curation queue + the fitness-gated curation model resolver / brain-curation\n * drain + compaction pre-digest, the legacy context-gc pass and its packed-artifact reference\n * release, the tool-output artifact store, and the current-context token estimate.\n *\n * Extracted verbatim from agent-session.ts (god-file decomposition). Owns the latest\n * audit/policy/correlation/enforcement/gc reports, the {@link BrainCurator} sidecar and its last\n * skip reasons, and the lazily-built tool-artifact store. Everything else it needs — the turn\n * index, the session/settings managers, the model registry, agent/workspace dirs, the active tool\n * names, the disposed flag, the isolated-completion primitive, spawned-usage accounting, and the\n * live {@link MemoryManager} — is reached through narrow deps accessors rather than the whole\n * AgentSession.\n *\n * Context-transform boundary (deliberate): the per-turn stages ({@link estimateCurrentContextTokens},\n * {@link runContextAudit}, {@link runPromptPolicyPlanning}, {@link applyContextGc},\n * {@link correlatePromptPolicyWithContextGc}, {@link runPromptEnforcement},\n * {@link enqueueRelevanceCuration}, {@link maybeDrainBrainCuration}) are invoked from the session's\n * context transform as one-line delegations, so the transform stays the single owner of the pass\n * ordering. This controller reaches {@link MemoryController} functionality only through\n * {@link ContextPipelineDeps.getMemoryManager} (never imports it), and MemoryController never imports\n * the pipeline — keeping the transform the one place the two subsystems meet.\n */\n\nimport { join } from \"node:path\";\nimport type { AgentMessage } from \"@caupulican/pi-agent-core\";\nimport {\n\testimateContextTokens,\n\tgetLatestCompactionEntry,\n\ttype SessionManager,\n\tTokenBudget,\n} from \"@caupulican/pi-agent-core/node\";\nimport type { Api, AssistantMessage, Model, Usage } from \"@caupulican/pi-ai\";\nimport type { IsolatedCompletionOptions, IsolatedCompletionResult } from \"./agent-session.ts\";\nimport { BrainCurator, type CurationTelemetrySnapshot, preDigestConversationText } from \"./context/brain-curator.ts\";\nimport { type ArtifactStore, createFileArtifactStore } from \"./context/context-artifacts.ts\";\nimport { type ContextAuditReport, runContextAudit } from \"./context/context-audit.ts\";\nimport { enforcePromptPolicy, type PromptEnforcementReport } from \"./context/context-prompt-enforcement.ts\";\nimport {\n\tcorrelateWithContextGc,\n\ttype PromptPolicyGcCorrelationReport,\n\ttype PromptPolicyShadowReport,\n\tplanPromptPolicy,\n} from \"./context/context-prompt-policy.ts\";\nimport { applyContextGc, type ContextGcReport } from \"./context-gc.ts\";\nimport type { MemoryManager } from \"./memory/memory-manager.ts\";\nimport type { ModelRegistry } from \"./model-registry.ts\";\nimport { resolveCliModel } from \"./model-resolver.ts\";\nimport { evaluateSurfaceFitness } from \"./model-router/fitness-gate.ts\";\nimport { FitnessStore } from \"./models/fitness-store.ts\";\nimport type { SettingsManager } from \"./settings-manager.ts\";\n\n/** Latest user prompt text in the provider-visible array (curation goal line; bounded by caller). */\nexport function latestUserPromptText(messages: AgentMessage[]): string {\n\tfor (let index = messages.length - 1; index >= 0; index--) {\n\t\tconst message = messages[index];\n\t\tif (!message || message.role !== \"user\") continue;\n\t\tif (typeof message.content === \"string\") return message.content;\n\t\tconst text = message.content\n\t\t\t.filter((part): part is { type: \"text\"; text: string } => (part as { type?: string }).type === \"text\")\n\t\t\t.map((part) => part.text)\n\t\t\t.join(\"\\n\");\n\t\tif (text.length > 0) return text;\n\t}\n\treturn \"\";\n}\n\n/** Read a packed grep/find tool result's `details.artifactId`, if present, without `any`. */\nfunction extractArtifactId(message: AgentMessage | undefined): string | undefined {\n\tif (!message || message.role !== \"toolResult\") return undefined;\n\tconst details = (message as { details?: unknown }).details;\n\tif (typeof details !== \"object\" || details === null) return undefined;\n\tconst artifactId = (details as { artifactId?: unknown }).artifactId;\n\treturn typeof artifactId === \"string\" ? artifactId : undefined;\n}\n\nexport interface ContextPipelineDeps {\n\t/** Current turn index, stamped into audit/policy/enforcement reports. */\n\tgetTurnIndex(): number;\n\t/** Session log: audit lookup, gc/artifact storage dirs, curation entries, token-estimate compaction anchor. */\n\tgetSessionManager(): SessionManager;\n\t/** Context-gc / prompt-enforcement / curation settings (all opt-in gates). */\n\tgetSettingsManager(): SettingsManager;\n\t/** Resolves a configured curation model pattern against configured auth. */\n\tgetModelRegistry(): ModelRegistry;\n\t/** Root dir the host-keyed {@link FitnessStore} and per-session gc/artifact storage live under. */\n\tgetAgentDir(): string;\n\t/** Workspace root, passed to the context-gc pass. */\n\tgetCwd(): string;\n\t/** Currently-active tool names — enforcement checks whether artifact_retrieve is a live affordance. */\n\tgetActiveToolNames(): string[];\n\t/** A disposed session must never persist a curation/pre-digest entry. */\n\tisDisposed(): boolean;\n\t/** The live memory manager — the active providers' page markers feed the semantic-gc scan. */\n\tgetMemoryManager(): MemoryManager;\n\t/** Roll a curation drain's spawned usage into session accounting (idempotent per reportId). */\n\taddSpawnedUsage(\n\t\tusage: Usage,\n\t\topts?: { label?: string; sourceSessionId?: string; reportId?: string },\n\t): string | undefined;\n\t/** One-shot LLM call fully isolated from the main session — the curation/pre-digest execution primitive. */\n\trunIsolatedCompletion(opts: IsolatedCompletionOptions): Promise<IsolatedCompletionResult>;\n}\n\nexport class ContextPipeline {\n\tprivate _latestContextGcReport: ContextGcReport | undefined = undefined;\n\t/** Brain-curation sidecar (design: brain-context-curation-design.md). Inert unless the\n\t * contextPolicy.curation setting is enabled AND the model passes the digest fitness gate. */\n\tprivate readonly _brainCurator = new BrainCurator();\n\tprivate _lastCurationSkipReason: string | undefined = undefined;\n\tprivate _lastPreDigestSkipReason: string | undefined = undefined;\n\tprivate _toolArtifactStore: ArtifactStore | undefined = undefined;\n\tprivate _latestContextAuditReport: ContextAuditReport | undefined = undefined;\n\tprivate _latestPromptPolicyReport: PromptPolicyShadowReport | undefined = undefined;\n\tprivate _latestPromptPolicyGcCorrelation: PromptPolicyGcCorrelationReport | undefined = undefined;\n\tprivate _latestPromptEnforcementReport: PromptEnforcementReport | undefined = undefined;\n\tprivate readonly _tokenBudget = new TokenBudget();\n\n\tprivate readonly deps: ContextPipelineDeps;\n\n\tconstructor(deps: ContextPipelineDeps) {\n\t\tthis.deps = deps;\n\t}\n\n\tprivate _contextGcStorageDir(): string {\n\t\treturn join(this.deps.getAgentDir(), \"context-gc\", this.deps.getSessionManager().getSessionId());\n\t}\n\n\tprivate _toolArtifactsDir(): string {\n\t\treturn join(this.deps.getAgentDir(), \"context-artifacts\", this.deps.getSessionManager().getSessionId());\n\t}\n\n\t/**\n\t * Session-scoped, filesystem-backed artifact store for first-capture-then-bound tool\n\t * output (grep/find only, for now -- see tool-output-artifacts.md). Lazily created and\n\t * cached so every tool construction in this session shares one store instance.\n\t *\n\t * `packToolOutput()` registers a reference (the packing tool call's id) at pack time\n\t * and fails closed, so packed artifacts are never prematurely collected.\n\t * `_releaseGcPackedArtifactReferences()` (called from `applyContextGc()`) releases\n\t * that reference once context-gc packs the result out of live context, and\n\t * opportunistically reclaims now-unreferenced artifacts via `cleanup()`.\n\t * Remaining carry-forward gap: cleanup() now also runs at dispose(), but only reclaims\n\t * already-released (zero-reference) artifacts. A session that ends before context-gc\n\t * ever evicts a result never releases that reference, so its artifact stays on disk by\n\t * design (resolvable on resume). Reclaiming those requires an explicit cross-session\n\t * expiry/liveness policy, not just a sweep.\n\t */\n\tgetToolArtifactStore(): ArtifactStore {\n\t\tthis._toolArtifactStore ??= createFileArtifactStore({ baseDir: this._toolArtifactsDir() });\n\t\treturn this._toolArtifactStore;\n\t}\n\n\t/**\n\t * Best-effort final sweep of any already-released (zero-reference) tool-output artifact at\n\t * session dispose. Reads the field (not the getter) so a session that never packed anything\n\t * doesn't force-create a store/dir just to sweep it.\n\t */\n\tcleanupToolArtifactStoreOnDispose(): void {\n\t\tthis._toolArtifactStore?.cleanup();\n\t}\n\n\t/**\n\t * One pass over the current branch, mapping each toolResult's toolCallId to its\n\t * persisted session-entry id. Rebuilt every audit pass (O(branch) per turn), so this is\n\t * O(n^2) over a long session. Fine at current scale; after the artifact-read fix this is\n\t * the next per-turn audit cost to optimize if it ever matters (e.g. cache/incrementally\n\t * update instead of a full rebuild).\n\t */\n\tprivate _buildSessionEntryIdLookup(): (toolCallId: string) => string | undefined {\n\t\tconst map = new Map<string, string>();\n\t\tfor (const entry of this.deps.getSessionManager().getBranch()) {\n\t\t\tif (entry.type === \"message\" && entry.message.role === \"toolResult\") {\n\t\t\t\tmap.set(entry.message.toolCallId, entry.id);\n\t\t\t}\n\t\t}\n\t\treturn (toolCallId: string) => map.get(toolCallId);\n\t}\n\n\t/**\n\t * Phase 1 observe-only audit pass (see context/context-audit.ts): converts live\n\t * toolResult messages into ContextItems and runs the existing retention/hard-constraint\n\t * evaluators over them, storing the latest deterministic report for tests/debugging.\n\t * Read-only with respect to messages, the transcript, and artifact references -- uses\n\t * `_toolArtifactStore` (the field), not `getToolArtifactStore()` (the getter), so a\n\t * session that never packed anything doesn't force-create a store/dir just to audit.\n\t * Never throws into a live turn: any failure degrades to an empty report.\n\t */\n\trunContextAudit(messages: AgentMessage[]): ContextAuditReport {\n\t\ttry {\n\t\t\tconst report = runContextAudit(messages, {\n\t\t\t\tturnIndex: this.deps.getTurnIndex(),\n\t\t\t\tartifactStore: this._toolArtifactStore,\n\t\t\t\tsessionEntryIdForToolCallId: this._buildSessionEntryIdLookup(),\n\t\t\t});\n\t\t\tthis._latestContextAuditReport = report;\n\t\t\treturn report;\n\t\t} catch {\n\t\t\tconst report: ContextAuditReport = { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t\t\tthis._latestContextAuditReport = report;\n\t\t\treturn report;\n\t\t}\n\t}\n\n\t/**\n\t * Read-only inspection of the context audit. With `messages`, recomputes fresh against\n\t * the given array (still no mutation of messages/transcript/artifact refs); without,\n\t * returns the last report computed during a real transform pass.\n\t */\n\tgetContextAuditReport(messages?: AgentMessage[]): ContextAuditReport {\n\t\tif (messages) return this.runContextAudit(messages);\n\t\treturn this._latestContextAuditReport ?? { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t}\n\n\t/**\n\t * Observe-first shadow/planning pass (see context/context-prompt-policy.ts): re-shapes\n\t * the audit report into a per-item policy plan whose `appliedAction` is always\n\t * \"keep_raw\" -- this never enforces anything, it only records what the policy engine\n\t * would say. Never throws into a live turn: any failure degrades to an empty report.\n\t */\n\trunPromptPolicyPlanning(auditReport: ContextAuditReport): PromptPolicyShadowReport {\n\t\ttry {\n\t\t\tconst report = planPromptPolicy(auditReport);\n\t\t\tthis._latestPromptPolicyReport = report;\n\t\t\treturn report;\n\t\t} catch {\n\t\t\tconst report: PromptPolicyShadowReport = { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t\t\tthis._latestPromptPolicyReport = report;\n\t\t\treturn report;\n\t\t}\n\t}\n\n\t/**\n\t * Read-only inspection of the shadow policy plan. With `messages`, recomputes fresh\n\t * (audit + plan) against the given array; without, returns the last plan computed\n\t * during a real transform pass. Never mutates messages/transcript/artifact refs.\n\t */\n\tgetPromptPolicyReport(messages?: AgentMessage[]): PromptPolicyShadowReport {\n\t\tif (messages) return this.runPromptPolicyPlanning(this.runContextAudit(messages));\n\t\treturn this._latestPromptPolicyReport ?? { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t}\n\n\t/**\n\t * Report-only correlation between the shadow plan just computed this turn and what the\n\t * legacy context-gc pass actually packed. Runs after `applyContextGc()` has already\n\t * produced its report; never influences context-gc itself. Never throws into a live\n\t * turn: any failure degrades to an empty correlation.\n\t */\n\tcorrelatePromptPolicyWithContextGc(gcReport: ContextGcReport): void {\n\t\tconst shadowReport = this._latestPromptPolicyReport ?? { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t\ttry {\n\t\t\tthis._latestPromptPolicyGcCorrelation = correlateWithContextGc(shadowReport, gcReport);\n\t\t} catch {\n\t\t\tthis._latestPromptPolicyGcCorrelation = { turnIndex: this.deps.getTurnIndex(), entries: [] };\n\t\t}\n\t}\n\n\t/** Read-only inspection of the latest shadow-plan/legacy-gc correlation, for tests/debugging. */\n\tgetPromptPolicyGcCorrelation(): PromptPolicyGcCorrelationReport {\n\t\treturn this._latestPromptPolicyGcCorrelation ?? { turnIndex: this.deps.getTurnIndex(), entries: [] };\n\t}\n\n\t/**\n\t * First enforcement pilot (see context/context-prompt-enforcement.ts): opt-in,\n\t * default-disabled stub-in-place of stale artifact-backed tool_output results in the\n\t * provider-visible message array only. Runs on `messages` AFTER context-gc has already\n\t * produced its own result, so legacy context-gc's own packing/reporting is completely\n\t * unaffected by this pass -- it only ever acts on messages gc left untouched this turn.\n\t * Never throws into a live turn: any failure degrades to returning `messages` unchanged.\n\t */\n\trunPromptEnforcement(\n\t\tmessages: AgentMessage[],\n\t\tshadowReport: PromptPolicyShadowReport,\n\t): { messages: AgentMessage[]; report: PromptEnforcementReport } {\n\t\ttry {\n\t\t\tconst persistedSettings = this.deps.getSettingsManager().getContextPromptEnforcementSettings();\n\t\t\tconst curationEnabled = this.deps.getSettingsManager().getContextCurationSettings().enabled;\n\t\t\tconst settings = {\n\t\t\t\t...persistedSettings,\n\t\t\t\t// Runtime fact, never assumed: artifact_retrieve is a companion affordance\n\t\t\t\t// (auto-activated alongside grep/find), not a default/global tool, so active\n\t\t\t\t// tools can differ turn to turn -- see context-prompt-enforcement.ts's doc\n\t\t\t\t// comment on why this is checked separately from hasAvailableRetrievalPath.\n\t\t\t\tretrievalToolAvailable: this.deps.getActiveToolNames().includes(\"artifact_retrieve\"),\n\t\t\t\tbrainRelevance: curationEnabled ? (itemId: string) => this._brainCurator.getRelevance(itemId) : undefined,\n\t\t\t};\n\t\t\tconst result = enforcePromptPolicy(messages, shadowReport, settings);\n\t\t\tthis._latestPromptEnforcementReport = result.report;\n\t\t\treturn result;\n\t\t} catch {\n\t\t\tconst report: PromptEnforcementReport = { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t\t\tthis._latestPromptEnforcementReport = report;\n\t\t\treturn { messages, report };\n\t\t}\n\t}\n\n\t/**\n\t * Enqueue relevance-scoring jobs for stale, artifact-backed tool outputs the enforcement\n\t * pilot could act on. Pure queueing — the verdicts only ever take effect through the\n\t * asymmetric advisory lever inside enforcePromptPolicy. Never throws into a turn.\n\t */\n\tenqueueRelevanceCuration(messages: AgentMessage[], shadowReport: PromptPolicyShadowReport): void {\n\t\ttry {\n\t\t\tconst settings = this.deps.getSettingsManager().getContextCurationSettings();\n\t\t\tif (!settings.enabled) return;\n\t\t\tconst goal = latestUserPromptText(messages).slice(0, 400);\n\t\t\tfor (const item of shadowReport.items) {\n\t\t\t\tif (!item.hasAvailableRetrievalPath) continue;\n\t\t\t\tconst message = messages[item.messageIndex];\n\t\t\t\tif (!message || message.role !== \"toolResult\" || message.toolCallId !== item.toolCallId) continue;\n\t\t\t\tif (message.isError) continue;\n\t\t\t\tconst details = message.details as\n\t\t\t\t\t| { contextGc?: { packed?: unknown }; promptPolicy?: { enforced?: unknown } }\n\t\t\t\t\t| undefined;\n\t\t\t\tif (details?.contextGc?.packed === true || details?.promptPolicy?.enforced === true) continue;\n\t\t\t\tconst text = message.content\n\t\t\t\t\t.filter((part): part is { type: \"text\"; text: string } => part.type === \"text\")\n\t\t\t\t\t.map((part) => part.text)\n\t\t\t\t\t.join(\"\\n\");\n\t\t\t\tif (text.length === 0) continue;\n\t\t\t\tthis._brainCurator.enqueue({ kind: \"relevance\", key: item.itemId, content: text.slice(0, 4000), goal });\n\t\t\t}\n\t\t} catch {\n\t\t\t// curation is a sidecar; it must never disrupt a turn\n\t\t}\n\t}\n\n\t/**\n\t * Drain gate: settings on, model configured+authed, and the model has PASSED the digest\n\t * fitness probe on THIS host (design: unfit or unprobed models are refused with a visible\n\t * reason, never silently degraded). Fire-and-forget; never throws into a turn.\n\t */\n\t/**\n\t * Resolve the curation model IFF every gate passes: setting enabled, model configured,\n\t * resolvable+authed, and digest-fitness-proven on THIS host (canonical \"provider/id\" ref —\n\t * runModelFitness stores reports under it, while settings.model may be a bare id or pattern).\n\t * Sets _lastCurationSkipReason on refusal; never throws.\n\t */\n\tresolveCurationModelIfFit(): Model<Api> | undefined {\n\t\tconst settings = this.deps.getSettingsManager().getContextCurationSettings();\n\t\tif (!settings.enabled) {\n\t\t\t// Never surface a stale refusal reason for a feature the user has since disabled.\n\t\t\tthis._lastCurationSkipReason = undefined;\n\t\t\treturn undefined;\n\t\t}\n\t\tif (!settings.model) {\n\t\t\tthis._lastCurationSkipReason = \"curation_model_unset\";\n\t\t\treturn undefined;\n\t\t}\n\t\tconst resolved = resolveCliModel({ cliModel: settings.model, modelRegistry: this.deps.getModelRegistry() });\n\t\tif (!resolved.model || !this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) {\n\t\t\tthis._lastCurationSkipReason = \"curation_model_unresolved\";\n\t\t\treturn undefined;\n\t\t}\n\t\tconst canonicalRef = `${resolved.model.provider}/${resolved.model.id}`;\n\t\tconst fitness = FitnessStore.forAgentDir(this.deps.getAgentDir())\n\t\t\t.getForHost()\n\t\t\t.find((entry) => entry.model === canonicalRef);\n\t\tconst verdict = evaluateSurfaceFitness(\"curation\", fitness?.report);\n\t\tif (!verdict.fit) {\n\t\t\tthis._lastCurationSkipReason =\n\t\t\t\tverdict.reason === \"unprobed\" ? \"curation_model_unprobed\" : \"curation_model_digest_unfit\";\n\t\t\treturn undefined;\n\t\t}\n\t\tthis._lastCurationSkipReason = undefined;\n\t\treturn resolved.model;\n\t}\n\n\tmaybeDrainBrainCuration(): void {\n\t\ttry {\n\t\t\tif (!this._brainCurator.hasWork() || this._brainCurator.isDraining) return;\n\t\t\tconst model = this.resolveCurationModelIfFit();\n\t\t\tif (!model) return;\n\t\t\tconst settings = this.deps.getSettingsManager().getContextCurationSettings();\n\t\t\tvoid this._drainBrainCuration(model, settings.maxJobsPerTurn);\n\t\t} catch {\n\t\t\t// curation is a sidecar; it must never disrupt a turn\n\t\t}\n\t}\n\n\t/**\n\t * Compaction pre-digest gate (design surface 3): everything the drain gate requires PLUS a\n\t * RUNTIME reliability proof — the curator must have run >=5 jobs on this session with a parse\n\t * failure rate <=5% before it is trusted to pre-digest compaction input. Returns undefined\n\t * (verbatim compaction, byte-for-byte today's behavior) whenever any gate refuses.\n\t */\n\tbuildCompactionPreDigest(): ((text: string, signal?: AbortSignal) => Promise<string>) | undefined {\n\t\ttry {\n\t\t\tconst model = this.resolveCurationModelIfFit();\n\t\t\tif (!model) return undefined;\n\t\t\tconst telemetry = this._brainCurator.telemetry();\n\t\t\tif (telemetry.jobsRun < 5 || telemetry.parseFailures / telemetry.jobsRun > 0.05) {\n\t\t\t\tthis._lastPreDigestSkipReason = \"curation_predigest_reliability_unproven\";\n\t\t\t\treturn undefined;\n\t\t\t}\n\t\t\tthis._lastPreDigestSkipReason = undefined;\n\t\t\treturn async (text, signal) => {\n\t\t\t\tconst result = await preDigestConversationText({\n\t\t\t\t\ttext,\n\t\t\t\t\tsignal,\n\t\t\t\t\tcomplete: async ({ systemPrompt, userPrompt, signal: chunkSignal }) => {\n\t\t\t\t\t\tconst completion = await this.deps.runIsolatedCompletion({\n\t\t\t\t\t\t\tsystemPrompt,\n\t\t\t\t\t\t\tmessages: [{ role: \"user\", content: [{ type: \"text\", text: userPrompt }], timestamp: Date.now() }],\n\t\t\t\t\t\t\tmodel,\n\t\t\t\t\t\t\tthinkingLevel: \"off\",\n\t\t\t\t\t\t\tmaxTokens: 512,\n\t\t\t\t\t\t\tsignal: chunkSignal,\n\t\t\t\t\t\t\tcacheRetention: \"short\",\n\t\t\t\t\t\t});\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttext: completion.text,\n\t\t\t\t\t\t\tcostUsd: completion.usage.cost.total,\n\t\t\t\t\t\t\tstopReason: String(completion.stopReason),\n\t\t\t\t\t\t};\n\t\t\t\t\t},\n\t\t\t\t});\n\t\t\t\tif (!this.deps.isDisposed() && result.totalChunks > 0) {\n\t\t\t\t\tthis.deps.getSessionManager().appendCustomEntry(\"brain-curation-predigest\", {\n\t\t\t\t\t\tversion: 1,\n\t\t\t\t\t\ttotalChunks: result.totalChunks,\n\t\t\t\t\t\tdigested: result.digested,\n\t\t\t\t\t\tfailed: result.failed,\n\t\t\t\t\t\tcharsBefore: text.length,\n\t\t\t\t\t\tcharsAfter: result.text.length,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t\treturn result.text;\n\t\t\t};\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate async _drainBrainCuration(model: Model<Api>, maxJobs: number): Promise<void> {\n\t\ttry {\n\t\t\t// ACCUMULATE across all drained jobs (the drain runs the completer once PER job) —\n\t\t\t// keeping only the last job's usage would under-report every multi-job drain.\n\t\t\tlet spentUsage: AssistantMessage[\"usage\"] | undefined;\n\t\t\tconst results = await this._brainCurator.drain({\n\t\t\t\tmaxJobs,\n\t\t\t\tcomplete: async ({ systemPrompt, userPrompt, signal }) => {\n\t\t\t\t\tconst completion = await this.deps.runIsolatedCompletion({\n\t\t\t\t\t\tsystemPrompt,\n\t\t\t\t\t\tmessages: [{ role: \"user\", content: [{ type: \"text\", text: userPrompt }], timestamp: Date.now() }],\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t\tthinkingLevel: \"off\",\n\t\t\t\t\t\tmaxTokens: 256,\n\t\t\t\t\t\tsignal,\n\t\t\t\t\t\t// Both curation system prompts are static — the provider can cache the prefix.\n\t\t\t\t\t\tcacheRetention: \"short\",\n\t\t\t\t\t});\n\t\t\t\t\tconst usage = completion.usage;\n\t\t\t\t\tif (!spentUsage) {\n\t\t\t\t\t\tspentUsage = structuredClone(usage);\n\t\t\t\t\t} else {\n\t\t\t\t\t\tspentUsage.input += usage.input;\n\t\t\t\t\t\tspentUsage.output += usage.output;\n\t\t\t\t\t\tspentUsage.cacheRead += usage.cacheRead;\n\t\t\t\t\t\tspentUsage.cacheWrite += usage.cacheWrite;\n\t\t\t\t\t\tspentUsage.totalTokens += usage.totalTokens;\n\t\t\t\t\t\tspentUsage.cost.input += usage.cost.input;\n\t\t\t\t\t\tspentUsage.cost.output += usage.cost.output;\n\t\t\t\t\t\tspentUsage.cost.cacheRead += usage.cost.cacheRead;\n\t\t\t\t\t\tspentUsage.cost.cacheWrite += usage.cost.cacheWrite;\n\t\t\t\t\t\tspentUsage.cost.total += usage.cost.total;\n\t\t\t\t\t}\n\t\t\t\t\treturn {\n\t\t\t\t\t\ttext: completion.text,\n\t\t\t\t\t\tcostUsd: completion.usage.cost.total,\n\t\t\t\t\t\tstopReason: String(completion.stopReason),\n\t\t\t\t\t};\n\t\t\t\t},\n\t\t\t});\n\t\t\t// Honest accounting even for free local models: token visibility is the contract.\n\t\t\tif (spentUsage && (spentUsage.cost.total > 0 || spentUsage.totalTokens > 0)) {\n\t\t\t\tthis.deps.addSpawnedUsage(spentUsage, { label: \"context-curator\" });\n\t\t\t}\n\t\t\tif (this.deps.isDisposed() || results.length === 0) return;\n\t\t\tthis.deps.getSessionManager().appendCustomEntry(\"brain-curation\", {\n\t\t\t\tversion: 1,\n\t\t\t\tresults: results.map((result) => ({\n\t\t\t\t\tkey: result.key,\n\t\t\t\t\tkind: result.kind,\n\t\t\t\t\tok: result.ok,\n\t\t\t\t\tms: result.ms,\n\t\t\t\t\t...(result.digest !== undefined ? { digest: result.digest } : {}),\n\t\t\t\t\t...(result.relevant !== undefined ? { relevant: result.relevant, confidence: result.confidence } : {}),\n\t\t\t\t})),\n\t\t\t\ttelemetry: this._brainCurator.telemetry(),\n\t\t\t});\n\t\t} catch {\n\t\t\t// curation is a sidecar; it must never disrupt a turn\n\t\t}\n\t}\n\n\t/** Curation status for diagnostics/dashboard: settings, live telemetry, last refusal reason. */\n\tgetContextCurationStatus(): {\n\t\tenabled: boolean;\n\t\tmodel?: string;\n\t\ttelemetry: CurationTelemetrySnapshot;\n\t\tlastSkipReason?: string;\n\t\tlastPreDigestSkipReason?: string;\n\t} {\n\t\tconst settings = this.deps.getSettingsManager().getContextCurationSettings();\n\t\treturn {\n\t\t\tenabled: settings.enabled,\n\t\t\tmodel: settings.model,\n\t\t\ttelemetry: this._brainCurator.telemetry(),\n\t\t\tlastSkipReason: this._lastCurationSkipReason,\n\t\t\tlastPreDigestSkipReason: this._lastPreDigestSkipReason,\n\t\t};\n\t}\n\n\t/** Read-only inspection of the latest prompt-enforcement report, for tests/debugging. */\n\tgetPromptEnforcementReport(): PromptEnforcementReport {\n\t\treturn this._latestPromptEnforcementReport ?? { turnIndex: this.deps.getTurnIndex(), items: [] };\n\t}\n\n\tapplyContextGc(\n\t\tmessages: AgentMessage[],\n\t\twritePayloads: boolean,\n\t): { messages: AgentMessage[]; report: ContextGcReport } {\n\t\ttry {\n\t\t\tconst settings = this.deps.getSettingsManager().getContextGcSettings();\n\t\t\t// Merge the ACTIVE memory providers' own page markers (e.g. transcript-recall's\n\t\t\t// \"<memory_context\") into the semantic-memory marker list. The settings default is\n\t\t\t// provider-agnostic and non-empty, so without this merge the recall pages the bundled\n\t\t\t// default provider actually emits are never recognized as semantic-memory pages and\n\t\t\t// accumulate raw for the life of the session — the exact growth Bug #7 GC exists to stop.\n\t\t\tconst providerMarkers = this.deps.getMemoryManager().getContextMarkers();\n\t\t\tconst curationSettings = this.deps.getSettingsManager().getContextCurationSettings();\n\t\t\tconst result = applyContextGc(messages, {\n\t\t\t\t...settings,\n\t\t\t\tsemanticMemory: {\n\t\t\t\t\t...settings.semanticMemory,\n\t\t\t\t\tmarkers: [...new Set([...settings.semanticMemory.markers, ...providerMarkers])],\n\t\t\t\t},\n\t\t\t\tcwd: this.deps.getCwd(),\n\t\t\t\tstorageDir: this._contextGcStorageDir(),\n\t\t\t\twritePayloads,\n\t\t\t\tcuration: curationSettings.enabled\n\t\t\t\t\t? {\n\t\t\t\t\t\t\tresolveDigest: (digestKey) => {\n\t\t\t\t\t\t\t\tconst digest = this._brainCurator.getDigest(digestKey);\n\t\t\t\t\t\t\t\t// Count serves on the REAL per-turn pass only, never the report path.\n\t\t\t\t\t\t\t\tif (digest !== undefined && writePayloads) this._brainCurator.noteDigestServed();\n\t\t\t\t\t\t\t\treturn digest;\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t// Only the real per-turn pass enqueues work; the read-only report path\n\t\t\t\t\t\t\t// (writePayloads=false) stays side-effect free.\n\t\t\t\t\t\t\tonPacked: writePayloads\n\t\t\t\t\t\t\t\t? (record, originalText) => {\n\t\t\t\t\t\t\t\t\t\tthis._brainCurator.enqueue({\n\t\t\t\t\t\t\t\t\t\t\tkind: \"stub_digest\",\n\t\t\t\t\t\t\t\t\t\t\tkey: record.key ?? record.toolCallId,\n\t\t\t\t\t\t\t\t\t\t\tcontent: originalText,\n\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t: undefined,\n\t\t\t\t\t\t}\n\t\t\t\t\t: undefined,\n\t\t\t});\n\t\t\tthis._latestContextGcReport = result.report;\n\t\t\t// Only release/reclaim on the real per-turn pass (writePayloads=true), never on\n\t\t\t// the read-only status-report path (getContextGcReport with writePayloads=false),\n\t\t\t// so merely inspecting the report can't have side effects.\n\t\t\tif (writePayloads && result.report.packedCount > 0) {\n\t\t\t\tthis._releaseGcPackedArtifactReferences(messages, result.report);\n\t\t\t}\n\t\t\treturn result;\n\t\t} catch {\n\t\t\tconst report: ContextGcReport = {\n\t\t\t\tenabled: false,\n\t\t\t\tpackedCount: 0,\n\t\t\t\toriginalTokens: 0,\n\t\t\t\tpackedTokens: 0,\n\t\t\t\tsavedTokens: 0,\n\t\t\t\trecords: [],\n\t\t\t};\n\t\t\tthis._latestContextGcReport = report;\n\t\t\treturn { messages, report };\n\t\t}\n\t}\n\n\t/**\n\t * Reference-release + cleanup lifecycle: once context-gc has packed a grep/find tool\n\t * result out of the live prompt (the message is no longer current/active working\n\t * context -- see contracts-and-retention.md's \"ephemeral\"/\"expired\" retention\n\t * classes), release the pack-time reference `packToolOutput()` registered for it, and\n\t * opportunistically reclaim now-unreferenced artifacts. This is the other half of the\n\t * D2b-1 gate: artifacts were being registered but never released, so they accumulated\n\t * for the life of the session.\n\t *\n\t * `record.toolCallId` (from context-gc's packed record) is exactly the holder id\n\t * `packToolOutput()` used when it called `addReference()` -- both trace back to the\n\t * same tool call's id -- so no separate bookkeeping is needed to find it.\n\t */\n\tprivate _releaseGcPackedArtifactReferences(messages: AgentMessage[], report: ContextGcReport): void {\n\t\tconst store = this._toolArtifactStore;\n\t\tif (!store) return; // no store was ever constructed, so nothing could have been packed to one\n\n\t\tlet releasedAny = false;\n\t\tfor (const record of report.records) {\n\t\t\tif (record.toolName !== \"grep\" && record.toolName !== \"find\") continue;\n\t\t\tconst artifactId = extractArtifactId(messages[record.messageIndex]);\n\t\t\tif (!artifactId) continue;\n\t\t\tif (store.removeReference(artifactId, record.toolCallId)) releasedAny = true;\n\t\t}\n\t\t// Cleanup only runs immediately after a release actually happened in this pass, so\n\t\t// a long session doesn't re-scan the artifact directory on every turn once nothing\n\t\t// new became eligible for release.\n\t\tif (releasedAny) store.cleanup();\n\t}\n\n\tgetContextGcReport(messages?: AgentMessage[]): ContextGcReport {\n\t\tif (messages) return this.applyContextGc(messages, false).report;\n\t\treturn (\n\t\t\tthis._latestContextGcReport ?? {\n\t\t\t\tenabled: this.deps.getSettingsManager().getContextGcSettings().enabled,\n\t\t\t\tpackedCount: 0,\n\t\t\t\toriginalTokens: 0,\n\t\t\t\tpackedTokens: 0,\n\t\t\t\tsavedTokens: 0,\n\t\t\t\trecords: [],\n\t\t\t}\n\t\t);\n\t}\n\n\testimateCurrentContextTokens(messages: AgentMessage[]): number {\n\t\tconst estimate = estimateContextTokens(messages);\n\t\tif (estimate.lastUsageIndex === null) {\n\t\t\treturn this._tokenBudget.estimateDelta(estimateConversationChars(messages));\n\t\t}\n\n\t\tconst usageMessage = messages[estimate.lastUsageIndex];\n\t\tconst compactionEntry = getLatestCompactionEntry(this.deps.getSessionManager().getBranch());\n\t\tif (usageMessage?.role !== \"assistant\" || !compactionEntry) {\n\t\t\treturn estimate.tokens;\n\t\t}\n\t\tconst usageTimestamp = (usageMessage as AssistantMessage).timestamp;\n\t\tconst compactionTimestamp = new Date(compactionEntry.timestamp).getTime();\n\t\tif (usageTimestamp <= compactionTimestamp) {\n\t\t\treturn this._tokenBudget.estimateDelta(estimateConversationChars(messages));\n\t\t}\n\n\t\tconst coveredChars = estimateConversationChars(messages, 0, estimate.lastUsageIndex + 1);\n\t\tconst deltaChars = estimateConversationChars(messages, estimate.lastUsageIndex + 1, messages.length);\n\t\tthis._tokenBudget.anchor(estimate.usageTokens, coveredChars);\n\t\treturn this._tokenBudget.current(deltaChars, estimate.tokens);\n\t}\n}\n\n/**\n * Estimate covered characters from the provider-visible context messages.\n */\nfunction estimateConversationChars(messages: AgentMessage[], start = 0, end = messages.length): number {\n\tlet total = 0;\n\tfor (let index = Math.max(0, start); index < Math.min(messages.length, end); index++) {\n\t\tconst message = messages[index];\n\t\tif (!message) continue;\n\t\tconst rawContent = (message as { content?: unknown }).content;\n\t\tif (typeof rawContent === \"string\") {\n\t\t\ttotal += rawContent.length;\n\t\t\tcontinue;\n\t\t}\n\t\tif (!Array.isArray(rawContent)) {\n\t\t\tcontinue;\n\t\t}\n\t\tfor (const block of rawContent) {\n\t\t\tif (typeof block === \"string\") {\n\t\t\t\ttotal += block.length;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (!block || typeof block !== \"object\") {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (typeof (block as { text?: unknown }).text === \"string\") {\n\t\t\t\ttotal += (block as { text?: string }).text?.length ?? 0;\n\t\t\t}\n\t\t\tif (typeof (block as { thinking?: unknown }).thinking === \"string\") {\n\t\t\t\ttotal += (block as { thinking?: string }).thinking?.length ?? 0;\n\t\t\t}\n\t\t}\n\t}\n\treturn total;\n}\n"]}

package/dist/core/context-pipeline.js

@@ -23,7 +23,7 @@
  * the pipeline — keeping the transform the one place the two subsystems meet.
  */
 import { join } from "node:path";
-import { estimateContextTokens, getLatestCompactionEntry } from "@caupulican/pi-agent-core/node";
+import { estimateContextTokens, getLatestCompactionEntry, TokenBudget, } from "@caupulican/pi-agent-core/node";
 import { BrainCurator, preDigestConversationText } from "./context/brain-curator.js";
 import { createFileArtifactStore } from "./context/context-artifacts.js";
 import { runContextAudit } from "./context/context-audit.js";
@@ -31,6 +31,7 @@ import { enforcePromptPolicy } from "./context/context-prompt-enforcement.js";
 import { correlateWithContextGc, planPromptPolicy, } from "./context/context-prompt-policy.js";
 import { applyContextGc } from "./context-gc.js";
 import { resolveCliModel } from "./model-resolver.js";
+import { evaluateSurfaceFitness } from "./model-router/fitness-gate.js";
 import { FitnessStore } from "./models/fitness-store.js";
 /** Latest user prompt text in the provider-visible array (curation goal line; bounded by caller). */
 export function latestUserPromptText(messages) {
@@ -71,6 +72,7 @@ export class ContextPipeline {
     _latestPromptPolicyReport = undefined;
     _latestPromptPolicyGcCorrelation = undefined;
     _latestPromptEnforcementReport = undefined;
+    _tokenBudget = new TokenBudget();
     deps;
     constructor(deps) {
         this.deps = deps;
@@ -304,13 +306,10 @@ export class ContextPipeline {
         const fitness = FitnessStore.forAgentDir(this.deps.getAgentDir())
             .getForHost()
             .find((entry) => entry.model === canonicalRef);
-        const digestScore = fitness?.report.digest;
-        if (!digestScore) {
-            this._lastCurationSkipReason = "curation_model_unprobed";
-            return undefined;
-        }
-        if (digestScore.succeeded < Math.ceil(digestScore.total * (2 / 3))) {
-            this._lastCurationSkipReason = "curation_model_digest_unfit";
+        const verdict = evaluateSurfaceFitness("curation", fitness?.report);
+        if (!verdict.fit) {
+            this._lastCurationSkipReason =
+                verdict.reason === "unprobed" ? "curation_model_unprobed" : "curation_model_digest_unfit";
             return undefined;
         }
         this._lastCurationSkipReason = undefined;
@@ -575,20 +574,58 @@ export class ContextPipeline {
     }
     estimateCurrentContextTokens(messages) {
         const estimate = estimateContextTokens(messages);
-        const compactionEntry = getLatestCompactionEntry(this.deps.getSessionManager().getBranch());
-        if (estimate.lastUsageIndex === null || !compactionEntry) {
-            return estimate.tokens;
+        if (estimate.lastUsageIndex === null) {
+            return this._tokenBudget.estimateDelta(estimateConversationChars(messages));
         }
         const usageMessage = messages[estimate.lastUsageIndex];
-        if (usageMessage?.role !== "assistant") {
+        const compactionEntry = getLatestCompactionEntry(this.deps.getSessionManager().getBranch());
+        if (usageMessage?.role !== "assistant" || !compactionEntry) {
             return estimate.tokens;
         }
         const usageTimestamp = usageMessage.timestamp;
         const compactionTimestamp = new Date(compactionEntry.timestamp).getTime();
         if (usageTimestamp <= compactionTimestamp) {
-            return estimate.trailingTokens;
+            return this._tokenBudget.estimateDelta(estimateConversationChars(messages));
+        }
+        const coveredChars = estimateConversationChars(messages, 0, estimate.lastUsageIndex + 1);
+        const deltaChars = estimateConversationChars(messages, estimate.lastUsageIndex + 1, messages.length);
+        this._tokenBudget.anchor(estimate.usageTokens, coveredChars);
+        return this._tokenBudget.current(deltaChars, estimate.tokens);
+    }
+}
+/**
+ * Estimate covered characters from the provider-visible context messages.
+ */
+function estimateConversationChars(messages, start = 0, end = messages.length) {
+    let total = 0;
+    for (let index = Math.max(0, start); index < Math.min(messages.length, end); index++) {
+        const message = messages[index];
+        if (!message)
+            continue;
+        const rawContent = message.content;
+        if (typeof rawContent === "string") {
+            total += rawContent.length;
+            continue;
+        }
+        if (!Array.isArray(rawContent)) {
+            continue;
+        }
+        for (const block of rawContent) {
+            if (typeof block === "string") {
+                total += block.length;
+                continue;
+            }
+            if (!block || typeof block !== "object") {
+                continue;
+            }
+            if (typeof block.text === "string") {
+                total += block.text?.length ?? 0;
+            }
+            if (typeof block.thinking === "string") {
+                total += block.thinking?.length ?? 0;
+            }
         }
-        return estimate.tokens;
     }
+    return total;
 }
 //# sourceMappingURL=context-pipeline.js.map