@caupulican/pi-adaptative 0.81.37 → 0.81.39

package/dist/core/tool-selection/tool-performance-store.js

@@ -1,13 +1,19 @@
 import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
-import { dirname, join } from "node:path";
+import { dirname } from "node:path";
+import { stateFile } from "../agent-paths.js";
 import { currentHostFingerprint } from "../models/fitness-store.js";
+import { isWorkerSession } from "../session-role.js";
 const STORE_VERSION = 1;
 const MAX_STATS_PER_HOST = 500;
 const MAX_OBSERVATIONS_PER_HOST = 1_000;
+const MAX_INTENT_AGREEMENT_PER_HOST = 500;
 const EWMA_ALPHA = 0.25;
 function statKey(key) {
     return `${key.modelRef}\0${key.intentClass}\0${key.tool}`;
 }
+function intentAgreementKey(modelRef, intentClass) {
+    return `${modelRef}\0${intentClass}`;
+}
 function isRecord(value) {
     return typeof value === "object" && value !== null && !Array.isArray(value);
 }
@@ -55,6 +61,16 @@ function isObservation(value) {
         (value.inputTokenEstimate === undefined || typeof value.inputTokenEstimate === "number") &&
         (value.outputTokenEstimate === undefined || typeof value.outputTokenEstimate === "number"));
 }
+function isIntentAgreement(value) {
+    return (isRecord(value) &&
+        typeof value.modelRef === "string" &&
+        isIntentClass(value.intentClass) &&
+        typeof value.sampleCount === "number" &&
+        typeof value.agreementCount === "number" &&
+        typeof value.hintActiveSampleCount === "number" &&
+        typeof value.hintActiveAgreementCount === "number" &&
+        typeof value.lastUpdatedAt === "string");
+}
 function emptyStats(key, at) {
     return {
         ...key,
@@ -67,6 +83,17 @@ function emptyStats(key, at) {
         lastUsedAt: at,
     };
 }
+function emptyIntentAgreement(modelRef, intentClass, at) {
+    return {
+        modelRef,
+        intentClass,
+        sampleCount: 0,
+        agreementCount: 0,
+        hintActiveSampleCount: 0,
+        hintActiveAgreementCount: 0,
+        lastUpdatedAt: at,
+    };
+}
 function finiteNonNegative(value) {
     return typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : undefined;
 }
@@ -90,10 +117,14 @@ function parseHost(value) {
         typeof host.cores !== "number" ||
         typeof host.totalMemGb !== "number")
         return undefined;
+    // intentAgreement is a purely additive field (older store files predate it) — tolerate absence
+    // rather than bumping STORE_VERSION, same as any other backward-compatible default-empty field.
+    const intentAgreementRaw = isRecord(value.intentAgreement) ? value.intentAgreement : {};
     return {
         host: { id: host.id, cpu: host.cpu, cores: host.cores, totalMemGb: host.totalMemGb },
         stats: Object.fromEntries(Object.entries(value.stats).filter((entry) => isStats(entry[1]))),
         observations: value.observations.filter(isObservation),
+        intentAgreement: Object.fromEntries(Object.entries(intentAgreementRaw).filter((entry) => isIntentAgreement(entry[1]))),
     };
 }
 function parseFile(value) {
@@ -115,15 +146,25 @@ function trimStats(stats) {
         .sort(([, left], [, right]) => Date.parse(right.lastUsedAt) - Date.parse(left.lastUsedAt))
         .slice(0, MAX_STATS_PER_HOST));
 }
+function trimIntentAgreement(records) {
+    const entries = Object.entries(records);
+    if (entries.length <= MAX_INTENT_AGREEMENT_PER_HOST)
+        return records;
+    return Object.fromEntries(entries
+        .sort(([, left], [, right]) => Date.parse(right.lastUpdatedAt) - Date.parse(left.lastUpdatedAt))
+        .slice(0, MAX_INTENT_AGREEMENT_PER_HOST));
+}
 export class ToolPerformanceStore {
     filePath;
     fingerprint;
+    readOnly;
     constructor(filePath, options = {}) {
         this.filePath = filePath;
         this.fingerprint = options.fingerprint ?? currentHostFingerprint;
+        this.readOnly = options.readOnly ?? isWorkerSession();
     }
     static forAgentDir(agentDir, options = {}) {
-        return new ToolPerformanceStore(join(agentDir, "state", "tool-performance.json"), options);
+        return new ToolPerformanceStore(stateFile(agentDir, "tool-performance.json"), options);
     }
     load() {
         try {
@@ -136,6 +177,10 @@ export class ToolPerformanceStore {
         }
     }
     save(file) {
+        // Zero-footprint (worker session): never create the state dir, lock, or file -- the caller
+        // still gets its normally-computed return value from the in-memory `file`.
+        if (this.readOnly)
+            return;
         mkdirSync(dirname(this.filePath), { recursive: true });
         writeFileSync(this.filePath, `${JSON.stringify(file, null, "\t")}\n`, "utf8");
     }
@@ -144,7 +189,7 @@ export class ToolPerformanceStore {
         const existing = file.hosts[fingerprint.id];
         if (existing)
             return existing;
-        const created = { host: fingerprint, stats: {}, observations: [] };
+        const created = { host: fingerprint, stats: {}, observations: [], intentAgreement: {} };
         file.hosts[fingerprint.id] = created;
         return created;
     }
@@ -154,6 +199,28 @@ export class ToolPerformanceStore {
         const stats = host?.stats[statKey(key)];
         return stats ? { ...stats } : emptyStats(key, new Date(0).toISOString());
     }
+    /** Every per-tool track record recorded for a (model,intent) bucket — the promotion.ts input. */
+    getStatsForIntent(modelRef, intentClass) {
+        const host = this.load().hosts[this.fingerprint().id];
+        if (!host)
+            return [];
+        return Object.values(host.stats)
+            .filter((stats) => stats.modelRef === modelRef && stats.intentClass === intentClass)
+            .map((stats) => ({ ...stats }));
+    }
+    /** Durable observe-mode agreement for one (model,intent) bucket (see {@link ToolSelectionIntentAgreement}). */
+    getIntentAgreement(modelRef, intentClass) {
+        const host = this.load().hosts[this.fingerprint().id];
+        const record = host?.intentAgreement[intentAgreementKey(modelRef, intentClass)];
+        return record ? { ...record } : emptyIntentAgreement(modelRef, intentClass, new Date(0).toISOString());
+    }
+    /** All recorded (model,intent) agreement buckets, optionally scoped to one model — report input. */
+    getAllIntentAgreements(modelRef) {
+        const host = this.load().hosts[this.fingerprint().id];
+        return Object.values(host?.intentAgreement ?? {})
+            .filter((record) => modelRef === undefined || record.modelRef === modelRef)
+            .map((record) => ({ ...record }));
+    }
     recordValidation(key, outcome, at = new Date().toISOString()) {
         const file = this.load();
         const host = this.hostData(file);
@@ -209,6 +276,23 @@ export class ToolPerformanceStore {
         if (host.observations.length > MAX_OBSERVATIONS_PER_HOST) {
             host.observations = host.observations.slice(-MAX_OBSERVATIONS_PER_HOST);
         }
+        // Observe-mode agreement: did the RAW ranking's top pick (before any actual-tool-only
+        // eligibility gating) match what the model actually called? Recorded durably per
+        // (model,intent), separate from the capped `observations` log above, so it survives trimming.
+        const predictedBest = observation.selection.ranked[0]?.tool;
+        const agreed = predictedBest !== undefined && predictedBest === observation.key.tool;
+        const agreementKey = intentAgreementKey(observation.key.modelRef, observation.key.intentClass);
+        const currentAgreement = host.intentAgreement[agreementKey] ??
+            emptyIntentAgreement(observation.key.modelRef, observation.key.intentClass, at);
+        host.intentAgreement[agreementKey] = {
+            ...currentAgreement,
+            sampleCount: currentAgreement.sampleCount + 1,
+            agreementCount: currentAgreement.agreementCount + (agreed ? 1 : 0),
+            hintActiveSampleCount: currentAgreement.hintActiveSampleCount + (observation.hintActiveAtCallTime ? 1 : 0),
+            hintActiveAgreementCount: currentAgreement.hintActiveAgreementCount + (observation.hintActiveAtCallTime && agreed ? 1 : 0),
+            lastUpdatedAt: at,
+        };
+        host.intentAgreement = trimIntentAgreement(host.intentAgreement);
         this.save(file);
         return { ...next };
     }

package/dist/core/tool-selection/tool-performance-store.js.map

@@ -1 +1 @@
-{"version":3,"file":"tool-performance-store.js","sourceRoot":"","sources":["../../../src/core/tool-selection/tool-performance-store.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,UAAU,EAAE,SAAS,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,SAAS,CAAC;AAC7E,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAC1C,OAAO,EAAE,sBAAsB,EAAwB,MAAM,4BAA4B,CAAC;AAE1F,MAAM,aAAa,GAAG,CAAC,CAAC;AACxB,MAAM,kBAAkB,GAAG,GAAG,CAAC;AAC/B,MAAM,yBAAyB,GAAG,KAAK,CAAC;AACxC,MAAM,UAAU,GAAG,IAAI,CAAC;AA6ExB,SAAS,OAAO,CAAC,GAAuB,EAAU;IACjD,OAAO,GAAG,GAAG,CAAC,QAAQ,KAAK,GAAG,CAAC,WAAW,KAAK,GAAG,CAAC,IAAI,EAAE,CAAC;AAAA,CAC1D;AAED,SAAS,QAAQ,CAAC,KAAc,EAAoC;IACnE,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAAA,CAC5E;AAED,SAAS,aAAa,CAAC,KAAc,EAAqC;IACzE,OAAO,CACN,KAAK,KAAK,MAAM;QAChB,KAAK,KAAK,QAAQ;QAClB,KAAK,KAAK,SAAS;QACnB,KAAK,KAAK,OAAO;QACjB,KAAK,KAAK,UAAU;QACpB,KAAK,KAAK,SAAS;QACnB,KAAK,KAAK,OAAO,CACjB,CAAC;AAAA,CACF;AAED,SAAS,OAAO,CAAC,KAAc,EAAiC;IAC/D,OAAO,CACN,QAAQ,CAAC,KAAK,CAAC;QACf,OAAO,KAAK,CAAC,QAAQ,KAAK,QAAQ;QAClC,aAAa,CAAC,KAAK,CAAC,WAAW,CAAC;QAChC,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ;QAC9B,OAAO,KAAK,CAAC,KAAK,KAAK,QAAQ;QAC/B,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ;QAC9B,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ;QACrC,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ;QACrC,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ;QACrC,OAAO,KAAK,CAAC,YAAY,KAAK,QAAQ;QACtC,OAAO,KAAK,CAAC,UAAU,KAAK,QAAQ,CACpC,CAAC;AAAA,CACF;AAED,SAAS,aAAa,CAAC,KAAc,EAAqC;IACzE,OAAO,CACN,QAAQ,CAAC,KAAK,CAAC;QACf,OAAO,KAAK,CAAC,EAAE,KAAK,QAAQ;QAC5B,OAAO,KAAK,CAAC,QAAQ,KAAK,QAAQ;QAClC,aAAa,CAAC,KAAK,CAAC,WAAW,CAAC;QAChC,OAAO,KAAK,CAAC,UAAU,KAAK,QAAQ;QACpC,OAAO,KAAK,CAAC,SAAS,KAAK,SAAS;QACpC,OAAO,KAAK,CAAC,SAAS,KAAK,SAAS;QACpC,CAAC,KAAK,CAAC,WAAW,KAAK,WAAW,IAAI,KAAK,CAAC,WAAW,KAAK,WAAW,IAAI,KAAK,CAAC,WAAW,KAAK,SAAS,CAAC;QAC3G,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAC9B,KAAK,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,OAAO,IAAI,KAAK,QAAQ,CAAC;QACzD,OAAO,KAAK,CAAC,OAAO,KAAK,QAAQ;QACjC,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ;QAChC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,MAAM,CAAC;QAC3B,KAAK,CAAC,MAAM,CAAC,KAAK,CACjB,CAAC,KAAK,EAAE,EAAE,CACT,QAAQ,CAAC,KAAK,CAAC;YACf,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ;YAC9B,OAAO,KAAK,CAAC,OAAO,KAAK,QAAQ;YACjC,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ,CACtC;QACD,CAAC,KAAK,CAAC,SAAS,KAAK,SAAS,IAAI,OAAO,KAAK,CAAC,SAAS,KAAK,QAAQ,CAAC;QACtE,CAAC,KAAK,CAAC,kBAAkB,KAAK,SAAS,IAAI,OAAO,KAAK,CAAC,kBAAkB,KAAK,QAAQ,CAAC;QACxF,CAAC,KAAK,CAAC,mBAAmB,KAAK,SAAS,IAAI,OAAO,KAAK,CAAC,mBAAmB,KAAK,QAAQ,CAAC,CAC1F,CAAC;AAAA,CACF;AAED,SAAS,UAAU,CAAC,GAAuB,EAAE,EAAU,EAAwB;IAC9E,OAAO;QACN,GAAG,GAAG;QACN,KAAK,EAAE,CAAC;QACR,IAAI,EAAE,CAAC;QACP,WAAW,EAAE,CAAC;QACd,WAAW,EAAE,CAAC;QACd,WAAW,EAAE,CAAC;QACd,YAAY,EAAE,CAAC;QACf,UAAU,EAAE,EAAE;KACd,CAAC;AAAA,CACF;AAED,SAAS,iBAAiB,CAAC,KAAyB,EAAsB;IACzE,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AAAA,CAC7F;AAED,SAAS,UAAU,CAAC,QAA4B,EAAE,IAAwB,EAAsB;IAC/F,IAAI,IAAI,KAAK,SAAS;QAAE,OAAO,QAAQ,CAAC;IACxC,OAAO,QAAQ,KAAK,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,QAAQ,GAAG,UAAU,GAAG,CAAC,IAAI,GAAG,QAAQ,CAAC,CAAC;AAAA,CACjF;AAED,SAAS,eAAe,CACvB,aAAiC,EACjC,iBAAqC,EACrC,IAAwB,EACH;IACrB,IAAI,IAAI,KAAK,SAAS;QAAE,OAAO,iBAAiB,CAAC;IACjD,MAAM,SAAS,GAAG,aAAa,KAAK,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,GAAG,aAAa,CAAC,CAAC;IACnF,OAAO,UAAU,CAAC,iBAAiB,EAAE,SAAS,CAAC,CAAC;AAAA,CAChD;AAED,SAAS,SAAS,CAAC,KAAc,EAAuC;IACvE,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,YAAY,CAAC;QAC5G,OAAO,SAAS,CAAC;IAClB,MAAM,IAAI,GAAG,KAAK,CAAC,IAAI,CAAC;IACxB,IACC,OAAO,IAAI,CAAC,EAAE,KAAK,QAAQ;QAC3B,OAAO,IAAI,CAAC,GAAG,KAAK,QAAQ;QAC5B,OAAO,IAAI,CAAC,KAAK,KAAK,QAAQ;QAC9B,OAAO,IAAI,CAAC,UAAU,KAAK,QAAQ;QAEnC,OAAO,SAAS,CAAC;IAClB,OAAO;QACN,IAAI,EAAE,EAAE,EAAE,EAAE,IAAI,CAAC,EAAE,EAAE,GAAG,EAAE,IAAI,CAAC,GAAG,EAAE,KAAK,EAAE,IAAI,CAAC,KAAK,EAAE,UAAU,EAAE,IAAI,CAAC,UAAU,EAAE;QACpF,KAAK,EAAE,MAAM,CAAC,WAAW,CACxB,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,CAAC,KAAK,EAA2C,EAAE,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CACzG;QACD,YAAY,EAAE,KAAK,CAAC,YAAY,CAAC,MAAM,CAAC,aAAa,CAAC;KACtD,CAAC;AAAA,CACF;AAED,SAAS,SAAS,CAAC,KAAc,EAA4B;IAC5D,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,OAAO,KAAK,aAAa,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC;QACnF,OAAO,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC;IAC9C,CAAC;IACD,OAAO;QACN,OAAO,EAAE,aAAa;QACtB,KAAK,EAAE,MAAM,CAAC,WAAW,CACxB,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,KAAK,CAAC;aACzB,GAAG,CAAC,CAAC,CAAC,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC,MAAM,EAAE,SAAS,CAAC,IAAI,CAAC,CAAU,CAAC;aAC3D,MAAM,CAAC,CAAC,KAAK,EAAuD,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,KAAK,SAAS,CAAC,CAChG;KACD,CAAC;AAAA,CACF;AAED,SAAS,SAAS,CAAC,KAA2C,EAAwC;IACrG,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;IACtC,IAAI,OAAO,CAAC,MAAM,IAAI,kBAAkB;QAAE,OAAO,KAAK,CAAC;IACvD,OAAO,MAAM,CAAC,WAAW,CACxB,OAAO;SACL,IAAI,CAAC,CAAC,CAAC,EAAE,IAAI,CAAC,EAAE,CAAC,EAAE,KAAK,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,UAAU,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC;SACzF,KAAK,CAAC,CAAC,EAAE,kBAAkB,CAAC,CAC9B,CAAC;AAAA,CACF;AAED,MAAM,OAAO,oBAAoB;IACf,QAAQ,CAAS;IACjB,WAAW,CAAwB;IAEpD,YAAY,QAAgB,EAAE,OAAO,GAA4C,EAAE,EAAE;QACpF,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;QACzB,IAAI,CAAC,WAAW,GAAG,OAAO,CAAC,WAAW,IAAI,sBAAsB,CAAC;IAAA,CACjE;IAED,MAAM,CAAC,WAAW,CAAC,QAAgB,EAAE,OAAO,GAA4C,EAAE,EAAwB;QACjH,OAAO,IAAI,oBAAoB,CAAC,IAAI,CAAC,QAAQ,EAAE,OAAO,EAAE,uBAAuB,CAAC,EAAE,OAAO,CAAC,CAAC;IAAA,CAC3F;IAEO,IAAI,GAA6B;QACxC,IAAI,CAAC;YACJ,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,QAAQ,CAAC;gBAAE,OAAO,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC;YAC7E,OAAO,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,IAAI,CAAC,QAAQ,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC;QACnE,CAAC;QAAC,MAAM,CAAC;YACR,OAAO,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC;QAC9C,CAAC;IAAA,CACD;IAEO,IAAI,CAAC,IAA8B,EAAQ;QAClD,SAAS,CAAC,OAAO,CAAC,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;QACvD,aAAa,CAAC,IAAI,CAAC,QAAQ,EAAE,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;IAAA,CAC9E;IAEO,QAAQ,CAAC,IAA8B,EAA2B;QACzE,MAAM,WAAW,GAAG,IAAI,CAAC,WAAW,EAAE,CAAC;QACvC,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC;QAC5C,IAAI,QAAQ;YAAE,OAAO,QAAQ,CAAC;QAC9B,MAAM,OAAO,GAA4B,EAAE,IAAI,EAAE,WAAW,EAAE,KAAK,EAAE,EAAE,EAAE,YAAY,EAAE,EAAE,EAAE,CAAC;QAC5F,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC;QACrC,OAAO,OAAO,CAAC;IAAA,CACf;IAED,GAAG,CAAC,GAAuB,EAAwB;QAClD,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;QACzB,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC,CAAC;QAC/C,MAAM,KAAK,GAAG,IAAI,EAAE,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;QACxC,OAAO,KAAK,CAAC,CAAC,CAAC,EAAE,GAAG,KAAK,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC,GAAG,EAAE,IAAI,IAAI,CAAC,CAAC,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC;IAAA,CACzE;IAED,gBAAgB,CACf,GAAuB,EACvB,OAA+B,EAC/B,EAAE,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,EACN;QACvB,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;QACzB,MAAM,IAAI,GAAG,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;QACjC,MAAM,UAAU,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC;QAChC,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,IAAI,UAAU,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;QAC9D,MAAM,IAAI,GAAyB;YAClC,GAAG,OAAO;YACV,WAAW,EAAE,OAAO,CAAC,WAAW,GAAG,CAAC,OAAO,KAAK,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACnE,WAAW,EAAE,OAAO,CAAC,WAAW,GAAG,CAAC,OAAO,KAAK,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAClE,UAAU,EAAE,EAAE;SACd,CAAC;QACF,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,GAAG,IAAI,CAAC;QAC9B,IAAI,CAAC,KAAK,GAAG,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACnC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAChB,OAAO,EAAE,GAAG,IAAI,EAAE,CAAC;IAAA,CACnB;IAED,eAAe,CAAC,WAAqC,EAAwB;QAC5E,MAAM,EAAE,GAAG,WAAW,CAAC,EAAE,IAAI,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;QACtD,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;QACzB,MAAM,IAAI,GAAG,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;QACjC,MAAM,UAAU,GAAG,OAAO,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;QAC5C,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,IAAI,UAAU,CAAC,WAAW,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;QAC1E,MAAM,SAAS,GAAG,iBAAiB,CAAC,WAAW,CAAC,SAAS,CAAC,CAAC;QAC3D,MAAM,kBAAkB,GAAG,iBAAiB,CAAC,WAAW,CAAC,kBAAkB,CAAC,CAAC;QAC7E,MAAM,mBAAmB,GAAG,iBAAiB,CAAC,WAAW,CAAC,mBAAmB,CAAC,CAAC;QAC/E,MAAM,IAAI,GAAyB;YAClC,GAAG,OAAO;YACV,KAAK,EAAE,OAAO,CAAC,KAAK,GAAG,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACpD,IAAI,EAAE,OAAO,CAAC,IAAI,GAAG,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAClD,WAAW,EAAE,OAAO,CAAC,WAAW,GAAG,CAAC;YACpC,aAAa,EAAE,UAAU,CAAC,OAAO,CAAC,aAAa,EAAE,SAAS,CAAC;YAC3D,sBAAsB,EAAE,eAAe,CAAC,OAAO,CAAC,aAAa,EAAE,OAAO,CAAC,sBAAsB,EAAE,SAAS,CAAC;YACzG,sBAAsB,EAAE,UAAU,CAAC,OAAO,CAAC,sBAAsB,EAAE,kBAAkB,CAAC;YACtF,uBAAuB,EAAE,UAAU,CAAC,OAAO,CAAC,uBAAuB,EAAE,mBAAmB,CAAC;YACzF,YAAY,EAAE,OAAO,CAAC,YAAY,GAAG,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAClE,UAAU,EAAE,EAAE;SACd,CAAC;QACF,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,GAAG,IAAI,CAAC;QAC9B,IAAI,CAAC,KAAK,GAAG,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACnC,IAAI,CAAC,YAAY,CAAC,IAAI,CAAC;YACtB,GAAG,WAAW,CAAC,SAAS;YACxB,EAAE;YACF,QAAQ,EAAE,WAAW,CAAC,GAAG,CAAC,QAAQ;YAClC,WAAW,EAAE,WAAW,CAAC,GAAG,CAAC,WAAW;YACxC,UAAU,EAAE,WAAW,CAAC,GAAG,CAAC,IAAI;YAChC,SAAS,EAAE,WAAW,CAAC,OAAO;YAC9B,MAAM,EAAE,WAAW,CAAC,SAAS,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC;YAChD,SAAS,EAAE,WAAW,CAAC,SAAS,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC;YACtD,SAAS;YACT,kBAAkB;YAClB,mBAAmB;SACnB,CAAC,CAAC;QACH,IAAI,IAAI,CAAC,YAAY,CAAC,MAAM,GAAG,yBAAyB,EAAE,CAAC;YAC1D,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC,yBAAyB,CAAC,CAAC;QACzE,CAAC;QACD,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAChB,OAAO,EAAE,GAAG,IAAI,EAAE,CAAC;IAAA,CACnB;IAED,UAAU,CAAC,QAAiB,EAAwB;QACnD,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC,CAAC;QACtD,MAAM,YAAY,GAAG,CAAC,IAAI,EAAE,YAAY,IAAI,EAAE,CAAC,CAAC,MAAM,CACrD,CAAC,WAAW,EAAE,EAAE,CAAC,QAAQ,KAAK,SAAS,IAAI,WAAW,CAAC,QAAQ,KAAK,QAAQ,CAC5E,CAAC;QACF,MAAM,UAAU,GAAG,YAAY,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,SAAS,CAAC,CAAC;QAC/E,MAAM,WAAW,GAAG,YAAY,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,WAAW,KAAK,WAAW,CAAC,CAAC;QAClG,MAAM,OAAO,GAAG,CAAC,MAAiC,EAAsB,EAAE,CAAC;YAC1E,MAAM,OAAO,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,KAAK,EAAmB,EAAE,CAAC,KAAK,KAAK,SAAS,CAAC,CAAC;YAC/E,OAAO,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,KAAK,EAAE,EAAE,CAAC,GAAG,GAAG,KAAK,EAAE,CAAC,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC;QAAA,CACxG,CAAC;QACF,OAAO;YACN,iBAAiB,EAAE,UAAU,CAAC,MAAM;YACpC,kBAAkB,EAAE,UAAU,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,SAAS,CAAC,CAAC,MAAM;YACpF,uBAAuB,EAAE,YAAY,CAAC,MAAM,CAC3C,CAAC,WAAW,EAAE,EAAE,CACf,CAAC,WAAW,CAAC,SAAS;gBACtB,CAAC,WAAW,CAAC,WAAW,KAAK,WAAW,IAAI,WAAW,CAAC,cAAc,KAAK,WAAW,CAAC,UAAU,CAAC,CACnG,CAAC,MAAM;YACR,mBAAmB,EAAE,WAAW,CAAC,MAAM;YACvC,0BAA0B,EAAE,WAAW,CAAC,MAAM,CAC7C,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,cAAc,KAAK,WAAW,CAAC,UAAU,CACtE,CAAC,MAAM;YACR,cAAc,EAAE,YAAY,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,WAAW,KAAK,WAAW,CAAC,CAAC,MAAM;YACpG,eAAe,EAAE,YAAY,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,WAAW,KAAK,SAAS,CAAC,CAAC,MAAM;YACnG,gBAAgB,EAAE,OAAO,CAAC,YAAY,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,SAAS,CAAC,CAAC;YACnF,yBAAyB,EAAE,OAAO,CAAC,YAAY,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,kBAAkB,CAAC,CAAC;YACrG,0BAA0B,EAAE,OAAO,CAAC,YAAY,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,mBAAmB,CAAC,CAAC;SACvG,CAAC;IAAA,CACF;IAED,eAAe,CAAC,QAAiB,EAA8B;QAC9D,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC,CAAC;QACtD,OAAO,CAAC,IAAI,EAAE,YAAY,IAAI,EAAE,CAAC;aAC/B,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,QAAQ,KAAK,SAAS,IAAI,WAAW,CAAC,QAAQ,KAAK,QAAQ,CAAC;aACpF,GAAG,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,CAAC;YACtB,GAAG,WAAW;YACd,SAAS,EAAE,CAAC,GAAG,WAAW,CAAC,SAAS,CAAC;YACrC,MAAM,EAAE,WAAW,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,SAAS,EAAE,CAAC,CAAC;SACjE,CAAC,CAAC,CAAC;IAAA,CACL;CACD","sourcesContent":["import { existsSync, mkdirSync, readFileSync, writeFileSync } from \"node:fs\";\nimport { dirname, join } from \"node:path\";\nimport { currentHostFingerprint, type HostFingerprint } from \"../models/fitness-store.ts\";\n\nconst STORE_VERSION = 1;\nconst MAX_STATS_PER_HOST = 500;\nconst MAX_OBSERVATIONS_PER_HOST = 1_000;\nconst EWMA_ALPHA = 0.25;\n\nexport type ToolSelectionIntentClass = \"read\" | \"search\" | \"execute\" | \"write\" | \"retrieve\" | \"explain\" | \"other\";\n\nexport interface ToolPerformanceKey {\n\tmodelRef: string;\n\tintentClass: ToolSelectionIntentClass;\n\ttool: string;\n}\n\nexport interface ToolPerformanceStats extends ToolPerformanceKey {\n\talpha: number;\n\tbeta: number;\n\tsampleCount: number;\n\tlatencyEwmaMs?: number;\n\tlatencyDeviationEwmaMs?: number;\n\tinputTokenEstimateEwma?: number;\n\toutputTokenEstimateEwma?: number;\n\trepairCount: number;\n\tbounceCount: number;\n\tfailureCount: number;\n\tlastUsedAt: string;\n}\n\nexport interface ToolSelectionObservation {\n\tat: string;\n\tmodelRef: string;\n\tintentClass: ToolSelectionIntentClass;\n\tactualTool: string;\n\tfirstTool: boolean;\n\tsucceeded: boolean;\n\tdisposition: \"recommend\" | \"shortlist\" | \"abstain\";\n\trecommendation?: string;\n\tshortlist: string[];\n\tentropy: number;\n\tmargin: number;\n\t/** Redacted ranking: names and numeric scores only, never prompts, arguments, paths, or output. */\n\tranked: Array<{ tool: string; utility: number; probability: number }>;\n\tlatencyMs?: number;\n\tinputTokenEstimate?: number;\n\toutputTokenEstimate?: number;\n}\n\nexport interface ToolSelectionMetrics {\n\tfirstToolAttempts: number;\n\tfirstToolSuccesses: number;\n\twrongToolOrFailureCount: number;\n\trecommendationCount: number;\n\trecommendationMatchedCount: number;\n\tshortlistCount: number;\n\tabstentionCount: number;\n\taverageLatencyMs?: number;\n\taverageInputTokenEstimate?: number;\n\taverageOutputTokenEstimate?: number;\n}\n\ninterface HostToolPerformanceData {\n\thost: HostFingerprint;\n\tstats: Record<string, ToolPerformanceStats>;\n\tobservations: ToolSelectionObservation[];\n}\n\ninterface ToolPerformanceStoreFile {\n\tversion: 1;\n\thosts: Record<string, HostToolPerformanceData>;\n}\n\nexport interface ToolExecutionObservation {\n\tkey: ToolPerformanceKey;\n\tsuccess: boolean;\n\tlatencyMs: number;\n\tinputTokenEstimate?: number;\n\toutputTokenEstimate?: number;\n\tselection: Omit<ToolSelectionObservation, \"at\" | \"modelRef\" | \"intentClass\" | \"actualTool\" | \"succeeded\">;\n\tat?: string;\n}\n\nfunction statKey(key: ToolPerformanceKey): string {\n\treturn `${key.modelRef}\\0${key.intentClass}\\0${key.tool}`;\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n\treturn typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction isIntentClass(value: unknown): value is ToolSelectionIntentClass {\n\treturn (\n\t\tvalue === \"read\" ||\n\t\tvalue === \"search\" ||\n\t\tvalue === \"execute\" ||\n\t\tvalue === \"write\" ||\n\t\tvalue === \"retrieve\" ||\n\t\tvalue === \"explain\" ||\n\t\tvalue === \"other\"\n\t);\n}\n\nfunction isStats(value: unknown): value is ToolPerformanceStats {\n\treturn (\n\t\tisRecord(value) &&\n\t\ttypeof value.modelRef === \"string\" &&\n\t\tisIntentClass(value.intentClass) &&\n\t\ttypeof value.tool === \"string\" &&\n\t\ttypeof value.alpha === \"number\" &&\n\t\ttypeof value.beta === \"number\" &&\n\t\ttypeof value.sampleCount === \"number\" &&\n\t\ttypeof value.repairCount === \"number\" &&\n\t\ttypeof value.bounceCount === \"number\" &&\n\t\ttypeof value.failureCount === \"number\" &&\n\t\ttypeof value.lastUsedAt === \"string\"\n\t);\n}\n\nfunction isObservation(value: unknown): value is ToolSelectionObservation {\n\treturn (\n\t\tisRecord(value) &&\n\t\ttypeof value.at === \"string\" &&\n\t\ttypeof value.modelRef === \"string\" &&\n\t\tisIntentClass(value.intentClass) &&\n\t\ttypeof value.actualTool === \"string\" &&\n\t\ttypeof value.firstTool === \"boolean\" &&\n\t\ttypeof value.succeeded === \"boolean\" &&\n\t\t(value.disposition === \"recommend\" || value.disposition === \"shortlist\" || value.disposition === \"abstain\") &&\n\t\tArray.isArray(value.shortlist) &&\n\t\tvalue.shortlist.every((tool) => typeof tool === \"string\") &&\n\t\ttypeof value.entropy === \"number\" &&\n\t\ttypeof value.margin === \"number\" &&\n\t\tArray.isArray(value.ranked) &&\n\t\tvalue.ranked.every(\n\t\t\t(entry) =>\n\t\t\t\tisRecord(entry) &&\n\t\t\t\ttypeof entry.tool === \"string\" &&\n\t\t\t\ttypeof entry.utility === \"number\" &&\n\t\t\t\ttypeof entry.probability === \"number\",\n\t\t) &&\n\t\t(value.latencyMs === undefined || typeof value.latencyMs === \"number\") &&\n\t\t(value.inputTokenEstimate === undefined || typeof value.inputTokenEstimate === \"number\") &&\n\t\t(value.outputTokenEstimate === undefined || typeof value.outputTokenEstimate === \"number\")\n\t);\n}\n\nfunction emptyStats(key: ToolPerformanceKey, at: string): ToolPerformanceStats {\n\treturn {\n\t\t...key,\n\t\talpha: 1,\n\t\tbeta: 1,\n\t\tsampleCount: 0,\n\t\trepairCount: 0,\n\t\tbounceCount: 0,\n\t\tfailureCount: 0,\n\t\tlastUsedAt: at,\n\t};\n}\n\nfunction finiteNonNegative(value: number | undefined): number | undefined {\n\treturn typeof value === \"number\" && Number.isFinite(value) && value >= 0 ? value : undefined;\n}\n\nfunction updateEwma(previous: number | undefined, next: number | undefined): number | undefined {\n\tif (next === undefined) return previous;\n\treturn previous === undefined ? next : previous + EWMA_ALPHA * (next - previous);\n}\n\nfunction updateDeviation(\n\tpreviousValue: number | undefined,\n\tpreviousDeviation: number | undefined,\n\tnext: number | undefined,\n): number | undefined {\n\tif (next === undefined) return previousDeviation;\n\tconst deviation = previousValue === undefined ? 0 : Math.abs(next - previousValue);\n\treturn updateEwma(previousDeviation, deviation);\n}\n\nfunction parseHost(value: unknown): HostToolPerformanceData | undefined {\n\tif (!isRecord(value) || !isRecord(value.host) || !isRecord(value.stats) || !Array.isArray(value.observations))\n\t\treturn undefined;\n\tconst host = value.host;\n\tif (\n\t\ttypeof host.id !== \"string\" ||\n\t\ttypeof host.cpu !== \"string\" ||\n\t\ttypeof host.cores !== \"number\" ||\n\t\ttypeof host.totalMemGb !== \"number\"\n\t)\n\t\treturn undefined;\n\treturn {\n\t\thost: { id: host.id, cpu: host.cpu, cores: host.cores, totalMemGb: host.totalMemGb },\n\t\tstats: Object.fromEntries(\n\t\t\tObject.entries(value.stats).filter((entry): entry is [string, ToolPerformanceStats] => isStats(entry[1])),\n\t\t),\n\t\tobservations: value.observations.filter(isObservation),\n\t};\n}\n\nfunction parseFile(value: unknown): ToolPerformanceStoreFile {\n\tif (!isRecord(value) || value.version !== STORE_VERSION || !isRecord(value.hosts)) {\n\t\treturn { version: STORE_VERSION, hosts: {} };\n\t}\n\treturn {\n\t\tversion: STORE_VERSION,\n\t\thosts: Object.fromEntries(\n\t\t\tObject.entries(value.hosts)\n\t\t\t\t.map(([hostId, host]) => [hostId, parseHost(host)] as const)\n\t\t\t\t.filter((entry): entry is readonly [string, HostToolPerformanceData] => entry[1] !== undefined),\n\t\t),\n\t};\n}\n\nfunction trimStats(stats: Record<string, ToolPerformanceStats>): Record<string, ToolPerformanceStats> {\n\tconst entries = Object.entries(stats);\n\tif (entries.length <= MAX_STATS_PER_HOST) return stats;\n\treturn Object.fromEntries(\n\t\tentries\n\t\t\t.sort(([, left], [, right]) => Date.parse(right.lastUsedAt) - Date.parse(left.lastUsedAt))\n\t\t\t.slice(0, MAX_STATS_PER_HOST),\n\t);\n}\n\nexport class ToolPerformanceStore {\n\tprivate readonly filePath: string;\n\tprivate readonly fingerprint: () => HostFingerprint;\n\n\tconstructor(filePath: string, options: { fingerprint?: () => HostFingerprint } = {}) {\n\t\tthis.filePath = filePath;\n\t\tthis.fingerprint = options.fingerprint ?? currentHostFingerprint;\n\t}\n\n\tstatic forAgentDir(agentDir: string, options: { fingerprint?: () => HostFingerprint } = {}): ToolPerformanceStore {\n\t\treturn new ToolPerformanceStore(join(agentDir, \"state\", \"tool-performance.json\"), options);\n\t}\n\n\tprivate load(): ToolPerformanceStoreFile {\n\t\ttry {\n\t\t\tif (!existsSync(this.filePath)) return { version: STORE_VERSION, hosts: {} };\n\t\t\treturn parseFile(JSON.parse(readFileSync(this.filePath, \"utf8\")));\n\t\t} catch {\n\t\t\treturn { version: STORE_VERSION, hosts: {} };\n\t\t}\n\t}\n\n\tprivate save(file: ToolPerformanceStoreFile): void {\n\t\tmkdirSync(dirname(this.filePath), { recursive: true });\n\t\twriteFileSync(this.filePath, `${JSON.stringify(file, null, \"\\t\")}\\n`, \"utf8\");\n\t}\n\n\tprivate hostData(file: ToolPerformanceStoreFile): HostToolPerformanceData {\n\t\tconst fingerprint = this.fingerprint();\n\t\tconst existing = file.hosts[fingerprint.id];\n\t\tif (existing) return existing;\n\t\tconst created: HostToolPerformanceData = { host: fingerprint, stats: {}, observations: [] };\n\t\tfile.hosts[fingerprint.id] = created;\n\t\treturn created;\n\t}\n\n\tget(key: ToolPerformanceKey): ToolPerformanceStats {\n\t\tconst file = this.load();\n\t\tconst host = file.hosts[this.fingerprint().id];\n\t\tconst stats = host?.stats[statKey(key)];\n\t\treturn stats ? { ...stats } : emptyStats(key, new Date(0).toISOString());\n\t}\n\n\trecordValidation(\n\t\tkey: ToolPerformanceKey,\n\t\toutcome: \"repaired\" | \"bounced\",\n\t\tat = new Date().toISOString(),\n\t): ToolPerformanceStats {\n\t\tconst file = this.load();\n\t\tconst host = this.hostData(file);\n\t\tconst storageKey = statKey(key);\n\t\tconst current = host.stats[storageKey] ?? emptyStats(key, at);\n\t\tconst next: ToolPerformanceStats = {\n\t\t\t...current,\n\t\t\trepairCount: current.repairCount + (outcome === \"repaired\" ? 1 : 0),\n\t\t\tbounceCount: current.bounceCount + (outcome === \"bounced\" ? 1 : 0),\n\t\t\tlastUsedAt: at,\n\t\t};\n\t\thost.stats[storageKey] = next;\n\t\thost.stats = trimStats(host.stats);\n\t\tthis.save(file);\n\t\treturn { ...next };\n\t}\n\n\trecordExecution(observation: ToolExecutionObservation): ToolPerformanceStats {\n\t\tconst at = observation.at ?? new Date().toISOString();\n\t\tconst file = this.load();\n\t\tconst host = this.hostData(file);\n\t\tconst storageKey = statKey(observation.key);\n\t\tconst current = host.stats[storageKey] ?? emptyStats(observation.key, at);\n\t\tconst latencyMs = finiteNonNegative(observation.latencyMs);\n\t\tconst inputTokenEstimate = finiteNonNegative(observation.inputTokenEstimate);\n\t\tconst outputTokenEstimate = finiteNonNegative(observation.outputTokenEstimate);\n\t\tconst next: ToolPerformanceStats = {\n\t\t\t...current,\n\t\t\talpha: current.alpha + (observation.success ? 1 : 0),\n\t\t\tbeta: current.beta + (observation.success ? 0 : 1),\n\t\t\tsampleCount: current.sampleCount + 1,\n\t\t\tlatencyEwmaMs: updateEwma(current.latencyEwmaMs, latencyMs),\n\t\t\tlatencyDeviationEwmaMs: updateDeviation(current.latencyEwmaMs, current.latencyDeviationEwmaMs, latencyMs),\n\t\t\tinputTokenEstimateEwma: updateEwma(current.inputTokenEstimateEwma, inputTokenEstimate),\n\t\t\toutputTokenEstimateEwma: updateEwma(current.outputTokenEstimateEwma, outputTokenEstimate),\n\t\t\tfailureCount: current.failureCount + (observation.success ? 0 : 1),\n\t\t\tlastUsedAt: at,\n\t\t};\n\t\thost.stats[storageKey] = next;\n\t\thost.stats = trimStats(host.stats);\n\t\thost.observations.push({\n\t\t\t...observation.selection,\n\t\t\tat,\n\t\t\tmodelRef: observation.key.modelRef,\n\t\t\tintentClass: observation.key.intentClass,\n\t\t\tactualTool: observation.key.tool,\n\t\t\tsucceeded: observation.success,\n\t\t\tranked: observation.selection.ranked.slice(0, 6),\n\t\t\tshortlist: observation.selection.shortlist.slice(0, 3),\n\t\t\tlatencyMs,\n\t\t\tinputTokenEstimate,\n\t\t\toutputTokenEstimate,\n\t\t});\n\t\tif (host.observations.length > MAX_OBSERVATIONS_PER_HOST) {\n\t\t\thost.observations = host.observations.slice(-MAX_OBSERVATIONS_PER_HOST);\n\t\t}\n\t\tthis.save(file);\n\t\treturn { ...next };\n\t}\n\n\tgetMetrics(modelRef?: string): ToolSelectionMetrics {\n\t\tconst host = this.load().hosts[this.fingerprint().id];\n\t\tconst observations = (host?.observations ?? []).filter(\n\t\t\t(observation) => modelRef === undefined || observation.modelRef === modelRef,\n\t\t);\n\t\tconst firstTools = observations.filter((observation) => observation.firstTool);\n\t\tconst recommended = observations.filter((observation) => observation.disposition === \"recommend\");\n\t\tconst average = (values: Array<number | undefined>): number | undefined => {\n\t\t\tconst present = values.filter((value): value is number => value !== undefined);\n\t\t\treturn present.length > 0 ? present.reduce((sum, value) => sum + value, 0) / present.length : undefined;\n\t\t};\n\t\treturn {\n\t\t\tfirstToolAttempts: firstTools.length,\n\t\t\tfirstToolSuccesses: firstTools.filter((observation) => observation.succeeded).length,\n\t\t\twrongToolOrFailureCount: observations.filter(\n\t\t\t\t(observation) =>\n\t\t\t\t\t!observation.succeeded ||\n\t\t\t\t\t(observation.disposition === \"recommend\" && observation.recommendation !== observation.actualTool),\n\t\t\t).length,\n\t\t\trecommendationCount: recommended.length,\n\t\t\trecommendationMatchedCount: recommended.filter(\n\t\t\t\t(observation) => observation.recommendation === observation.actualTool,\n\t\t\t).length,\n\t\t\tshortlistCount: observations.filter((observation) => observation.disposition === \"shortlist\").length,\n\t\t\tabstentionCount: observations.filter((observation) => observation.disposition === \"abstain\").length,\n\t\t\taverageLatencyMs: average(observations.map((observation) => observation.latencyMs)),\n\t\t\taverageInputTokenEstimate: average(observations.map((observation) => observation.inputTokenEstimate)),\n\t\t\taverageOutputTokenEstimate: average(observations.map((observation) => observation.outputTokenEstimate)),\n\t\t};\n\t}\n\n\tgetObservations(modelRef?: string): ToolSelectionObservation[] {\n\t\tconst host = this.load().hosts[this.fingerprint().id];\n\t\treturn (host?.observations ?? [])\n\t\t\t.filter((observation) => modelRef === undefined || observation.modelRef === modelRef)\n\t\t\t.map((observation) => ({\n\t\t\t\t...observation,\n\t\t\t\tshortlist: [...observation.shortlist],\n\t\t\t\tranked: observation.ranked.map((candidate) => ({ ...candidate })),\n\t\t\t}));\n\t}\n}\n"]}
+{"version":3,"file":"tool-performance-store.js","sourceRoot":"","sources":["../../../src/core/tool-selection/tool-performance-store.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,UAAU,EAAE,SAAS,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,SAAS,CAAC;AAC7E,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AACpC,OAAO,EAAE,SAAS,EAAE,MAAM,mBAAmB,CAAC;AAC9C,OAAO,EAAE,sBAAsB,EAAwB,MAAM,4BAA4B,CAAC;AAC1F,OAAO,EAAE,eAAe,EAAE,MAAM,oBAAoB,CAAC;AAErD,MAAM,aAAa,GAAG,CAAC,CAAC;AACxB,MAAM,kBAAkB,GAAG,GAAG,CAAC;AAC/B,MAAM,yBAAyB,GAAG,KAAK,CAAC;AACxC,MAAM,6BAA6B,GAAG,GAAG,CAAC;AAC1C,MAAM,UAAU,GAAG,IAAI,CAAC;AA2GxB,SAAS,OAAO,CAAC,GAAuB,EAAU;IACjD,OAAO,GAAG,GAAG,CAAC,QAAQ,KAAK,GAAG,CAAC,WAAW,KAAK,GAAG,CAAC,IAAI,EAAE,CAAC;AAAA,CAC1D;AAED,SAAS,kBAAkB,CAAC,QAAgB,EAAE,WAAqC,EAAU;IAC5F,OAAO,GAAG,QAAQ,KAAK,WAAW,EAAE,CAAC;AAAA,CACrC;AAED,SAAS,QAAQ,CAAC,KAAc,EAAoC;IACnE,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAAA,CAC5E;AAED,SAAS,aAAa,CAAC,KAAc,EAAqC;IACzE,OAAO,CACN,KAAK,KAAK,MAAM;QAChB,KAAK,KAAK,QAAQ;QAClB,KAAK,KAAK,SAAS;QACnB,KAAK,KAAK,OAAO;QACjB,KAAK,KAAK,UAAU;QACpB,KAAK,KAAK,SAAS;QACnB,KAAK,KAAK,OAAO,CACjB,CAAC;AAAA,CACF;AAED,SAAS,OAAO,CAAC,KAAc,EAAiC;IAC/D,OAAO,CACN,QAAQ,CAAC,KAAK,CAAC;QACf,OAAO,KAAK,CAAC,QAAQ,KAAK,QAAQ;QAClC,aAAa,CAAC,KAAK,CAAC,WAAW,CAAC;QAChC,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ;QAC9B,OAAO,KAAK,CAAC,KAAK,KAAK,QAAQ;QAC/B,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ;QAC9B,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ;QACrC,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ;QACrC,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ;QACrC,OAAO,KAAK,CAAC,YAAY,KAAK,QAAQ;QACtC,OAAO,KAAK,CAAC,UAAU,KAAK,QAAQ,CACpC,CAAC;AAAA,CACF;AAED,SAAS,aAAa,CAAC,KAAc,EAAqC;IACzE,OAAO,CACN,QAAQ,CAAC,KAAK,CAAC;QACf,OAAO,KAAK,CAAC,EAAE,KAAK,QAAQ;QAC5B,OAAO,KAAK,CAAC,QAAQ,KAAK,QAAQ;QAClC,aAAa,CAAC,KAAK,CAAC,WAAW,CAAC;QAChC,OAAO,KAAK,CAAC,UAAU,KAAK,QAAQ;QACpC,OAAO,KAAK,CAAC,SAAS,KAAK,SAAS;QACpC,OAAO,KAAK,CAAC,SAAS,KAAK,SAAS;QACpC,CAAC,KAAK,CAAC,WAAW,KAAK,WAAW,IAAI,KAAK,CAAC,WAAW,KAAK,WAAW,IAAI,KAAK,CAAC,WAAW,KAAK,SAAS,CAAC;QAC3G,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAC9B,KAAK,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,OAAO,IAAI,KAAK,QAAQ,CAAC;QACzD,OAAO,KAAK,CAAC,OAAO,KAAK,QAAQ;QACjC,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ;QAChC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,MAAM,CAAC;QAC3B,KAAK,CAAC,MAAM,CAAC,KAAK,CACjB,CAAC,KAAK,EAAE,EAAE,CACT,QAAQ,CAAC,KAAK,CAAC;YACf,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ;YAC9B,OAAO,KAAK,CAAC,OAAO,KAAK,QAAQ;YACjC,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ,CACtC;QACD,CAAC,KAAK,CAAC,SAAS,KAAK,SAAS,IAAI,OAAO,KAAK,CAAC,SAAS,KAAK,QAAQ,CAAC;QACtE,CAAC,KAAK,CAAC,kBAAkB,KAAK,SAAS,IAAI,OAAO,KAAK,CAAC,kBAAkB,KAAK,QAAQ,CAAC;QACxF,CAAC,KAAK,CAAC,mBAAmB,KAAK,SAAS,IAAI,OAAO,KAAK,CAAC,mBAAmB,KAAK,QAAQ,CAAC,CAC1F,CAAC;AAAA,CACF;AAED,SAAS,iBAAiB,CAAC,KAAc,EAAyC;IACjF,OAAO,CACN,QAAQ,CAAC,KAAK,CAAC;QACf,OAAO,KAAK,CAAC,QAAQ,KAAK,QAAQ;QAClC,aAAa,CAAC,KAAK,CAAC,WAAW,CAAC;QAChC,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ;QACrC,OAAO,KAAK,CAAC,cAAc,KAAK,QAAQ;QACxC,OAAO,KAAK,CAAC,qBAAqB,KAAK,QAAQ;QAC/C,OAAO,KAAK,CAAC,wBAAwB,KAAK,QAAQ;QAClD,OAAO,KAAK,CAAC,aAAa,KAAK,QAAQ,CACvC,CAAC;AAAA,CACF;AAED,SAAS,UAAU,CAAC,GAAuB,EAAE,EAAU,EAAwB;IAC9E,OAAO;QACN,GAAG,GAAG;QACN,KAAK,EAAE,CAAC;QACR,IAAI,EAAE,CAAC;QACP,WAAW,EAAE,CAAC;QACd,WAAW,EAAE,CAAC;QACd,WAAW,EAAE,CAAC;QACd,YAAY,EAAE,CAAC;QACf,UAAU,EAAE,EAAE;KACd,CAAC;AAAA,CACF;AAED,SAAS,oBAAoB,CAC5B,QAAgB,EAChB,WAAqC,EACrC,EAAU,EACqB;IAC/B,OAAO;QACN,QAAQ;QACR,WAAW;QACX,WAAW,EAAE,CAAC;QACd,cAAc,EAAE,CAAC;QACjB,qBAAqB,EAAE,CAAC;QACxB,wBAAwB,EAAE,CAAC;QAC3B,aAAa,EAAE,EAAE;KACjB,CAAC;AAAA,CACF;AAED,SAAS,iBAAiB,CAAC,KAAyB,EAAsB;IACzE,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AAAA,CAC7F;AAED,SAAS,UAAU,CAAC,QAA4B,EAAE,IAAwB,EAAsB;IAC/F,IAAI,IAAI,KAAK,SAAS;QAAE,OAAO,QAAQ,CAAC;IACxC,OAAO,QAAQ,KAAK,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,QAAQ,GAAG,UAAU,GAAG,CAAC,IAAI,GAAG,QAAQ,CAAC,CAAC;AAAA,CACjF;AAED,SAAS,eAAe,CACvB,aAAiC,EACjC,iBAAqC,EACrC,IAAwB,EACH;IACrB,IAAI,IAAI,KAAK,SAAS;QAAE,OAAO,iBAAiB,CAAC;IACjD,MAAM,SAAS,GAAG,aAAa,KAAK,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,GAAG,aAAa,CAAC,CAAC;IACnF,OAAO,UAAU,CAAC,iBAAiB,EAAE,SAAS,CAAC,CAAC;AAAA,CAChD;AAED,SAAS,SAAS,CAAC,KAAc,EAAuC;IACvE,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,YAAY,CAAC;QAC5G,OAAO,SAAS,CAAC;IAClB,MAAM,IAAI,GAAG,KAAK,CAAC,IAAI,CAAC;IACxB,IACC,OAAO,IAAI,CAAC,EAAE,KAAK,QAAQ;QAC3B,OAAO,IAAI,CAAC,GAAG,KAAK,QAAQ;QAC5B,OAAO,IAAI,CAAC,KAAK,KAAK,QAAQ;QAC9B,OAAO,IAAI,CAAC,UAAU,KAAK,QAAQ;QAEnC,OAAO,SAAS,CAAC;IAClB,iGAA+F;IAC/F,gGAAgG;IAChG,MAAM,kBAAkB,GAAG,QAAQ,CAAC,KAAK,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,eAAe,CAAC,CAAC,CAAC,EAAE,CAAC;IACxF,OAAO;QACN,IAAI,EAAE,EAAE,EAAE,EAAE,IAAI,CAAC,EAAE,EAAE,GAAG,EAAE,IAAI,CAAC,GAAG,EAAE,KAAK,EAAE,IAAI,CAAC,KAAK,EAAE,UAAU,EAAE,IAAI,CAAC,UAAU,EAAE;QACpF,KAAK,EAAE,MAAM,CAAC,WAAW,CACxB,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,CAAC,KAAK,EAA2C,EAAE,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CACzG;QACD,YAAY,EAAE,KAAK,CAAC,YAAY,CAAC,MAAM,CAAC,aAAa,CAAC;QACtD,eAAe,EAAE,MAAM,CAAC,WAAW,CAClC,MAAM,CAAC,OAAO,CAAC,kBAAkB,CAAC,CAAC,MAAM,CAAC,CAAC,KAAK,EAAmD,EAAE,CACpG,iBAAiB,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAC3B,CACD;KACD,CAAC;AAAA,CACF;AAED,SAAS,SAAS,CAAC,KAAc,EAA4B;IAC5D,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,OAAO,KAAK,aAAa,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC;QACnF,OAAO,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC;IAC9C,CAAC;IACD,OAAO;QACN,OAAO,EAAE,aAAa;QACtB,KAAK,EAAE,MAAM,CAAC,WAAW,CACxB,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,KAAK,CAAC;aACzB,GAAG,CAAC,CAAC,CAAC,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC,MAAM,EAAE,SAAS,CAAC,IAAI,CAAC,CAAU,CAAC;aAC3D,MAAM,CAAC,CAAC,KAAK,EAAuD,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,KAAK,SAAS,CAAC,CAChG;KACD,CAAC;AAAA,CACF;AAED,SAAS,SAAS,CAAC,KAA2C,EAAwC;IACrG,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;IACtC,IAAI,OAAO,CAAC,MAAM,IAAI,kBAAkB;QAAE,OAAO,KAAK,CAAC;IACvD,OAAO,MAAM,CAAC,WAAW,CACxB,OAAO;SACL,IAAI,CAAC,CAAC,CAAC,EAAE,IAAI,CAAC,EAAE,CAAC,EAAE,KAAK,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,UAAU,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC;SACzF,KAAK,CAAC,CAAC,EAAE,kBAAkB,CAAC,CAC9B,CAAC;AAAA,CACF;AAED,SAAS,mBAAmB,CAC3B,OAAqD,EACN;IAC/C,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IACxC,IAAI,OAAO,CAAC,MAAM,IAAI,6BAA6B;QAAE,OAAO,OAAO,CAAC;IACpE,OAAO,MAAM,CAAC,WAAW,CACxB,OAAO;SACL,IAAI,CAAC,CAAC,CAAC,EAAE,IAAI,CAAC,EAAE,CAAC,EAAE,KAAK,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,aAAa,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC;SAC/F,KAAK,CAAC,CAAC,EAAE,6BAA6B,CAAC,CACzC,CAAC;AAAA,CACF;AAED,MAAM,OAAO,oBAAoB;IACf,QAAQ,CAAS;IACjB,WAAW,CAAwB;IACnC,QAAQ,CAAU;IAEnC,YAAY,QAAgB,EAAE,OAAO,GAAgE,EAAE,EAAE;QACxG,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;QACzB,IAAI,CAAC,WAAW,GAAG,OAAO,CAAC,WAAW,IAAI,sBAAsB,CAAC;QACjE,IAAI,CAAC,QAAQ,GAAG,OAAO,CAAC,QAAQ,IAAI,eAAe,EAAE,CAAC;IAAA,CACtD;IAED,MAAM,CAAC,WAAW,CACjB,QAAgB,EAChB,OAAO,GAAgE,EAAE,EAClD;QACvB,OAAO,IAAI,oBAAoB,CAAC,SAAS,CAAC,QAAQ,EAAE,uBAAuB,CAAC,EAAE,OAAO,CAAC,CAAC;IAAA,CACvF;IAEO,IAAI,GAA6B;QACxC,IAAI,CAAC;YACJ,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,QAAQ,CAAC;gBAAE,OAAO,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC;YAC7E,OAAO,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,IAAI,CAAC,QAAQ,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC;QACnE,CAAC;QAAC,MAAM,CAAC;YACR,OAAO,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC;QAC9C,CAAC;IAAA,CACD;IAEO,IAAI,CAAC,IAA8B,EAAQ;QAClD,2FAA2F;QAC3F,2EAA2E;QAC3E,IAAI,IAAI,CAAC,QAAQ;YAAE,OAAO;QAC1B,SAAS,CAAC,OAAO,CAAC,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;QACvD,aAAa,CAAC,IAAI,CAAC,QAAQ,EAAE,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;IAAA,CAC9E;IAEO,QAAQ,CAAC,IAA8B,EAA2B;QACzE,MAAM,WAAW,GAAG,IAAI,CAAC,WAAW,EAAE,CAAC;QACvC,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC;QAC5C,IAAI,QAAQ;YAAE,OAAO,QAAQ,CAAC;QAC9B,MAAM,OAAO,GAA4B,EAAE,IAAI,EAAE,WAAW,EAAE,KAAK,EAAE,EAAE,EAAE,YAAY,EAAE,EAAE,EAAE,eAAe,EAAE,EAAE,EAAE,CAAC;QACjH,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC;QACrC,OAAO,OAAO,CAAC;IAAA,CACf;IAED,GAAG,CAAC,GAAuB,EAAwB;QAClD,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;QACzB,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC,CAAC;QAC/C,MAAM,KAAK,GAAG,IAAI,EAAE,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;QACxC,OAAO,KAAK,CAAC,CAAC,CAAC,EAAE,GAAG,KAAK,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC,GAAG,EAAE,IAAI,IAAI,CAAC,CAAC,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC;IAAA,CACzE;IAED,mGAAiG;IACjG,iBAAiB,CAAC,QAAgB,EAAE,WAAqC,EAA0B;QAClG,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC,CAAC;QACtD,IAAI,CAAC,IAAI;YAAE,OAAO,EAAE,CAAC;QACrB,OAAO,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC;aAC9B,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,QAAQ,KAAK,QAAQ,IAAI,KAAK,CAAC,WAAW,KAAK,WAAW,CAAC;aACnF,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,KAAK,EAAE,CAAC,CAAC,CAAC;IAAA,CACjC;IAED,+GAA+G;IAC/G,kBAAkB,CAAC,QAAgB,EAAE,WAAqC,EAAgC;QACzG,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC,CAAC;QACtD,MAAM,MAAM,GAAG,IAAI,EAAE,eAAe,CAAC,kBAAkB,CAAC,QAAQ,EAAE,WAAW,CAAC,CAAC,CAAC;QAChF,OAAO,MAAM,CAAC,CAAC,CAAC,EAAE,GAAG,MAAM,EAAE,CAAC,CAAC,CAAC,oBAAoB,CAAC,QAAQ,EAAE,WAAW,EAAE,IAAI,IAAI,CAAC,CAAC,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC;IAAA,CACvG;IAED,sGAAoG;IACpG,sBAAsB,CAAC,QAAiB,EAAkC;QACzE,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC,CAAC;QACtD,OAAO,MAAM,CAAC,MAAM,CAAC,IAAI,EAAE,eAAe,IAAI,EAAE,CAAC;aAC/C,MAAM,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,QAAQ,KAAK,SAAS,IAAI,MAAM,CAAC,QAAQ,KAAK,QAAQ,CAAC;aAC1E,GAAG,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,MAAM,EAAE,CAAC,CAAC,CAAC;IAAA,CACnC;IAED,gBAAgB,CACf,GAAuB,EACvB,OAA+B,EAC/B,EAAE,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,EACN;QACvB,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;QACzB,MAAM,IAAI,GAAG,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;QACjC,MAAM,UAAU,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC;QAChC,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,IAAI,UAAU,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;QAC9D,MAAM,IAAI,GAAyB;YAClC,GAAG,OAAO;YACV,WAAW,EAAE,OAAO,CAAC,WAAW,GAAG,CAAC,OAAO,KAAK,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACnE,WAAW,EAAE,OAAO,CAAC,WAAW,GAAG,CAAC,OAAO,KAAK,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAClE,UAAU,EAAE,EAAE;SACd,CAAC;QACF,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,GAAG,IAAI,CAAC;QAC9B,IAAI,CAAC,KAAK,GAAG,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACnC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAChB,OAAO,EAAE,GAAG,IAAI,EAAE,CAAC;IAAA,CACnB;IAED,eAAe,CAAC,WAAqC,EAAwB;QAC5E,MAAM,EAAE,GAAG,WAAW,CAAC,EAAE,IAAI,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;QACtD,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;QACzB,MAAM,IAAI,GAAG,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;QACjC,MAAM,UAAU,GAAG,OAAO,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;QAC5C,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,IAAI,UAAU,CAAC,WAAW,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;QAC1E,MAAM,SAAS,GAAG,iBAAiB,CAAC,WAAW,CAAC,SAAS,CAAC,CAAC;QAC3D,MAAM,kBAAkB,GAAG,iBAAiB,CAAC,WAAW,CAAC,kBAAkB,CAAC,CAAC;QAC7E,MAAM,mBAAmB,GAAG,iBAAiB,CAAC,WAAW,CAAC,mBAAmB,CAAC,CAAC;QAC/E,MAAM,IAAI,GAAyB;YAClC,GAAG,OAAO;YACV,KAAK,EAAE,OAAO,CAAC,KAAK,GAAG,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACpD,IAAI,EAAE,OAAO,CAAC,IAAI,GAAG,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAClD,WAAW,EAAE,OAAO,CAAC,WAAW,GAAG,CAAC;YACpC,aAAa,EAAE,UAAU,CAAC,OAAO,CAAC,aAAa,EAAE,SAAS,CAAC;YAC3D,sBAAsB,EAAE,eAAe,CAAC,OAAO,CAAC,aAAa,EAAE,OAAO,CAAC,sBAAsB,EAAE,SAAS,CAAC;YACzG,sBAAsB,EAAE,UAAU,CAAC,OAAO,CAAC,sBAAsB,EAAE,kBAAkB,CAAC;YACtF,uBAAuB,EAAE,UAAU,CAAC,OAAO,CAAC,uBAAuB,EAAE,mBAAmB,CAAC;YACzF,YAAY,EAAE,OAAO,CAAC,YAAY,GAAG,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAClE,UAAU,EAAE,EAAE;SACd,CAAC;QACF,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,GAAG,IAAI,CAAC;QAC9B,IAAI,CAAC,KAAK,GAAG,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACnC,IAAI,CAAC,YAAY,CAAC,IAAI,CAAC;YACtB,GAAG,WAAW,CAAC,SAAS;YACxB,EAAE;YACF,QAAQ,EAAE,WAAW,CAAC,GAAG,CAAC,QAAQ;YAClC,WAAW,EAAE,WAAW,CAAC,GAAG,CAAC,WAAW;YACxC,UAAU,EAAE,WAAW,CAAC,GAAG,CAAC,IAAI;YAChC,SAAS,EAAE,WAAW,CAAC,OAAO;YAC9B,MAAM,EAAE,WAAW,CAAC,SAAS,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC;YAChD,SAAS,EAAE,WAAW,CAAC,SAAS,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC;YACtD,SAAS;YACT,kBAAkB;YAClB,mBAAmB;SACnB,CAAC,CAAC;QACH,IAAI,IAAI,CAAC,YAAY,CAAC,MAAM,GAAG,yBAAyB,EAAE,CAAC;YAC1D,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC,yBAAyB,CAAC,CAAC;QACzE,CAAC;QAED,sFAAsF;QACtF,iFAAiF;QACjF,8FAA8F;QAC9F,MAAM,aAAa,GAAG,WAAW,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,IAAI,CAAC;QAC5D,MAAM,MAAM,GAAG,aAAa,KAAK,SAAS,IAAI,aAAa,KAAK,WAAW,CAAC,GAAG,CAAC,IAAI,CAAC;QACrF,MAAM,YAAY,GAAG,kBAAkB,CAAC,WAAW,CAAC,GAAG,CAAC,QAAQ,EAAE,WAAW,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC;QAC/F,MAAM,gBAAgB,GACrB,IAAI,CAAC,eAAe,CAAC,YAAY,CAAC;YAClC,oBAAoB,CAAC,WAAW,CAAC,GAAG,CAAC,QAAQ,EAAE,WAAW,CAAC,GAAG,CAAC,WAAW,EAAE,EAAE,CAAC,CAAC;QACjF,IAAI,CAAC,eAAe,CAAC,YAAY,CAAC,GAAG;YACpC,GAAG,gBAAgB;YACnB,WAAW,EAAE,gBAAgB,CAAC,WAAW,GAAG,CAAC;YAC7C,cAAc,EAAE,gBAAgB,CAAC,cAAc,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAClE,qBAAqB,EAAE,gBAAgB,CAAC,qBAAqB,GAAG,CAAC,WAAW,CAAC,oBAAoB,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAC1G,wBAAwB,EACvB,gBAAgB,CAAC,wBAAwB,GAAG,CAAC,WAAW,CAAC,oBAAoB,IAAI,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACjG,aAAa,EAAE,EAAE;SACjB,CAAC;QACF,IAAI,CAAC,eAAe,GAAG,mBAAmB,CAAC,IAAI,CAAC,eAAe,CAAC,CAAC;QAEjE,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAChB,OAAO,EAAE,GAAG,IAAI,EAAE,CAAC;IAAA,CACnB;IAED,UAAU,CAAC,QAAiB,EAAwB;QACnD,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC,CAAC;QACtD,MAAM,YAAY,GAAG,CAAC,IAAI,EAAE,YAAY,IAAI,EAAE,CAAC,CAAC,MAAM,CACrD,CAAC,WAAW,EAAE,EAAE,CAAC,QAAQ,KAAK,SAAS,IAAI,WAAW,CAAC,QAAQ,KAAK,QAAQ,CAC5E,CAAC;QACF,MAAM,UAAU,GAAG,YAAY,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,SAAS,CAAC,CAAC;QAC/E,MAAM,WAAW,GAAG,YAAY,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,WAAW,KAAK,WAAW,CAAC,CAAC;QAClG,MAAM,OAAO,GAAG,CAAC,MAAiC,EAAsB,EAAE,CAAC;YAC1E,MAAM,OAAO,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,KAAK,EAAmB,EAAE,CAAC,KAAK,KAAK,SAAS,CAAC,CAAC;YAC/E,OAAO,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,KAAK,EAAE,EAAE,CAAC,GAAG,GAAG,KAAK,EAAE,CAAC,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC;QAAA,CACxG,CAAC;QACF,OAAO;YACN,iBAAiB,EAAE,UAAU,CAAC,MAAM;YACpC,kBAAkB,EAAE,UAAU,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,SAAS,CAAC,CAAC,MAAM;YACpF,uBAAuB,EAAE,YAAY,CAAC,MAAM,CAC3C,CAAC,WAAW,EAAE,EAAE,CACf,CAAC,WAAW,CAAC,SAAS;gBACtB,CAAC,WAAW,CAAC,WAAW,KAAK,WAAW,IAAI,WAAW,CAAC,cAAc,KAAK,WAAW,CAAC,UAAU,CAAC,CACnG,CAAC,MAAM;YACR,mBAAmB,EAAE,WAAW,CAAC,MAAM;YACvC,0BAA0B,EAAE,WAAW,CAAC,MAAM,CAC7C,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,cAAc,KAAK,WAAW,CAAC,UAAU,CACtE,CAAC,MAAM;YACR,cAAc,EAAE,YAAY,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,WAAW,KAAK,WAAW,CAAC,CAAC,MAAM;YACpG,eAAe,EAAE,YAAY,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,WAAW,KAAK,SAAS,CAAC,CAAC,MAAM;YACnG,gBAAgB,EAAE,OAAO,CAAC,YAAY,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,SAAS,CAAC,CAAC;YACnF,yBAAyB,EAAE,OAAO,CAAC,YAAY,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,kBAAkB,CAAC,CAAC;YACrG,0BAA0B,EAAE,OAAO,CAAC,YAAY,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,WAAW,CAAC,mBAAmB,CAAC,CAAC;SACvG,CAAC;IAAA,CACF;IAED,eAAe,CAAC,QAAiB,EAA8B;QAC9D,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC,CAAC;QACtD,OAAO,CAAC,IAAI,EAAE,YAAY,IAAI,EAAE,CAAC;aAC/B,MAAM,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,QAAQ,KAAK,SAAS,IAAI,WAAW,CAAC,QAAQ,KAAK,QAAQ,CAAC;aACpF,GAAG,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC,CAAC;YACtB,GAAG,WAAW;YACd,SAAS,EAAE,CAAC,GAAG,WAAW,CAAC,SAAS,CAAC;YACrC,MAAM,EAAE,WAAW,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,SAAS,EAAE,CAAC,CAAC;SACjE,CAAC,CAAC,CAAC;IAAA,CACL;CACD","sourcesContent":["import { existsSync, mkdirSync, readFileSync, writeFileSync } from \"node:fs\";\nimport { dirname } from \"node:path\";\nimport { stateFile } from \"../agent-paths.ts\";\nimport { currentHostFingerprint, type HostFingerprint } from \"../models/fitness-store.ts\";\nimport { isWorkerSession } from \"../session-role.ts\";\n\nconst STORE_VERSION = 1;\nconst MAX_STATS_PER_HOST = 500;\nconst MAX_OBSERVATIONS_PER_HOST = 1_000;\nconst MAX_INTENT_AGREEMENT_PER_HOST = 500;\nconst EWMA_ALPHA = 0.25;\n\nexport type ToolSelectionIntentClass = \"read\" | \"search\" | \"execute\" | \"write\" | \"retrieve\" | \"explain\" | \"other\";\n\nexport interface ToolPerformanceKey {\n\tmodelRef: string;\n\tintentClass: ToolSelectionIntentClass;\n\ttool: string;\n}\n\nexport interface ToolPerformanceStats extends ToolPerformanceKey {\n\talpha: number;\n\tbeta: number;\n\tsampleCount: number;\n\tlatencyEwmaMs?: number;\n\tlatencyDeviationEwmaMs?: number;\n\tinputTokenEstimateEwma?: number;\n\toutputTokenEstimateEwma?: number;\n\trepairCount: number;\n\tbounceCount: number;\n\tfailureCount: number;\n\tlastUsedAt: string;\n}\n\nexport interface ToolSelectionObservation {\n\tat: string;\n\tmodelRef: string;\n\tintentClass: ToolSelectionIntentClass;\n\tactualTool: string;\n\tfirstTool: boolean;\n\tsucceeded: boolean;\n\tdisposition: \"recommend\" | \"shortlist\" | \"abstain\";\n\trecommendation?: string;\n\tshortlist: string[];\n\tentropy: number;\n\tmargin: number;\n\t/** Redacted ranking: names and numeric scores only, never prompts, arguments, paths, or output. */\n\tranked: Array<{ tool: string; utility: number; probability: number }>;\n\tlatencyMs?: number;\n\tinputTokenEstimate?: number;\n\toutputTokenEstimate?: number;\n}\n\nexport interface ToolSelectionMetrics {\n\tfirstToolAttempts: number;\n\tfirstToolSuccesses: number;\n\twrongToolOrFailureCount: number;\n\trecommendationCount: number;\n\trecommendationMatchedCount: number;\n\tshortlistCount: number;\n\tabstentionCount: number;\n\taverageLatencyMs?: number;\n\taverageInputTokenEstimate?: number;\n\taverageOutputTokenEstimate?: number;\n}\n\n/**\n * Durable, per-(model,intent) aggregate of the observe-mode loop: does the raw expected-utility\n * ranking's top pick (`ToolSelectionObservation.ranked[0]`) match what the model actually called\n * (`ToolSelectionObservation.actualTool`)? Tracked separately from the capped, rolling\n * `observations` log so evidence for a given (model,intent) pair survives that log's trimming.\n * `hintActive*` fields are the SAME agreement measure, but restricted to calls made while an\n * evidence-gated promotion hint (see promotion.ts) was active for this bucket — the hint's own\n * efficacy trace: it never gates activation directly (that is always\n * recomputed live from `ToolPerformanceStats`), but it is the durable evidence a report can show\n * for \"is the hint still earning its keep\".\n */\nexport interface ToolSelectionIntentAgreement {\n\tmodelRef: string;\n\tintentClass: ToolSelectionIntentClass;\n\tsampleCount: number;\n\tagreementCount: number;\n\thintActiveSampleCount: number;\n\thintActiveAgreementCount: number;\n\tlastUpdatedAt: string;\n}\n\ninterface HostToolPerformanceData {\n\thost: HostFingerprint;\n\tstats: Record<string, ToolPerformanceStats>;\n\tobservations: ToolSelectionObservation[];\n\tintentAgreement: Record<string, ToolSelectionIntentAgreement>;\n}\n\ninterface ToolPerformanceStoreFile {\n\tversion: 1;\n\thosts: Record<string, HostToolPerformanceData>;\n}\n\nexport interface ToolExecutionObservation {\n\tkey: ToolPerformanceKey;\n\tsuccess: boolean;\n\tlatencyMs: number;\n\tinputTokenEstimate?: number;\n\toutputTokenEstimate?: number;\n\tselection: Omit<ToolSelectionObservation, \"at\" | \"modelRef\" | \"intentClass\" | \"actualTool\" | \"succeeded\">;\n\tat?: string;\n\t/**\n\t * Whether an evidence-gated promotion hint (promotion.ts) was already active for this\n\t * (model,intent) bucket BEFORE this call was recorded — captured by the controller at\n\t * `begin()` time, so it reflects evidence up to but not including this observation. Used only\n\t * to bucket the durable agreement stats (`ToolSelectionIntentAgreement.hintActive*`); never\n\t * changes what gets recorded, only how it is split for the efficacy report.\n\t */\n\thintActiveAtCallTime?: boolean;\n}\n\nfunction statKey(key: ToolPerformanceKey): string {\n\treturn `${key.modelRef}\\0${key.intentClass}\\0${key.tool}`;\n}\n\nfunction intentAgreementKey(modelRef: string, intentClass: ToolSelectionIntentClass): string {\n\treturn `${modelRef}\\0${intentClass}`;\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n\treturn typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction isIntentClass(value: unknown): value is ToolSelectionIntentClass {\n\treturn (\n\t\tvalue === \"read\" ||\n\t\tvalue === \"search\" ||\n\t\tvalue === \"execute\" ||\n\t\tvalue === \"write\" ||\n\t\tvalue === \"retrieve\" ||\n\t\tvalue === \"explain\" ||\n\t\tvalue === \"other\"\n\t);\n}\n\nfunction isStats(value: unknown): value is ToolPerformanceStats {\n\treturn (\n\t\tisRecord(value) &&\n\t\ttypeof value.modelRef === \"string\" &&\n\t\tisIntentClass(value.intentClass) &&\n\t\ttypeof value.tool === \"string\" &&\n\t\ttypeof value.alpha === \"number\" &&\n\t\ttypeof value.beta === \"number\" &&\n\t\ttypeof value.sampleCount === \"number\" &&\n\t\ttypeof value.repairCount === \"number\" &&\n\t\ttypeof value.bounceCount === \"number\" &&\n\t\ttypeof value.failureCount === \"number\" &&\n\t\ttypeof value.lastUsedAt === \"string\"\n\t);\n}\n\nfunction isObservation(value: unknown): value is ToolSelectionObservation {\n\treturn (\n\t\tisRecord(value) &&\n\t\ttypeof value.at === \"string\" &&\n\t\ttypeof value.modelRef === \"string\" &&\n\t\tisIntentClass(value.intentClass) &&\n\t\ttypeof value.actualTool === \"string\" &&\n\t\ttypeof value.firstTool === \"boolean\" &&\n\t\ttypeof value.succeeded === \"boolean\" &&\n\t\t(value.disposition === \"recommend\" || value.disposition === \"shortlist\" || value.disposition === \"abstain\") &&\n\t\tArray.isArray(value.shortlist) &&\n\t\tvalue.shortlist.every((tool) => typeof tool === \"string\") &&\n\t\ttypeof value.entropy === \"number\" &&\n\t\ttypeof value.margin === \"number\" &&\n\t\tArray.isArray(value.ranked) &&\n\t\tvalue.ranked.every(\n\t\t\t(entry) =>\n\t\t\t\tisRecord(entry) &&\n\t\t\t\ttypeof entry.tool === \"string\" &&\n\t\t\t\ttypeof entry.utility === \"number\" &&\n\t\t\t\ttypeof entry.probability === \"number\",\n\t\t) &&\n\t\t(value.latencyMs === undefined || typeof value.latencyMs === \"number\") &&\n\t\t(value.inputTokenEstimate === undefined || typeof value.inputTokenEstimate === \"number\") &&\n\t\t(value.outputTokenEstimate === undefined || typeof value.outputTokenEstimate === \"number\")\n\t);\n}\n\nfunction isIntentAgreement(value: unknown): value is ToolSelectionIntentAgreement {\n\treturn (\n\t\tisRecord(value) &&\n\t\ttypeof value.modelRef === \"string\" &&\n\t\tisIntentClass(value.intentClass) &&\n\t\ttypeof value.sampleCount === \"number\" &&\n\t\ttypeof value.agreementCount === \"number\" &&\n\t\ttypeof value.hintActiveSampleCount === \"number\" &&\n\t\ttypeof value.hintActiveAgreementCount === \"number\" &&\n\t\ttypeof value.lastUpdatedAt === \"string\"\n\t);\n}\n\nfunction emptyStats(key: ToolPerformanceKey, at: string): ToolPerformanceStats {\n\treturn {\n\t\t...key,\n\t\talpha: 1,\n\t\tbeta: 1,\n\t\tsampleCount: 0,\n\t\trepairCount: 0,\n\t\tbounceCount: 0,\n\t\tfailureCount: 0,\n\t\tlastUsedAt: at,\n\t};\n}\n\nfunction emptyIntentAgreement(\n\tmodelRef: string,\n\tintentClass: ToolSelectionIntentClass,\n\tat: string,\n): ToolSelectionIntentAgreement {\n\treturn {\n\t\tmodelRef,\n\t\tintentClass,\n\t\tsampleCount: 0,\n\t\tagreementCount: 0,\n\t\thintActiveSampleCount: 0,\n\t\thintActiveAgreementCount: 0,\n\t\tlastUpdatedAt: at,\n\t};\n}\n\nfunction finiteNonNegative(value: number | undefined): number | undefined {\n\treturn typeof value === \"number\" && Number.isFinite(value) && value >= 0 ? value : undefined;\n}\n\nfunction updateEwma(previous: number | undefined, next: number | undefined): number | undefined {\n\tif (next === undefined) return previous;\n\treturn previous === undefined ? next : previous + EWMA_ALPHA * (next - previous);\n}\n\nfunction updateDeviation(\n\tpreviousValue: number | undefined,\n\tpreviousDeviation: number | undefined,\n\tnext: number | undefined,\n): number | undefined {\n\tif (next === undefined) return previousDeviation;\n\tconst deviation = previousValue === undefined ? 0 : Math.abs(next - previousValue);\n\treturn updateEwma(previousDeviation, deviation);\n}\n\nfunction parseHost(value: unknown): HostToolPerformanceData | undefined {\n\tif (!isRecord(value) || !isRecord(value.host) || !isRecord(value.stats) || !Array.isArray(value.observations))\n\t\treturn undefined;\n\tconst host = value.host;\n\tif (\n\t\ttypeof host.id !== \"string\" ||\n\t\ttypeof host.cpu !== \"string\" ||\n\t\ttypeof host.cores !== \"number\" ||\n\t\ttypeof host.totalMemGb !== \"number\"\n\t)\n\t\treturn undefined;\n\t// intentAgreement is a purely additive field (older store files predate it) — tolerate absence\n\t// rather than bumping STORE_VERSION, same as any other backward-compatible default-empty field.\n\tconst intentAgreementRaw = isRecord(value.intentAgreement) ? value.intentAgreement : {};\n\treturn {\n\t\thost: { id: host.id, cpu: host.cpu, cores: host.cores, totalMemGb: host.totalMemGb },\n\t\tstats: Object.fromEntries(\n\t\t\tObject.entries(value.stats).filter((entry): entry is [string, ToolPerformanceStats] => isStats(entry[1])),\n\t\t),\n\t\tobservations: value.observations.filter(isObservation),\n\t\tintentAgreement: Object.fromEntries(\n\t\t\tObject.entries(intentAgreementRaw).filter((entry): entry is [string, ToolSelectionIntentAgreement] =>\n\t\t\t\tisIntentAgreement(entry[1]),\n\t\t\t),\n\t\t),\n\t};\n}\n\nfunction parseFile(value: unknown): ToolPerformanceStoreFile {\n\tif (!isRecord(value) || value.version !== STORE_VERSION || !isRecord(value.hosts)) {\n\t\treturn { version: STORE_VERSION, hosts: {} };\n\t}\n\treturn {\n\t\tversion: STORE_VERSION,\n\t\thosts: Object.fromEntries(\n\t\t\tObject.entries(value.hosts)\n\t\t\t\t.map(([hostId, host]) => [hostId, parseHost(host)] as const)\n\t\t\t\t.filter((entry): entry is readonly [string, HostToolPerformanceData] => entry[1] !== undefined),\n\t\t),\n\t};\n}\n\nfunction trimStats(stats: Record<string, ToolPerformanceStats>): Record<string, ToolPerformanceStats> {\n\tconst entries = Object.entries(stats);\n\tif (entries.length <= MAX_STATS_PER_HOST) return stats;\n\treturn Object.fromEntries(\n\t\tentries\n\t\t\t.sort(([, left], [, right]) => Date.parse(right.lastUsedAt) - Date.parse(left.lastUsedAt))\n\t\t\t.slice(0, MAX_STATS_PER_HOST),\n\t);\n}\n\nfunction trimIntentAgreement(\n\trecords: Record<string, ToolSelectionIntentAgreement>,\n): Record<string, ToolSelectionIntentAgreement> {\n\tconst entries = Object.entries(records);\n\tif (entries.length <= MAX_INTENT_AGREEMENT_PER_HOST) return records;\n\treturn Object.fromEntries(\n\t\tentries\n\t\t\t.sort(([, left], [, right]) => Date.parse(right.lastUpdatedAt) - Date.parse(left.lastUpdatedAt))\n\t\t\t.slice(0, MAX_INTENT_AGREEMENT_PER_HOST),\n\t);\n}\n\nexport class ToolPerformanceStore {\n\tprivate readonly filePath: string;\n\tprivate readonly fingerprint: () => HostFingerprint;\n\tprivate readonly readOnly: boolean;\n\n\tconstructor(filePath: string, options: { fingerprint?: () => HostFingerprint; readOnly?: boolean } = {}) {\n\t\tthis.filePath = filePath;\n\t\tthis.fingerprint = options.fingerprint ?? currentHostFingerprint;\n\t\tthis.readOnly = options.readOnly ?? isWorkerSession();\n\t}\n\n\tstatic forAgentDir(\n\t\tagentDir: string,\n\t\toptions: { fingerprint?: () => HostFingerprint; readOnly?: boolean } = {},\n\t): ToolPerformanceStore {\n\t\treturn new ToolPerformanceStore(stateFile(agentDir, \"tool-performance.json\"), options);\n\t}\n\n\tprivate load(): ToolPerformanceStoreFile {\n\t\ttry {\n\t\t\tif (!existsSync(this.filePath)) return { version: STORE_VERSION, hosts: {} };\n\t\t\treturn parseFile(JSON.parse(readFileSync(this.filePath, \"utf8\")));\n\t\t} catch {\n\t\t\treturn { version: STORE_VERSION, hosts: {} };\n\t\t}\n\t}\n\n\tprivate save(file: ToolPerformanceStoreFile): void {\n\t\t// Zero-footprint (worker session): never create the state dir, lock, or file -- the caller\n\t\t// still gets its normally-computed return value from the in-memory `file`.\n\t\tif (this.readOnly) return;\n\t\tmkdirSync(dirname(this.filePath), { recursive: true });\n\t\twriteFileSync(this.filePath, `${JSON.stringify(file, null, \"\\t\")}\\n`, \"utf8\");\n\t}\n\n\tprivate hostData(file: ToolPerformanceStoreFile): HostToolPerformanceData {\n\t\tconst fingerprint = this.fingerprint();\n\t\tconst existing = file.hosts[fingerprint.id];\n\t\tif (existing) return existing;\n\t\tconst created: HostToolPerformanceData = { host: fingerprint, stats: {}, observations: [], intentAgreement: {} };\n\t\tfile.hosts[fingerprint.id] = created;\n\t\treturn created;\n\t}\n\n\tget(key: ToolPerformanceKey): ToolPerformanceStats {\n\t\tconst file = this.load();\n\t\tconst host = file.hosts[this.fingerprint().id];\n\t\tconst stats = host?.stats[statKey(key)];\n\t\treturn stats ? { ...stats } : emptyStats(key, new Date(0).toISOString());\n\t}\n\n\t/** Every per-tool track record recorded for a (model,intent) bucket — the promotion.ts input. */\n\tgetStatsForIntent(modelRef: string, intentClass: ToolSelectionIntentClass): ToolPerformanceStats[] {\n\t\tconst host = this.load().hosts[this.fingerprint().id];\n\t\tif (!host) return [];\n\t\treturn Object.values(host.stats)\n\t\t\t.filter((stats) => stats.modelRef === modelRef && stats.intentClass === intentClass)\n\t\t\t.map((stats) => ({ ...stats }));\n\t}\n\n\t/** Durable observe-mode agreement for one (model,intent) bucket (see {@link ToolSelectionIntentAgreement}). */\n\tgetIntentAgreement(modelRef: string, intentClass: ToolSelectionIntentClass): ToolSelectionIntentAgreement {\n\t\tconst host = this.load().hosts[this.fingerprint().id];\n\t\tconst record = host?.intentAgreement[intentAgreementKey(modelRef, intentClass)];\n\t\treturn record ? { ...record } : emptyIntentAgreement(modelRef, intentClass, new Date(0).toISOString());\n\t}\n\n\t/** All recorded (model,intent) agreement buckets, optionally scoped to one model — report input. */\n\tgetAllIntentAgreements(modelRef?: string): ToolSelectionIntentAgreement[] {\n\t\tconst host = this.load().hosts[this.fingerprint().id];\n\t\treturn Object.values(host?.intentAgreement ?? {})\n\t\t\t.filter((record) => modelRef === undefined || record.modelRef === modelRef)\n\t\t\t.map((record) => ({ ...record }));\n\t}\n\n\trecordValidation(\n\t\tkey: ToolPerformanceKey,\n\t\toutcome: \"repaired\" | \"bounced\",\n\t\tat = new Date().toISOString(),\n\t): ToolPerformanceStats {\n\t\tconst file = this.load();\n\t\tconst host = this.hostData(file);\n\t\tconst storageKey = statKey(key);\n\t\tconst current = host.stats[storageKey] ?? emptyStats(key, at);\n\t\tconst next: ToolPerformanceStats = {\n\t\t\t...current,\n\t\t\trepairCount: current.repairCount + (outcome === \"repaired\" ? 1 : 0),\n\t\t\tbounceCount: current.bounceCount + (outcome === \"bounced\" ? 1 : 0),\n\t\t\tlastUsedAt: at,\n\t\t};\n\t\thost.stats[storageKey] = next;\n\t\thost.stats = trimStats(host.stats);\n\t\tthis.save(file);\n\t\treturn { ...next };\n\t}\n\n\trecordExecution(observation: ToolExecutionObservation): ToolPerformanceStats {\n\t\tconst at = observation.at ?? new Date().toISOString();\n\t\tconst file = this.load();\n\t\tconst host = this.hostData(file);\n\t\tconst storageKey = statKey(observation.key);\n\t\tconst current = host.stats[storageKey] ?? emptyStats(observation.key, at);\n\t\tconst latencyMs = finiteNonNegative(observation.latencyMs);\n\t\tconst inputTokenEstimate = finiteNonNegative(observation.inputTokenEstimate);\n\t\tconst outputTokenEstimate = finiteNonNegative(observation.outputTokenEstimate);\n\t\tconst next: ToolPerformanceStats = {\n\t\t\t...current,\n\t\t\talpha: current.alpha + (observation.success ? 1 : 0),\n\t\t\tbeta: current.beta + (observation.success ? 0 : 1),\n\t\t\tsampleCount: current.sampleCount + 1,\n\t\t\tlatencyEwmaMs: updateEwma(current.latencyEwmaMs, latencyMs),\n\t\t\tlatencyDeviationEwmaMs: updateDeviation(current.latencyEwmaMs, current.latencyDeviationEwmaMs, latencyMs),\n\t\t\tinputTokenEstimateEwma: updateEwma(current.inputTokenEstimateEwma, inputTokenEstimate),\n\t\t\toutputTokenEstimateEwma: updateEwma(current.outputTokenEstimateEwma, outputTokenEstimate),\n\t\t\tfailureCount: current.failureCount + (observation.success ? 0 : 1),\n\t\t\tlastUsedAt: at,\n\t\t};\n\t\thost.stats[storageKey] = next;\n\t\thost.stats = trimStats(host.stats);\n\t\thost.observations.push({\n\t\t\t...observation.selection,\n\t\t\tat,\n\t\t\tmodelRef: observation.key.modelRef,\n\t\t\tintentClass: observation.key.intentClass,\n\t\t\tactualTool: observation.key.tool,\n\t\t\tsucceeded: observation.success,\n\t\t\tranked: observation.selection.ranked.slice(0, 6),\n\t\t\tshortlist: observation.selection.shortlist.slice(0, 3),\n\t\t\tlatencyMs,\n\t\t\tinputTokenEstimate,\n\t\t\toutputTokenEstimate,\n\t\t});\n\t\tif (host.observations.length > MAX_OBSERVATIONS_PER_HOST) {\n\t\t\thost.observations = host.observations.slice(-MAX_OBSERVATIONS_PER_HOST);\n\t\t}\n\n\t\t// Observe-mode agreement: did the RAW ranking's top pick (before any actual-tool-only\n\t\t// eligibility gating) match what the model actually called? Recorded durably per\n\t\t// (model,intent), separate from the capped `observations` log above, so it survives trimming.\n\t\tconst predictedBest = observation.selection.ranked[0]?.tool;\n\t\tconst agreed = predictedBest !== undefined && predictedBest === observation.key.tool;\n\t\tconst agreementKey = intentAgreementKey(observation.key.modelRef, observation.key.intentClass);\n\t\tconst currentAgreement =\n\t\t\thost.intentAgreement[agreementKey] ??\n\t\t\temptyIntentAgreement(observation.key.modelRef, observation.key.intentClass, at);\n\t\thost.intentAgreement[agreementKey] = {\n\t\t\t...currentAgreement,\n\t\t\tsampleCount: currentAgreement.sampleCount + 1,\n\t\t\tagreementCount: currentAgreement.agreementCount + (agreed ? 1 : 0),\n\t\t\thintActiveSampleCount: currentAgreement.hintActiveSampleCount + (observation.hintActiveAtCallTime ? 1 : 0),\n\t\t\thintActiveAgreementCount:\n\t\t\t\tcurrentAgreement.hintActiveAgreementCount + (observation.hintActiveAtCallTime && agreed ? 1 : 0),\n\t\t\tlastUpdatedAt: at,\n\t\t};\n\t\thost.intentAgreement = trimIntentAgreement(host.intentAgreement);\n\n\t\tthis.save(file);\n\t\treturn { ...next };\n\t}\n\n\tgetMetrics(modelRef?: string): ToolSelectionMetrics {\n\t\tconst host = this.load().hosts[this.fingerprint().id];\n\t\tconst observations = (host?.observations ?? []).filter(\n\t\t\t(observation) => modelRef === undefined || observation.modelRef === modelRef,\n\t\t);\n\t\tconst firstTools = observations.filter((observation) => observation.firstTool);\n\t\tconst recommended = observations.filter((observation) => observation.disposition === \"recommend\");\n\t\tconst average = (values: Array<number | undefined>): number | undefined => {\n\t\t\tconst present = values.filter((value): value is number => value !== undefined);\n\t\t\treturn present.length > 0 ? present.reduce((sum, value) => sum + value, 0) / present.length : undefined;\n\t\t};\n\t\treturn {\n\t\t\tfirstToolAttempts: firstTools.length,\n\t\t\tfirstToolSuccesses: firstTools.filter((observation) => observation.succeeded).length,\n\t\t\twrongToolOrFailureCount: observations.filter(\n\t\t\t\t(observation) =>\n\t\t\t\t\t!observation.succeeded ||\n\t\t\t\t\t(observation.disposition === \"recommend\" && observation.recommendation !== observation.actualTool),\n\t\t\t).length,\n\t\t\trecommendationCount: recommended.length,\n\t\t\trecommendationMatchedCount: recommended.filter(\n\t\t\t\t(observation) => observation.recommendation === observation.actualTool,\n\t\t\t).length,\n\t\t\tshortlistCount: observations.filter((observation) => observation.disposition === \"shortlist\").length,\n\t\t\tabstentionCount: observations.filter((observation) => observation.disposition === \"abstain\").length,\n\t\t\taverageLatencyMs: average(observations.map((observation) => observation.latencyMs)),\n\t\t\taverageInputTokenEstimate: average(observations.map((observation) => observation.inputTokenEstimate)),\n\t\t\taverageOutputTokenEstimate: average(observations.map((observation) => observation.outputTokenEstimate)),\n\t\t};\n\t}\n\n\tgetObservations(modelRef?: string): ToolSelectionObservation[] {\n\t\tconst host = this.load().hosts[this.fingerprint().id];\n\t\treturn (host?.observations ?? [])\n\t\t\t.filter((observation) => modelRef === undefined || observation.modelRef === modelRef)\n\t\t\t.map((observation) => ({\n\t\t\t\t...observation,\n\t\t\t\tshortlist: [...observation.shortlist],\n\t\t\t\tranked: observation.ranked.map((candidate) => ({ ...candidate })),\n\t\t\t}));\n\t}\n}\n"]}

package/dist/core/tool-selection/tool-selection-controller.d.ts

@@ -1,3 +1,4 @@
+import { type ToolSelectionHint } from "./promotion.ts";
 import type { ToolPerformanceKey, ToolPerformanceStore, ToolSelectionIntentClass, ToolSelectionObservation } from "./tool-performance-store.ts";
 export interface ToolSelectionTool {
     name: string;
@@ -14,6 +15,13 @@ export interface ToolSelectionControllerDeps {
     getModelRef: () => string;
     getActiveTools: () => readonly ToolSelectionTool[];
     isCandidateAllowed?: (toolName: string) => boolean;
+    /**
+     * Env-var source for the kill switches, injected for hermetic testing; defaults to
+     * `process.env`. `PI_TOOL_SELECTION_OBSERVE=0` disables recording (the observe/stats layer,
+     * default ON); `PI_TOOL_SELECTION_HINTS=0` disables surfacing the evidence-gated prompt hint
+     * (default ON, but the hint itself only ever activates once evidence thresholds are met).
+     */
+    env?: Record<string, string | undefined>;
 }
 export interface ToolSelectionPendingObservation {
     id: string;
@@ -22,6 +30,18 @@ export interface ToolSelectionPendingObservation {
     startedAt: number;
     inputTokenEstimate?: number;
     selection: Omit<ToolSelectionObservation, "at" | "modelRef" | "intentClass" | "actualTool" | "succeeded">;
+    /** Was a promotion hint already active for this (model,intent) bucket before this call? */
+    hintActiveAtCallTime: boolean;
+}
+/** Report row for the observe/agreement/promotion loop — see {@link ToolSelectionController.getReport}. */
+export interface ToolSelectionReportEntry {
+    modelRef: string;
+    intentClass: ToolSelectionIntentClass;
+    sampleCount: number;
+    agreementRate?: number;
+    hintTool?: string;
+    hintSampleCount: number;
+    hintAgreementRate?: number;
 }
 declare function classifyToolIntent(tool: ToolSelectionTool): ToolSelectionIntentClass;
 declare function estimateTokens(value: unknown): number | undefined;
@@ -30,12 +50,37 @@ export declare class ToolSelectionController {
     private readonly deps;
     private readonly pending;
     private firstToolInTurn;
+    /** Kill switch: observe/stats recording, default ON. `PI_TOOL_SELECTION_OBSERVE=0` disables it. */
+    private readonly observeEnabled;
     constructor(deps: ToolSelectionControllerDeps);
     startTurn(): void;
     begin(toolCallId: string, toolName: string, args: unknown): ToolSelectionPendingObservation;
     complete(toolCallId: string, succeeded: boolean, content?: readonly unknown[]): void;
     recordValidation(toolName: string, outcome: "repaired" | "bounced"): void;
+    /**
+     * Evidence-gated promotion for one (model,intent) bucket — see promotion.ts. Pure read, no
+     * mutation; used both to stamp `hintActiveAtCallTime` and to build the live hint list below.
+     */
+    private evaluatePromotion;
+    /**
+     * The currently active evidence-gated hints for the session's current model — one per
+     * (intentClass) at most, only where accumulated evidence clears the promotion gate. Consumed by
+     * `system-prompt-builder.ts` to render a compact, cache-stable prompt block. Empty when the
+     * hint kill switch (`PI_TOOL_SELECTION_HINTS=0`) is set, observing is disabled, or no bucket has
+     * cleared the gate yet.
+     */
+    getActiveHints(modelRef?: string): ToolSelectionHint[];
+    /**
+     * Report surface for the observe/agreement/promotion loop: per (model,intent), the durable
+     * agreement rate (did the raw ranking's top pick match what was actually called), plus the
+     * currently active hint (if any) and its own efficacy (agreement rate while it has been active).
+     * A read-only diagnostic — never used to gate behavior. Render with
+     * {@link formatToolSelectionReport}.
+     */
+    getReport(modelRef?: string): ToolSelectionReportEntry[];
     private selectionSnapshot;
 }
+/** Renders {@link ToolSelectionController.getReport} rows into the /toolhealth-style diagnostic text. */
+export declare function formatToolSelectionReport(entries: readonly ToolSelectionReportEntry[]): string;
 export { classifyToolIntent, estimateContentTokens, estimateTokens };
 //# sourceMappingURL=tool-selection-controller.d.ts.map

package/dist/core/tool-selection/tool-selection-controller.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"tool-selection-controller.d.ts","sourceRoot":"","sources":["../../../src/core/tool-selection/tool-selection-controller.ts"],"names":[],"mappings":"AAKA,OAAO,KAAK,EAEX,kBAAkB,EAClB,oBAAoB,EACpB,wBAAwB,EACxB,wBAAwB,EACxB,MAAM,6BAA6B,CAAC;AAErC,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,UAAU,CAAC,EAAE,OAAO,CAAC;IACrB,cAAc,CAAC,EAAE,OAAO,CAAC;IACzB,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,WAAW,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,2BAA2B;IAC3C,KAAK,EAAE,oBAAoB,CAAC;IAC5B,WAAW,EAAE,MAAM,MAAM,CAAC;IAC1B,cAAc,EAAE,MAAM,SAAS,iBAAiB,EAAE,CAAC;IACnD,kBAAkB,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,OAAO,CAAC;CACnD;AAED,MAAM,WAAW,+BAA+B;IAC/C,EAAE,EAAE,MAAM,CAAC;IACX,GAAG,EAAE,kBAAkB,CAAC;IACxB,SAAS,EAAE,OAAO,CAAC;IACnB,SAAS,EAAE,MAAM,CAAC;IAClB,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,SAAS,EAAE,IAAI,CAAC,wBAAwB,EAAE,IAAI,GAAG,UAAU,GAAG,aAAa,GAAG,YAAY,GAAG,WAAW,CAAC,CAAC;CAC1G;AA6BD,iBAAS,kBAAkB,CAAC,IAAI,EAAE,iBAAiB,GAAG,wBAAwB,CAa7E;AAyBD,iBAAS,cAAc,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,GAAG,SAAS,CAQ1D;AAED,iBAAS,qBAAqB,CAAC,OAAO,EAAE,SAAS,OAAO,EAAE,GAAG,MAAM,GAAG,SAAS,CAS9E;AAiCD,qBAAa,uBAAuB;IACnC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAA8B;IACnD,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAsD;IAC9E,OAAO,CAAC,eAAe,CAAQ;IAE/B,YAAY,IAAI,EAAE,2BAA2B,EAE5C;IAED,SAAS,IAAI,IAAI,CAEhB;IAED,KAAK,CAAC,UAAU,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,OAAO,GAAG,+BAA+B,CAqD1F;IAED,QAAQ,CAAC,UAAU,EAAE,MAAM,EAAE,SAAS,EAAE,OAAO,EAAE,OAAO,GAAE,SAAS,OAAO,EAAO,GAAG,IAAI,CAavF;IAED,gBAAgB,CAAC,QAAQ,EAAE,MAAM,EAAE,OAAO,EAAE,UAAU,GAAG,SAAS,GAAG,IAAI,CAIxE;IAED,OAAO,CAAC,iBAAiB;CAkBzB;AAED,OAAO,EAAE,kBAAkB,EAAE,qBAAqB,EAAE,cAAc,EAAE,CAAC","sourcesContent":["import {\n\tdecideExpectedUtility,\n\ttype ExpectedUtilityCandidate,\n\ttype ToolSelectionDecision,\n} from \"./expected-utility.ts\";\nimport type {\n\tToolExecutionObservation,\n\tToolPerformanceKey,\n\tToolPerformanceStore,\n\tToolSelectionIntentClass,\n\tToolSelectionObservation,\n} from \"./tool-performance-store.ts\";\n\nexport interface ToolSelectionTool {\n\tname: string;\n\tdescription?: string;\n\tparameters?: unknown;\n\tprofileAllowed?: boolean;\n\tcapabilityAllowed?: boolean;\n\tpathValidated?: boolean;\n\triskCost?: number;\n\tcontextCost?: number;\n}\n\nexport interface ToolSelectionControllerDeps {\n\tstore: ToolPerformanceStore;\n\tgetModelRef: () => string;\n\tgetActiveTools: () => readonly ToolSelectionTool[];\n\tisCandidateAllowed?: (toolName: string) => boolean;\n}\n\nexport interface ToolSelectionPendingObservation {\n\tid: string;\n\tkey: ToolPerformanceKey;\n\tfirstTool: boolean;\n\tstartedAt: number;\n\tinputTokenEstimate?: number;\n\tselection: Omit<ToolSelectionObservation, \"at\" | \"modelRef\" | \"intentClass\" | \"actualTool\" | \"succeeded\">;\n}\n\nconst INTENT_CLASSES: readonly ToolSelectionIntentClass[] = [\n\t\"read\",\n\t\"search\",\n\t\"execute\",\n\t\"write\",\n\t\"retrieve\",\n\t\"explain\",\n\t\"other\",\n];\n\nconst INTENT_WORDS: Record<ToolSelectionIntentClass, readonly string[]> = {\n\tread: [\"read\", \"cat\", \"ls\", \"list\", \"stat\", \"head\", \"tail\"],\n\tsearch: [\"search\", \"grep\", \"find\", \"glob\", \"ripgrep\", \"query\"],\n\texecute: [\"bash\", \"python\", \"powershell\", \"shell\", \"exec\", \"run\", \"process\", \"command\", \"script\"],\n\twrite: [\"write\", \"edit\", \"patch\", \"apply\", \"delete\", \"mkdir\", \"move\", \"rename\"],\n\tretrieve: [\"retrieve\", \"artifact\", \"recall\", \"memory\", \"fetch\", \"download\"],\n\texplain: [\"explain\", \"help\", \"describe\", \"inspect\", \"context\"],\n\tother: [],\n};\n\nfunction tokenize(value: string): string[] {\n\treturn value\n\t\t.toLowerCase()\n\t\t.split(/[^a-z0-9]+/u)\n\t\t.filter(Boolean);\n}\n\nfunction classifyToolIntent(tool: ToolSelectionTool): ToolSelectionIntentClass {\n\tconst tokens = tokenize(`${tool.name} ${tool.description ?? \"\"}`);\n\tlet best: ToolSelectionIntentClass = \"other\";\n\tlet bestScore = 0;\n\tfor (const intent of INTENT_CLASSES) {\n\t\tif (intent === \"other\") continue;\n\t\tconst score = INTENT_WORDS[intent].reduce((total, word) => total + (tokens.includes(word) ? 1 : 0), 0);\n\t\tif (score > bestScore) {\n\t\t\tbest = intent;\n\t\t\tbestScore = score;\n\t\t}\n\t}\n\treturn best;\n}\n\nfunction intentValue(intent: ToolSelectionIntentClass, toolIntent: ToolSelectionIntentClass): number {\n\tif (toolIntent === intent) return 0.85;\n\tif (intent === \"other\") return 0.35;\n\tif (toolIntent === \"other\") return 0.2;\n\treturn 0.12;\n}\n\nfunction noToolValue(intent: ToolSelectionIntentClass): number {\n\tswitch (intent) {\n\t\tcase \"explain\":\n\t\t\treturn 0.35;\n\t\tcase \"read\":\n\t\tcase \"search\":\n\t\tcase \"retrieve\":\n\t\t\treturn 0.12;\n\t\tcase \"write\":\n\t\tcase \"execute\":\n\t\t\treturn 0.05;\n\t\tdefault:\n\t\t\treturn 0.2;\n\t}\n}\n\nfunction estimateTokens(value: unknown): number | undefined {\n\ttry {\n\t\tconst serialized = typeof value === \"string\" ? value : JSON.stringify(value);\n\t\tif (!serialized) return undefined;\n\t\treturn Math.max(1, Math.ceil(serialized.length / 4));\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction estimateContentTokens(content: readonly unknown[]): number | undefined {\n\tconst text = content\n\t\t.map((block) => {\n\t\t\tif (!block || typeof block !== \"object\") return \"\";\n\t\t\tconst candidate = block as { type?: unknown; text?: unknown };\n\t\t\treturn candidate.type === \"text\" && typeof candidate.text === \"string\" ? candidate.text : \"\";\n\t\t})\n\t\t.join(\"\\n\");\n\treturn estimateTokens(text);\n}\n\nfunction modelToolKey(modelRef: string, intentClass: ToolSelectionIntentClass, tool: string): ToolPerformanceKey {\n\treturn { modelRef, intentClass, tool };\n}\n\nfunction candidateFor(\n\tstore: ToolPerformanceStore,\n\tmodelRef: string,\n\tintentClass: ToolSelectionIntentClass,\n\ttool: ToolSelectionTool,\n\tactualTool: string,\n): ExpectedUtilityCandidate {\n\tconst key = modelToolKey(modelRef, intentClass, tool.name);\n\tconst stats = store.get(key);\n\treturn {\n\t\ttool: tool.name,\n\t\tvalue: intentValue(intentClass, classifyToolIntent(tool)),\n\t\talpha: stats.alpha,\n\t\tbeta: stats.beta,\n\t\tsampleCount: stats.sampleCount,\n\t\tlatencyMs: stats.latencyEwmaMs,\n\t\ttokenEstimate:\n\t\t\tstats.inputTokenEstimateEwma === undefined && stats.outputTokenEstimateEwma === undefined\n\t\t\t\t? undefined\n\t\t\t\t: (stats.inputTokenEstimateEwma ?? 0) + (stats.outputTokenEstimateEwma ?? 0),\n\t\triskCost: tool.riskCost,\n\t\tcontextCost: tool.contextCost,\n\t\tdeterministicMatch: tool.name === actualTool && tool.pathValidated !== false,\n\t\tautomaticEligible: tool.pathValidated !== false,\n\t};\n}\n\nexport class ToolSelectionController {\n\tprivate readonly deps: ToolSelectionControllerDeps;\n\tprivate readonly pending = new Map<string, ToolSelectionPendingObservation>();\n\tprivate firstToolInTurn = true;\n\n\tconstructor(deps: ToolSelectionControllerDeps) {\n\t\tthis.deps = deps;\n\t}\n\n\tstartTurn(): void {\n\t\tthis.firstToolInTurn = true;\n\t}\n\n\tbegin(toolCallId: string, toolName: string, args: unknown): ToolSelectionPendingObservation {\n\t\tconst modelRef = this.deps.getModelRef();\n\t\tconst activeTools = this.deps\n\t\t\t.getActiveTools()\n\t\t\t.filter(\n\t\t\t\t(tool) =>\n\t\t\t\t\ttool.profileAllowed !== false &&\n\t\t\t\t\ttool.capabilityAllowed !== false &&\n\t\t\t\t\t(this.deps.isCandidateAllowed?.(tool.name) ?? true),\n\t\t\t);\n\t\tconst actualTool = activeTools.find((tool) => tool.name === toolName) ?? {\n\t\t\tname: toolName,\n\t\t\tpathValidated: true,\n\t\t};\n\t\tconst intentClass = classifyToolIntent(actualTool);\n\t\tconst candidates = activeTools.map((tool) =>\n\t\t\tcandidateFor(\n\t\t\t\tthis.deps.store,\n\t\t\t\tmodelRef,\n\t\t\t\tintentClass,\n\t\t\t\ttool.name === toolName ? tool : { ...tool, pathValidated: false },\n\t\t\t\ttoolName,\n\t\t\t),\n\t\t);\n\t\tif (!candidates.some((candidate) => candidate.tool === toolName)) {\n\t\t\tcandidates.push(candidateFor(this.deps.store, modelRef, intentClass, actualTool, toolName));\n\t\t}\n\t\tcandidates.push({\n\t\t\ttool: \"no_tool\",\n\t\t\tvalue: noToolValue(intentClass),\n\t\t\talpha: 1,\n\t\t\tbeta: 1,\n\t\t\tsampleCount: 0,\n\t\t\tsuccessProbability: 1,\n\t\t\tlatencyMs: 0,\n\t\t\ttokenEstimate: 0,\n\t\t\triskCost: 0,\n\t\t\tcontextCost: 0,\n\t\t});\n\t\tconst decision = decideExpectedUtility(candidates);\n\t\tconst firstTool = this.firstToolInTurn;\n\t\tconst selection = this.selectionSnapshot(decision, firstTool);\n\t\tconst pending: ToolSelectionPendingObservation = {\n\t\t\tid: toolCallId,\n\t\t\tkey: modelToolKey(modelRef, intentClass, toolName),\n\t\t\tfirstTool: this.firstToolInTurn,\n\t\t\tstartedAt: Date.now(),\n\t\t\tinputTokenEstimate: estimateTokens(args),\n\t\t\tselection,\n\t\t};\n\t\tthis.firstToolInTurn = false;\n\t\tthis.pending.set(toolCallId, pending);\n\t\treturn pending;\n\t}\n\n\tcomplete(toolCallId: string, succeeded: boolean, content: readonly unknown[] = []): void {\n\t\tconst pending = this.pending.get(toolCallId);\n\t\tif (!pending) return;\n\t\tthis.pending.delete(toolCallId);\n\t\tconst execution: ToolExecutionObservation = {\n\t\t\tkey: pending.key,\n\t\t\tsuccess: succeeded,\n\t\t\tlatencyMs: Math.max(0, Date.now() - pending.startedAt),\n\t\t\tinputTokenEstimate: pending.inputTokenEstimate,\n\t\t\toutputTokenEstimate: estimateContentTokens(content),\n\t\t\tselection: pending.selection,\n\t\t};\n\t\tthis.deps.store.recordExecution(execution);\n\t}\n\n\trecordValidation(toolName: string, outcome: \"repaired\" | \"bounced\"): void {\n\t\tconst modelRef = this.deps.getModelRef();\n\t\tconst tool = this.deps.getActiveTools().find((candidate) => candidate.name === toolName) ?? { name: toolName };\n\t\tthis.deps.store.recordValidation(modelToolKey(modelRef, classifyToolIntent(tool), toolName), outcome);\n\t}\n\n\tprivate selectionSnapshot(\n\t\tdecision: ToolSelectionDecision,\n\t\tfirstTool: boolean,\n\t): ToolSelectionPendingObservation[\"selection\"] {\n\t\treturn {\n\t\t\tfirstTool,\n\t\t\tdisposition: decision.disposition,\n\t\t\trecommendation: decision.recommendation,\n\t\t\tshortlist: decision.shortlist,\n\t\t\tentropy: decision.entropy,\n\t\t\tmargin: decision.margin,\n\t\t\tranked: decision.ranked.map((candidate) => ({\n\t\t\t\ttool: candidate.tool,\n\t\t\t\tutility: candidate.utility,\n\t\t\t\tprobability: candidate.probability,\n\t\t\t})),\n\t\t};\n\t}\n}\n\nexport { classifyToolIntent, estimateContentTokens, estimateTokens };\n"]}
+{"version":3,"file":"tool-selection-controller.d.ts","sourceRoot":"","sources":["../../../src/core/tool-selection/tool-selection-controller.ts"],"names":[],"mappings":"AAKA,OAAO,EAAyB,KAAK,iBAAiB,EAAE,MAAM,gBAAgB,CAAC;AAC/E,OAAO,KAAK,EAEX,kBAAkB,EAClB,oBAAoB,EACpB,wBAAwB,EACxB,wBAAwB,EACxB,MAAM,6BAA6B,CAAC;AAErC,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,UAAU,CAAC,EAAE,OAAO,CAAC;IACrB,cAAc,CAAC,EAAE,OAAO,CAAC;IACzB,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,WAAW,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,2BAA2B;IAC3C,KAAK,EAAE,oBAAoB,CAAC;IAC5B,WAAW,EAAE,MAAM,MAAM,CAAC;IAC1B,cAAc,EAAE,MAAM,SAAS,iBAAiB,EAAE,CAAC;IACnD,kBAAkB,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,OAAO,CAAC;IACnD;;;;;OAKG;IACH,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,GAAG,SAAS,CAAC,CAAC;CACzC;AAED,MAAM,WAAW,+BAA+B;IAC/C,EAAE,EAAE,MAAM,CAAC;IACX,GAAG,EAAE,kBAAkB,CAAC;IACxB,SAAS,EAAE,OAAO,CAAC;IACnB,SAAS,EAAE,MAAM,CAAC;IAClB,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,SAAS,EAAE,IAAI,CAAC,wBAAwB,EAAE,IAAI,GAAG,UAAU,GAAG,aAAa,GAAG,YAAY,GAAG,WAAW,CAAC,CAAC;IAC1G,2FAA2F;IAC3F,oBAAoB,EAAE,OAAO,CAAC;CAC9B;AAED,6GAA2G;AAC3G,MAAM,WAAW,wBAAwB;IACxC,QAAQ,EAAE,MAAM,CAAC;IACjB,WAAW,EAAE,wBAAwB,CAAC;IACtC,WAAW,EAAE,MAAM,CAAC;IACpB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,eAAe,EAAE,MAAM,CAAC;IACxB,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC3B;AA6BD,iBAAS,kBAAkB,CAAC,IAAI,EAAE,iBAAiB,GAAG,wBAAwB,CAa7E;AAyBD,iBAAS,cAAc,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,GAAG,SAAS,CAQ1D;AAED,iBAAS,qBAAqB,CAAC,OAAO,EAAE,SAAS,OAAO,EAAE,GAAG,MAAM,GAAG,SAAS,CAS9E;AAiCD,qBAAa,uBAAuB;IACnC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAA8B;IACnD,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAsD;IAC9E,OAAO,CAAC,eAAe,CAAQ;IAC/B,mGAAmG;IACnG,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAU;IAEzC,YAAY,IAAI,EAAE,2BAA2B,EAG5C;IAED,SAAS,IAAI,IAAI,CAEhB;IAED,KAAK,CAAC,UAAU,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,OAAO,GAAG,+BAA+B,CA4D1F;IAED,QAAQ,CAAC,UAAU,EAAE,MAAM,EAAE,SAAS,EAAE,OAAO,EAAE,OAAO,GAAE,SAAS,OAAO,EAAO,GAAG,IAAI,CAevF;IAED,gBAAgB,CAAC,QAAQ,EAAE,MAAM,EAAE,OAAO,EAAE,UAAU,GAAG,SAAS,GAAG,IAAI,CAKxE;IAED;;;OAGG;IACH,OAAO,CAAC,iBAAiB;IAIzB;;;;;;OAMG;IACH,cAAc,CAAC,QAAQ,GAAE,MAAgC,GAAG,iBAAiB,EAAE,CAiB9E;IAED;;;;;;OAMG;IACH,SAAS,CAAC,QAAQ,GAAE,MAAgC,GAAG,wBAAwB,EAAE,CAiBhF;IAED,OAAO,CAAC,iBAAiB;CAkBzB;AAED,yGAAyG;AACzG,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,SAAS,wBAAwB,EAAE,GAAG,MAAM,CAuB9F;AAED,OAAO,EAAE,kBAAkB,EAAE,qBAAqB,EAAE,cAAc,EAAE,CAAC","sourcesContent":["import {\n\tdecideExpectedUtility,\n\ttype ExpectedUtilityCandidate,\n\ttype ToolSelectionDecision,\n} from \"./expected-utility.ts\";\nimport { evaluateToolPromotion, type ToolSelectionHint } from \"./promotion.ts\";\nimport type {\n\tToolExecutionObservation,\n\tToolPerformanceKey,\n\tToolPerformanceStore,\n\tToolSelectionIntentClass,\n\tToolSelectionObservation,\n} from \"./tool-performance-store.ts\";\n\nexport interface ToolSelectionTool {\n\tname: string;\n\tdescription?: string;\n\tparameters?: unknown;\n\tprofileAllowed?: boolean;\n\tcapabilityAllowed?: boolean;\n\tpathValidated?: boolean;\n\triskCost?: number;\n\tcontextCost?: number;\n}\n\nexport interface ToolSelectionControllerDeps {\n\tstore: ToolPerformanceStore;\n\tgetModelRef: () => string;\n\tgetActiveTools: () => readonly ToolSelectionTool[];\n\tisCandidateAllowed?: (toolName: string) => boolean;\n\t/**\n\t * Env-var source for the kill switches, injected for hermetic testing; defaults to\n\t * `process.env`. `PI_TOOL_SELECTION_OBSERVE=0` disables recording (the observe/stats layer,\n\t * default ON); `PI_TOOL_SELECTION_HINTS=0` disables surfacing the evidence-gated prompt hint\n\t * (default ON, but the hint itself only ever activates once evidence thresholds are met).\n\t */\n\tenv?: Record<string, string | undefined>;\n}\n\nexport interface ToolSelectionPendingObservation {\n\tid: string;\n\tkey: ToolPerformanceKey;\n\tfirstTool: boolean;\n\tstartedAt: number;\n\tinputTokenEstimate?: number;\n\tselection: Omit<ToolSelectionObservation, \"at\" | \"modelRef\" | \"intentClass\" | \"actualTool\" | \"succeeded\">;\n\t/** Was a promotion hint already active for this (model,intent) bucket before this call? */\n\thintActiveAtCallTime: boolean;\n}\n\n/** Report row for the observe/agreement/promotion loop — see {@link ToolSelectionController.getReport}. */\nexport interface ToolSelectionReportEntry {\n\tmodelRef: string;\n\tintentClass: ToolSelectionIntentClass;\n\tsampleCount: number;\n\tagreementRate?: number;\n\thintTool?: string;\n\thintSampleCount: number;\n\thintAgreementRate?: number;\n}\n\nconst INTENT_CLASSES: readonly ToolSelectionIntentClass[] = [\n\t\"read\",\n\t\"search\",\n\t\"execute\",\n\t\"write\",\n\t\"retrieve\",\n\t\"explain\",\n\t\"other\",\n];\n\nconst INTENT_WORDS: Record<ToolSelectionIntentClass, readonly string[]> = {\n\tread: [\"read\", \"cat\", \"ls\", \"list\", \"stat\", \"head\", \"tail\"],\n\tsearch: [\"search\", \"grep\", \"find\", \"glob\", \"ripgrep\", \"query\"],\n\texecute: [\"bash\", \"python\", \"powershell\", \"shell\", \"exec\", \"run\", \"process\", \"command\", \"script\"],\n\twrite: [\"write\", \"edit\", \"patch\", \"apply\", \"delete\", \"mkdir\", \"move\", \"rename\"],\n\tretrieve: [\"retrieve\", \"artifact\", \"recall\", \"memory\", \"fetch\", \"download\"],\n\texplain: [\"explain\", \"help\", \"describe\", \"inspect\", \"context\"],\n\tother: [],\n};\n\nfunction tokenize(value: string): string[] {\n\treturn value\n\t\t.toLowerCase()\n\t\t.split(/[^a-z0-9]+/u)\n\t\t.filter(Boolean);\n}\n\nfunction classifyToolIntent(tool: ToolSelectionTool): ToolSelectionIntentClass {\n\tconst tokens = tokenize(`${tool.name} ${tool.description ?? \"\"}`);\n\tlet best: ToolSelectionIntentClass = \"other\";\n\tlet bestScore = 0;\n\tfor (const intent of INTENT_CLASSES) {\n\t\tif (intent === \"other\") continue;\n\t\tconst score = INTENT_WORDS[intent].reduce((total, word) => total + (tokens.includes(word) ? 1 : 0), 0);\n\t\tif (score > bestScore) {\n\t\t\tbest = intent;\n\t\t\tbestScore = score;\n\t\t}\n\t}\n\treturn best;\n}\n\nfunction intentValue(intent: ToolSelectionIntentClass, toolIntent: ToolSelectionIntentClass): number {\n\tif (toolIntent === intent) return 0.85;\n\tif (intent === \"other\") return 0.35;\n\tif (toolIntent === \"other\") return 0.2;\n\treturn 0.12;\n}\n\nfunction noToolValue(intent: ToolSelectionIntentClass): number {\n\tswitch (intent) {\n\t\tcase \"explain\":\n\t\t\treturn 0.35;\n\t\tcase \"read\":\n\t\tcase \"search\":\n\t\tcase \"retrieve\":\n\t\t\treturn 0.12;\n\t\tcase \"write\":\n\t\tcase \"execute\":\n\t\t\treturn 0.05;\n\t\tdefault:\n\t\t\treturn 0.2;\n\t}\n}\n\nfunction estimateTokens(value: unknown): number | undefined {\n\ttry {\n\t\tconst serialized = typeof value === \"string\" ? value : JSON.stringify(value);\n\t\tif (!serialized) return undefined;\n\t\treturn Math.max(1, Math.ceil(serialized.length / 4));\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction estimateContentTokens(content: readonly unknown[]): number | undefined {\n\tconst text = content\n\t\t.map((block) => {\n\t\t\tif (!block || typeof block !== \"object\") return \"\";\n\t\t\tconst candidate = block as { type?: unknown; text?: unknown };\n\t\t\treturn candidate.type === \"text\" && typeof candidate.text === \"string\" ? candidate.text : \"\";\n\t\t})\n\t\t.join(\"\\n\");\n\treturn estimateTokens(text);\n}\n\nfunction modelToolKey(modelRef: string, intentClass: ToolSelectionIntentClass, tool: string): ToolPerformanceKey {\n\treturn { modelRef, intentClass, tool };\n}\n\nfunction candidateFor(\n\tstore: ToolPerformanceStore,\n\tmodelRef: string,\n\tintentClass: ToolSelectionIntentClass,\n\ttool: ToolSelectionTool,\n\tactualTool: string,\n): ExpectedUtilityCandidate {\n\tconst key = modelToolKey(modelRef, intentClass, tool.name);\n\tconst stats = store.get(key);\n\treturn {\n\t\ttool: tool.name,\n\t\tvalue: intentValue(intentClass, classifyToolIntent(tool)),\n\t\talpha: stats.alpha,\n\t\tbeta: stats.beta,\n\t\tsampleCount: stats.sampleCount,\n\t\tlatencyMs: stats.latencyEwmaMs,\n\t\ttokenEstimate:\n\t\t\tstats.inputTokenEstimateEwma === undefined && stats.outputTokenEstimateEwma === undefined\n\t\t\t\t? undefined\n\t\t\t\t: (stats.inputTokenEstimateEwma ?? 0) + (stats.outputTokenEstimateEwma ?? 0),\n\t\triskCost: tool.riskCost,\n\t\tcontextCost: tool.contextCost,\n\t\tdeterministicMatch: tool.name === actualTool && tool.pathValidated !== false,\n\t\tautomaticEligible: tool.pathValidated !== false,\n\t};\n}\n\nexport class ToolSelectionController {\n\tprivate readonly deps: ToolSelectionControllerDeps;\n\tprivate readonly pending = new Map<string, ToolSelectionPendingObservation>();\n\tprivate firstToolInTurn = true;\n\t/** Kill switch: observe/stats recording, default ON. `PI_TOOL_SELECTION_OBSERVE=0` disables it. */\n\tprivate readonly observeEnabled: boolean;\n\n\tconstructor(deps: ToolSelectionControllerDeps) {\n\t\tthis.deps = deps;\n\t\tthis.observeEnabled = (deps.env ?? process.env).PI_TOOL_SELECTION_OBSERVE !== \"0\";\n\t}\n\n\tstartTurn(): void {\n\t\tthis.firstToolInTurn = true;\n\t}\n\n\tbegin(toolCallId: string, toolName: string, args: unknown): ToolSelectionPendingObservation {\n\t\tconst modelRef = this.deps.getModelRef();\n\t\tconst activeTools = this.deps\n\t\t\t.getActiveTools()\n\t\t\t.filter(\n\t\t\t\t(tool) =>\n\t\t\t\t\ttool.profileAllowed !== false &&\n\t\t\t\t\ttool.capabilityAllowed !== false &&\n\t\t\t\t\t(this.deps.isCandidateAllowed?.(tool.name) ?? true),\n\t\t\t);\n\t\tconst actualTool = activeTools.find((tool) => tool.name === toolName) ?? {\n\t\t\tname: toolName,\n\t\t\tpathValidated: true,\n\t\t};\n\t\tconst intentClass = classifyToolIntent(actualTool);\n\t\tconst candidates = activeTools.map((tool) =>\n\t\t\tcandidateFor(\n\t\t\t\tthis.deps.store,\n\t\t\t\tmodelRef,\n\t\t\t\tintentClass,\n\t\t\t\ttool.name === toolName ? tool : { ...tool, pathValidated: false },\n\t\t\t\ttoolName,\n\t\t\t),\n\t\t);\n\t\tif (!candidates.some((candidate) => candidate.tool === toolName)) {\n\t\t\tcandidates.push(candidateFor(this.deps.store, modelRef, intentClass, actualTool, toolName));\n\t\t}\n\t\tcandidates.push({\n\t\t\ttool: \"no_tool\",\n\t\t\tvalue: noToolValue(intentClass),\n\t\t\talpha: 1,\n\t\t\tbeta: 1,\n\t\t\tsampleCount: 0,\n\t\t\tsuccessProbability: 1,\n\t\t\tlatencyMs: 0,\n\t\t\ttokenEstimate: 0,\n\t\t\triskCost: 0,\n\t\t\tcontextCost: 0,\n\t\t});\n\t\tconst decision = decideExpectedUtility(candidates);\n\t\tconst firstTool = this.firstToolInTurn;\n\t\tconst selection = this.selectionSnapshot(decision, firstTool);\n\t\t// Evaluated BEFORE this call is recorded, so it reflects evidence up to (not including) this\n\t\t// observation — captured now because complete() (later, async) can no longer distinguish\n\t\t// \"before\" from \"after\" once the store has been written.\n\t\tconst hintActiveAtCallTime = this.observeEnabled\n\t\t\t? this.evaluatePromotion(modelRef, intentClass).tool !== undefined\n\t\t\t: false;\n\t\tconst pending: ToolSelectionPendingObservation = {\n\t\t\tid: toolCallId,\n\t\t\tkey: modelToolKey(modelRef, intentClass, toolName),\n\t\t\tfirstTool: this.firstToolInTurn,\n\t\t\tstartedAt: Date.now(),\n\t\t\tinputTokenEstimate: estimateTokens(args),\n\t\t\tselection,\n\t\t\thintActiveAtCallTime,\n\t\t};\n\t\tthis.firstToolInTurn = false;\n\t\tthis.pending.set(toolCallId, pending);\n\t\treturn pending;\n\t}\n\n\tcomplete(toolCallId: string, succeeded: boolean, content: readonly unknown[] = []): void {\n\t\tconst pending = this.pending.get(toolCallId);\n\t\tif (!pending) return;\n\t\tthis.pending.delete(toolCallId);\n\t\tif (!this.observeEnabled) return;\n\t\tconst execution: ToolExecutionObservation = {\n\t\t\tkey: pending.key,\n\t\t\tsuccess: succeeded,\n\t\t\tlatencyMs: Math.max(0, Date.now() - pending.startedAt),\n\t\t\tinputTokenEstimate: pending.inputTokenEstimate,\n\t\t\toutputTokenEstimate: estimateContentTokens(content),\n\t\t\tselection: pending.selection,\n\t\t\thintActiveAtCallTime: pending.hintActiveAtCallTime,\n\t\t};\n\t\tthis.deps.store.recordExecution(execution);\n\t}\n\n\trecordValidation(toolName: string, outcome: \"repaired\" | \"bounced\"): void {\n\t\tif (!this.observeEnabled) return;\n\t\tconst modelRef = this.deps.getModelRef();\n\t\tconst tool = this.deps.getActiveTools().find((candidate) => candidate.name === toolName) ?? { name: toolName };\n\t\tthis.deps.store.recordValidation(modelToolKey(modelRef, classifyToolIntent(tool), toolName), outcome);\n\t}\n\n\t/**\n\t * Evidence-gated promotion for one (model,intent) bucket — see promotion.ts. Pure read, no\n\t * mutation; used both to stamp `hintActiveAtCallTime` and to build the live hint list below.\n\t */\n\tprivate evaluatePromotion(modelRef: string, intentClass: ToolSelectionIntentClass) {\n\t\treturn evaluateToolPromotion(this.deps.store.getStatsForIntent(modelRef, intentClass));\n\t}\n\n\t/**\n\t * The currently active evidence-gated hints for the session's current model — one per\n\t * (intentClass) at most, only where accumulated evidence clears the promotion gate. Consumed by\n\t * `system-prompt-builder.ts` to render a compact, cache-stable prompt block. Empty when the\n\t * hint kill switch (`PI_TOOL_SELECTION_HINTS=0`) is set, observing is disabled, or no bucket has\n\t * cleared the gate yet.\n\t */\n\tgetActiveHints(modelRef: string = this.deps.getModelRef()): ToolSelectionHint[] {\n\t\tif ((this.deps.env ?? process.env).PI_TOOL_SELECTION_HINTS === \"0\") return [];\n\t\tif (!this.observeEnabled) return [];\n\t\tconst hints: ToolSelectionHint[] = [];\n\t\tfor (const intentClass of INTENT_CLASSES) {\n\t\t\tconst promotion = this.evaluatePromotion(modelRef, intentClass);\n\t\t\tif (!promotion.tool) continue;\n\t\t\thints.push({\n\t\t\t\tmodelRef,\n\t\t\t\tintentClass,\n\t\t\t\ttool: promotion.tool,\n\t\t\t\tsampleCount: promotion.sampleCount,\n\t\t\t\tmargin: promotion.margin,\n\t\t\t\tentropy: promotion.entropy,\n\t\t\t});\n\t\t}\n\t\treturn hints;\n\t}\n\n\t/**\n\t * Report surface for the observe/agreement/promotion loop: per (model,intent), the durable\n\t * agreement rate (did the raw ranking's top pick match what was actually called), plus the\n\t * currently active hint (if any) and its own efficacy (agreement rate while it has been active).\n\t * A read-only diagnostic — never used to gate behavior. Render with\n\t * {@link formatToolSelectionReport}.\n\t */\n\tgetReport(modelRef: string = this.deps.getModelRef()): ToolSelectionReportEntry[] {\n\t\tconst activeHints = new Map(this.getActiveHints(modelRef).map((hint) => [hint.intentClass, hint]));\n\t\treturn this.deps.store\n\t\t\t.getAllIntentAgreements(modelRef)\n\t\t\t.filter((agreement) => agreement.sampleCount > 0)\n\t\t\t.map((agreement) => ({\n\t\t\t\tmodelRef: agreement.modelRef,\n\t\t\t\tintentClass: agreement.intentClass,\n\t\t\t\tsampleCount: agreement.sampleCount,\n\t\t\t\tagreementRate: agreement.agreementCount / agreement.sampleCount,\n\t\t\t\thintTool: activeHints.get(agreement.intentClass)?.tool,\n\t\t\t\thintSampleCount: agreement.hintActiveSampleCount,\n\t\t\t\thintAgreementRate:\n\t\t\t\t\tagreement.hintActiveSampleCount > 0\n\t\t\t\t\t\t? agreement.hintActiveAgreementCount / agreement.hintActiveSampleCount\n\t\t\t\t\t\t: undefined,\n\t\t\t}));\n\t}\n\n\tprivate selectionSnapshot(\n\t\tdecision: ToolSelectionDecision,\n\t\tfirstTool: boolean,\n\t): ToolSelectionPendingObservation[\"selection\"] {\n\t\treturn {\n\t\t\tfirstTool,\n\t\t\tdisposition: decision.disposition,\n\t\t\trecommendation: decision.recommendation,\n\t\t\tshortlist: decision.shortlist,\n\t\t\tentropy: decision.entropy,\n\t\t\tmargin: decision.margin,\n\t\t\tranked: decision.ranked.map((candidate) => ({\n\t\t\t\ttool: candidate.tool,\n\t\t\t\tutility: candidate.utility,\n\t\t\t\tprobability: candidate.probability,\n\t\t\t})),\n\t\t};\n\t}\n}\n\n/** Renders {@link ToolSelectionController.getReport} rows into the /toolhealth-style diagnostic text. */\nexport function formatToolSelectionReport(entries: readonly ToolSelectionReportEntry[]): string {\n\tif (entries.length === 0) {\n\t\treturn \"Tool-selection loop: no observations recorded yet for this host.\";\n\t}\n\tconst sorted = [...entries].sort(\n\t\t(left, right) => left.modelRef.localeCompare(right.modelRef) || left.intentClass.localeCompare(right.intentClass),\n\t);\n\tconst lines = [\"Tool-selection loop (observe -> agreement -> evidence-gated hint)\"];\n\tfor (const entry of sorted) {\n\t\tconst agreementText =\n\t\t\tentry.agreementRate === undefined\n\t\t\t\t? \"n/a\"\n\t\t\t\t: `${Math.round(entry.agreementRate * 100)}% (n=${entry.sampleCount})`;\n\t\tlines.push(`  ${entry.modelRef} / ${entry.intentClass}: agreement ${agreementText}`);\n\t\tif (entry.hintTool) {\n\t\t\tconst hintAgreementText =\n\t\t\t\tentry.hintAgreementRate === undefined\n\t\t\t\t\t? \"n/a\"\n\t\t\t\t\t: `${Math.round(entry.hintAgreementRate * 100)}% (n=${entry.hintSampleCount})`;\n\t\t\tlines.push(`    hint active: prefer \\`${entry.hintTool}\\` — agreement while active ${hintAgreementText}`);\n\t\t}\n\t}\n\treturn lines.join(\"\\n\");\n}\n\nexport { classifyToolIntent, estimateContentTokens, estimateTokens };\n"]}

package/dist/core/tool-selection/tool-selection-controller.js

@@ -1,4 +1,5 @@
 import { decideExpectedUtility, } from "./expected-utility.js";
+import { evaluateToolPromotion } from "./promotion.js";
 const INTENT_CLASSES = [
     "read",
     "search",
@@ -110,8 +111,11 @@ export class ToolSelectionController {
     deps;
     pending = new Map();
     firstToolInTurn = true;
+    /** Kill switch: observe/stats recording, default ON. `PI_TOOL_SELECTION_OBSERVE=0` disables it. */
+    observeEnabled;
     constructor(deps) {
         this.deps = deps;
+        this.observeEnabled = (deps.env ?? process.env).PI_TOOL_SELECTION_OBSERVE !== "0";
     }
     startTurn() {
         this.firstToolInTurn = true;
@@ -147,6 +151,12 @@ export class ToolSelectionController {
         const decision = decideExpectedUtility(candidates);
         const firstTool = this.firstToolInTurn;
         const selection = this.selectionSnapshot(decision, firstTool);
+        // Evaluated BEFORE this call is recorded, so it reflects evidence up to (not including) this
+        // observation — captured now because complete() (later, async) can no longer distinguish
+        // "before" from "after" once the store has been written.
+        const hintActiveAtCallTime = this.observeEnabled
+            ? this.evaluatePromotion(modelRef, intentClass).tool !== undefined
+            : false;
         const pending = {
             id: toolCallId,
             key: modelToolKey(modelRef, intentClass, toolName),
@@ -154,6 +164,7 @@ export class ToolSelectionController {
             startedAt: Date.now(),
             inputTokenEstimate: estimateTokens(args),
             selection,
+            hintActiveAtCallTime,
         };
         this.firstToolInTurn = false;
         this.pending.set(toolCallId, pending);
@@ -164,6 +175,8 @@ export class ToolSelectionController {
         if (!pending)
             return;
         this.pending.delete(toolCallId);
+        if (!this.observeEnabled)
+            return;
         const execution = {
             key: pending.key,
             success: succeeded,
@@ -171,14 +184,76 @@ export class ToolSelectionController {
             inputTokenEstimate: pending.inputTokenEstimate,
             outputTokenEstimate: estimateContentTokens(content),
             selection: pending.selection,
+            hintActiveAtCallTime: pending.hintActiveAtCallTime,
         };
         this.deps.store.recordExecution(execution);
     }
     recordValidation(toolName, outcome) {
+        if (!this.observeEnabled)
+            return;
         const modelRef = this.deps.getModelRef();
         const tool = this.deps.getActiveTools().find((candidate) => candidate.name === toolName) ?? { name: toolName };
         this.deps.store.recordValidation(modelToolKey(modelRef, classifyToolIntent(tool), toolName), outcome);
     }
+    /**
+     * Evidence-gated promotion for one (model,intent) bucket — see promotion.ts. Pure read, no
+     * mutation; used both to stamp `hintActiveAtCallTime` and to build the live hint list below.
+     */
+    evaluatePromotion(modelRef, intentClass) {
+        return evaluateToolPromotion(this.deps.store.getStatsForIntent(modelRef, intentClass));
+    }
+    /**
+     * The currently active evidence-gated hints for the session's current model — one per
+     * (intentClass) at most, only where accumulated evidence clears the promotion gate. Consumed by
+     * `system-prompt-builder.ts` to render a compact, cache-stable prompt block. Empty when the
+     * hint kill switch (`PI_TOOL_SELECTION_HINTS=0`) is set, observing is disabled, or no bucket has
+     * cleared the gate yet.
+     */
+    getActiveHints(modelRef = this.deps.getModelRef()) {
+        if ((this.deps.env ?? process.env).PI_TOOL_SELECTION_HINTS === "0")
+            return [];
+        if (!this.observeEnabled)
+            return [];
+        const hints = [];
+        for (const intentClass of INTENT_CLASSES) {
+            const promotion = this.evaluatePromotion(modelRef, intentClass);
+            if (!promotion.tool)
+                continue;
+            hints.push({
+                modelRef,
+                intentClass,
+                tool: promotion.tool,
+                sampleCount: promotion.sampleCount,
+                margin: promotion.margin,
+                entropy: promotion.entropy,
+            });
+        }
+        return hints;
+    }
+    /**
+     * Report surface for the observe/agreement/promotion loop: per (model,intent), the durable
+     * agreement rate (did the raw ranking's top pick match what was actually called), plus the
+     * currently active hint (if any) and its own efficacy (agreement rate while it has been active).
+     * A read-only diagnostic — never used to gate behavior. Render with
+     * {@link formatToolSelectionReport}.
+     */
+    getReport(modelRef = this.deps.getModelRef()) {
+        const activeHints = new Map(this.getActiveHints(modelRef).map((hint) => [hint.intentClass, hint]));
+        return this.deps.store
+            .getAllIntentAgreements(modelRef)
+            .filter((agreement) => agreement.sampleCount > 0)
+            .map((agreement) => ({
+            modelRef: agreement.modelRef,
+            intentClass: agreement.intentClass,
+            sampleCount: agreement.sampleCount,
+            agreementRate: agreement.agreementCount / agreement.sampleCount,
+            hintTool: activeHints.get(agreement.intentClass)?.tool,
+            hintSampleCount: agreement.hintActiveSampleCount,
+            hintAgreementRate: agreement.hintActiveSampleCount > 0
+                ? agreement.hintActiveAgreementCount / agreement.hintActiveSampleCount
+                : undefined,
+        }));
+    }
     selectionSnapshot(decision, firstTool) {
         return {
             firstTool,
@@ -195,5 +270,26 @@ export class ToolSelectionController {
         };
     }
 }
+/** Renders {@link ToolSelectionController.getReport} rows into the /toolhealth-style diagnostic text. */
+export function formatToolSelectionReport(entries) {
+    if (entries.length === 0) {
+        return "Tool-selection loop: no observations recorded yet for this host.";
+    }
+    const sorted = [...entries].sort((left, right) => left.modelRef.localeCompare(right.modelRef) || left.intentClass.localeCompare(right.intentClass));
+    const lines = ["Tool-selection loop (observe -> agreement -> evidence-gated hint)"];
+    for (const entry of sorted) {
+        const agreementText = entry.agreementRate === undefined
+            ? "n/a"
+            : `${Math.round(entry.agreementRate * 100)}% (n=${entry.sampleCount})`;
+        lines.push(`  ${entry.modelRef} / ${entry.intentClass}: agreement ${agreementText}`);
+        if (entry.hintTool) {
+            const hintAgreementText = entry.hintAgreementRate === undefined
+                ? "n/a"
+                : `${Math.round(entry.hintAgreementRate * 100)}% (n=${entry.hintSampleCount})`;
+            lines.push(`    hint active: prefer \`${entry.hintTool}\` — agreement while active ${hintAgreementText}`);
+        }
+    }
+    return lines.join("\n");
+}
 export { classifyToolIntent, estimateContentTokens, estimateTokens };
 //# sourceMappingURL=tool-selection-controller.js.map