@caupulican/pi-adaptative 0.81.39 → 0.81.41

package/dist/core/research/evidence-bundle.js

@@ -1,3 +1,4 @@
+import { isPlainRecord } from "../util/value-guards.js";
 function cloneJsonValue(value) {
     if (Array.isArray(value)) {
         return value.map(cloneJsonValue);
@@ -10,12 +11,6 @@ function cloneJsonValue(value) {
 function cloneJsonObject(value) {
     return Object.fromEntries(Object.entries(value).map(([key, nested]) => [key, cloneJsonValue(nested)]));
 }
-function isPlainRecord(value) {
-    if (!value || typeof value !== "object" || Array.isArray(value))
-        return false;
-    const prototype = Object.getPrototypeOf(value);
-    return prototype === Object.prototype || prototype === null;
-}
 function isJsonValue(value) {
     if (value === null || typeof value === "string" || typeof value === "boolean") {
         return true;

package/dist/core/research/evidence-bundle.js.map

@@ -1 +1 @@
-{"version":3,"file":"evidence-bundle.js","sourceRoot":"","sources":["../../../src/core/research/evidence-bundle.ts"],"names":[],"mappings":"AAEA,SAAS,cAAc,CAAC,KAAgB,EAAa;IACpD,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;QAC1B,OAAO,KAAK,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC;IAClC,CAAC;IACD,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;QACjD,OAAO,MAAM,CAAC,WAAW,CAAC,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,EAAE,MAAM,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,EAAE,cAAc,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;IACxG,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAED,SAAS,eAAe,CAAC,KAAiB,EAAc;IACvD,OAAO,MAAM,CAAC,WAAW,CAAC,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,EAAE,MAAM,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,EAAE,cAAc,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;AAAA,CACvG;AAED,SAAS,aAAa,CAAC,KAAc,EAAoC;IACxE,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IAC9E,MAAM,SAAS,GAAG,MAAM,CAAC,cAAc,CAAC,KAAK,CAAC,CAAC;IAC/C,OAAO,SAAS,KAAK,MAAM,CAAC,SAAS,IAAI,SAAS,KAAK,IAAI,CAAC;AAAA,CAC5D;AAED,SAAS,WAAW,CAAC,KAAc,EAAsB;IACxD,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,OAAO,KAAK,KAAK,SAAS,EAAE,CAAC;QAC/E,OAAO,IAAI,CAAC;IACb,CAAC;IACD,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;QAC/B,OAAO,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IAC/B,CAAC;IACD,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;QAC1B,OAAO,KAAK,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;IACjC,CAAC;IACD,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;QAC/B,OAAO,YAAY,CAAC,KAAK,CAAC,CAAC;IAC5B,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAED,SAAS,YAAY,CAAC,KAAc,EAAuB;IAC1D,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC,EAAE,CAAC;QAC3B,OAAO,KAAK,CAAC;IACd,CAAC;IACD,OAAO,MAAM,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;AAAA,CAC/C;AAED,SAAS,gBAAgB,CAAC,MAAmB,EAAe;IAC3D,OAAO;QACN,GAAG,MAAM;QACT,QAAQ,EAAE,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,eAAe,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,SAAS;KACxE,CAAC;AAAA,CACF;AAED,SAAS,YAAY,CAAC,OAAgB,EAAW;IAChD,OAAO;QACN,GAAG,OAAO;QACV,WAAW,EAAE,CAAC,GAAG,OAAO,CAAC,WAAW,CAAC;KACrC,CAAC;AAAA,CACF;AAED,MAAM,UAAU,oBAAoB,CAAC,IAKpC,EAAkB;IAClB,OAAO;QACN,KAAK,EAAE,IAAI,CAAC,KAAK;QACjB,OAAO,EAAE,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,gBAAgB,CAAC;QAC3C,QAAQ,EAAE,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,YAAY,CAAC;QACzC,SAAS,EAAE,IAAI,CAAC,GAAG;KACnB,CAAC;AAAA,CACF;AAED,MAAM,UAAU,6BAA6B,CAAC,MAAsB,EAAkB;IACrF,OAAO;QACN,KAAK,EAAE,MAAM,CAAC,KAAK;QACnB,OAAO,EAAE,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,gBAAgB,CAAC;QAC7C,QAAQ,EAAE,MAAM,CAAC,QAAQ,CAAC,GAAG,CAAC,YAAY,CAAC;QAC3C,SAAS,EAAE,MAAM,CAAC,SAAS;KAC3B,CAAC;AAAA,CACF;AAED,MAAM,UAAU,gBAAgB,CAAC,KAAc,EAA2B;IACzE,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IACxC,MAAM,MAAM,GAAG,KAAgC,CAAC;IAEhD,IAAI,OAAO,MAAM,CAAC,KAAK,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IACnD,IAAI,MAAM,CAAC,SAAS,KAAK,SAAS,IAAI,OAAO,MAAM,CAAC,SAAS,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IAEzF,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,OAAO,CAAC;QAAE,OAAO,KAAK,CAAC;IACjD,KAAK,MAAM,MAAM,IAAI,MAAM,CAAC,OAAO,EAAE,CAAC;QACrC,IAAI,CAAC,aAAa,CAAC,MAAM,CAAC;YAAE,OAAO,KAAK,CAAC;QACzC,MAAM,GAAG,GAAG,MAAiC,CAAC;QAC9C,IAAI,OAAO,GAAG,CAAC,EAAE,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QAC7C,IACC,OAAO,GAAG,CAAC,IAAI,KAAK,QAAQ;YAC5B,CAAC,CAAC,WAAW,EAAE,YAAY,EAAE,UAAU,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,CAAC,CAAC,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,EACjF,CAAC;YACF,OAAO,KAAK,CAAC;QACd,CAAC;QACD,IAAI,OAAO,GAAG,CAAC,OAAO,KAAK,SAAS;YAAE,OAAO,KAAK,CAAC;QACnD,IAAI,GAAG,CAAC,KAAK,KAAK,SAAS,IAAI,OAAO,GAAG,CAAC,KAAK,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QAC3E,IAAI,GAAG,CAAC,GAAG,KAAK,SAAS,IAAI,OAAO,GAAG,CAAC,GAAG,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QACvE,IAAI,GAAG,CAAC,OAAO,KAAK,SAAS,IAAI,OAAO,GAAG,CAAC,OAAO,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QAC/E,IAAI,GAAG,CAAC,QAAQ,KAAK,SAAS,IAAI,CAAC,YAAY,CAAC,GAAG,CAAC,QAAQ,CAAC;YAAE,OAAO,KAAK,CAAC;IAC7E,CAAC;IAED,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC;QAAE,OAAO,KAAK,CAAC;IAClD,KAAK,MAAM,OAAO,IAAI,MAAM,CAAC,QAAQ,EAAE,CAAC;QACvC,IAAI,CAAC,aAAa,CAAC,OAAO,CAAC;YAAE,OAAO,KAAK,CAAC;QAC1C,MAAM,CAAC,GAAG,OAAkC,CAAC;QAC7C,IAAI,OAAO,CAAC,CAAC,EAAE,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QAC3C,IAAI,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QAChD,IAAI,CAAC,CAAC,UAAU,KAAK,SAAS,IAAI,CAAC,OAAO,CAAC,CAAC,UAAU,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC;YACrG,OAAO,KAAK,CAAC;QACd,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC,CAAC,WAAW,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,OAAO,EAAE,KAAK,QAAQ,CAAC;YAAE,OAAO,KAAK,CAAC;IACzG,CAAC;IAED,OAAO,IAAI,CAAC;AAAA,CACZ","sourcesContent":["import type { EvidenceBundle, EvidenceRef, Finding, JsonObject, JsonValue } from \"../autonomy/contracts.ts\";\n\nfunction cloneJsonValue(value: JsonValue): JsonValue {\n\tif (Array.isArray(value)) {\n\t\treturn value.map(cloneJsonValue);\n\t}\n\tif (value !== null && typeof value === \"object\") {\n\t\treturn Object.fromEntries(Object.entries(value).map(([key, nested]) => [key, cloneJsonValue(nested)]));\n\t}\n\treturn value;\n}\n\nfunction cloneJsonObject(value: JsonObject): JsonObject {\n\treturn Object.fromEntries(Object.entries(value).map(([key, nested]) => [key, cloneJsonValue(nested)]));\n}\n\nfunction isPlainRecord(value: unknown): value is Record<string, unknown> {\n\tif (!value || typeof value !== \"object\" || Array.isArray(value)) return false;\n\tconst prototype = Object.getPrototypeOf(value);\n\treturn prototype === Object.prototype || prototype === null;\n}\n\nfunction isJsonValue(value: unknown): value is JsonValue {\n\tif (value === null || typeof value === \"string\" || typeof value === \"boolean\") {\n\t\treturn true;\n\t}\n\tif (typeof value === \"number\") {\n\t\treturn Number.isFinite(value);\n\t}\n\tif (Array.isArray(value)) {\n\t\treturn value.every(isJsonValue);\n\t}\n\tif (typeof value === \"object\") {\n\t\treturn isJsonObject(value);\n\t}\n\treturn false;\n}\n\nfunction isJsonObject(value: unknown): value is JsonObject {\n\tif (!isPlainRecord(value)) {\n\t\treturn false;\n\t}\n\treturn Object.values(value).every(isJsonValue);\n}\n\nfunction cloneEvidenceRef(source: EvidenceRef): EvidenceRef {\n\treturn {\n\t\t...source,\n\t\tmetadata: source.metadata ? cloneJsonObject(source.metadata) : undefined,\n\t};\n}\n\nfunction cloneFinding(finding: Finding): Finding {\n\treturn {\n\t\t...finding,\n\t\tevidenceIds: [...finding.evidenceIds],\n\t};\n}\n\nexport function createEvidenceBundle(args: {\n\tquery: string;\n\tsources: readonly EvidenceRef[];\n\tfindings: readonly Finding[];\n\tnow?: string;\n}): EvidenceBundle {\n\treturn {\n\t\tquery: args.query,\n\t\tsources: args.sources.map(cloneEvidenceRef),\n\t\tfindings: args.findings.map(cloneFinding),\n\t\tcreatedAt: args.now,\n\t};\n}\n\nexport function cloneEvidenceBundleForStorage(bundle: EvidenceBundle): EvidenceBundle {\n\treturn {\n\t\tquery: bundle.query,\n\t\tsources: bundle.sources.map(cloneEvidenceRef),\n\t\tfindings: bundle.findings.map(cloneFinding),\n\t\tcreatedAt: bundle.createdAt,\n\t};\n}\n\nexport function isEvidenceBundle(value: unknown): value is EvidenceBundle {\n\tif (!isPlainRecord(value)) return false;\n\tconst bundle = value as Record<string, unknown>;\n\n\tif (typeof bundle.query !== \"string\") return false;\n\tif (bundle.createdAt !== undefined && typeof bundle.createdAt !== \"string\") return false;\n\n\tif (!Array.isArray(bundle.sources)) return false;\n\tfor (const source of bundle.sources) {\n\t\tif (!isPlainRecord(source)) return false;\n\t\tconst ref = source as Record<string, unknown>;\n\t\tif (typeof ref.id !== \"string\") return false;\n\t\tif (\n\t\t\ttypeof ref.kind !== \"string\" ||\n\t\t\t![\"workspace\", \"transcript\", \"automata\", \"web\", \"user\", \"tool\"].includes(ref.kind)\n\t\t) {\n\t\t\treturn false;\n\t\t}\n\t\tif (typeof ref.trusted !== \"boolean\") return false;\n\t\tif (ref.title !== undefined && typeof ref.title !== \"string\") return false;\n\t\tif (ref.uri !== undefined && typeof ref.uri !== \"string\") return false;\n\t\tif (ref.excerpt !== undefined && typeof ref.excerpt !== \"string\") return false;\n\t\tif (ref.metadata !== undefined && !isJsonObject(ref.metadata)) return false;\n\t}\n\n\tif (!Array.isArray(bundle.findings)) return false;\n\tfor (const finding of bundle.findings) {\n\t\tif (!isPlainRecord(finding)) return false;\n\t\tconst f = finding as Record<string, unknown>;\n\t\tif (typeof f.id !== \"string\") return false;\n\t\tif (typeof f.summary !== \"string\") return false;\n\t\tif (f.confidence !== undefined && (typeof f.confidence !== \"number\" || !Number.isFinite(f.confidence)))\n\t\t\treturn false;\n\t\tif (!Array.isArray(f.evidenceIds) || !f.evidenceIds.every((id) => typeof id === \"string\")) return false;\n\t}\n\n\treturn true;\n}\n"]}
+{"version":3,"file":"evidence-bundle.js","sourceRoot":"","sources":["../../../src/core/research/evidence-bundle.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,aAAa,EAAE,MAAM,yBAAyB,CAAC;AAExD,SAAS,cAAc,CAAC,KAAgB,EAAa;IACpD,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;QAC1B,OAAO,KAAK,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC;IAClC,CAAC;IACD,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;QACjD,OAAO,MAAM,CAAC,WAAW,CAAC,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,EAAE,MAAM,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,EAAE,cAAc,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;IACxG,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAED,SAAS,eAAe,CAAC,KAAiB,EAAc;IACvD,OAAO,MAAM,CAAC,WAAW,CAAC,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,EAAE,MAAM,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,EAAE,cAAc,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;AAAA,CACvG;AAED,SAAS,WAAW,CAAC,KAAc,EAAsB;IACxD,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,OAAO,KAAK,KAAK,SAAS,EAAE,CAAC;QAC/E,OAAO,IAAI,CAAC;IACb,CAAC;IACD,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;QAC/B,OAAO,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IAC/B,CAAC;IACD,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;QAC1B,OAAO,KAAK,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;IACjC,CAAC;IACD,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;QAC/B,OAAO,YAAY,CAAC,KAAK,CAAC,CAAC;IAC5B,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAED,SAAS,YAAY,CAAC,KAAc,EAAuB;IAC1D,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC,EAAE,CAAC;QAC3B,OAAO,KAAK,CAAC;IACd,CAAC;IACD,OAAO,MAAM,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;AAAA,CAC/C;AAED,SAAS,gBAAgB,CAAC,MAAmB,EAAe;IAC3D,OAAO;QACN,GAAG,MAAM;QACT,QAAQ,EAAE,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,eAAe,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,SAAS;KACxE,CAAC;AAAA,CACF;AAED,SAAS,YAAY,CAAC,OAAgB,EAAW;IAChD,OAAO;QACN,GAAG,OAAO;QACV,WAAW,EAAE,CAAC,GAAG,OAAO,CAAC,WAAW,CAAC;KACrC,CAAC;AAAA,CACF;AAED,MAAM,UAAU,oBAAoB,CAAC,IAKpC,EAAkB;IAClB,OAAO;QACN,KAAK,EAAE,IAAI,CAAC,KAAK;QACjB,OAAO,EAAE,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,gBAAgB,CAAC;QAC3C,QAAQ,EAAE,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,YAAY,CAAC;QACzC,SAAS,EAAE,IAAI,CAAC,GAAG;KACnB,CAAC;AAAA,CACF;AAED,MAAM,UAAU,6BAA6B,CAAC,MAAsB,EAAkB;IACrF,OAAO;QACN,KAAK,EAAE,MAAM,CAAC,KAAK;QACnB,OAAO,EAAE,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,gBAAgB,CAAC;QAC7C,QAAQ,EAAE,MAAM,CAAC,QAAQ,CAAC,GAAG,CAAC,YAAY,CAAC;QAC3C,SAAS,EAAE,MAAM,CAAC,SAAS;KAC3B,CAAC;AAAA,CACF;AAED,MAAM,UAAU,gBAAgB,CAAC,KAAc,EAA2B;IACzE,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IACxC,MAAM,MAAM,GAAG,KAAgC,CAAC;IAEhD,IAAI,OAAO,MAAM,CAAC,KAAK,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IACnD,IAAI,MAAM,CAAC,SAAS,KAAK,SAAS,IAAI,OAAO,MAAM,CAAC,SAAS,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IAEzF,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,OAAO,CAAC;QAAE,OAAO,KAAK,CAAC;IACjD,KAAK,MAAM,MAAM,IAAI,MAAM,CAAC,OAAO,EAAE,CAAC;QACrC,IAAI,CAAC,aAAa,CAAC,MAAM,CAAC;YAAE,OAAO,KAAK,CAAC;QACzC,MAAM,GAAG,GAAG,MAAiC,CAAC;QAC9C,IAAI,OAAO,GAAG,CAAC,EAAE,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QAC7C,IACC,OAAO,GAAG,CAAC,IAAI,KAAK,QAAQ;YAC5B,CAAC,CAAC,WAAW,EAAE,YAAY,EAAE,UAAU,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,CAAC,CAAC,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,EACjF,CAAC;YACF,OAAO,KAAK,CAAC;QACd,CAAC;QACD,IAAI,OAAO,GAAG,CAAC,OAAO,KAAK,SAAS;YAAE,OAAO,KAAK,CAAC;QACnD,IAAI,GAAG,CAAC,KAAK,KAAK,SAAS,IAAI,OAAO,GAAG,CAAC,KAAK,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QAC3E,IAAI,GAAG,CAAC,GAAG,KAAK,SAAS,IAAI,OAAO,GAAG,CAAC,GAAG,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QACvE,IAAI,GAAG,CAAC,OAAO,KAAK,SAAS,IAAI,OAAO,GAAG,CAAC,OAAO,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QAC/E,IAAI,GAAG,CAAC,QAAQ,KAAK,SAAS,IAAI,CAAC,YAAY,CAAC,GAAG,CAAC,QAAQ,CAAC;YAAE,OAAO,KAAK,CAAC;IAC7E,CAAC;IAED,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC;QAAE,OAAO,KAAK,CAAC;IAClD,KAAK,MAAM,OAAO,IAAI,MAAM,CAAC,QAAQ,EAAE,CAAC;QACvC,IAAI,CAAC,aAAa,CAAC,OAAO,CAAC;YAAE,OAAO,KAAK,CAAC;QAC1C,MAAM,CAAC,GAAG,OAAkC,CAAC;QAC7C,IAAI,OAAO,CAAC,CAAC,EAAE,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QAC3C,IAAI,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QAChD,IAAI,CAAC,CAAC,UAAU,KAAK,SAAS,IAAI,CAAC,OAAO,CAAC,CAAC,UAAU,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC;YACrG,OAAO,KAAK,CAAC;QACd,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC,CAAC,WAAW,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,OAAO,EAAE,KAAK,QAAQ,CAAC;YAAE,OAAO,KAAK,CAAC;IACzG,CAAC;IAED,OAAO,IAAI,CAAC;AAAA,CACZ","sourcesContent":["import type { EvidenceBundle, EvidenceRef, Finding, JsonObject, JsonValue } from \"../autonomy/contracts.ts\";\nimport { isPlainRecord } from \"../util/value-guards.ts\";\n\nfunction cloneJsonValue(value: JsonValue): JsonValue {\n\tif (Array.isArray(value)) {\n\t\treturn value.map(cloneJsonValue);\n\t}\n\tif (value !== null && typeof value === \"object\") {\n\t\treturn Object.fromEntries(Object.entries(value).map(([key, nested]) => [key, cloneJsonValue(nested)]));\n\t}\n\treturn value;\n}\n\nfunction cloneJsonObject(value: JsonObject): JsonObject {\n\treturn Object.fromEntries(Object.entries(value).map(([key, nested]) => [key, cloneJsonValue(nested)]));\n}\n\nfunction isJsonValue(value: unknown): value is JsonValue {\n\tif (value === null || typeof value === \"string\" || typeof value === \"boolean\") {\n\t\treturn true;\n\t}\n\tif (typeof value === \"number\") {\n\t\treturn Number.isFinite(value);\n\t}\n\tif (Array.isArray(value)) {\n\t\treturn value.every(isJsonValue);\n\t}\n\tif (typeof value === \"object\") {\n\t\treturn isJsonObject(value);\n\t}\n\treturn false;\n}\n\nfunction isJsonObject(value: unknown): value is JsonObject {\n\tif (!isPlainRecord(value)) {\n\t\treturn false;\n\t}\n\treturn Object.values(value).every(isJsonValue);\n}\n\nfunction cloneEvidenceRef(source: EvidenceRef): EvidenceRef {\n\treturn {\n\t\t...source,\n\t\tmetadata: source.metadata ? cloneJsonObject(source.metadata) : undefined,\n\t};\n}\n\nfunction cloneFinding(finding: Finding): Finding {\n\treturn {\n\t\t...finding,\n\t\tevidenceIds: [...finding.evidenceIds],\n\t};\n}\n\nexport function createEvidenceBundle(args: {\n\tquery: string;\n\tsources: readonly EvidenceRef[];\n\tfindings: readonly Finding[];\n\tnow?: string;\n}): EvidenceBundle {\n\treturn {\n\t\tquery: args.query,\n\t\tsources: args.sources.map(cloneEvidenceRef),\n\t\tfindings: args.findings.map(cloneFinding),\n\t\tcreatedAt: args.now,\n\t};\n}\n\nexport function cloneEvidenceBundleForStorage(bundle: EvidenceBundle): EvidenceBundle {\n\treturn {\n\t\tquery: bundle.query,\n\t\tsources: bundle.sources.map(cloneEvidenceRef),\n\t\tfindings: bundle.findings.map(cloneFinding),\n\t\tcreatedAt: bundle.createdAt,\n\t};\n}\n\nexport function isEvidenceBundle(value: unknown): value is EvidenceBundle {\n\tif (!isPlainRecord(value)) return false;\n\tconst bundle = value as Record<string, unknown>;\n\n\tif (typeof bundle.query !== \"string\") return false;\n\tif (bundle.createdAt !== undefined && typeof bundle.createdAt !== \"string\") return false;\n\n\tif (!Array.isArray(bundle.sources)) return false;\n\tfor (const source of bundle.sources) {\n\t\tif (!isPlainRecord(source)) return false;\n\t\tconst ref = source as Record<string, unknown>;\n\t\tif (typeof ref.id !== \"string\") return false;\n\t\tif (\n\t\t\ttypeof ref.kind !== \"string\" ||\n\t\t\t![\"workspace\", \"transcript\", \"automata\", \"web\", \"user\", \"tool\"].includes(ref.kind)\n\t\t) {\n\t\t\treturn false;\n\t\t}\n\t\tif (typeof ref.trusted !== \"boolean\") return false;\n\t\tif (ref.title !== undefined && typeof ref.title !== \"string\") return false;\n\t\tif (ref.uri !== undefined && typeof ref.uri !== \"string\") return false;\n\t\tif (ref.excerpt !== undefined && typeof ref.excerpt !== \"string\") return false;\n\t\tif (ref.metadata !== undefined && !isJsonObject(ref.metadata)) return false;\n\t}\n\n\tif (!Array.isArray(bundle.findings)) return false;\n\tfor (const finding of bundle.findings) {\n\t\tif (!isPlainRecord(finding)) return false;\n\t\tconst f = finding as Record<string, unknown>;\n\t\tif (typeof f.id !== \"string\") return false;\n\t\tif (typeof f.summary !== \"string\") return false;\n\t\tif (f.confidence !== undefined && (typeof f.confidence !== \"number\" || !Number.isFinite(f.confidence)))\n\t\t\treturn false;\n\t\tif (!Array.isArray(f.evidenceIds) || !f.evidenceIds.every((id) => typeof id === \"string\")) return false;\n\t}\n\n\treturn true;\n}\n"]}

package/dist/core/research/lane-model-resolver.d.ts

@@ -0,0 +1,33 @@
+import type { Api, Model } from "@caupulican/pi-ai";
+import { type ModelCapabilityProfile } from "../model-capability.ts";
+import type { ModelRegistry } from "../model-registry.ts";
+import type { NormalizedProfile } from "../profile-registry.ts";
+import type { SettingsManager } from "../settings-manager.ts";
+export interface LaneModelResolverDeps {
+    getCwd(): string;
+    getModel(): Model<Api> | undefined;
+    getModelRegistry(): ModelRegistry;
+    getSettingsManager(): SettingsManager;
+    isModelExhausted(model: Model<Api>): boolean;
+}
+export type LaneShipment = {
+    ok: true;
+    model: Model<Api>;
+    laneProfile?: NormalizedProfile;
+} | {
+    ok: false;
+    skipReason: string;
+};
+export declare function clampLaneMaxUsd(settingsMaxUsd: number, foregroundMaxEstimatedUsd?: number): number;
+/** Shared provider-neutral model/profile resolution for research and fitness lanes. */
+export declare class LaneModelResolver {
+    private readonly deps;
+    constructor(deps: LaneModelResolverDeps);
+    capabilityProfile(model: Model<Api>): ModelCapabilityProfile;
+    resolveModel(configuredPattern: string | undefined): Model<Api> | undefined;
+    resolveShipment(laneSettings: {
+        model?: string;
+        profile?: string;
+    }, missingModelReason: string): LaneShipment;
+}
+//# sourceMappingURL=lane-model-resolver.d.ts.map

package/dist/core/research/lane-model-resolver.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"lane-model-resolver.d.ts","sourceRoot":"","sources":["../../../src/core/research/lane-model-resolver.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,MAAM,mBAAmB,CAAC;AACpD,OAAO,EAAgC,KAAK,sBAAsB,EAAE,MAAM,wBAAwB,CAAC;AACnG,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,sBAAsB,CAAC;AAE1D,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,wBAAwB,CAAC;AAChE,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,wBAAwB,CAAC;AAE9D,MAAM,WAAW,qBAAqB;IACrC,MAAM,IAAI,MAAM,CAAC;IACjB,QAAQ,IAAI,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CAAC;IACnC,gBAAgB,IAAI,aAAa,CAAC;IAClC,kBAAkB,IAAI,eAAe,CAAC;IACtC,gBAAgB,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,OAAO,CAAC;CAC7C;AAED,MAAM,MAAM,YAAY,GACrB;IAAE,EAAE,EAAE,IAAI,CAAC;IAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;IAAC,WAAW,CAAC,EAAE,iBAAiB,CAAA;CAAE,GAChE;IAAE,EAAE,EAAE,KAAK,CAAC;IAAC,UAAU,EAAE,MAAM,CAAA;CAAE,CAAC;AAErC,wBAAgB,eAAe,CAAC,cAAc,EAAE,MAAM,EAAE,yBAAyB,CAAC,EAAE,MAAM,GAAG,MAAM,CAElG;AAED,uFAAuF;AACvF,qBAAa,iBAAiB;IAC7B,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAwB;IAE7C,YAAY,IAAI,EAAE,qBAAqB,EAEtC;IAED,iBAAiB,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,sBAAsB,CAK3D;IAED,YAAY,CAAC,iBAAiB,EAAE,MAAM,GAAG,SAAS,GAAG,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CAY1E;IAED,eAAe,CAAC,YAAY,EAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAA;KAAE,EAAE,kBAAkB,EAAE,MAAM,GAAG,YAAY,CA2B5G;CACD","sourcesContent":["import type { Api, Model } from \"@caupulican/pi-ai\";\nimport { deriveModelCapabilityProfile, type ModelCapabilityProfile } from \"../model-capability.ts\";\nimport type { ModelRegistry } from \"../model-registry.ts\";\nimport { resolveCliModel } from \"../model-resolver.ts\";\nimport type { NormalizedProfile } from \"../profile-registry.ts\";\nimport type { SettingsManager } from \"../settings-manager.ts\";\n\nexport interface LaneModelResolverDeps {\n\tgetCwd(): string;\n\tgetModel(): Model<Api> | undefined;\n\tgetModelRegistry(): ModelRegistry;\n\tgetSettingsManager(): SettingsManager;\n\tisModelExhausted(model: Model<Api>): boolean;\n}\n\nexport type LaneShipment =\n\t| { ok: true; model: Model<Api>; laneProfile?: NormalizedProfile }\n\t| { ok: false; skipReason: string };\n\nexport function clampLaneMaxUsd(settingsMaxUsd: number, foregroundMaxEstimatedUsd?: number): number {\n\treturn Math.min(settingsMaxUsd, foregroundMaxEstimatedUsd ?? Number.POSITIVE_INFINITY);\n}\n\n/** Shared provider-neutral model/profile resolution for research and fitness lanes. */\nexport class LaneModelResolver {\n\tprivate readonly deps: LaneModelResolverDeps;\n\n\tconstructor(deps: LaneModelResolverDeps) {\n\t\tthis.deps = deps;\n\t}\n\n\tcapabilityProfile(model: Model<Api>): ModelCapabilityProfile {\n\t\treturn deriveModelCapabilityProfile({\n\t\t\tcontextWindow: model.contextWindow,\n\t\t\tmode: this.deps.getSettingsManager().getModelCapabilitySettings().mode,\n\t\t});\n\t}\n\n\tresolveModel(configuredPattern: string | undefined): Model<Api> | undefined {\n\t\tif (configuredPattern) {\n\t\t\tconst resolved = resolveCliModel({\n\t\t\t\tcliModel: configuredPattern,\n\t\t\t\tmodelRegistry: this.deps.getModelRegistry(),\n\t\t\t});\n\t\t\tif (resolved.model && this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) {\n\t\t\t\treturn resolved.model;\n\t\t\t}\n\t\t\treturn undefined;\n\t\t}\n\t\treturn this.deps.getModel() ?? undefined;\n\t}\n\n\tresolveShipment(laneSettings: { model?: string; profile?: string }, missingModelReason: string): LaneShipment {\n\t\tlet laneProfile: NormalizedProfile | undefined;\n\t\tif (laneSettings.profile) {\n\t\t\tconst profileRef = laneSettings.profile.trim();\n\t\t\tconst registry = this.deps.getSettingsManager().getProfileRegistry();\n\t\t\tlaneProfile =\n\t\t\t\tprofileRef.startsWith(\"./\") || profileRef.startsWith(\"../\")\n\t\t\t\t\t? registry.resolveProfileRef(profileRef, this.deps.getCwd())\n\t\t\t\t\t: registry.getProfile(profileRef);\n\t\t\tif (!laneProfile) return { ok: false, skipReason: \"lane_profile_not_found\" };\n\t\t}\n\n\t\tlet model: Model<Api> | undefined;\n\t\tif (laneSettings.model) {\n\t\t\tmodel = this.resolveModel(laneSettings.model);\n\t\t\tif (!model) return { ok: false, skipReason: missingModelReason };\n\t\t} else if (laneProfile?.model) {\n\t\t\tmodel = this.resolveModel(laneProfile.model);\n\t\t\tif (!model) return { ok: false, skipReason: \"no_lane_profile_model\" };\n\t\t} else {\n\t\t\tmodel = this.deps.getModel() ?? undefined;\n\t\t\tif (!model) return { ok: false, skipReason: missingModelReason };\n\t\t}\n\t\tif (this.deps.isModelExhausted(model)) {\n\t\t\treturn { ok: false, skipReason: `${model.provider}/${model.id} model exhausted: quota` };\n\t\t}\n\t\treturn { ok: true, model, laneProfile };\n\t}\n}\n"]}

package/dist/core/research/lane-model-resolver.js

@@ -0,0 +1,65 @@
+import { deriveModelCapabilityProfile } from "../model-capability.js";
+import { resolveCliModel } from "../model-resolver.js";
+export function clampLaneMaxUsd(settingsMaxUsd, foregroundMaxEstimatedUsd) {
+    return Math.min(settingsMaxUsd, foregroundMaxEstimatedUsd ?? Number.POSITIVE_INFINITY);
+}
+/** Shared provider-neutral model/profile resolution for research and fitness lanes. */
+export class LaneModelResolver {
+    deps;
+    constructor(deps) {
+        this.deps = deps;
+    }
+    capabilityProfile(model) {
+        return deriveModelCapabilityProfile({
+            contextWindow: model.contextWindow,
+            mode: this.deps.getSettingsManager().getModelCapabilitySettings().mode,
+        });
+    }
+    resolveModel(configuredPattern) {
+        if (configuredPattern) {
+            const resolved = resolveCliModel({
+                cliModel: configuredPattern,
+                modelRegistry: this.deps.getModelRegistry(),
+            });
+            if (resolved.model && this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) {
+                return resolved.model;
+            }
+            return undefined;
+        }
+        return this.deps.getModel() ?? undefined;
+    }
+    resolveShipment(laneSettings, missingModelReason) {
+        let laneProfile;
+        if (laneSettings.profile) {
+            const profileRef = laneSettings.profile.trim();
+            const registry = this.deps.getSettingsManager().getProfileRegistry();
+            laneProfile =
+                profileRef.startsWith("./") || profileRef.startsWith("../")
+                    ? registry.resolveProfileRef(profileRef, this.deps.getCwd())
+                    : registry.getProfile(profileRef);
+            if (!laneProfile)
+                return { ok: false, skipReason: "lane_profile_not_found" };
+        }
+        let model;
+        if (laneSettings.model) {
+            model = this.resolveModel(laneSettings.model);
+            if (!model)
+                return { ok: false, skipReason: missingModelReason };
+        }
+        else if (laneProfile?.model) {
+            model = this.resolveModel(laneProfile.model);
+            if (!model)
+                return { ok: false, skipReason: "no_lane_profile_model" };
+        }
+        else {
+            model = this.deps.getModel() ?? undefined;
+            if (!model)
+                return { ok: false, skipReason: missingModelReason };
+        }
+        if (this.deps.isModelExhausted(model)) {
+            return { ok: false, skipReason: `${model.provider}/${model.id} model exhausted: quota` };
+        }
+        return { ok: true, model, laneProfile };
+    }
+}
+//# sourceMappingURL=lane-model-resolver.js.map

package/dist/core/research/lane-model-resolver.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"lane-model-resolver.js","sourceRoot":"","sources":["../../../src/core/research/lane-model-resolver.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,4BAA4B,EAA+B,MAAM,wBAAwB,CAAC;AAEnG,OAAO,EAAE,eAAe,EAAE,MAAM,sBAAsB,CAAC;AAgBvD,MAAM,UAAU,eAAe,CAAC,cAAsB,EAAE,yBAAkC,EAAU;IACnG,OAAO,IAAI,CAAC,GAAG,CAAC,cAAc,EAAE,yBAAyB,IAAI,MAAM,CAAC,iBAAiB,CAAC,CAAC;AAAA,CACvF;AAED,uFAAuF;AACvF,MAAM,OAAO,iBAAiB;IACZ,IAAI,CAAwB;IAE7C,YAAY,IAA2B,EAAE;QACxC,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;IAAA,CACjB;IAED,iBAAiB,CAAC,KAAiB,EAA0B;QAC5D,OAAO,4BAA4B,CAAC;YACnC,aAAa,EAAE,KAAK,CAAC,aAAa;YAClC,IAAI,EAAE,IAAI,CAAC,IAAI,CAAC,kBAAkB,EAAE,CAAC,0BAA0B,EAAE,CAAC,IAAI;SACtE,CAAC,CAAC;IAAA,CACH;IAED,YAAY,CAAC,iBAAqC,EAA0B;QAC3E,IAAI,iBAAiB,EAAE,CAAC;YACvB,MAAM,QAAQ,GAAG,eAAe,CAAC;gBAChC,QAAQ,EAAE,iBAAiB;gBAC3B,aAAa,EAAE,IAAI,CAAC,IAAI,CAAC,gBAAgB,EAAE;aAC3C,CAAC,CAAC;YACH,IAAI,QAAQ,CAAC,KAAK,IAAI,IAAI,CAAC,IAAI,CAAC,gBAAgB,EAAE,CAAC,iBAAiB,CAAC,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC;gBACtF,OAAO,QAAQ,CAAC,KAAK,CAAC;YACvB,CAAC;YACD,OAAO,SAAS,CAAC;QAClB,CAAC;QACD,OAAO,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,IAAI,SAAS,CAAC;IAAA,CACzC;IAED,eAAe,CAAC,YAAkD,EAAE,kBAA0B,EAAgB;QAC7G,IAAI,WAA0C,CAAC;QAC/C,IAAI,YAAY,CAAC,OAAO,EAAE,CAAC;YAC1B,MAAM,UAAU,GAAG,YAAY,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC;YAC/C,MAAM,QAAQ,GAAG,IAAI,CAAC,IAAI,CAAC,kBAAkB,EAAE,CAAC,kBAAkB,EAAE,CAAC;YACrE,WAAW;gBACV,UAAU,CAAC,UAAU,CAAC,IAAI,CAAC,IAAI,UAAU,CAAC,UAAU,CAAC,KAAK,CAAC;oBAC1D,CAAC,CAAC,QAAQ,CAAC,iBAAiB,CAAC,UAAU,EAAE,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC;oBAC5D,CAAC,CAAC,QAAQ,CAAC,UAAU,CAAC,UAAU,CAAC,CAAC;YACpC,IAAI,CAAC,WAAW;gBAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,wBAAwB,EAAE,CAAC;QAC9E,CAAC;QAED,IAAI,KAA6B,CAAC;QAClC,IAAI,YAAY,CAAC,KAAK,EAAE,CAAC;YACxB,KAAK,GAAG,IAAI,CAAC,YAAY,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC;YAC9C,IAAI,CAAC,KAAK;gBAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,kBAAkB,EAAE,CAAC;QAClE,CAAC;aAAM,IAAI,WAAW,EAAE,KAAK,EAAE,CAAC;YAC/B,KAAK,GAAG,IAAI,CAAC,YAAY,CAAC,WAAW,CAAC,KAAK,CAAC,CAAC;YAC7C,IAAI,CAAC,KAAK;gBAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,uBAAuB,EAAE,CAAC;QACvE,CAAC;aAAM,CAAC;YACP,KAAK,GAAG,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,IAAI,SAAS,CAAC;YAC1C,IAAI,CAAC,KAAK;gBAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,kBAAkB,EAAE,CAAC;QAClE,CAAC;QACD,IAAI,IAAI,CAAC,IAAI,CAAC,gBAAgB,CAAC,KAAK,CAAC,EAAE,CAAC;YACvC,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,GAAG,KAAK,CAAC,QAAQ,IAAI,KAAK,CAAC,EAAE,yBAAyB,EAAE,CAAC;QAC1F,CAAC;QACD,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,KAAK,EAAE,WAAW,EAAE,CAAC;IAAA,CACxC;CACD","sourcesContent":["import type { Api, Model } from \"@caupulican/pi-ai\";\nimport { deriveModelCapabilityProfile, type ModelCapabilityProfile } from \"../model-capability.ts\";\nimport type { ModelRegistry } from \"../model-registry.ts\";\nimport { resolveCliModel } from \"../model-resolver.ts\";\nimport type { NormalizedProfile } from \"../profile-registry.ts\";\nimport type { SettingsManager } from \"../settings-manager.ts\";\n\nexport interface LaneModelResolverDeps {\n\tgetCwd(): string;\n\tgetModel(): Model<Api> | undefined;\n\tgetModelRegistry(): ModelRegistry;\n\tgetSettingsManager(): SettingsManager;\n\tisModelExhausted(model: Model<Api>): boolean;\n}\n\nexport type LaneShipment =\n\t| { ok: true; model: Model<Api>; laneProfile?: NormalizedProfile }\n\t| { ok: false; skipReason: string };\n\nexport function clampLaneMaxUsd(settingsMaxUsd: number, foregroundMaxEstimatedUsd?: number): number {\n\treturn Math.min(settingsMaxUsd, foregroundMaxEstimatedUsd ?? Number.POSITIVE_INFINITY);\n}\n\n/** Shared provider-neutral model/profile resolution for research and fitness lanes. */\nexport class LaneModelResolver {\n\tprivate readonly deps: LaneModelResolverDeps;\n\n\tconstructor(deps: LaneModelResolverDeps) {\n\t\tthis.deps = deps;\n\t}\n\n\tcapabilityProfile(model: Model<Api>): ModelCapabilityProfile {\n\t\treturn deriveModelCapabilityProfile({\n\t\t\tcontextWindow: model.contextWindow,\n\t\t\tmode: this.deps.getSettingsManager().getModelCapabilitySettings().mode,\n\t\t});\n\t}\n\n\tresolveModel(configuredPattern: string | undefined): Model<Api> | undefined {\n\t\tif (configuredPattern) {\n\t\t\tconst resolved = resolveCliModel({\n\t\t\t\tcliModel: configuredPattern,\n\t\t\t\tmodelRegistry: this.deps.getModelRegistry(),\n\t\t\t});\n\t\t\tif (resolved.model && this.deps.getModelRegistry().hasConfiguredAuth(resolved.model)) {\n\t\t\t\treturn resolved.model;\n\t\t\t}\n\t\t\treturn undefined;\n\t\t}\n\t\treturn this.deps.getModel() ?? undefined;\n\t}\n\n\tresolveShipment(laneSettings: { model?: string; profile?: string }, missingModelReason: string): LaneShipment {\n\t\tlet laneProfile: NormalizedProfile | undefined;\n\t\tif (laneSettings.profile) {\n\t\t\tconst profileRef = laneSettings.profile.trim();\n\t\t\tconst registry = this.deps.getSettingsManager().getProfileRegistry();\n\t\t\tlaneProfile =\n\t\t\t\tprofileRef.startsWith(\"./\") || profileRef.startsWith(\"../\")\n\t\t\t\t\t? registry.resolveProfileRef(profileRef, this.deps.getCwd())\n\t\t\t\t\t: registry.getProfile(profileRef);\n\t\t\tif (!laneProfile) return { ok: false, skipReason: \"lane_profile_not_found\" };\n\t\t}\n\n\t\tlet model: Model<Api> | undefined;\n\t\tif (laneSettings.model) {\n\t\t\tmodel = this.resolveModel(laneSettings.model);\n\t\t\tif (!model) return { ok: false, skipReason: missingModelReason };\n\t\t} else if (laneProfile?.model) {\n\t\t\tmodel = this.resolveModel(laneProfile.model);\n\t\t\tif (!model) return { ok: false, skipReason: \"no_lane_profile_model\" };\n\t\t} else {\n\t\t\tmodel = this.deps.getModel() ?? undefined;\n\t\t\tif (!model) return { ok: false, skipReason: missingModelReason };\n\t\t}\n\t\tif (this.deps.isModelExhausted(model)) {\n\t\t\treturn { ok: false, skipReason: `${model.provider}/${model.id} model exhausted: quota` };\n\t\t}\n\t\treturn { ok: true, model, laneProfile };\n\t}\n}\n"]}

package/dist/core/research/model-fitness-controller.d.ts

@@ -0,0 +1,38 @@
+import type { Usage } from "@caupulican/pi-ai";
+import type { IsolatedCompletionOptions, IsolatedCompletionResult } from "../agent-session.ts";
+import { type StoredFitnessReport } from "../models/fitness-store.ts";
+import type { LaneModelResolver } from "./lane-model-resolver.ts";
+import { type ModelFitnessReport } from "./model-fitness.ts";
+export interface ModelFitnessControllerDeps {
+    isDisposed(): boolean;
+    getSessionId(): string;
+    getAgentDir(): string;
+    addSpawnedUsage(usage: Usage, opts: {
+        label?: string;
+        sourceSessionId?: string;
+        reportId: string;
+    }): string | undefined;
+    runIsolatedCompletion(opts: IsolatedCompletionOptions): Promise<IsolatedCompletionResult>;
+}
+/** Owns model-fitness execution, cancellation, accounting, and host-keyed persistence. */
+export declare class ModelFitnessController {
+    private readonly abortController;
+    private readonly deps;
+    private readonly models;
+    constructor(deps: ModelFitnessControllerDeps, models: LaneModelResolver);
+    abort(): void;
+    run(args: {
+        model: string;
+        trials?: number;
+        toolCallId?: string;
+    }): Promise<{
+        started: true;
+        model: string;
+        report: ModelFitnessReport;
+    } | {
+        started: false;
+        skipReason: string;
+    }>;
+    getStoredReports(): StoredFitnessReport[];
+}
+//# sourceMappingURL=model-fitness-controller.d.ts.map

package/dist/core/research/model-fitness-controller.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-fitness-controller.d.ts","sourceRoot":"","sources":["../../../src/core/research/model-fitness-controller.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,KAAK,EAAE,MAAM,mBAAmB,CAAC;AAC/C,OAAO,KAAK,EAAE,yBAAyB,EAAE,wBAAwB,EAAE,MAAM,qBAAqB,CAAC;AAC/F,OAAO,EAAgB,KAAK,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AAEpF,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,0BAA0B,CAAC;AAClE,OAAO,EAAE,KAAK,kBAAkB,EAAwB,MAAM,oBAAoB,CAAC;AAEnF,MAAM,WAAW,0BAA0B;IAC1C,UAAU,IAAI,OAAO,CAAC;IACtB,YAAY,IAAI,MAAM,CAAC;IACvB,WAAW,IAAI,MAAM,CAAC;IACtB,eAAe,CACd,KAAK,EAAE,KAAK,EACZ,IAAI,EAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,eAAe,CAAC,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,GAClE,MAAM,GAAG,SAAS,CAAC;IACtB,qBAAqB,CAAC,IAAI,EAAE,yBAAyB,GAAG,OAAO,CAAC,wBAAwB,CAAC,CAAC;CAC1F;AAOD,0FAA0F;AAC1F,qBAAa,sBAAsB;IAClC,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAyB;IACzD,OAAO,CAAC,QAAQ,CAAC,IAAI,CAA6B;IAClD,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAoB;IAE3C,YAAY,IAAI,EAAE,0BAA0B,EAAE,MAAM,EAAE,iBAAiB,EAGtE;IAED,KAAK,IAAI,IAAI,CAEZ;IAEK,GAAG,CAAC,IAAI,EAAE;QACf,KAAK,EAAE,MAAM,CAAC;QACd,MAAM,CAAC,EAAE,MAAM,CAAC;QAChB,UAAU,CAAC,EAAE,MAAM,CAAC;KACpB,GAAG,OAAO,CAAC;QAAE,OAAO,EAAE,IAAI,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,kBAAkB,CAAA;KAAE,GAAG;QAAE,OAAO,EAAE,KAAK,CAAC;QAAC,UAAU,EAAE,MAAM,CAAA;KAAE,CAAC,CA2EjH;IAED,gBAAgB,IAAI,mBAAmB,EAAE,CAMxC;CACD","sourcesContent":["import { createHash } from \"node:crypto\";\nimport type { Usage } from \"@caupulican/pi-ai\";\nimport type { IsolatedCompletionOptions, IsolatedCompletionResult } from \"../agent-session.ts\";\nimport { FitnessStore, type StoredFitnessReport } from \"../models/fitness-store.ts\";\nimport { registerInFlightWork } from \"../reload-blockers.ts\";\nimport type { LaneModelResolver } from \"./lane-model-resolver.ts\";\nimport { type ModelFitnessReport, runModelFitnessProbe } from \"./model-fitness.ts\";\n\nexport interface ModelFitnessControllerDeps {\n\tisDisposed(): boolean;\n\tgetSessionId(): string;\n\tgetAgentDir(): string;\n\taddSpawnedUsage(\n\t\tusage: Usage,\n\t\topts: { label?: string; sourceSessionId?: string; reportId: string },\n\t): string | undefined;\n\trunIsolatedCompletion(opts: IsolatedCompletionOptions): Promise<IsolatedCompletionResult>;\n}\n\nfunction deriveSpawnedUsageReportId(kind: string, sessionId: string, identity: string): string {\n\tconst digest = createHash(\"sha256\").update(identity).digest(\"hex\").slice(0, 16);\n\treturn `${kind}:${sessionId}:${digest}`;\n}\n\n/** Owns model-fitness execution, cancellation, accounting, and host-keyed persistence. */\nexport class ModelFitnessController {\n\tprivate readonly abortController = new AbortController();\n\tprivate readonly deps: ModelFitnessControllerDeps;\n\tprivate readonly models: LaneModelResolver;\n\n\tconstructor(deps: ModelFitnessControllerDeps, models: LaneModelResolver) {\n\t\tthis.deps = deps;\n\t\tthis.models = models;\n\t}\n\n\tabort(): void {\n\t\tthis.abortController.abort();\n\t}\n\n\tasync run(args: {\n\t\tmodel: string;\n\t\ttrials?: number;\n\t\ttoolCallId?: string;\n\t}): Promise<{ started: true; model: string; report: ModelFitnessReport } | { started: false; skipReason: string }> {\n\t\tif (this.deps.isDisposed()) return { started: false, skipReason: \"session_disposed\" };\n\t\tconst resolved = this.models.resolveModel(args.model.trim() || undefined);\n\t\tif (!resolved) return { started: false, skipReason: \"model_unresolved_or_unauthenticated\" };\n\t\tconst capability = this.models.capabilityProfile(resolved);\n\t\tconst deregisterInFlight = registerInFlightWork(\n\t\t\tthis.deps.getAgentDir(),\n\t\t\t\"lane\",\n\t\t\t`fitness:${resolved.provider}/${resolved.id}`,\n\t\t);\n\t\ttry {\n\t\t\tconst spent: Usage = {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t};\n\t\t\tconst report = await runModelFitnessProbe({\n\t\t\t\ttrials: args.trials,\n\t\t\t\tsignal: this.abortController.signal,\n\t\t\t\tcapacityProbe:\n\t\t\t\t\tresolved.provider === \"ollama\" && resolved.contextWindow > 0\n\t\t\t\t\t\t? { registeredContextWindow: resolved.contextWindow }\n\t\t\t\t\t\t: undefined,\n\t\t\t\tcomplete: async ({ systemPrompt, userPrompt, signal }) => {\n\t\t\t\t\tconst callStarted = Date.now();\n\t\t\t\t\tconst completion = await this.deps.runIsolatedCompletion({\n\t\t\t\t\t\tsystemPrompt,\n\t\t\t\t\t\tmessages: [{ role: \"user\", content: [{ type: \"text\", text: userPrompt }], timestamp: Date.now() }],\n\t\t\t\t\t\tmodel: resolved,\n\t\t\t\t\t\tthinkingLevel: \"off\",\n\t\t\t\t\t\tmaxTokens: capability.laneMaxOutputTokens,\n\t\t\t\t\t\tsignal,\n\t\t\t\t\t\tcacheRetention: \"short\",\n\t\t\t\t\t\tlaneKind: \"fitness\",\n\t\t\t\t\t});\n\t\t\t\t\tconst callMs = Date.now() - callStarted;\n\t\t\t\t\tspent.input += completion.usage.input;\n\t\t\t\t\tspent.output += completion.usage.output;\n\t\t\t\t\tspent.cacheRead += completion.usage.cacheRead;\n\t\t\t\t\tspent.cacheWrite += completion.usage.cacheWrite;\n\t\t\t\t\tspent.totalTokens += completion.usage.totalTokens;\n\t\t\t\t\tspent.cost.input += completion.usage.cost.input;\n\t\t\t\t\tspent.cost.output += completion.usage.cost.output;\n\t\t\t\t\tspent.cost.cacheRead += completion.usage.cost.cacheRead;\n\t\t\t\t\tspent.cost.cacheWrite += completion.usage.cost.cacheWrite;\n\t\t\t\t\tspent.cost.total += completion.usage.cost.total;\n\t\t\t\t\treturn {\n\t\t\t\t\t\ttext: completion.text,\n\t\t\t\t\t\tcostUsd: completion.usage.cost.total,\n\t\t\t\t\t\tstopReason: String(completion.stopReason),\n\t\t\t\t\t\toutputTokens: completion.usage.output,\n\t\t\t\t\t\tevalMs: callMs,\n\t\t\t\t\t};\n\t\t\t\t},\n\t\t\t});\n\t\t\tconst modelRef = `${resolved.provider}/${resolved.id}`;\n\t\t\tif (!this.deps.isDisposed() && (spent.cost.total > 0 || spent.totalTokens > 0)) {\n\t\t\t\tconst identity = args.toolCallId\n\t\t\t\t\t? `toolcall:${args.toolCallId}`\n\t\t\t\t\t: `${modelRef} ${args.trials ?? \"default\"}`;\n\t\t\t\tconst reportId = deriveSpawnedUsageReportId(\"model-fitness\", this.deps.getSessionId(), identity);\n\t\t\t\tthis.deps.addSpawnedUsage(spent, { label: \"model-fitness\", reportId });\n\t\t\t}\n\t\t\ttry {\n\t\t\t\tif (!this.deps.isDisposed()) FitnessStore.forAgentDir(this.deps.getAgentDir()).save(modelRef, report);\n\t\t\t} catch {\n\t\t\t\t// Best-effort persistence must not fail the probe.\n\t\t\t}\n\t\t\treturn { started: true, model: modelRef, report };\n\t\t} finally {\n\t\t\tderegisterInFlight();\n\t\t}\n\t}\n\n\tgetStoredReports(): StoredFitnessReport[] {\n\t\ttry {\n\t\t\treturn FitnessStore.forAgentDir(this.deps.getAgentDir()).getForHost();\n\t\t} catch {\n\t\t\treturn [];\n\t\t}\n\t}\n}\n"]}

package/dist/core/research/model-fitness-controller.js

@@ -0,0 +1,106 @@
+import { createHash } from "node:crypto";
+import { FitnessStore } from "../models/fitness-store.js";
+import { registerInFlightWork } from "../reload-blockers.js";
+import { runModelFitnessProbe } from "./model-fitness.js";
+function deriveSpawnedUsageReportId(kind, sessionId, identity) {
+    const digest = createHash("sha256").update(identity).digest("hex").slice(0, 16);
+    return `${kind}:${sessionId}:${digest}`;
+}
+/** Owns model-fitness execution, cancellation, accounting, and host-keyed persistence. */
+export class ModelFitnessController {
+    abortController = new AbortController();
+    deps;
+    models;
+    constructor(deps, models) {
+        this.deps = deps;
+        this.models = models;
+    }
+    abort() {
+        this.abortController.abort();
+    }
+    async run(args) {
+        if (this.deps.isDisposed())
+            return { started: false, skipReason: "session_disposed" };
+        const resolved = this.models.resolveModel(args.model.trim() || undefined);
+        if (!resolved)
+            return { started: false, skipReason: "model_unresolved_or_unauthenticated" };
+        const capability = this.models.capabilityProfile(resolved);
+        const deregisterInFlight = registerInFlightWork(this.deps.getAgentDir(), "lane", `fitness:${resolved.provider}/${resolved.id}`);
+        try {
+            const spent = {
+                input: 0,
+                output: 0,
+                cacheRead: 0,
+                cacheWrite: 0,
+                totalTokens: 0,
+                cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+            };
+            const report = await runModelFitnessProbe({
+                trials: args.trials,
+                signal: this.abortController.signal,
+                capacityProbe: resolved.provider === "ollama" && resolved.contextWindow > 0
+                    ? { registeredContextWindow: resolved.contextWindow }
+                    : undefined,
+                complete: async ({ systemPrompt, userPrompt, signal }) => {
+                    const callStarted = Date.now();
+                    const completion = await this.deps.runIsolatedCompletion({
+                        systemPrompt,
+                        messages: [{ role: "user", content: [{ type: "text", text: userPrompt }], timestamp: Date.now() }],
+                        model: resolved,
+                        thinkingLevel: "off",
+                        maxTokens: capability.laneMaxOutputTokens,
+                        signal,
+                        cacheRetention: "short",
+                        laneKind: "fitness",
+                    });
+                    const callMs = Date.now() - callStarted;
+                    spent.input += completion.usage.input;
+                    spent.output += completion.usage.output;
+                    spent.cacheRead += completion.usage.cacheRead;
+                    spent.cacheWrite += completion.usage.cacheWrite;
+                    spent.totalTokens += completion.usage.totalTokens;
+                    spent.cost.input += completion.usage.cost.input;
+                    spent.cost.output += completion.usage.cost.output;
+                    spent.cost.cacheRead += completion.usage.cost.cacheRead;
+                    spent.cost.cacheWrite += completion.usage.cost.cacheWrite;
+                    spent.cost.total += completion.usage.cost.total;
+                    return {
+                        text: completion.text,
+                        costUsd: completion.usage.cost.total,
+                        stopReason: String(completion.stopReason),
+                        outputTokens: completion.usage.output,
+                        evalMs: callMs,
+                    };
+                },
+            });
+            const modelRef = `${resolved.provider}/${resolved.id}`;
+            if (!this.deps.isDisposed() && (spent.cost.total > 0 || spent.totalTokens > 0)) {
+                const identity = args.toolCallId
+                    ? `toolcall:${args.toolCallId}`
+                    : `${modelRef} ${args.trials ?? "default"}`;
+                const reportId = deriveSpawnedUsageReportId("model-fitness", this.deps.getSessionId(), identity);
+                this.deps.addSpawnedUsage(spent, { label: "model-fitness", reportId });
+            }
+            try {
+                if (!this.deps.isDisposed())
+                    FitnessStore.forAgentDir(this.deps.getAgentDir()).save(modelRef, report);
+            }
+            catch {
+                // Best-effort persistence must not fail the probe.
+            }
+            return { started: true, model: modelRef, report };
+        }
+        finally {
+            deregisterInFlight();
+        }
+    }
+    getStoredReports() {
+        try {
+            return FitnessStore.forAgentDir(this.deps.getAgentDir()).getForHost();
+        }
+        catch {
+            return [];
+        }
+    }
+}
+//# sourceMappingURL=model-fitness-controller.js.map

package/dist/core/research/model-fitness-controller.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-fitness-controller.js","sourceRoot":"","sources":["../../../src/core/research/model-fitness-controller.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAGzC,OAAO,EAAE,YAAY,EAA4B,MAAM,4BAA4B,CAAC;AACpF,OAAO,EAAE,oBAAoB,EAAE,MAAM,uBAAuB,CAAC;AAE7D,OAAO,EAA2B,oBAAoB,EAAE,MAAM,oBAAoB,CAAC;AAanF,SAAS,0BAA0B,CAAC,IAAY,EAAE,SAAiB,EAAE,QAAgB,EAAU;IAC9F,MAAM,MAAM,GAAG,UAAU,CAAC,QAAQ,CAAC,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;IAChF,OAAO,GAAG,IAAI,IAAI,SAAS,IAAI,MAAM,EAAE,CAAC;AAAA,CACxC;AAED,0FAA0F;AAC1F,MAAM,OAAO,sBAAsB;IACjB,eAAe,GAAG,IAAI,eAAe,EAAE,CAAC;IACxC,IAAI,CAA6B;IACjC,MAAM,CAAoB;IAE3C,YAAY,IAAgC,EAAE,MAAyB,EAAE;QACxE,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IAAA,CACrB;IAED,KAAK,GAAS;QACb,IAAI,CAAC,eAAe,CAAC,KAAK,EAAE,CAAC;IAAA,CAC7B;IAED,KAAK,CAAC,GAAG,CAAC,IAIT,EAAkH;QAClH,IAAI,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE;YAAE,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,UAAU,EAAE,kBAAkB,EAAE,CAAC;QACtF,MAAM,QAAQ,GAAG,IAAI,CAAC,MAAM,CAAC,YAAY,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,EAAE,IAAI,SAAS,CAAC,CAAC;QAC1E,IAAI,CAAC,QAAQ;YAAE,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,UAAU,EAAE,qCAAqC,EAAE,CAAC;QAC5F,MAAM,UAAU,GAAG,IAAI,CAAC,MAAM,CAAC,iBAAiB,CAAC,QAAQ,CAAC,CAAC;QAC3D,MAAM,kBAAkB,GAAG,oBAAoB,CAC9C,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,EACvB,MAAM,EACN,WAAW,QAAQ,CAAC,QAAQ,IAAI,QAAQ,CAAC,EAAE,EAAE,CAC7C,CAAC;QACF,IAAI,CAAC;YACJ,MAAM,KAAK,GAAU;gBACpB,KAAK,EAAE,CAAC;gBACR,MAAM,EAAE,CAAC;gBACT,SAAS,EAAE,CAAC;gBACZ,UAAU,EAAE,CAAC;gBACb,WAAW,EAAE,CAAC;gBACd,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE;aACpE,CAAC;YACF,MAAM,MAAM,GAAG,MAAM,oBAAoB,CAAC;gBACzC,MAAM,EAAE,IAAI,CAAC,MAAM;gBACnB,MAAM,EAAE,IAAI,CAAC,eAAe,CAAC,MAAM;gBACnC,aAAa,EACZ,QAAQ,CAAC,QAAQ,KAAK,QAAQ,IAAI,QAAQ,CAAC,aAAa,GAAG,CAAC;oBAC3D,CAAC,CAAC,EAAE,uBAAuB,EAAE,QAAQ,CAAC,aAAa,EAAE;oBACrD,CAAC,CAAC,SAAS;gBACb,QAAQ,EAAE,KAAK,EAAE,EAAE,YAAY,EAAE,UAAU,EAAE,MAAM,EAAE,EAAE,EAAE,CAAC;oBACzD,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;oBAC/B,MAAM,UAAU,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,qBAAqB,CAAC;wBACxD,YAAY;wBACZ,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,UAAU,EAAE,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE,EAAE,CAAC;wBAClG,KAAK,EAAE,QAAQ;wBACf,aAAa,EAAE,KAAK;wBACpB,SAAS,EAAE,UAAU,CAAC,mBAAmB;wBACzC,MAAM;wBACN,cAAc,EAAE,OAAO;wBACvB,QAAQ,EAAE,SAAS;qBACnB,CAAC,CAAC;oBACH,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,WAAW,CAAC;oBACxC,KAAK,CAAC,KAAK,IAAI,UAAU,CAAC,KAAK,CAAC,KAAK,CAAC;oBACtC,KAAK,CAAC,MAAM,IAAI,UAAU,CAAC,KAAK,CAAC,MAAM,CAAC;oBACxC,KAAK,CAAC,SAAS,IAAI,UAAU,CAAC,KAAK,CAAC,SAAS,CAAC;oBAC9C,KAAK,CAAC,UAAU,IAAI,UAAU,CAAC,KAAK,CAAC,UAAU,CAAC;oBAChD,KAAK,CAAC,WAAW,IAAI,UAAU,CAAC,KAAK,CAAC,WAAW,CAAC;oBAClD,KAAK,CAAC,IAAI,CAAC,KAAK,IAAI,UAAU,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC;oBAChD,KAAK,CAAC,IAAI,CAAC,MAAM,IAAI,UAAU,CAAC,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;oBAClD,KAAK,CAAC,IAAI,CAAC,SAAS,IAAI,UAAU,CAAC,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC;oBACxD,KAAK,CAAC,IAAI,CAAC,UAAU,IAAI,UAAU,CAAC,KAAK,CAAC,IAAI,CAAC,UAAU,CAAC;oBAC1D,KAAK,CAAC,IAAI,CAAC,KAAK,IAAI,UAAU,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC;oBAChD,OAAO;wBACN,IAAI,EAAE,UAAU,CAAC,IAAI;wBACrB,OAAO,EAAE,UAAU,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK;wBACpC,UAAU,EAAE,MAAM,CAAC,UAAU,CAAC,UAAU,CAAC;wBACzC,YAAY,EAAE,UAAU,CAAC,KAAK,CAAC,MAAM;wBACrC,MAAM,EAAE,MAAM;qBACd,CAAC;gBAAA,CACF;aACD,CAAC,CAAC;YACH,MAAM,QAAQ,GAAG,GAAG,QAAQ,CAAC,QAAQ,IAAI,QAAQ,CAAC,EAAE,EAAE,CAAC;YACvD,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,IAAI,KAAK,CAAC,WAAW,GAAG,CAAC,CAAC,EAAE,CAAC;gBAChF,MAAM,QAAQ,GAAG,IAAI,CAAC,UAAU;oBAC/B,CAAC,CAAC,YAAY,IAAI,CAAC,UAAU,EAAE;oBAC/B,CAAC,CAAC,GAAG,QAAQ,IAAI,IAAI,CAAC,MAAM,IAAI,SAAS,EAAE,CAAC;gBAC7C,MAAM,QAAQ,GAAG,0BAA0B,CAAC,eAAe,EAAE,IAAI,CAAC,IAAI,CAAC,YAAY,EAAE,EAAE,QAAQ,CAAC,CAAC;gBACjG,IAAI,CAAC,IAAI,CAAC,eAAe,CAAC,KAAK,EAAE,EAAE,KAAK,EAAE,eAAe,EAAE,QAAQ,EAAE,CAAC,CAAC;YACxE,CAAC;YACD,IAAI,CAAC;gBACJ,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE;oBAAE,YAAY,CAAC,WAAW,CAAC,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,CAAC,IAAI,CAAC,QAAQ,EAAE,MAAM,CAAC,CAAC;YACvG,CAAC;YAAC,MAAM,CAAC;gBACR,mDAAmD;YACpD,CAAC;YACD,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,QAAQ,EAAE,MAAM,EAAE,CAAC;QACnD,CAAC;gBAAS,CAAC;YACV,kBAAkB,EAAE,CAAC;QACtB,CAAC;IAAA,CACD;IAED,gBAAgB,GAA0B;QACzC,IAAI,CAAC;YACJ,OAAO,YAAY,CAAC,WAAW,CAAC,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,CAAC,UAAU,EAAE,CAAC;QACvE,CAAC;QAAC,MAAM,CAAC;YACR,OAAO,EAAE,CAAC;QACX,CAAC;IAAA,CACD;CACD","sourcesContent":["import { createHash } from \"node:crypto\";\nimport type { Usage } from \"@caupulican/pi-ai\";\nimport type { IsolatedCompletionOptions, IsolatedCompletionResult } from \"../agent-session.ts\";\nimport { FitnessStore, type StoredFitnessReport } from \"../models/fitness-store.ts\";\nimport { registerInFlightWork } from \"../reload-blockers.ts\";\nimport type { LaneModelResolver } from \"./lane-model-resolver.ts\";\nimport { type ModelFitnessReport, runModelFitnessProbe } from \"./model-fitness.ts\";\n\nexport interface ModelFitnessControllerDeps {\n\tisDisposed(): boolean;\n\tgetSessionId(): string;\n\tgetAgentDir(): string;\n\taddSpawnedUsage(\n\t\tusage: Usage,\n\t\topts: { label?: string; sourceSessionId?: string; reportId: string },\n\t): string | undefined;\n\trunIsolatedCompletion(opts: IsolatedCompletionOptions): Promise<IsolatedCompletionResult>;\n}\n\nfunction deriveSpawnedUsageReportId(kind: string, sessionId: string, identity: string): string {\n\tconst digest = createHash(\"sha256\").update(identity).digest(\"hex\").slice(0, 16);\n\treturn `${kind}:${sessionId}:${digest}`;\n}\n\n/** Owns model-fitness execution, cancellation, accounting, and host-keyed persistence. */\nexport class ModelFitnessController {\n\tprivate readonly abortController = new AbortController();\n\tprivate readonly deps: ModelFitnessControllerDeps;\n\tprivate readonly models: LaneModelResolver;\n\n\tconstructor(deps: ModelFitnessControllerDeps, models: LaneModelResolver) {\n\t\tthis.deps = deps;\n\t\tthis.models = models;\n\t}\n\n\tabort(): void {\n\t\tthis.abortController.abort();\n\t}\n\n\tasync run(args: {\n\t\tmodel: string;\n\t\ttrials?: number;\n\t\ttoolCallId?: string;\n\t}): Promise<{ started: true; model: string; report: ModelFitnessReport } | { started: false; skipReason: string }> {\n\t\tif (this.deps.isDisposed()) return { started: false, skipReason: \"session_disposed\" };\n\t\tconst resolved = this.models.resolveModel(args.model.trim() || undefined);\n\t\tif (!resolved) return { started: false, skipReason: \"model_unresolved_or_unauthenticated\" };\n\t\tconst capability = this.models.capabilityProfile(resolved);\n\t\tconst deregisterInFlight = registerInFlightWork(\n\t\t\tthis.deps.getAgentDir(),\n\t\t\t\"lane\",\n\t\t\t`fitness:${resolved.provider}/${resolved.id}`,\n\t\t);\n\t\ttry {\n\t\t\tconst spent: Usage = {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t};\n\t\t\tconst report = await runModelFitnessProbe({\n\t\t\t\ttrials: args.trials,\n\t\t\t\tsignal: this.abortController.signal,\n\t\t\t\tcapacityProbe:\n\t\t\t\t\tresolved.provider === \"ollama\" && resolved.contextWindow > 0\n\t\t\t\t\t\t? { registeredContextWindow: resolved.contextWindow }\n\t\t\t\t\t\t: undefined,\n\t\t\t\tcomplete: async ({ systemPrompt, userPrompt, signal }) => {\n\t\t\t\t\tconst callStarted = Date.now();\n\t\t\t\t\tconst completion = await this.deps.runIsolatedCompletion({\n\t\t\t\t\t\tsystemPrompt,\n\t\t\t\t\t\tmessages: [{ role: \"user\", content: [{ type: \"text\", text: userPrompt }], timestamp: Date.now() }],\n\t\t\t\t\t\tmodel: resolved,\n\t\t\t\t\t\tthinkingLevel: \"off\",\n\t\t\t\t\t\tmaxTokens: capability.laneMaxOutputTokens,\n\t\t\t\t\t\tsignal,\n\t\t\t\t\t\tcacheRetention: \"short\",\n\t\t\t\t\t\tlaneKind: \"fitness\",\n\t\t\t\t\t});\n\t\t\t\t\tconst callMs = Date.now() - callStarted;\n\t\t\t\t\tspent.input += completion.usage.input;\n\t\t\t\t\tspent.output += completion.usage.output;\n\t\t\t\t\tspent.cacheRead += completion.usage.cacheRead;\n\t\t\t\t\tspent.cacheWrite += completion.usage.cacheWrite;\n\t\t\t\t\tspent.totalTokens += completion.usage.totalTokens;\n\t\t\t\t\tspent.cost.input += completion.usage.cost.input;\n\t\t\t\t\tspent.cost.output += completion.usage.cost.output;\n\t\t\t\t\tspent.cost.cacheRead += completion.usage.cost.cacheRead;\n\t\t\t\t\tspent.cost.cacheWrite += completion.usage.cost.cacheWrite;\n\t\t\t\t\tspent.cost.total += completion.usage.cost.total;\n\t\t\t\t\treturn {\n\t\t\t\t\t\ttext: completion.text,\n\t\t\t\t\t\tcostUsd: completion.usage.cost.total,\n\t\t\t\t\t\tstopReason: String(completion.stopReason),\n\t\t\t\t\t\toutputTokens: completion.usage.output,\n\t\t\t\t\t\tevalMs: callMs,\n\t\t\t\t\t};\n\t\t\t\t},\n\t\t\t});\n\t\t\tconst modelRef = `${resolved.provider}/${resolved.id}`;\n\t\t\tif (!this.deps.isDisposed() && (spent.cost.total > 0 || spent.totalTokens > 0)) {\n\t\t\t\tconst identity = args.toolCallId\n\t\t\t\t\t? `toolcall:${args.toolCallId}`\n\t\t\t\t\t: `${modelRef} ${args.trials ?? \"default\"}`;\n\t\t\t\tconst reportId = deriveSpawnedUsageReportId(\"model-fitness\", this.deps.getSessionId(), identity);\n\t\t\t\tthis.deps.addSpawnedUsage(spent, { label: \"model-fitness\", reportId });\n\t\t\t}\n\t\t\ttry {\n\t\t\t\tif (!this.deps.isDisposed()) FitnessStore.forAgentDir(this.deps.getAgentDir()).save(modelRef, report);\n\t\t\t} catch {\n\t\t\t\t// Best-effort persistence must not fail the probe.\n\t\t\t}\n\t\t\treturn { started: true, model: modelRef, report };\n\t\t} finally {\n\t\t\tderegisterInFlight();\n\t\t}\n\t}\n\n\tgetStoredReports(): StoredFitnessReport[] {\n\t\ttry {\n\t\t\treturn FitnessStore.forAgentDir(this.deps.getAgentDir()).getForHost();\n\t\t} catch {\n\t\t\treturn [];\n\t\t}\n\t}\n}\n"]}

package/dist/core/research/model-fitness.d.ts

@@ -1,6 +1,6 @@
 /**
  * Model fitness probe: measures whether a candidate model can actually drive the harness's
- * subagent contracts — the research lane, the scout-worker lane, and the routing judge — by
+ * subagent contracts — the research lane, the delegated-worker lane, and the routing judge — by
  * running each real runner against the model and scoring parse/success rates plus judge
  * discrimination. Provider-free: the completion executor is injected, so this works against any
  * registered model (local Ollama models included) and against faux providers in tests.

package/dist/core/research/model-fitness.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"model-fitness.d.ts","sourceRoot":"","sources":["../../../src/core/research/model-fitness.ts"],"names":[],"mappings":"AAOA;;;;;;GAMG;AAEH,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,iGAAiG;IACjG,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,kGAAkG;IAClG,MAAM,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,MAAM,eAAe,GAAG,CAAC,IAAI,EAAE;IACpC,YAAY,EAAE,MAAM,CAAC;IACrB,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB,KAAK,OAAO,CAAC,iBAAiB,CAAC,CAAC;AAEjC,MAAM,WAAW,kBAAkB;IAClC,MAAM,EAAE,MAAM,CAAC;IACf,0EAA0E;IAC1E,QAAQ,EAAE,OAAO,CAAC;CAClB;AAED,qFAAqF;AACrF,eAAO,MAAM,6BAA6B,EAAE,SAAS,kBAAkB,EAOtE,CAAC;AAEF,MAAM,WAAW,oBAAoB;IACpC,uBAAuB,EAAE,MAAM,CAAC;IAChC,4FAA4F;IAC5F,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,mBAAmB;IACnC,QAAQ,EAAE,eAAe,CAAC;IAC1B,0CAA0C;IAC1C,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,wDAAwD;IACxD,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,YAAY,CAAC,EAAE,SAAS,kBAAkB,EAAE,CAAC;IAC7C,uFAAuF;IACvF,aAAa,CAAC,EAAE,oBAAoB,CAAC;IACrC,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,gFAAgF;IAChF,GAAG,CAAC,EAAE,MAAM,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,gBAAgB;IAChC,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,yFAAyF;IACzF,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,iBAAiB;IACjC,MAAM,EAAE,MAAM,CAAC;IACf,gBAAgB,EAAE,MAAM,CAAC;IACzB,aAAa,EAAE,MAAM,CAAC;IACtB,YAAY,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,qFAAqF;IACrF,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,oBAAoB;IACpC,uBAAuB,EAAE,MAAM,CAAC;IAChC,mBAAmB,EAAE,MAAM,CAAC;IAC5B,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,kBAAkB;IAClC,MAAM,EAAE,MAAM,CAAC;IACf,yFAAyF;IACzF,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,MAAM,EAAE,gBAAgB,CAAC;IACzB,KAAK,EAAE,iBAAiB,CAAC;IACzB,8EAA8E;IAC9E,MAAM,EAAE,gBAAgB,CAAC;IACzB,yFAAyF;IACzF,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,qGAAqG;IACrG,MAAM,EAAE,gBAAgB,CAAC;IACzB,qFAAqF;IACrF,QAAQ,CAAC,EAAE,oBAAoB,CAAC;IAChC,YAAY,EAAE,MAAM,CAAC;CACrB;AAED,yFAAyF;AACzF,eAAO,MAAM,0BAA0B,QAK3B,CAAC;AAEb,eAAO,MAAM,6BAA6B,QAK9B,CAAC;AAEb,eAAO,MAAM,4BAA4B,QAI7B,CAAC;AASb,OAAO,EAAE,6BAA6B,IAAI,0BAA0B,EAAE,MAAM,6BAA6B,CAAC;AA0K1G,wBAAsB,oBAAoB,CAAC,OAAO,EAAE,mBAAmB,GAAG,OAAO,CAAC,kBAAkB,CAAC,CA+KpG;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,kBAAkB,GAAG,OAAO,CAQpE;AAED,kGAAkG;AAClG,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,kBAAkB,GAAG,MAAM,CAsB1F","sourcesContent":["import { runBoundedCompletion } from \"../autonomy/bounded-completion.ts\";\nimport type { CapabilityEnvelope } from \"../autonomy/contracts.ts\";\nimport { CURATION_DIGEST_SYSTEM_PROMPT } from \"../context/brain-curator.ts\";\nimport { runWorker } from \"../delegation/worker-runner.ts\";\nimport { runRouteJudge } from \"../model-router/route-judge.ts\";\nimport { runResearch } from \"./research-runner.ts\";\n\n/**\n * Model fitness probe: measures whether a candidate model can actually drive the harness's\n * subagent contracts — the research lane, the scout-worker lane, and the routing judge — by\n * running each real runner against the model and scoring parse/success rates plus judge\n * discrimination. Provider-free: the completion executor is injected, so this works against any\n * registered model (local Ollama models included) and against faux providers in tests.\n */\n\nexport interface FitnessCompletion {\n\ttext: string;\n\tcostUsd: number;\n\tstopReason: string;\n\t/** Output tokens generated (for tok/s). Optional: providers that don't report it are skipped. */\n\toutputTokens?: number;\n\t/** Pure generation time in ms (e.g. Ollama eval_duration). Falls back to wall-clock if absent. */\n\tevalMs?: number;\n}\n\nexport type FitnessComplete = (args: {\n\tsystemPrompt: string;\n\tuserPrompt: string;\n\tsignal?: AbortSignal;\n}) => Promise<FitnessCompletion>;\n\nexport interface JudgeFitnessPrompt {\n\tprompt: string;\n\t/** True when the prompt is planning-shaped and must never route cheap. */\n\tplanning: boolean;\n}\n\n/** Default judge probe set: three planning-shaped prompts, three trivial lookups. */\nexport const DEFAULT_JUDGE_FITNESS_PROMPTS: readonly JudgeFitnessPrompt[] = [\n\t{ prompt: \"how should we plan the migration of the session storage layer?\", planning: true },\n\t{ prompt: \"design an approach for splitting the settings manager\", planning: true },\n\t{ prompt: \"draft a roadmap for the autonomy rework\", planning: true },\n\t{ prompt: \"what does the resolvePath function return?\", planning: false },\n\t{ prompt: \"list the files in the delegation module\", planning: false },\n\t{ prompt: \"why is this test flaky?\", planning: false },\n];\n\nexport interface CapacityProbeOptions {\n\tregisteredContextWindow: number;\n\t/** Smallest candidate window to try before declaring the capacity unknown. Default 1024. */\n\tminContextWindow?: number;\n}\n\nexport interface ModelFitnessOptions {\n\tcomplete: FitnessComplete;\n\t/** Trials per lane surface. Default 3. */\n\ttrials?: number;\n\t/** Wall-clock budget per call in ms. Default 120000. */\n\tmaxWallClockMs?: number;\n\tjudgePrompts?: readonly JudgeFitnessPrompt[];\n\t/** Optional local-model capacity lane: measures the actually served context window. */\n\tcapacityProbe?: CapacityProbeOptions;\n\tsignal?: AbortSignal;\n\t/** Injected clock for latency measurement (test seam). Defaults to Date.now. */\n\tnow?: () => number;\n}\n\nexport interface LaneFitnessScore {\n\tsucceeded: number;\n\ttotal: number;\n\toutcomes: string[];\n\tmeanMs: number;\n\t/** Mean output tokens/second across the surface's calls; undefined when not reported. */\n\ttokensPerSecond?: number;\n}\n\nexport interface JudgeFitnessScore {\n\tparsed: number;\n\tplanningElevated: number;\n\tplanningTotal: number;\n\ttrivialCheap: number;\n\ttrivialTotal: number;\n\ttotal: number;\n\toutcomes: string[];\n\tmeanMs: number;\n\t/** Mean output tokens/second across the judge calls; undefined when not reported. */\n\ttokensPerSecond?: number;\n}\n\nexport interface CapacityFitnessScore {\n\tregisteredContextWindow: number;\n\tservedContextWindow: number;\n\toutcomes: string[];\n\tmeanMs: number;\n}\n\nexport interface ModelFitnessReport {\n\ttrials: number;\n\t/** Aggregate output tokens/second across ALL probe calls (the headline speed number). */\n\ttokensPerSecond?: number;\n\tresearch: LaneFitnessScore;\n\tworker: LaneFitnessScore;\n\tjudge: JudgeFitnessScore;\n\t/** Heavy-lifter surface: can the model formulate a structured search plan? */\n\tsearch: LaneFitnessScore;\n\t/** Heavy-lifter surface: can the model emit a well-formed tool call against a schema? */\n\ttoolCall: LaneFitnessScore;\n\t/** Curator surface: can the model digest a context chunk to strict JSON WITHOUT losing key facts? */\n\tdigest: LaneFitnessScore;\n\t/** Local-model capacity lane: measured context window the server actually serves. */\n\tcapacity?: CapacityFitnessScore;\n\ttotalCostUsd: number;\n}\n\n/** Static prompts for the heavy-lifter surfaces (stable for provider prompt caching). */\nexport const SEARCH_PROBE_SYSTEM_PROMPT = [\n\t\"You plan code searches for a coding agent. You never answer the question yourself.\",\n\t\"Given a question about a codebase, respond with STRICT JSON only - no prose:\",\n\t'{\"queries\":[{\"pattern\":\"<regex or literal to grep>\",\"glob\":\"<file glob like **/*.ts>\"}]}',\n\t\"Return 1 to 4 queries, most specific first.\",\n].join(\"\\n\");\n\nexport const TOOL_CALL_PROBE_SYSTEM_PROMPT = [\n\t\"You operate tools for a coding agent. You have exactly one tool:\",\n\t\"grep(pattern: string, path: string) - search files under a path for a pattern.\",\n\t\"Respond to every task with STRICT JSON only - no prose:\",\n\t'{\"tool\":\"grep\",\"arguments\":{\"pattern\":\"<pattern>\",\"path\":\"<path>\"}}',\n].join(\"\\n\");\n\nexport const CAPACITY_PROBE_SYSTEM_PROMPT = [\n\t\"You are a context-window capacity probe for a local model server.\",\n\t\"Find the unique NEEDLE token in the user's text and echo that token only.\",\n\t\"Do not summarize, explain, or add punctuation.\",\n].join(\"\\n\");\n\nconst SEARCH_PROBE_TASKS: readonly string[] = [\n\t\"Where is the retry/backoff logic for HTTP requests implemented?\",\n\t\"Which files define the settings for background research?\",\n\t\"Find where session entries of type custom are appended.\",\n];\n\n// The probe measures the REAL curation contract — same prompt the BrainCurator ships.\nexport { CURATION_DIGEST_SYSTEM_PROMPT as DIGEST_PROBE_SYSTEM_PROMPT } from \"../context/brain-curator.ts\";\n\n/**\n * Digest probe chunks each carry a NONCE identifier that cannot be guessed from the\n * instructions: acceptance requires the digest to RETAIN the nonce verbatim, so the score\n * measures extraction fidelity, not narration (a model cannot pass by paraphrasing).\n */\nconst DIGEST_PROBE_TASKS: readonly { chunk: string; nonce: string }[] = [\n\t{\n\t\tnonce: \"retryWithJitter_zx41\",\n\t\tchunk: [\n\t\t\t\"grep results for 'retry' under src/http:\",\n\t\t\t\"src/http/client.ts:88: export function retryWithJitter_zx41(fn, attempts = 3) {\",\n\t\t\t\"src/http/client.ts:112:   // exponential backoff capped at 30s\",\n\t\t\t\"src/http/pool.ts:41:   client.retry = false\",\n\t\t].join(\"\\n\"),\n\t},\n\t{\n\t\tnonce: \"ERR_QM_7734\",\n\t\tchunk: [\n\t\t\t\"$ npm run migrate\",\n\t\t\t\"migrating 14 files...\",\n\t\t\t\"error ERR_QM_7734: column 'owner_id' missing on table sessions (migration 0009)\",\n\t\t\t\"exit code 1\",\n\t\t].join(\"\\n\"),\n\t},\n\t{\n\t\tnonce: \"v3.9.2-hotfix.1\",\n\t\tchunk: [\n\t\t\t\"read package.json (34 lines):\",\n\t\t\t'  \"name\": \"acme-billing\",',\n\t\t\t'  \"version\": \"v3.9.2-hotfix.1\",',\n\t\t\t'  \"engines\": { \"node\": \">=22\" },',\n\t\t].join(\"\\n\"),\n\t},\n];\n\nfunction parseDigest(text: string, nonce: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst digest = (parsed as { digest?: unknown }).digest;\n\tif (typeof digest !== \"string\") return false;\n\tconst trimmed = digest.trim();\n\t// Bounded and faithful: short enough to be a stub annotation, still carrying the nonce fact.\n\treturn trimmed.length > 0 && trimmed.length <= 240 && trimmed.includes(nonce);\n}\n\nconst TOOL_CALL_PROBE_TASKS: readonly string[] = [\n\t\"Find usages of the function resolveCliModel under src/.\",\n\t\"Search for the string 'budget_exhausted' in the core directory.\",\n\t\"Locate where LaneTracker is instantiated under src/core.\",\n];\n\nfunction parseSearchPlan(text: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst queries = (parsed as { queries?: unknown }).queries;\n\tif (!Array.isArray(queries) || queries.length === 0 || queries.length > 8) return false;\n\treturn queries.every(\n\t\t(query) =>\n\t\t\tquery &&\n\t\t\ttypeof query === \"object\" &&\n\t\t\ttypeof (query as { pattern?: unknown }).pattern === \"string\" &&\n\t\t\t(query as { pattern: string }).pattern.trim().length > 0,\n\t);\n}\n\nfunction parseToolCall(text: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst record = parsed as { tool?: unknown; arguments?: unknown };\n\tif (record.tool !== \"grep\") return false;\n\tconst args = record.arguments;\n\tif (!args || typeof args !== \"object\" || Array.isArray(args)) return false;\n\tconst pattern = (args as { pattern?: unknown }).pattern;\n\tconst path = (args as { path?: unknown }).path;\n\treturn (\n\t\ttypeof pattern === \"string\" && pattern.trim().length > 0 && typeof path === \"string\" && path.trim().length > 0\n\t);\n}\n\nfunction extractJsonObject(text: string): unknown | undefined {\n\tconst trimmed = text.trim();\n\tconst candidates: string[] = [trimmed];\n\tconst fenced = /```(?:json)?\\s*([\\s\\S]*?)```/.exec(trimmed);\n\tif (fenced?.[1]) candidates.push(fenced[1].trim());\n\tconst start = trimmed.indexOf(\"{\");\n\tconst end = trimmed.lastIndexOf(\"}\");\n\tif (start >= 0 && end > start) candidates.push(trimmed.slice(start, end + 1));\n\tfor (const candidate of candidates) {\n\t\ttry {\n\t\t\tconst parsed = JSON.parse(candidate);\n\t\t\tif (parsed && typeof parsed === \"object\" && !Array.isArray(parsed)) return parsed;\n\t\t} catch {\n\t\t\t// try next candidate\n\t\t}\n\t}\n\treturn undefined;\n}\n\nfunction fitnessEnvelope(): CapabilityEnvelope {\n\treturn {\n\t\tid: \"model-fitness-probe\",\n\t\tcapabilities: [\"research\", \"read_files\", \"memory_read\"],\n\t\tmaxEstimatedUsd: 1,\n\t\tcreatedAt: new Date().toISOString(),\n\t};\n}\n\nfunction buildCapacityProbePrompt(targetTokens: number, startNeedle: string, endNeedle: string): string {\n\tconst targetChars = Math.max(\n\t\tstartNeedle.length + endNeedle.length + 64,\n\t\ttargetTokens * 4 - CAPACITY_PROBE_SYSTEM_PROMPT.length,\n\t);\n\tconst prefix = `CAPACITY PROBE START\\n${startNeedle}\\n`;\n\tconst suffix = `\\n${endNeedle}\\nCAPACITY PROBE END`;\n\tconst fillerLength = Math.max(0, targetChars - prefix.length - suffix.length);\n\treturn `${prefix}${\"x \".repeat(Math.ceil(fillerLength / 2)).slice(0, fillerLength)}${suffix}`;\n}\n\nasync function measureServedContextWindow(args: {\n\tprobe: CapacityProbeOptions;\n\tcomplete: FitnessComplete;\n\tmaxWallClockMs: number;\n\tsignal?: AbortSignal;\n\tnow: () => number;\n}): Promise<{ score: CapacityFitnessScore; costUsd: number }> {\n\tconst registered = Math.max(1, Math.floor(args.probe.registeredContextWindow));\n\tconst minWindow = Math.max(1, Math.floor(args.probe.minContextWindow ?? 1024));\n\tconst nonce = `${registered.toString(16).toUpperCase()}_${minWindow.toString(16).toUpperCase()}`;\n\tconst startNeedle = `NEEDLE_START_${nonce}`;\n\tconst endNeedle = `NEEDLE_END_${nonce}`;\n\tconst score: CapacityFitnessScore = {\n\t\tregisteredContextWindow: registered,\n\t\tservedContextWindow: 0,\n\t\toutcomes: [],\n\t\tmeanMs: 0,\n\t};\n\tlet candidate = registered;\n\tlet costUsd = 0;\n\tlet calls = 0;\n\twhile (candidate >= minWindow) {\n\t\tconst started = args.now();\n\t\tcalls++;\n\t\tconst bounded = await runBoundedCompletion({\n\t\t\tmaxWallClockMs: args.maxWallClockMs,\n\t\t\tsignal: args.signal,\n\t\t\texecute: (signal) =>\n\t\t\t\targs.complete({\n\t\t\t\t\tsystemPrompt: CAPACITY_PROBE_SYSTEM_PROMPT,\n\t\t\t\t\tuserPrompt: buildCapacityProbePrompt(candidate, startNeedle, endNeedle),\n\t\t\t\t\tsignal,\n\t\t\t\t}),\n\t\t});\n\t\tscore.meanMs += args.now() - started;\n\t\tif (bounded.completion) costUsd += bounded.completion.costUsd;\n\t\tconst output = bounded.completion?.text.trim() ?? \"\";\n\t\tconst recalled = !bounded.failure && output.includes(startNeedle) && output.includes(endNeedle);\n\t\tscore.outcomes.push(`${candidate}:${recalled ? \"ok\" : (bounded.failure?.status ?? \"miss\")}`);\n\t\tif (recalled) {\n\t\t\tscore.servedContextWindow = candidate;\n\t\t\tbreak;\n\t\t}\n\t\tcandidate = Math.floor(candidate / 2);\n\t}\n\tif (score.servedContextWindow === 0) score.servedContextWindow = minWindow;\n\tscore.meanMs = calls > 0 ? Math.round(score.meanMs / calls) : 0;\n\treturn { score, costUsd };\n}\n\nexport async function runModelFitnessProbe(options: ModelFitnessOptions): Promise<ModelFitnessReport> {\n\tconst trials = Math.max(1, Math.min(options.trials ?? 3, 20));\n\tconst maxWallClockMs = options.maxWallClockMs ?? 120_000;\n\tconst judgePrompts = options.judgePrompts ?? DEFAULT_JUDGE_FITNESS_PROMPTS;\n\tconst now = options.now ?? Date.now;\n\tlet totalCostUsd = 0;\n\n\t// Token-speed instrumentation: the lane runners' own contracts carry text/cost only, so the\n\t// completer is wrapped once here and generation stats are accumulated per surface.\n\tconst overallSpeed = { tokens: 0, evalMs: 0 };\n\tlet surfaceSpeed = { tokens: 0, evalMs: 0 };\n\tconst complete: FitnessComplete = async (args) => {\n\t\tconst completion = await options.complete(args);\n\t\tconst tokens = completion.outputTokens ?? 0;\n\t\tconst evalMs = completion.evalMs ?? 0;\n\t\tif (tokens > 0 && evalMs > 0) {\n\t\t\tsurfaceSpeed.tokens += tokens;\n\t\t\tsurfaceSpeed.evalMs += evalMs;\n\t\t\toverallSpeed.tokens += tokens;\n\t\t\toverallSpeed.evalMs += evalMs;\n\t\t}\n\t\treturn completion;\n\t};\n\tconst takeSurfaceSpeed = (): number | undefined => {\n\t\tconst speed =\n\t\t\tsurfaceSpeed.evalMs > 0 ? Math.round((surfaceSpeed.tokens / surfaceSpeed.evalMs) * 1000) : undefined;\n\t\tsurfaceSpeed = { tokens: 0, evalMs: 0 };\n\t\treturn speed;\n\t};\n\n\tconst research: LaneFitnessScore = { succeeded: 0, total: trials, outcomes: [], meanMs: 0 };\n\tfor (let i = 0; i < trials; i++) {\n\t\tconst started = now();\n\t\tconst result = await runResearch({\n\t\t\tquery: `fitness:probe requirements:req-${i}`,\n\t\t\tcontext: [\n\t\t\t\t\"Goal: add a retry helper to the HTTP client module\",\n\t\t\t\t\"Open requirements:\",\n\t\t\t\t\"- Find what retry/backoff conventions the codebase already uses\",\n\t\t\t\t\"- Identify which call sites would adopt the helper\",\n\t\t\t].join(\"\\n\"),\n\t\t\tenvelope: fitnessEnvelope(),\n\t\t\tmaxUsd: 1,\n\t\t\tmaxSources: 8,\n\t\t\tmaxFindings: 5,\n\t\t\tmaxWallClockMs,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tresearch.meanMs += now() - started;\n\t\ttotalCostUsd += result.costUsd;\n\t\tif (result.status === \"succeeded\") research.succeeded++;\n\t\tresearch.outcomes.push(`${result.status}/${result.reasonCode}`);\n\t}\n\tresearch.meanMs = Math.round(research.meanMs / trials);\n\tresearch.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst worker: LaneFitnessScore = { succeeded: 0, total: trials, outcomes: [], meanMs: 0 };\n\tfor (let i = 0; i < trials; i++) {\n\t\tconst started = now();\n\t\tconst outcome = await runWorker({\n\t\t\trequest: {\n\t\t\t\tid: `fitness-worker-${i}`,\n\t\t\t\tinstructions:\n\t\t\t\t\t\"Summarize in two sentences what a capability envelope is: a declared set of allowed tools, paths, and capability names that bounds what a delegated worker may do.\",\n\t\t\t\troute: { tier: \"cheap\", risk: \"read-only\", confidence: 1, reasonCode: \"fitness_probe\", reasons: [] },\n\t\t\t\tenvelope: { id: `fitness-env-${i}`, capabilities: [\"read_files\"], maxEstimatedUsd: 1 },\n\t\t\t\tmaxEstimatedUsd: 1,\n\t\t\t},\n\t\t\tmaxUsd: 1,\n\t\t\tmaxWallClockMs,\n\t\t\tusageReportId: `fitness:${i}`,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tworker.meanMs += now() - started;\n\t\ttotalCostUsd += outcome.costUsd;\n\t\tif (outcome.result.status === \"completed\" && outcome.accepted) worker.succeeded++;\n\t\tworker.outcomes.push(`${outcome.result.status}/${outcome.reasonCode}`);\n\t}\n\tworker.meanMs = Math.round(worker.meanMs / trials);\n\tworker.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst judge: JudgeFitnessScore = {\n\t\tparsed: 0,\n\t\tplanningElevated: 0,\n\t\tplanningTotal: judgePrompts.filter((entry) => entry.planning).length,\n\t\ttrivialCheap: 0,\n\t\ttrivialTotal: judgePrompts.filter((entry) => !entry.planning).length,\n\t\ttotal: judgePrompts.length,\n\t\toutcomes: [],\n\t\tmeanMs: 0,\n\t};\n\tfor (const entry of judgePrompts) {\n\t\tconst started = now();\n\t\tconst result = await runRouteJudge({\n\t\t\tprompt: entry.prompt,\n\t\t\tbaseline: { tier: \"cheap\", risk: \"read-only\", confidence: 0.5, reasonCode: \"fitness_probe\", reasons: [] },\n\t\t\tmaxWallClockMs,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tjudge.meanMs += now() - started;\n\t\ttotalCostUsd += result.costUsd;\n\t\tconst tier = result.decision.tier;\n\t\tif (result.verdict) {\n\t\t\tjudge.parsed++;\n\t\t\t// A useful judge must both keep planning off the cheap tier AND actually send trivial\n\t\t\t// prompts there — all-medium verdicts are safe but save nothing.\n\t\t\tif (entry.planning && tier !== \"cheap\") judge.planningElevated++;\n\t\t\tif (!entry.planning && tier === \"cheap\") judge.trivialCheap++;\n\t\t}\n\t\tjudge.outcomes.push(\n\t\t\t`\"${entry.prompt.slice(0, 40)}\" -> ${tier}${result.fallbackReason ? ` (${result.fallbackReason})` : \"\"}`,\n\t\t);\n\t}\n\tjudge.meanMs = judgePrompts.length > 0 ? Math.round(judge.meanMs / judgePrompts.length) : 0;\n\tjudge.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst probeSurface = async (\n\t\tsystemPrompt: string,\n\t\ttasks: readonly string[],\n\t\taccepts: (text: string, taskIndex: number) => boolean,\n\t): Promise<LaneFitnessScore> => {\n\t\tconst score: LaneFitnessScore = { succeeded: 0, total: tasks.length, outcomes: [], meanMs: 0 };\n\t\tfor (const [taskIndex, task] of tasks.entries()) {\n\t\t\tconst started = now();\n\t\t\t// Same wall-clock envelope as the lane surfaces — a hung model must not hang the probe.\n\t\t\tconst bounded = await runBoundedCompletion({\n\t\t\t\tmaxWallClockMs,\n\t\t\t\tsignal: options.signal,\n\t\t\t\texecute: (signal) => complete({ systemPrompt, userPrompt: task, signal }),\n\t\t\t});\n\t\t\tif (bounded.completion) totalCostUsd += bounded.completion.costUsd;\n\t\t\tif (bounded.failure || !bounded.completion) {\n\t\t\t\tscore.outcomes.push(bounded.failure ? bounded.failure.status : \"completion_error\");\n\t\t\t} else {\n\t\t\t\tconst ok = accepts(bounded.completion.text, taskIndex);\n\t\t\t\tif (ok) score.succeeded++;\n\t\t\t\tscore.outcomes.push(ok ? \"ok\" : \"unparseable_output\");\n\t\t\t}\n\t\t\tscore.meanMs += now() - started;\n\t\t}\n\t\tscore.meanMs = tasks.length > 0 ? Math.round(score.meanMs / tasks.length) : 0;\n\t\treturn score;\n\t};\n\n\tconst search = await probeSurface(SEARCH_PROBE_SYSTEM_PROMPT, SEARCH_PROBE_TASKS, parseSearchPlan);\n\tsearch.tokensPerSecond = takeSurfaceSpeed();\n\tconst toolCall = await probeSurface(TOOL_CALL_PROBE_SYSTEM_PROMPT, TOOL_CALL_PROBE_TASKS, parseToolCall);\n\ttoolCall.tokensPerSecond = takeSurfaceSpeed();\n\tconst digest = await probeSurface(\n\t\tCURATION_DIGEST_SYSTEM_PROMPT,\n\t\tDIGEST_PROBE_TASKS.map((task) => task.chunk),\n\t\t(text, taskIndex) => parseDigest(text, DIGEST_PROBE_TASKS[taskIndex]!.nonce),\n\t);\n\tdigest.tokensPerSecond = takeSurfaceSpeed();\n\n\tlet capacity: CapacityFitnessScore | undefined;\n\tif (options.capacityProbe) {\n\t\tconst result = await measureServedContextWindow({\n\t\t\tprobe: options.capacityProbe,\n\t\t\tcomplete,\n\t\t\tmaxWallClockMs,\n\t\t\tsignal: options.signal,\n\t\t\tnow,\n\t\t});\n\t\tcapacity = result.score;\n\t\ttotalCostUsd += result.costUsd;\n\t}\n\n\tconst tokensPerSecond =\n\t\toverallSpeed.evalMs > 0 ? Math.round((overallSpeed.tokens / overallSpeed.evalMs) * 1000) : undefined;\n\n\treturn { trials, tokensPerSecond, research, worker, judge, search, toolCall, digest, capacity, totalCostUsd };\n}\n\n/**\n * Pure verdict: true when the probe found ZERO successes on every LANE surface it actually graded\n * AND the judge (if it ran) also failed. A lane/judge with total 0 (i.e. never run) carries no\n * evidence and is excluded from the lane check — but at least one lane must actually have been\n * graded for an all-failed verdict at all: `gradedLanes.every(...)` is vacuously true over an\n * empty array, so a report where only the judge ran (every research/worker/search/toolCall/digest\n * lane is ungraded) is excluded explicitly rather than misread as \"all lanes failed\" on zero lane\n * evidence. An empty/degenerate report (nothing graded at all, lanes AND judge) is likewise never\n * mistaken for a failed one. This is the gate adoption flows must consult before assigning a role —\n * see `isProbeAllFailed` callers in interactive-mode.ts and agent-session.ts.\n */\nexport function isProbeAllFailed(report: ModelFitnessReport): boolean {\n\tconst lanes = [report.research, report.worker, report.search, report.toolCall, report.digest];\n\tconst gradedLanes = lanes.filter((lane) => lane.total > 0);\n\tconst judgeGraded = report.judge.total > 0;\n\tif (gradedLanes.length === 0 && !judgeGraded) return false;\n\tconst lanesAllFailed = gradedLanes.length > 0 && gradedLanes.every((lane) => lane.succeeded === 0);\n\tconst judgeFailed = !judgeGraded || report.judge.parsed === 0;\n\treturn lanesAllFailed && judgeFailed;\n}\n\n/** Compact human-readable report for tool output / interactive display. Bounded, no raw dumps. */\nexport function formatModelFitnessReport(model: string, report: ModelFitnessReport): string {\n\tconst speed = (tokensPerSecond: number | undefined) =>\n\t\ttokensPerSecond !== undefined ? `, ~${tokensPerSecond} tok/s` : \"\";\n\tconst lines = [\n\t\t`Model fitness: ${model} (${report.trials} trials/lane${speed(report.tokensPerSecond)})`,\n\t\t`- research lane: ${report.research.succeeded}/${report.research.total} succeeded, mean ${report.research.meanMs}ms${speed(report.research.tokensPerSecond)} [${report.research.outcomes.join(\", \")}]`,\n\t\t`- worker lane:   ${report.worker.succeeded}/${report.worker.total} completed+accepted, mean ${report.worker.meanMs}ms${speed(report.worker.tokensPerSecond)} [${report.worker.outcomes.join(\", \")}]`,\n\t\t`- search plans:  ${report.search.succeeded}/${report.search.total} well-formed, mean ${report.search.meanMs}ms${speed(report.search.tokensPerSecond)}`,\n\t\t`- tool calls:    ${report.toolCall.succeeded}/${report.toolCall.total} well-formed, mean ${report.toolCall.meanMs}ms${speed(report.toolCall.tokensPerSecond)}`,\n\t\t`- digests:       ${report.digest.succeeded}/${report.digest.total} faithful, mean ${report.digest.meanMs}ms${speed(report.digest.tokensPerSecond)}`,\n\t\t...(report.capacity\n\t\t\t? [\n\t\t\t\t\t`- capacity:      served window ${report.capacity.servedContextWindow}/${report.capacity.registeredContextWindow} tokens, mean ${report.capacity.meanMs}ms`,\n\t\t\t\t]\n\t\t\t: []),\n\t\t`- route judge:   parsed ${report.judge.parsed}/${report.judge.total}, planning-elevated ${report.judge.planningElevated}/${report.judge.planningTotal}, trivial-cheap ${report.judge.trivialCheap}/${report.judge.trivialTotal}, mean ${report.judge.meanMs}ms${speed(report.judge.tokensPerSecond)}`,\n\t\t...report.judge.outcomes.map((outcome) => `    ${outcome}`),\n\t];\n\tif (report.totalCostUsd > 0) {\n\t\tlines.push(`- probe cost: $${report.totalCostUsd.toFixed(4)}`);\n\t}\n\treturn lines.join(\"\\n\");\n}\n"]}
+{"version":3,"file":"model-fitness.d.ts","sourceRoot":"","sources":["../../../src/core/research/model-fitness.ts"],"names":[],"mappings":"AAOA;;;;;;GAMG;AAEH,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,iGAAiG;IACjG,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,kGAAkG;IAClG,MAAM,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,MAAM,eAAe,GAAG,CAAC,IAAI,EAAE;IACpC,YAAY,EAAE,MAAM,CAAC;IACrB,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB,KAAK,OAAO,CAAC,iBAAiB,CAAC,CAAC;AAEjC,MAAM,WAAW,kBAAkB;IAClC,MAAM,EAAE,MAAM,CAAC;IACf,0EAA0E;IAC1E,QAAQ,EAAE,OAAO,CAAC;CAClB;AAED,qFAAqF;AACrF,eAAO,MAAM,6BAA6B,EAAE,SAAS,kBAAkB,EAOtE,CAAC;AAEF,MAAM,WAAW,oBAAoB;IACpC,uBAAuB,EAAE,MAAM,CAAC;IAChC,4FAA4F;IAC5F,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,mBAAmB;IACnC,QAAQ,EAAE,eAAe,CAAC;IAC1B,0CAA0C;IAC1C,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,wDAAwD;IACxD,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,YAAY,CAAC,EAAE,SAAS,kBAAkB,EAAE,CAAC;IAC7C,uFAAuF;IACvF,aAAa,CAAC,EAAE,oBAAoB,CAAC;IACrC,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,gFAAgF;IAChF,GAAG,CAAC,EAAE,MAAM,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,gBAAgB;IAChC,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,yFAAyF;IACzF,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,iBAAiB;IACjC,MAAM,EAAE,MAAM,CAAC;IACf,gBAAgB,EAAE,MAAM,CAAC;IACzB,aAAa,EAAE,MAAM,CAAC;IACtB,YAAY,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,qFAAqF;IACrF,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,oBAAoB;IACpC,uBAAuB,EAAE,MAAM,CAAC;IAChC,mBAAmB,EAAE,MAAM,CAAC;IAC5B,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,kBAAkB;IAClC,MAAM,EAAE,MAAM,CAAC;IACf,yFAAyF;IACzF,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,MAAM,EAAE,gBAAgB,CAAC;IACzB,KAAK,EAAE,iBAAiB,CAAC;IACzB,8EAA8E;IAC9E,MAAM,EAAE,gBAAgB,CAAC;IACzB,yFAAyF;IACzF,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,qGAAqG;IACrG,MAAM,EAAE,gBAAgB,CAAC;IACzB,qFAAqF;IACrF,QAAQ,CAAC,EAAE,oBAAoB,CAAC;IAChC,YAAY,EAAE,MAAM,CAAC;CACrB;AAED,yFAAyF;AACzF,eAAO,MAAM,0BAA0B,QAK3B,CAAC;AAEb,eAAO,MAAM,6BAA6B,QAK9B,CAAC;AAEb,eAAO,MAAM,4BAA4B,QAI7B,CAAC;AASb,OAAO,EAAE,6BAA6B,IAAI,0BAA0B,EAAE,MAAM,6BAA6B,CAAC;AA0K1G,wBAAsB,oBAAoB,CAAC,OAAO,EAAE,mBAAmB,GAAG,OAAO,CAAC,kBAAkB,CAAC,CA+KpG;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,kBAAkB,GAAG,OAAO,CAQpE;AAED,kGAAkG;AAClG,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,kBAAkB,GAAG,MAAM,CAsB1F","sourcesContent":["import { runBoundedCompletion } from \"../autonomy/bounded-completion.ts\";\nimport type { CapabilityEnvelope } from \"../autonomy/contracts.ts\";\nimport { CURATION_DIGEST_SYSTEM_PROMPT } from \"../context/brain-curator.ts\";\nimport { runWorker } from \"../delegation/worker-runner.ts\";\nimport { runRouteJudge } from \"../model-router/route-judge.ts\";\nimport { runResearch } from \"./research-runner.ts\";\n\n/**\n * Model fitness probe: measures whether a candidate model can actually drive the harness's\n * subagent contracts — the research lane, the delegated-worker lane, and the routing judge — by\n * running each real runner against the model and scoring parse/success rates plus judge\n * discrimination. Provider-free: the completion executor is injected, so this works against any\n * registered model (local Ollama models included) and against faux providers in tests.\n */\n\nexport interface FitnessCompletion {\n\ttext: string;\n\tcostUsd: number;\n\tstopReason: string;\n\t/** Output tokens generated (for tok/s). Optional: providers that don't report it are skipped. */\n\toutputTokens?: number;\n\t/** Pure generation time in ms (e.g. Ollama eval_duration). Falls back to wall-clock if absent. */\n\tevalMs?: number;\n}\n\nexport type FitnessComplete = (args: {\n\tsystemPrompt: string;\n\tuserPrompt: string;\n\tsignal?: AbortSignal;\n}) => Promise<FitnessCompletion>;\n\nexport interface JudgeFitnessPrompt {\n\tprompt: string;\n\t/** True when the prompt is planning-shaped and must never route cheap. */\n\tplanning: boolean;\n}\n\n/** Default judge probe set: three planning-shaped prompts, three trivial lookups. */\nexport const DEFAULT_JUDGE_FITNESS_PROMPTS: readonly JudgeFitnessPrompt[] = [\n\t{ prompt: \"how should we plan the migration of the session storage layer?\", planning: true },\n\t{ prompt: \"design an approach for splitting the settings manager\", planning: true },\n\t{ prompt: \"draft a roadmap for the autonomy rework\", planning: true },\n\t{ prompt: \"what does the resolvePath function return?\", planning: false },\n\t{ prompt: \"list the files in the delegation module\", planning: false },\n\t{ prompt: \"why is this test flaky?\", planning: false },\n];\n\nexport interface CapacityProbeOptions {\n\tregisteredContextWindow: number;\n\t/** Smallest candidate window to try before declaring the capacity unknown. Default 1024. */\n\tminContextWindow?: number;\n}\n\nexport interface ModelFitnessOptions {\n\tcomplete: FitnessComplete;\n\t/** Trials per lane surface. Default 3. */\n\ttrials?: number;\n\t/** Wall-clock budget per call in ms. Default 120000. */\n\tmaxWallClockMs?: number;\n\tjudgePrompts?: readonly JudgeFitnessPrompt[];\n\t/** Optional local-model capacity lane: measures the actually served context window. */\n\tcapacityProbe?: CapacityProbeOptions;\n\tsignal?: AbortSignal;\n\t/** Injected clock for latency measurement (test seam). Defaults to Date.now. */\n\tnow?: () => number;\n}\n\nexport interface LaneFitnessScore {\n\tsucceeded: number;\n\ttotal: number;\n\toutcomes: string[];\n\tmeanMs: number;\n\t/** Mean output tokens/second across the surface's calls; undefined when not reported. */\n\ttokensPerSecond?: number;\n}\n\nexport interface JudgeFitnessScore {\n\tparsed: number;\n\tplanningElevated: number;\n\tplanningTotal: number;\n\ttrivialCheap: number;\n\ttrivialTotal: number;\n\ttotal: number;\n\toutcomes: string[];\n\tmeanMs: number;\n\t/** Mean output tokens/second across the judge calls; undefined when not reported. */\n\ttokensPerSecond?: number;\n}\n\nexport interface CapacityFitnessScore {\n\tregisteredContextWindow: number;\n\tservedContextWindow: number;\n\toutcomes: string[];\n\tmeanMs: number;\n}\n\nexport interface ModelFitnessReport {\n\ttrials: number;\n\t/** Aggregate output tokens/second across ALL probe calls (the headline speed number). */\n\ttokensPerSecond?: number;\n\tresearch: LaneFitnessScore;\n\tworker: LaneFitnessScore;\n\tjudge: JudgeFitnessScore;\n\t/** Heavy-lifter surface: can the model formulate a structured search plan? */\n\tsearch: LaneFitnessScore;\n\t/** Heavy-lifter surface: can the model emit a well-formed tool call against a schema? */\n\ttoolCall: LaneFitnessScore;\n\t/** Curator surface: can the model digest a context chunk to strict JSON WITHOUT losing key facts? */\n\tdigest: LaneFitnessScore;\n\t/** Local-model capacity lane: measured context window the server actually serves. */\n\tcapacity?: CapacityFitnessScore;\n\ttotalCostUsd: number;\n}\n\n/** Static prompts for the heavy-lifter surfaces (stable for provider prompt caching). */\nexport const SEARCH_PROBE_SYSTEM_PROMPT = [\n\t\"You plan code searches for a coding agent. You never answer the question yourself.\",\n\t\"Given a question about a codebase, respond with STRICT JSON only - no prose:\",\n\t'{\"queries\":[{\"pattern\":\"<regex or literal to grep>\",\"glob\":\"<file glob like **/*.ts>\"}]}',\n\t\"Return 1 to 4 queries, most specific first.\",\n].join(\"\\n\");\n\nexport const TOOL_CALL_PROBE_SYSTEM_PROMPT = [\n\t\"You operate tools for a coding agent. You have exactly one tool:\",\n\t\"grep(pattern: string, path: string) - search files under a path for a pattern.\",\n\t\"Respond to every task with STRICT JSON only - no prose:\",\n\t'{\"tool\":\"grep\",\"arguments\":{\"pattern\":\"<pattern>\",\"path\":\"<path>\"}}',\n].join(\"\\n\");\n\nexport const CAPACITY_PROBE_SYSTEM_PROMPT = [\n\t\"You are a context-window capacity probe for a local model server.\",\n\t\"Find the unique NEEDLE token in the user's text and echo that token only.\",\n\t\"Do not summarize, explain, or add punctuation.\",\n].join(\"\\n\");\n\nconst SEARCH_PROBE_TASKS: readonly string[] = [\n\t\"Where is the retry/backoff logic for HTTP requests implemented?\",\n\t\"Which files define the settings for background research?\",\n\t\"Find where session entries of type custom are appended.\",\n];\n\n// The probe measures the REAL curation contract — same prompt the BrainCurator ships.\nexport { CURATION_DIGEST_SYSTEM_PROMPT as DIGEST_PROBE_SYSTEM_PROMPT } from \"../context/brain-curator.ts\";\n\n/**\n * Digest probe chunks each carry a NONCE identifier that cannot be guessed from the\n * instructions: acceptance requires the digest to RETAIN the nonce verbatim, so the score\n * measures extraction fidelity, not narration (a model cannot pass by paraphrasing).\n */\nconst DIGEST_PROBE_TASKS: readonly { chunk: string; nonce: string }[] = [\n\t{\n\t\tnonce: \"retryWithJitter_zx41\",\n\t\tchunk: [\n\t\t\t\"grep results for 'retry' under src/http:\",\n\t\t\t\"src/http/client.ts:88: export function retryWithJitter_zx41(fn, attempts = 3) {\",\n\t\t\t\"src/http/client.ts:112:   // exponential backoff capped at 30s\",\n\t\t\t\"src/http/pool.ts:41:   client.retry = false\",\n\t\t].join(\"\\n\"),\n\t},\n\t{\n\t\tnonce: \"ERR_QM_7734\",\n\t\tchunk: [\n\t\t\t\"$ npm run migrate\",\n\t\t\t\"migrating 14 files...\",\n\t\t\t\"error ERR_QM_7734: column 'owner_id' missing on table sessions (migration 0009)\",\n\t\t\t\"exit code 1\",\n\t\t].join(\"\\n\"),\n\t},\n\t{\n\t\tnonce: \"v3.9.2-hotfix.1\",\n\t\tchunk: [\n\t\t\t\"read package.json (34 lines):\",\n\t\t\t'  \"name\": \"acme-billing\",',\n\t\t\t'  \"version\": \"v3.9.2-hotfix.1\",',\n\t\t\t'  \"engines\": { \"node\": \">=22\" },',\n\t\t].join(\"\\n\"),\n\t},\n];\n\nfunction parseDigest(text: string, nonce: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst digest = (parsed as { digest?: unknown }).digest;\n\tif (typeof digest !== \"string\") return false;\n\tconst trimmed = digest.trim();\n\t// Bounded and faithful: short enough to be a stub annotation, still carrying the nonce fact.\n\treturn trimmed.length > 0 && trimmed.length <= 240 && trimmed.includes(nonce);\n}\n\nconst TOOL_CALL_PROBE_TASKS: readonly string[] = [\n\t\"Find usages of the function resolveCliModel under src/.\",\n\t\"Search for the string 'budget_exhausted' in the core directory.\",\n\t\"Locate where LaneTracker is instantiated under src/core.\",\n];\n\nfunction parseSearchPlan(text: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst queries = (parsed as { queries?: unknown }).queries;\n\tif (!Array.isArray(queries) || queries.length === 0 || queries.length > 8) return false;\n\treturn queries.every(\n\t\t(query) =>\n\t\t\tquery &&\n\t\t\ttypeof query === \"object\" &&\n\t\t\ttypeof (query as { pattern?: unknown }).pattern === \"string\" &&\n\t\t\t(query as { pattern: string }).pattern.trim().length > 0,\n\t);\n}\n\nfunction parseToolCall(text: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst record = parsed as { tool?: unknown; arguments?: unknown };\n\tif (record.tool !== \"grep\") return false;\n\tconst args = record.arguments;\n\tif (!args || typeof args !== \"object\" || Array.isArray(args)) return false;\n\tconst pattern = (args as { pattern?: unknown }).pattern;\n\tconst path = (args as { path?: unknown }).path;\n\treturn (\n\t\ttypeof pattern === \"string\" && pattern.trim().length > 0 && typeof path === \"string\" && path.trim().length > 0\n\t);\n}\n\nfunction extractJsonObject(text: string): unknown | undefined {\n\tconst trimmed = text.trim();\n\tconst candidates: string[] = [trimmed];\n\tconst fenced = /```(?:json)?\\s*([\\s\\S]*?)```/.exec(trimmed);\n\tif (fenced?.[1]) candidates.push(fenced[1].trim());\n\tconst start = trimmed.indexOf(\"{\");\n\tconst end = trimmed.lastIndexOf(\"}\");\n\tif (start >= 0 && end > start) candidates.push(trimmed.slice(start, end + 1));\n\tfor (const candidate of candidates) {\n\t\ttry {\n\t\t\tconst parsed = JSON.parse(candidate);\n\t\t\tif (parsed && typeof parsed === \"object\" && !Array.isArray(parsed)) return parsed;\n\t\t} catch {\n\t\t\t// try next candidate\n\t\t}\n\t}\n\treturn undefined;\n}\n\nfunction fitnessEnvelope(): CapabilityEnvelope {\n\treturn {\n\t\tid: \"model-fitness-probe\",\n\t\tcapabilities: [\"research.execute\", \"filesystem.read\", \"memory.query\"],\n\t\tmaxEstimatedUsd: 1,\n\t\tcreatedAt: new Date().toISOString(),\n\t};\n}\n\nfunction buildCapacityProbePrompt(targetTokens: number, startNeedle: string, endNeedle: string): string {\n\tconst targetChars = Math.max(\n\t\tstartNeedle.length + endNeedle.length + 64,\n\t\ttargetTokens * 4 - CAPACITY_PROBE_SYSTEM_PROMPT.length,\n\t);\n\tconst prefix = `CAPACITY PROBE START\\n${startNeedle}\\n`;\n\tconst suffix = `\\n${endNeedle}\\nCAPACITY PROBE END`;\n\tconst fillerLength = Math.max(0, targetChars - prefix.length - suffix.length);\n\treturn `${prefix}${\"x \".repeat(Math.ceil(fillerLength / 2)).slice(0, fillerLength)}${suffix}`;\n}\n\nasync function measureServedContextWindow(args: {\n\tprobe: CapacityProbeOptions;\n\tcomplete: FitnessComplete;\n\tmaxWallClockMs: number;\n\tsignal?: AbortSignal;\n\tnow: () => number;\n}): Promise<{ score: CapacityFitnessScore; costUsd: number }> {\n\tconst registered = Math.max(1, Math.floor(args.probe.registeredContextWindow));\n\tconst minWindow = Math.max(1, Math.floor(args.probe.minContextWindow ?? 1024));\n\tconst nonce = `${registered.toString(16).toUpperCase()}_${minWindow.toString(16).toUpperCase()}`;\n\tconst startNeedle = `NEEDLE_START_${nonce}`;\n\tconst endNeedle = `NEEDLE_END_${nonce}`;\n\tconst score: CapacityFitnessScore = {\n\t\tregisteredContextWindow: registered,\n\t\tservedContextWindow: 0,\n\t\toutcomes: [],\n\t\tmeanMs: 0,\n\t};\n\tlet candidate = registered;\n\tlet costUsd = 0;\n\tlet calls = 0;\n\twhile (candidate >= minWindow) {\n\t\tconst started = args.now();\n\t\tcalls++;\n\t\tconst bounded = await runBoundedCompletion({\n\t\t\tmaxWallClockMs: args.maxWallClockMs,\n\t\t\tsignal: args.signal,\n\t\t\texecute: (signal) =>\n\t\t\t\targs.complete({\n\t\t\t\t\tsystemPrompt: CAPACITY_PROBE_SYSTEM_PROMPT,\n\t\t\t\t\tuserPrompt: buildCapacityProbePrompt(candidate, startNeedle, endNeedle),\n\t\t\t\t\tsignal,\n\t\t\t\t}),\n\t\t});\n\t\tscore.meanMs += args.now() - started;\n\t\tif (bounded.completion) costUsd += bounded.completion.costUsd;\n\t\tconst output = bounded.completion?.text.trim() ?? \"\";\n\t\tconst recalled = !bounded.failure && output.includes(startNeedle) && output.includes(endNeedle);\n\t\tscore.outcomes.push(`${candidate}:${recalled ? \"ok\" : (bounded.failure?.status ?? \"miss\")}`);\n\t\tif (recalled) {\n\t\t\tscore.servedContextWindow = candidate;\n\t\t\tbreak;\n\t\t}\n\t\tcandidate = Math.floor(candidate / 2);\n\t}\n\tif (score.servedContextWindow === 0) score.servedContextWindow = minWindow;\n\tscore.meanMs = calls > 0 ? Math.round(score.meanMs / calls) : 0;\n\treturn { score, costUsd };\n}\n\nexport async function runModelFitnessProbe(options: ModelFitnessOptions): Promise<ModelFitnessReport> {\n\tconst trials = Math.max(1, Math.min(options.trials ?? 3, 20));\n\tconst maxWallClockMs = options.maxWallClockMs ?? 120_000;\n\tconst judgePrompts = options.judgePrompts ?? DEFAULT_JUDGE_FITNESS_PROMPTS;\n\tconst now = options.now ?? Date.now;\n\tlet totalCostUsd = 0;\n\n\t// Token-speed instrumentation: the lane runners' own contracts carry text/cost only, so the\n\t// completer is wrapped once here and generation stats are accumulated per surface.\n\tconst overallSpeed = { tokens: 0, evalMs: 0 };\n\tlet surfaceSpeed = { tokens: 0, evalMs: 0 };\n\tconst complete: FitnessComplete = async (args) => {\n\t\tconst completion = await options.complete(args);\n\t\tconst tokens = completion.outputTokens ?? 0;\n\t\tconst evalMs = completion.evalMs ?? 0;\n\t\tif (tokens > 0 && evalMs > 0) {\n\t\t\tsurfaceSpeed.tokens += tokens;\n\t\t\tsurfaceSpeed.evalMs += evalMs;\n\t\t\toverallSpeed.tokens += tokens;\n\t\t\toverallSpeed.evalMs += evalMs;\n\t\t}\n\t\treturn completion;\n\t};\n\tconst takeSurfaceSpeed = (): number | undefined => {\n\t\tconst speed =\n\t\t\tsurfaceSpeed.evalMs > 0 ? Math.round((surfaceSpeed.tokens / surfaceSpeed.evalMs) * 1000) : undefined;\n\t\tsurfaceSpeed = { tokens: 0, evalMs: 0 };\n\t\treturn speed;\n\t};\n\n\tconst research: LaneFitnessScore = { succeeded: 0, total: trials, outcomes: [], meanMs: 0 };\n\tfor (let i = 0; i < trials; i++) {\n\t\tconst started = now();\n\t\tconst result = await runResearch({\n\t\t\tquery: `fitness:probe requirements:req-${i}`,\n\t\t\tcontext: [\n\t\t\t\t\"Goal: add a retry helper to the HTTP client module\",\n\t\t\t\t\"Open requirements:\",\n\t\t\t\t\"- Find what retry/backoff conventions the codebase already uses\",\n\t\t\t\t\"- Identify which call sites would adopt the helper\",\n\t\t\t].join(\"\\n\"),\n\t\t\tenvelope: fitnessEnvelope(),\n\t\t\tmaxUsd: 1,\n\t\t\tmaxSources: 8,\n\t\t\tmaxFindings: 5,\n\t\t\tmaxWallClockMs,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tresearch.meanMs += now() - started;\n\t\ttotalCostUsd += result.costUsd;\n\t\tif (result.status === \"succeeded\") research.succeeded++;\n\t\tresearch.outcomes.push(`${result.status}/${result.reasonCode}`);\n\t}\n\tresearch.meanMs = Math.round(research.meanMs / trials);\n\tresearch.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst worker: LaneFitnessScore = { succeeded: 0, total: trials, outcomes: [], meanMs: 0 };\n\tfor (let i = 0; i < trials; i++) {\n\t\tconst started = now();\n\t\tconst outcome = await runWorker({\n\t\t\trequest: {\n\t\t\t\tid: `fitness-worker-${i}`,\n\t\t\t\tinstructions:\n\t\t\t\t\t\"Summarize in two sentences what a capability envelope is: a declared set of allowed tools, paths, and capability names that bounds what a delegated worker may do.\",\n\t\t\t\troute: { tier: \"cheap\", risk: \"read-only\", confidence: 1, reasonCode: \"fitness_probe\", reasons: [] },\n\t\t\t\tenvelope: { id: `fitness-env-${i}`, capabilities: [\"filesystem.read\"], maxEstimatedUsd: 1 },\n\t\t\t\tmaxEstimatedUsd: 1,\n\t\t\t},\n\t\t\tmaxUsd: 1,\n\t\t\tmaxWallClockMs,\n\t\t\tusageReportId: `fitness:${i}`,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tworker.meanMs += now() - started;\n\t\ttotalCostUsd += outcome.costUsd;\n\t\tif (outcome.claim.status === \"completed\" && outcome.accepted) worker.succeeded++;\n\t\tworker.outcomes.push(`${outcome.claim.status}/${outcome.reasonCode}`);\n\t}\n\tworker.meanMs = Math.round(worker.meanMs / trials);\n\tworker.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst judge: JudgeFitnessScore = {\n\t\tparsed: 0,\n\t\tplanningElevated: 0,\n\t\tplanningTotal: judgePrompts.filter((entry) => entry.planning).length,\n\t\ttrivialCheap: 0,\n\t\ttrivialTotal: judgePrompts.filter((entry) => !entry.planning).length,\n\t\ttotal: judgePrompts.length,\n\t\toutcomes: [],\n\t\tmeanMs: 0,\n\t};\n\tfor (const entry of judgePrompts) {\n\t\tconst started = now();\n\t\tconst result = await runRouteJudge({\n\t\t\tprompt: entry.prompt,\n\t\t\tbaseline: { tier: \"cheap\", risk: \"read-only\", confidence: 0.5, reasonCode: \"fitness_probe\", reasons: [] },\n\t\t\tmaxWallClockMs,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tjudge.meanMs += now() - started;\n\t\ttotalCostUsd += result.costUsd;\n\t\tconst tier = result.decision.tier;\n\t\tif (result.verdict) {\n\t\t\tjudge.parsed++;\n\t\t\t// A useful judge must both keep planning off the cheap tier AND actually send trivial\n\t\t\t// prompts there — all-medium verdicts are safe but save nothing.\n\t\t\tif (entry.planning && tier !== \"cheap\") judge.planningElevated++;\n\t\t\tif (!entry.planning && tier === \"cheap\") judge.trivialCheap++;\n\t\t}\n\t\tjudge.outcomes.push(\n\t\t\t`\"${entry.prompt.slice(0, 40)}\" -> ${tier}${result.fallbackReason ? ` (${result.fallbackReason})` : \"\"}`,\n\t\t);\n\t}\n\tjudge.meanMs = judgePrompts.length > 0 ? Math.round(judge.meanMs / judgePrompts.length) : 0;\n\tjudge.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst probeSurface = async (\n\t\tsystemPrompt: string,\n\t\ttasks: readonly string[],\n\t\taccepts: (text: string, taskIndex: number) => boolean,\n\t): Promise<LaneFitnessScore> => {\n\t\tconst score: LaneFitnessScore = { succeeded: 0, total: tasks.length, outcomes: [], meanMs: 0 };\n\t\tfor (const [taskIndex, task] of tasks.entries()) {\n\t\t\tconst started = now();\n\t\t\t// Same wall-clock envelope as the lane surfaces — a hung model must not hang the probe.\n\t\t\tconst bounded = await runBoundedCompletion({\n\t\t\t\tmaxWallClockMs,\n\t\t\t\tsignal: options.signal,\n\t\t\t\texecute: (signal) => complete({ systemPrompt, userPrompt: task, signal }),\n\t\t\t});\n\t\t\tif (bounded.completion) totalCostUsd += bounded.completion.costUsd;\n\t\t\tif (bounded.failure || !bounded.completion) {\n\t\t\t\tscore.outcomes.push(bounded.failure ? bounded.failure.status : \"completion_error\");\n\t\t\t} else {\n\t\t\t\tconst ok = accepts(bounded.completion.text, taskIndex);\n\t\t\t\tif (ok) score.succeeded++;\n\t\t\t\tscore.outcomes.push(ok ? \"ok\" : \"unparseable_output\");\n\t\t\t}\n\t\t\tscore.meanMs += now() - started;\n\t\t}\n\t\tscore.meanMs = tasks.length > 0 ? Math.round(score.meanMs / tasks.length) : 0;\n\t\treturn score;\n\t};\n\n\tconst search = await probeSurface(SEARCH_PROBE_SYSTEM_PROMPT, SEARCH_PROBE_TASKS, parseSearchPlan);\n\tsearch.tokensPerSecond = takeSurfaceSpeed();\n\tconst toolCall = await probeSurface(TOOL_CALL_PROBE_SYSTEM_PROMPT, TOOL_CALL_PROBE_TASKS, parseToolCall);\n\ttoolCall.tokensPerSecond = takeSurfaceSpeed();\n\tconst digest = await probeSurface(\n\t\tCURATION_DIGEST_SYSTEM_PROMPT,\n\t\tDIGEST_PROBE_TASKS.map((task) => task.chunk),\n\t\t(text, taskIndex) => parseDigest(text, DIGEST_PROBE_TASKS[taskIndex]!.nonce),\n\t);\n\tdigest.tokensPerSecond = takeSurfaceSpeed();\n\n\tlet capacity: CapacityFitnessScore | undefined;\n\tif (options.capacityProbe) {\n\t\tconst result = await measureServedContextWindow({\n\t\t\tprobe: options.capacityProbe,\n\t\t\tcomplete,\n\t\t\tmaxWallClockMs,\n\t\t\tsignal: options.signal,\n\t\t\tnow,\n\t\t});\n\t\tcapacity = result.score;\n\t\ttotalCostUsd += result.costUsd;\n\t}\n\n\tconst tokensPerSecond =\n\t\toverallSpeed.evalMs > 0 ? Math.round((overallSpeed.tokens / overallSpeed.evalMs) * 1000) : undefined;\n\n\treturn { trials, tokensPerSecond, research, worker, judge, search, toolCall, digest, capacity, totalCostUsd };\n}\n\n/**\n * Pure verdict: true when the probe found ZERO successes on every LANE surface it actually graded\n * AND the judge (if it ran) also failed. A lane/judge with total 0 (i.e. never run) carries no\n * evidence and is excluded from the lane check — but at least one lane must actually have been\n * graded for an all-failed verdict at all: `gradedLanes.every(...)` is vacuously true over an\n * empty array, so a report where only the judge ran (every research/worker/search/toolCall/digest\n * lane is ungraded) is excluded explicitly rather than misread as \"all lanes failed\" on zero lane\n * evidence. An empty/degenerate report (nothing graded at all, lanes AND judge) is likewise never\n * mistaken for a failed one. This is the gate adoption flows must consult before assigning a role —\n * see `isProbeAllFailed` callers in interactive-mode.ts and agent-session.ts.\n */\nexport function isProbeAllFailed(report: ModelFitnessReport): boolean {\n\tconst lanes = [report.research, report.worker, report.search, report.toolCall, report.digest];\n\tconst gradedLanes = lanes.filter((lane) => lane.total > 0);\n\tconst judgeGraded = report.judge.total > 0;\n\tif (gradedLanes.length === 0 && !judgeGraded) return false;\n\tconst lanesAllFailed = gradedLanes.length > 0 && gradedLanes.every((lane) => lane.succeeded === 0);\n\tconst judgeFailed = !judgeGraded || report.judge.parsed === 0;\n\treturn lanesAllFailed && judgeFailed;\n}\n\n/** Compact human-readable report for tool output / interactive display. Bounded, no raw dumps. */\nexport function formatModelFitnessReport(model: string, report: ModelFitnessReport): string {\n\tconst speed = (tokensPerSecond: number | undefined) =>\n\t\ttokensPerSecond !== undefined ? `, ~${tokensPerSecond} tok/s` : \"\";\n\tconst lines = [\n\t\t`Model fitness: ${model} (${report.trials} trials/lane${speed(report.tokensPerSecond)})`,\n\t\t`- research lane: ${report.research.succeeded}/${report.research.total} succeeded, mean ${report.research.meanMs}ms${speed(report.research.tokensPerSecond)} [${report.research.outcomes.join(\", \")}]`,\n\t\t`- worker lane:   ${report.worker.succeeded}/${report.worker.total} completed+accepted, mean ${report.worker.meanMs}ms${speed(report.worker.tokensPerSecond)} [${report.worker.outcomes.join(\", \")}]`,\n\t\t`- search plans:  ${report.search.succeeded}/${report.search.total} well-formed, mean ${report.search.meanMs}ms${speed(report.search.tokensPerSecond)}`,\n\t\t`- tool calls:    ${report.toolCall.succeeded}/${report.toolCall.total} well-formed, mean ${report.toolCall.meanMs}ms${speed(report.toolCall.tokensPerSecond)}`,\n\t\t`- digests:       ${report.digest.succeeded}/${report.digest.total} faithful, mean ${report.digest.meanMs}ms${speed(report.digest.tokensPerSecond)}`,\n\t\t...(report.capacity\n\t\t\t? [\n\t\t\t\t\t`- capacity:      served window ${report.capacity.servedContextWindow}/${report.capacity.registeredContextWindow} tokens, mean ${report.capacity.meanMs}ms`,\n\t\t\t\t]\n\t\t\t: []),\n\t\t`- route judge:   parsed ${report.judge.parsed}/${report.judge.total}, planning-elevated ${report.judge.planningElevated}/${report.judge.planningTotal}, trivial-cheap ${report.judge.trivialCheap}/${report.judge.trivialTotal}, mean ${report.judge.meanMs}ms${speed(report.judge.tokensPerSecond)}`,\n\t\t...report.judge.outcomes.map((outcome) => `    ${outcome}`),\n\t];\n\tif (report.totalCostUsd > 0) {\n\t\tlines.push(`- probe cost: $${report.totalCostUsd.toFixed(4)}`);\n\t}\n\treturn lines.join(\"\\n\");\n}\n"]}

package/dist/core/research/model-fitness.js

@@ -138,7 +138,7 @@ function extractJsonObject(text) {
 function fitnessEnvelope() {
     return {
         id: "model-fitness-probe",
-        capabilities: ["research", "read_files", "memory_read"],
+        capabilities: ["research.execute", "filesystem.read", "memory.query"],
         maxEstimatedUsd: 1,
         createdAt: new Date().toISOString(),
     };
@@ -256,7 +256,7 @@ export async function runModelFitnessProbe(options) {
                 id: `fitness-worker-${i}`,
                 instructions: "Summarize in two sentences what a capability envelope is: a declared set of allowed tools, paths, and capability names that bounds what a delegated worker may do.",
                 route: { tier: "cheap", risk: "read-only", confidence: 1, reasonCode: "fitness_probe", reasons: [] },
-                envelope: { id: `fitness-env-${i}`, capabilities: ["read_files"], maxEstimatedUsd: 1 },
+                envelope: { id: `fitness-env-${i}`, capabilities: ["filesystem.read"], maxEstimatedUsd: 1 },
                 maxEstimatedUsd: 1,
             },
             maxUsd: 1,
@@ -267,9 +267,9 @@ export async function runModelFitnessProbe(options) {
         });
         worker.meanMs += now() - started;
         totalCostUsd += outcome.costUsd;
-        if (outcome.result.status === "completed" && outcome.accepted)
+        if (outcome.claim.status === "completed" && outcome.accepted)
             worker.succeeded++;
-        worker.outcomes.push(`${outcome.result.status}/${outcome.reasonCode}`);
+        worker.outcomes.push(`${outcome.claim.status}/${outcome.reasonCode}`);
     }
     worker.meanMs = Math.round(worker.meanMs / trials);
     worker.tokensPerSecond = takeSurfaceSpeed();