@relay-harness/coding-agent 0.0.0-stage → 1.0.3

package/dist/extensions/laya/memory.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"memory.d.ts","sourceRoot":"","sources":["../../../src/extensions/laya/memory.ts"],"names":[],"mappings":"AAAA,OAAO,EAAkB,KAAK,WAAW,EAAE,MAAM,eAAe,CAAC;AAEjE;;;;;;;;GAQG;AAEH,yEAAyE;AACzE,eAAO,MAAM,gBAAgB,MAAM,CAAC;AACpC,sEAAsE;AACtE,eAAO,MAAM,iBAAiB,OAAO,CAAC;AAEtC,MAAM,WAAW,WAAW;IAC3B,EAAE,EAAE,MAAM,CAAC;IACX,OAAO,EAAE,MAAM,CAAC;IAChB,QAAQ,EAAE,WAAW,CAAC,UAAU,CAAC,CAAC;IAClC,MAAM,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,WAAW;IAC3B,IAAI,EAAE,WAAW,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;CACnB;AAWD,iGAAiG;AACjG,wBAAgB,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,CAQ5C;AAID,qBAAa,UAAU;IACtB,QAAQ,CAAC,KAAK,EAAE,SAAS,WAAW,EAAE,CAAC;IACvC,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAsB;IAC1C,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAS;IACpC,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAW;IAEnC,0FAA0F;IAC1F,YAAY,KAAK,EAAE,SAAS,WAAW,EAAE,EAAE,MAAM,EAAE,SAAS,MAAM,EAAE,EAUnE;IAED,+EAA+E;IAC/E,MAAM,CAAC,QAAQ,CAAC,IAAI,EAAE,SAAS,WAAW,EAAE,GAAG,UAAU,CAQxD;IAED,IAAI,IAAI,IAAI,MAAM,CAEjB;IAED,OAAO,CAAC,MAAM;IAed,kFAAkF;IAClF,MAAM,CAAC,OAAO,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,KAAK,SAAI,GAAG,WAAW,EAAE,CAUvE;CACD;AAED,kGAAkG;AAClG,eAAO,MAAM,oBAAoB,iBAAiB,CAAC;AAEnD;;;GAGG;AACH,wBAAgB,aAAa,CAAC,OAAO,EAAE,SAAS,WAAW,EAAE,GAAG,MAAM,GAAG,SAAS,CAYjF"}

package/dist/extensions/laya/memory.js

@@ -0,0 +1,107 @@
+import { SESSION_SOURCE } from "./training.js";
+/**
+ * Memory of the tasks taught with `/laya learn`. Training takes minutes on a GPU and needs the Laya
+ * environment; the memory works from the next request. A request close to a learned task is routed
+ * with that task's labels, and the model that runs it gets the lessons of similar tasks.
+ *
+ * Similarity is the cosine of TF-IDF word vectors. Word weights come from every exercise of the
+ * workspace, so words common to all coding requests ("fix", "add", "file") count little and the
+ * specific ones ("pager", "off-by-one", "token expiry") decide.
+ */
+/** A learned task close enough for its labels to route a new request. */
+export const ROUTE_SIMILARITY = 0.6;
+/** A learned task close enough for its lesson to be worth reading. */
+export const LESSON_SIMILARITY = 0.45;
+const STOPWORDS = new Set(("the a an and or of to in on at for with from by is are be it this that these those i you we me my our your " +
+    "please can could would should will do does did not no so as if then than into about all any some " +
+    "o a os as um uma uns umas de do da dos das no na nos nas em para por com sem e ou que se nao isso este esta " +
+    "esse essa ele ela eu voce meu minha seu sua pelo pela ao aos mais mas como ja tambem favor").split(" "));
+/** Words of a request: identifiers split at case changes, accents removed, stopwords dropped. */
+export function words(text) {
+    return text
+        .replace(/([a-z])([A-Z])/g, "$1 $2")
+        .normalize("NFD")
+        .replace(/\p{Diacritic}/gu, "")
+        .toLowerCase()
+        .split(/[^a-z0-9]+/)
+        .filter((word) => word.length > 1 && !STOPWORDS.has(word));
+}
+export class TaskMemory {
+    tasks;
+    idf;
+    defaultIdf;
+    vectors;
+    /** `corpus` is every request the word weights are counted over; it includes the tasks. */
+    constructor(tasks, corpus) {
+        this.tasks = tasks;
+        const documents = corpus.length;
+        const frequency = new Map();
+        for (const request of corpus) {
+            for (const word of new Set(words(request)))
+                frequency.set(word, (frequency.get(word) ?? 0) + 1);
+        }
+        this.idf = new Map([...frequency].map(([word, count]) => [word, Math.log((documents + 1) / (count + 1)) + 1]));
+        this.defaultIdf = Math.log(documents + 1) + 1;
+        this.vectors = tasks.map((task) => this.vector(task.request));
+    }
+    /** Session tasks of a training dataset, weighted over all of its exercises. */
+    static fromRows(rows) {
+        const tasks = rows
+            .filter((row) => row.source === SESSION_SOURCE && row.state?.request)
+            .map((row) => ({ id: row.id, request: row.state.request, expected: row.expected, lesson: row.lesson }));
+        return new TaskMemory(tasks, rows.flatMap((row) => (row.state?.request ? [row.state.request] : [])));
+    }
+    get size() {
+        return this.tasks.length;
+    }
+    vector(text) {
+        const counts = new Map();
+        for (const word of words(text))
+            counts.set(word, (counts.get(word) ?? 0) + 1);
+        const vector = new Map();
+        let norm = 0;
+        for (const [word, count] of counts) {
+            const weight = (1 + Math.log(count)) * (this.idf.get(word) ?? this.defaultIdf);
+            vector.set(word, weight);
+            norm += weight * weight;
+        }
+        norm = Math.sqrt(norm);
+        for (const [word, weight] of vector)
+            vector.set(word, weight / norm);
+        return vector;
+    }
+    /** Learned tasks at least `minSimilarity` close to the request, closest first. */
+    search(request, minSimilarity, limit = 3) {
+        if (this.tasks.length === 0)
+            return [];
+        const query = this.vector(request);
+        const matches = [];
+        this.vectors.forEach((vector, index) => {
+            let similarity = 0;
+            for (const [word, weight] of query)
+                similarity += weight * (vector.get(word) ?? 0);
+            if (similarity >= minSimilarity)
+                matches.push({ task: this.tasks[index], similarity });
+        });
+        return matches.sort((a, b) => b.similarity - a.similarity).slice(0, limit);
+    }
+}
+/** Custom message with the lessons, sent next to the request when laya/auto does not route it. */
+export const LAYA_LESSONS_MESSAGE = "laya.lessons";
+/**
+ * Lessons of similar tasks for the model that runs the request, or undefined when none has one.
+ * The plan message includes the text as one item; other models get it with the `[laya:lessons]` prefix.
+ */
+export function renderLessons(matches) {
+    const clip = (text) => {
+        const flat = text.replace(/\s+/g, " ").trim();
+        return flat.length > 100 ? `${flat.slice(0, 100)}…` : flat;
+    };
+    const lines = matches.flatMap((match) => match.task.lesson
+        ? [`  - "${clip(match.task.request)}" (${Math.round(match.similarity * 100)}% similar): ${match.task.lesson}`]
+        : []);
+    if (lines.length === 0)
+        return undefined;
+    return ["Lessons from similar tasks done before in this harness (check they still apply):", ...lines].join("\n");
+}
+//# sourceMappingURL=memory.js.map

package/dist/extensions/laya/memory.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"memory.js","sourceRoot":"","sources":["../../../src/extensions/laya/memory.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,cAAc,EAAoB,MAAM,eAAe,CAAC;AAEjE;;;;;;;;GAQG;AAEH,yEAAyE;AACzE,MAAM,CAAC,MAAM,gBAAgB,GAAG,GAAG,CAAC;AACpC,sEAAsE;AACtE,MAAM,CAAC,MAAM,iBAAiB,GAAG,IAAI,CAAC;AActC,MAAM,SAAS,GAAG,IAAI,GAAG,CACxB,CACC,6GAA6G;IAC7G,mGAAmG;IACnG,8GAA8G;IAC9G,4FAA4F,CAC5F,CAAC,KAAK,CAAC,GAAG,CAAC,CACZ,CAAC;AAEF,iGAAiG;AACjG,MAAM,UAAU,KAAK,CAAC,IAAY;IACjC,OAAO,IAAI;SACT,OAAO,CAAC,iBAAiB,EAAE,OAAO,CAAC;SACnC,SAAS,CAAC,KAAK,CAAC;SAChB,OAAO,CAAC,iBAAiB,EAAE,EAAE,CAAC;SAC9B,WAAW,EAAE;SACb,KAAK,CAAC,YAAY,CAAC;SACnB,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC;AAC7D,CAAC;AAID,MAAM,OAAO,UAAU;IACb,KAAK,CAAyB;IACtB,GAAG,CAAsB;IACzB,UAAU,CAAS;IACnB,OAAO,CAAW;IAEnC,0FAA0F;IAC1F,YAAY,KAA6B,EAAE,MAAyB;QACnE,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,MAAM,SAAS,GAAG,MAAM,CAAC,MAAM,CAAC;QAChC,MAAM,SAAS,GAAG,IAAI,GAAG,EAAkB,CAAC;QAC5C,KAAK,MAAM,OAAO,IAAI,MAAM,EAAE,CAAC;YAC9B,KAAK,MAAM,IAAI,IAAI,IAAI,GAAG,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;gBAAE,SAAS,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QACjG,CAAC;QACD,IAAI,CAAC,GAAG,GAAG,IAAI,GAAG,CAAC,CAAC,GAAG,SAAS,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,KAAK,CAAC,EAAE,EAAE,CAAC,CAAC,IAAI,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,SAAS,GAAG,CAAC,CAAC,GAAG,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;QAC/G,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,SAAS,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC;QAC9C,IAAI,CAAC,OAAO,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC;IAC/D,CAAC;IAED,+EAA+E;IAC/E,MAAM,CAAC,QAAQ,CAAC,IAA4B;QAC3C,MAAM,KAAK,GAAG,IAAI;aAChB,MAAM,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,MAAM,KAAK,cAAc,IAAI,GAAG,CAAC,KAAK,EAAE,OAAO,CAAC;aACpE,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,EAAE,GAAG,CAAC,EAAE,EAAE,OAAO,EAAE,GAAG,CAAC,KAAK,CAAC,OAAO,EAAE,QAAQ,EAAE,GAAG,CAAC,QAAQ,EAAE,MAAM,EAAE,GAAG,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC;QACzG,OAAO,IAAI,UAAU,CACpB,KAAK,EACL,IAAI,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CACtE,CAAC;IACH,CAAC;IAED,IAAI,IAAI;QACP,OAAO,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC;IAC1B,CAAC;IAEO,MAAM,CAAC,IAAY;QAC1B,MAAM,MAAM,GAAG,IAAI,GAAG,EAAkB,CAAC;QACzC,KAAK,MAAM,IAAI,IAAI,KAAK,CAAC,IAAI,CAAC;YAAE,MAAM,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC,MAAM,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QAC9E,MAAM,MAAM,GAAW,IAAI,GAAG,EAAE,CAAC;QACjC,IAAI,IAAI,GAAG,CAAC,CAAC;QACb,KAAK,MAAM,CAAC,IAAI,EAAE,KAAK,CAAC,IAAI,MAAM,EAAE,CAAC;YACpC,MAAM,MAAM,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,IAAI,CAAC,UAAU,CAAC,CAAC;YAC/E,MAAM,CAAC,GAAG,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;YACzB,IAAI,IAAI,MAAM,GAAG,MAAM,CAAC;QACzB,CAAC;QACD,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACvB,KAAK,MAAM,CAAC,IAAI,EAAE,MAAM,CAAC,IAAI,MAAM;YAAE,MAAM,CAAC,GAAG,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC,CAAC;QACrE,OAAO,MAAM,CAAC;IACf,CAAC;IAED,kFAAkF;IAClF,MAAM,CAAC,OAAe,EAAE,aAAqB,EAAE,KAAK,GAAG,CAAC;QACvD,IAAI,IAAI,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC;YAAE,OAAO,EAAE,CAAC;QACvC,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;QACnC,MAAM,OAAO,GAAkB,EAAE,CAAC;QAClC,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,MAAM,EAAE,KAAK,EAAE,EAAE;YACtC,IAAI,UAAU,GAAG,CAAC,CAAC;YACnB,KAAK,MAAM,CAAC,IAAI,EAAE,MAAM,CAAC,IAAI,KAAK;gBAAE,UAAU,IAAI,MAAM,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC;YACnF,IAAI,UAAU,IAAI,aAAa;gBAAE,OAAO,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,UAAU,EAAE,CAAC,CAAC;QACxF,CAAC,CAAC,CAAC;QACH,OAAO,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,UAAU,GAAG,CAAC,CAAC,UAAU,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC;IAC5E,CAAC;CACD;AAED,kGAAkG;AAClG,MAAM,CAAC,MAAM,oBAAoB,GAAG,cAAc,CAAC;AAEnD;;;GAGG;AACH,MAAM,UAAU,aAAa,CAAC,OAA+B;IAC5D,MAAM,IAAI,GAAG,CAAC,IAAY,EAAE,EAAE;QAC7B,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC;QAC9C,OAAO,IAAI,CAAC,MAAM,GAAG,GAAG,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC;IAC5D,CAAC,CAAC;IACF,MAAM,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC,KAAK,EAAE,EAAE,CACvC,KAAK,CAAC,IAAI,CAAC,MAAM;QAChB,CAAC,CAAC,CAAC,QAAQ,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,UAAU,GAAG,GAAG,CAAC,eAAe,KAAK,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC;QAC9G,CAAC,CAAC,EAAE,CACL,CAAC;IACF,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,SAAS,CAAC;IACzC,OAAO,CAAC,kFAAkF,EAAE,GAAG,KAAK,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAClH,CAAC","sourcesContent":["import { SESSION_SOURCE, type TrainingRow } from \"./training.ts\";\n\n/**\n * Memory of the tasks taught with `/laya learn`. Training takes minutes on a GPU and needs the Laya\n * environment; the memory works from the next request. A request close to a learned task is routed\n * with that task's labels, and the model that runs it gets the lessons of similar tasks.\n *\n * Similarity is the cosine of TF-IDF word vectors. Word weights come from every exercise of the\n * workspace, so words common to all coding requests (\"fix\", \"add\", \"file\") count little and the\n * specific ones (\"pager\", \"off-by-one\", \"token expiry\") decide.\n */\n\n/** A learned task close enough for its labels to route a new request. */\nexport const ROUTE_SIMILARITY = 0.6;\n/** A learned task close enough for its lesson to be worth reading. */\nexport const LESSON_SIMILARITY = 0.45;\n\nexport interface LearnedTask {\n\tid: string;\n\trequest: string;\n\texpected: TrainingRow[\"expected\"];\n\tlesson?: string;\n}\n\nexport interface MemoryMatch {\n\ttask: LearnedTask;\n\tsimilarity: number;\n}\n\nconst STOPWORDS = new Set(\n\t(\n\t\t\"the a an and or of to in on at for with from by is are be it this that these those i you we me my our your \" +\n\t\t\"please can could would should will do does did not no so as if then than into about all any some \" +\n\t\t\"o a os as um uma uns umas de do da dos das no na nos nas em para por com sem e ou que se nao isso este esta \" +\n\t\t\"esse essa ele ela eu voce meu minha seu sua pelo pela ao aos mais mas como ja tambem favor\"\n\t).split(\" \"),\n);\n\n/** Words of a request: identifiers split at case changes, accents removed, stopwords dropped. */\nexport function words(text: string): string[] {\n\treturn text\n\t\t.replace(/([a-z])([A-Z])/g, \"$1 $2\")\n\t\t.normalize(\"NFD\")\n\t\t.replace(/\\p{Diacritic}/gu, \"\")\n\t\t.toLowerCase()\n\t\t.split(/[^a-z0-9]+/)\n\t\t.filter((word) => word.length > 1 && !STOPWORDS.has(word));\n}\n\ntype Vector = Map<string, number>;\n\nexport class TaskMemory {\n\treadonly tasks: readonly LearnedTask[];\n\tprivate readonly idf: Map<string, number>;\n\tprivate readonly defaultIdf: number;\n\tprivate readonly vectors: Vector[];\n\n\t/** `corpus` is every request the word weights are counted over; it includes the tasks. */\n\tconstructor(tasks: readonly LearnedTask[], corpus: readonly string[]) {\n\t\tthis.tasks = tasks;\n\t\tconst documents = corpus.length;\n\t\tconst frequency = new Map<string, number>();\n\t\tfor (const request of corpus) {\n\t\t\tfor (const word of new Set(words(request))) frequency.set(word, (frequency.get(word) ?? 0) + 1);\n\t\t}\n\t\tthis.idf = new Map([...frequency].map(([word, count]) => [word, Math.log((documents + 1) / (count + 1)) + 1]));\n\t\tthis.defaultIdf = Math.log(documents + 1) + 1;\n\t\tthis.vectors = tasks.map((task) => this.vector(task.request));\n\t}\n\n\t/** Session tasks of a training dataset, weighted over all of its exercises. */\n\tstatic fromRows(rows: readonly TrainingRow[]): TaskMemory {\n\t\tconst tasks = rows\n\t\t\t.filter((row) => row.source === SESSION_SOURCE && row.state?.request)\n\t\t\t.map((row) => ({ id: row.id, request: row.state.request, expected: row.expected, lesson: row.lesson }));\n\t\treturn new TaskMemory(\n\t\t\ttasks,\n\t\t\trows.flatMap((row) => (row.state?.request ? [row.state.request] : [])),\n\t\t);\n\t}\n\n\tget size(): number {\n\t\treturn this.tasks.length;\n\t}\n\n\tprivate vector(text: string): Vector {\n\t\tconst counts = new Map<string, number>();\n\t\tfor (const word of words(text)) counts.set(word, (counts.get(word) ?? 0) + 1);\n\t\tconst vector: Vector = new Map();\n\t\tlet norm = 0;\n\t\tfor (const [word, count] of counts) {\n\t\t\tconst weight = (1 + Math.log(count)) * (this.idf.get(word) ?? this.defaultIdf);\n\t\t\tvector.set(word, weight);\n\t\t\tnorm += weight * weight;\n\t\t}\n\t\tnorm = Math.sqrt(norm);\n\t\tfor (const [word, weight] of vector) vector.set(word, weight / norm);\n\t\treturn vector;\n\t}\n\n\t/** Learned tasks at least `minSimilarity` close to the request, closest first. */\n\tsearch(request: string, minSimilarity: number, limit = 3): MemoryMatch[] {\n\t\tif (this.tasks.length === 0) return [];\n\t\tconst query = this.vector(request);\n\t\tconst matches: MemoryMatch[] = [];\n\t\tthis.vectors.forEach((vector, index) => {\n\t\t\tlet similarity = 0;\n\t\t\tfor (const [word, weight] of query) similarity += weight * (vector.get(word) ?? 0);\n\t\t\tif (similarity >= minSimilarity) matches.push({ task: this.tasks[index], similarity });\n\t\t});\n\t\treturn matches.sort((a, b) => b.similarity - a.similarity).slice(0, limit);\n\t}\n}\n\n/** Custom message with the lessons, sent next to the request when laya/auto does not route it. */\nexport const LAYA_LESSONS_MESSAGE = \"laya.lessons\";\n\n/**\n * Lessons of similar tasks for the model that runs the request, or undefined when none has one.\n * The plan message includes the text as one item; other models get it with the `[laya:lessons]` prefix.\n */\nexport function renderLessons(matches: readonly MemoryMatch[]): string | undefined {\n\tconst clip = (text: string) => {\n\t\tconst flat = text.replace(/\\s+/g, \" \").trim();\n\t\treturn flat.length > 100 ? `${flat.slice(0, 100)}…` : flat;\n\t};\n\tconst lines = matches.flatMap((match) =>\n\t\tmatch.task.lesson\n\t\t\t? [`  - \"${clip(match.task.request)}\" (${Math.round(match.similarity * 100)}% similar): ${match.task.lesson}`]\n\t\t\t: [],\n\t);\n\tif (lines.length === 0) return undefined;\n\treturn [\"Lessons from similar tasks done before in this harness (check they still apply):\", ...lines].join(\"\\n\");\n}\n"]}

package/dist/extensions/laya/model-manifest.d.ts

@@ -0,0 +1,7 @@
+import type { LayaModelManifest } from "./runtime.ts";
+/**
+ * The trained Laya model baked into the Laya Docker image. Generated by `scripts/package-laya-model.mjs`;
+ * do not edit by hand. The files live at `https://huggingface.co/eltonssouza/relay-laya/resolve/deb313b03518f282a7c4c227dc776c9fe229c30e/`.
+ */
+export declare const LAYA_MODEL_MANIFEST: LayaModelManifest;
+//# sourceMappingURL=model-manifest.d.ts.map

package/dist/extensions/laya/model-manifest.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-manifest.d.ts","sourceRoot":"","sources":["../../../src/extensions/laya/model-manifest.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,cAAc,CAAC;AAEtD;;;GAGG;AACH,eAAO,MAAM,mBAAmB,EAAE,iBA0CjC,CAAC"}

package/dist/extensions/laya/model-manifest.js

@@ -0,0 +1,48 @@
+/**
+ * The trained Laya model baked into the Laya Docker image. Generated by `scripts/package-laya-model.mjs`;
+ * do not edit by hand. The files live at `https://huggingface.co/eltonssouza/relay-laya/resolve/deb313b03518f282a7c4c227dc776c9fe229c30e/`.
+ */
+export const LAYA_MODEL_MANIFEST = {
+    version: "v1",
+    layaPackage: "laya[serve]==0.3.27",
+    baseUrl: "https://huggingface.co/eltonssouza/relay-laya/resolve/deb313b03518f282a7c4c227dc776c9fe229c30e/",
+    files: [
+        {
+            path: "encoder/config.json",
+            asset: "encoder/config.json",
+            size: 2018,
+            sha256: "d69c2320a59a1dd825838a66792048566537c845a30127d2e4fef7ccc759d938",
+        },
+        {
+            path: "laya_trainer_meta.json",
+            asset: "laya_trainer_meta.json",
+            size: 904,
+            sha256: "8a8970d8339248e2358a5bb0764cf02f3524f467634575774f9370d7c288c589",
+        },
+        {
+            path: "model.safetensors",
+            asset: "model.safetensors",
+            size: 643835524,
+            sha256: "d950dc93a2bd0fcf12b7ec1575390c3507073d0f00896adb4727b94ceef192ad",
+        },
+        {
+            path: "rl_agent_config.json",
+            asset: "rl_agent_config.json",
+            size: 584,
+            sha256: "1c556afd40de99c56dd3bec53fbf65120461bfe4f7900b242dcaed10e79d2fad",
+        },
+        {
+            path: "tokenizer/tokenizer.json",
+            asset: "tokenizer/tokenizer.json",
+            size: 34363188,
+            sha256: "609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f",
+        },
+        {
+            path: "tokenizer/tokenizer_config.json",
+            asset: "tokenizer/tokenizer_config.json",
+            size: 648,
+            sha256: "5f36fa307ecb9e70c9b058ca0621530f6bd3e89b371c4d6c0279c48d9dbd2629",
+        },
+    ],
+};
+//# sourceMappingURL=model-manifest.js.map

package/dist/extensions/laya/model-manifest.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-manifest.js","sourceRoot":"","sources":["../../../src/extensions/laya/model-manifest.ts"],"names":[],"mappings":"AAEA;;;GAGG;AACH,MAAM,CAAC,MAAM,mBAAmB,GAAsB;IACrD,OAAO,EAAE,IAAI;IACb,WAAW,EAAE,qBAAqB;IAClC,OAAO,EAAE,iGAAiG;IAC1G,KAAK,EAAE;QACN;YACC,IAAI,EAAE,qBAAqB;YAC3B,KAAK,EAAE,qBAAqB;YAC5B,IAAI,EAAE,IAAI;YACV,MAAM,EAAE,kEAAkE;SAC1E;QACD;YACC,IAAI,EAAE,wBAAwB;YAC9B,KAAK,EAAE,wBAAwB;YAC/B,IAAI,EAAE,GAAG;YACT,MAAM,EAAE,kEAAkE;SAC1E;QACD;YACC,IAAI,EAAE,mBAAmB;YACzB,KAAK,EAAE,mBAAmB;YAC1B,IAAI,EAAE,SAAS;YACf,MAAM,EAAE,kEAAkE;SAC1E;QACD;YACC,IAAI,EAAE,sBAAsB;YAC5B,KAAK,EAAE,sBAAsB;YAC7B,IAAI,EAAE,GAAG;YACT,MAAM,EAAE,kEAAkE;SAC1E;QACD;YACC,IAAI,EAAE,0BAA0B;YAChC,KAAK,EAAE,0BAA0B;YACjC,IAAI,EAAE,QAAQ;YACd,MAAM,EAAE,kEAAkE;SAC1E;QACD;YACC,IAAI,EAAE,iCAAiC;YACvC,KAAK,EAAE,iCAAiC;YACxC,IAAI,EAAE,GAAG;YACT,MAAM,EAAE,kEAAkE;SAC1E;KACD;CACD,CAAC","sourcesContent":["import type { LayaModelManifest } from \"./runtime.ts\";\n\n/**\n * The trained Laya model baked into the Laya Docker image. Generated by `scripts/package-laya-model.mjs`;\n * do not edit by hand. The files live at `https://huggingface.co/eltonssouza/relay-laya/resolve/deb313b03518f282a7c4c227dc776c9fe229c30e/`.\n */\nexport const LAYA_MODEL_MANIFEST: LayaModelManifest = {\n\tversion: \"v1\",\n\tlayaPackage: \"laya[serve]==0.3.27\",\n\tbaseUrl: \"https://huggingface.co/eltonssouza/relay-laya/resolve/deb313b03518f282a7c4c227dc776c9fe229c30e/\",\n\tfiles: [\n\t\t{\n\t\t\tpath: \"encoder/config.json\",\n\t\t\tasset: \"encoder/config.json\",\n\t\t\tsize: 2018,\n\t\t\tsha256: \"d69c2320a59a1dd825838a66792048566537c845a30127d2e4fef7ccc759d938\",\n\t\t},\n\t\t{\n\t\t\tpath: \"laya_trainer_meta.json\",\n\t\t\tasset: \"laya_trainer_meta.json\",\n\t\t\tsize: 904,\n\t\t\tsha256: \"8a8970d8339248e2358a5bb0764cf02f3524f467634575774f9370d7c288c589\",\n\t\t},\n\t\t{\n\t\t\tpath: \"model.safetensors\",\n\t\t\tasset: \"model.safetensors\",\n\t\t\tsize: 643835524,\n\t\t\tsha256: \"d950dc93a2bd0fcf12b7ec1575390c3507073d0f00896adb4727b94ceef192ad\",\n\t\t},\n\t\t{\n\t\t\tpath: \"rl_agent_config.json\",\n\t\t\tasset: \"rl_agent_config.json\",\n\t\t\tsize: 584,\n\t\t\tsha256: \"1c556afd40de99c56dd3bec53fbf65120461bfe4f7900b242dcaed10e79d2fad\",\n\t\t},\n\t\t{\n\t\t\tpath: \"tokenizer/tokenizer.json\",\n\t\t\tasset: \"tokenizer/tokenizer.json\",\n\t\t\tsize: 34363188,\n\t\t\tsha256: \"609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f\",\n\t\t},\n\t\t{\n\t\t\tpath: \"tokenizer/tokenizer_config.json\",\n\t\t\tasset: \"tokenizer/tokenizer_config.json\",\n\t\t\tsize: 648,\n\t\t\tsha256: \"5f36fa307ecb9e70c9b058ca0621530f6bd3e89b371c4d6c0279c48d9dbd2629\",\n\t\t},\n\t],\n};\n"]}

package/dist/extensions/laya/policy.d.ts

@@ -0,0 +1,130 @@
+import type { ModelThinkingLevel } from "@relay-harness/ai";
+import type { TaskAssessment } from "./assessment.ts";
+import { type CapabilityTier, type ReasoningEffort, type ValidationLevel } from "./questions.ts";
+/**
+ * Policy engine: turns Laya's recommendation into a model choice the harness accepts.
+ *
+ * Laya recommends; the harness decides. Deterministic rules set floors that Laya cannot lower,
+ * the model registry maps abstract tiers to concrete models, the quota manager and performance
+ * history adjust each candidate, and a profile-weighted utility picks the cheapest model that is
+ * likely enough to succeed.
+ */
+export declare const POLICY_PROFILES: readonly ["economy", "balanced", "quality", "critical"];
+export type PolicyProfile = (typeof POLICY_PROFILES)[number];
+export interface ProfileWeights {
+    quality: number;
+    cost: number;
+    latency: number;
+    risk: number;
+    /** Minimum estimated success probability a model needs to be chosen. */
+    minSuccess: number;
+}
+export declare const PROFILES: Record<PolicyProfile, ProfileWeights>;
+/**
+ * Concrete models per tier as `provider/model`, in preference order. Only models whose provider
+ * has credentials are candidates. Override with the `laya.models` setting.
+ */
+export declare const DEFAULT_MODEL_REGISTRY: Record<CapabilityTier, string[]>;
+export declare const tierIndex: (tier: CapabilityTier) => number;
+export declare const maxTier: (a: CapabilityTier, b: CapabilityTier) => "balanced" | "fast" | "frontier" | "strong";
+export declare function nextTier(tier: CapabilityTier): CapabilityTier | undefined;
+/** Requests that must not run on a weak model, whatever Laya says. */
+export declare const DETERMINISTIC_RULES: Array<{
+    id: string;
+    pattern: RegExp;
+}>;
+/** The top effort maps to `xhigh`; routing never uses `max`. */
+export declare const EFFORT_TO_THINKING: Record<ReasoningEffort, ModelThinkingLevel>;
+export interface PolicyOptions {
+    /** Answers below this confidence raise the tier by one (`insufficient_confidence`). */
+    minConfidence: number;
+}
+export interface PolicyDecision {
+    /** Tier Laya recommended. */
+    recommendedTier: CapabilityTier;
+    /** Tier the policy requires: the recommendation raised by rules and low confidence. */
+    requiredTier: CapabilityTier;
+    /** Hard floor: no model below this tier is chosen. */
+    minTier: CapabilityTier;
+    effort: ReasoningEffort;
+    thinkingLevel: ModelThinkingLevel;
+    validation: {
+        required: boolean;
+        level: ValidationLevel;
+    };
+    /** Deterministic rules that fired. */
+    rules: string[];
+    /** Why the tier differs from Laya's recommendation. */
+    escalations: EscalationReason[];
+}
+export type EscalationReason = "insufficient_confidence" | "test_failure" | "missing_context" | "high_risk" | "repeated_failure" | "architecture_complexity" | "security_sensitive" | "provider_unavailable";
+/** Applies the deterministic rules and the confidence threshold to Laya's assessment. */
+export declare function applyPolicy(assessment: TaskAssessment, request: string, options: PolicyOptions): PolicyDecision;
+/**
+ * Relative availability of each provider's subscription quota, 0..1. Configured values come from
+ * the `laya.quota` setting; rate-limit and overload errors lower a provider's availability for
+ * ten minutes, so equivalent models of other providers win in the meantime.
+ */
+export declare class QuotaManager {
+    private configured;
+    private penalties;
+    private readonly now;
+    constructor(now?: () => number);
+    configure(quota: Record<string, number> | undefined): void;
+    /** Records a rate-limit or overload error from `provider`. */
+    recordPressure(provider: string): void;
+    available(provider: string): number;
+    snapshot(providers: Iterable<string>): Record<string, number>;
+    private penaltyOf;
+}
+/** History key: task type plus complexity level, e.g. `bug_fix:2`. */
+export declare function taskKey(assessment: Pick<TaskAssessment, "task">): string;
+/** Observed outcomes per task class and model, from telemetry. */
+export declare class PerformanceHistory {
+    private outcomes;
+    record(key: string, model: string, success: boolean): void;
+    get(key: string, model: string): {
+        runs: number;
+        successes: number;
+    } | undefined;
+    /** P(success | task, model): the prior blended with observed runs. */
+    successProbability(key: string, model: string, prior: number): number;
+}
+export interface Candidate {
+    /** `provider/model` */
+    ref: string;
+    provider: string;
+    tier: CapabilityTier;
+    /** Position in the registry; breaks ties in favor of the configured preference. */
+    order: number;
+}
+export interface ScoredCandidate extends Candidate {
+    success: number;
+    cost: number;
+    latency: number;
+    utility: number;
+    eligible: boolean;
+}
+export interface RankInput {
+    requiredTier: CapabilityTier;
+    minTier: CapabilityTier;
+    /** Task risk, 0..1. */
+    risk: number;
+    profile: PolicyProfile;
+    taskKey: string;
+    history: PerformanceHistory;
+    quota: QuotaManager;
+}
+/**
+ * Scores every candidate with the profile's utility
+ * `quality * P(success) - cost * quota_cost - latency * latency - risk * failure_risk`
+ * and sorts the best first. Candidates below the minimum tier are dropped. A candidate is eligible
+ * when its success probability reaches the profile minimum; when none is, the most likely one wins.
+ */
+export declare function rankCandidates(candidates: readonly Candidate[], input: RankInput): ScoredCandidate[];
+/** Parses `provider/model`; the model id may itself contain slashes. */
+export declare function parseModelRef(ref: string): {
+    provider: string;
+    id: string;
+} | undefined;
+//# sourceMappingURL=policy.d.ts.map

package/dist/extensions/laya/policy.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"policy.d.ts","sourceRoot":"","sources":["../../../src/extensions/laya/policy.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAC;AAC5D,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,iBAAiB,CAAC;AACtD,OAAO,EAEN,KAAK,cAAc,EACnB,KAAK,eAAe,EAEpB,KAAK,eAAe,EACpB,MAAM,gBAAgB,CAAC;AAExB;;;;;;;GAOG;AAEH,eAAO,MAAM,eAAe,YAAI,SAAS,EAAE,UAAU,EAAE,SAAS,EAAE,UAAU,CAAU,CAAC;AACvF,MAAM,MAAM,aAAa,GAAG,CAAC,OAAO,eAAe,CAAC,CAAC,MAAM,CAAC,CAAC;AAE7D,MAAM,WAAW,cAAc;IAC9B,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,EAAE,MAAM,CAAC;IACb,wEAAwE;IACxE,UAAU,EAAE,MAAM,CAAC;CACnB;AAED,eAAO,MAAM,QAAQ,EAAE,MAAM,CAAC,aAAa,EAAE,cAAc,CAK1D,CAAC;AAEF;;;GAGG;AACH,eAAO,MAAM,sBAAsB,EAAE,MAAM,CAAC,cAAc,EAAE,MAAM,EAAE,CAKnE,CAAC;AAqBF,eAAO,MAAM,SAAS,SAAU,cAAc,WAAmC,CAAC;AAClF,eAAO,MAAM,OAAO,MAAO,cAAc,KAAK,cAAc,gDAA2C,CAAC;AACxG,wBAAgB,QAAQ,CAAC,IAAI,EAAE,cAAc,GAAG,cAAc,GAAG,SAAS,CAEzE;AAED,sEAAsE;AACtE,eAAO,MAAM,mBAAmB,EAAE,KAAK,CAAC;IAAE,EAAE,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,CAStE,CAAC;AAEF,gEAAgE;AAChE,eAAO,MAAM,kBAAkB,EAAE,MAAM,CAAC,eAAe,EAAE,kBAAkB,CAM1E,CAAC;AAEF,MAAM,WAAW,aAAa;IAC7B,uFAAuF;IACvF,aAAa,EAAE,MAAM,CAAC;CACtB;AAED,MAAM,WAAW,cAAc;IAC9B,6BAA6B;IAC7B,eAAe,EAAE,cAAc,CAAC;IAChC,uFAAuF;IACvF,YAAY,EAAE,cAAc,CAAC;IAC7B,sDAAsD;IACtD,OAAO,EAAE,cAAc,CAAC;IACxB,MAAM,EAAE,eAAe,CAAC;IACxB,aAAa,EAAE,kBAAkB,CAAC;IAClC,UAAU,EAAE;QAAE,QAAQ,EAAE,OAAO,CAAC;QAAC,KAAK,EAAE,eAAe,CAAA;KAAE,CAAC;IAC1D,sCAAsC;IACtC,KAAK,EAAE,MAAM,EAAE,CAAC;IAChB,uDAAuD;IACvD,WAAW,EAAE,gBAAgB,EAAE,CAAC;CAChC;AAED,MAAM,MAAM,gBAAgB,GACzB,yBAAyB,GACzB,cAAc,GACd,iBAAiB,GACjB,WAAW,GACX,kBAAkB,GAClB,yBAAyB,GACzB,oBAAoB,GACpB,sBAAsB,CAAC;AAE1B,yFAAyF;AACzF,wBAAgB,WAAW,CAAC,UAAU,EAAE,cAAc,EAAE,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,aAAa,GAAG,cAAc,CA2C/G;AAED;;;;GAIG;AACH,qBAAa,YAAY;IACxB,OAAO,CAAC,UAAU,CAA8B;IAChD,OAAO,CAAC,SAAS,CAAqD;IACtE,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAe;IAEnC,YAAY,GAAG,GAAE,MAAM,MAAiB,EAEvC;IAED,SAAS,CAAC,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,GAAG,SAAS,GAAG,IAAI,CAEzD;IAED,8DAA8D;IAC9D,cAAc,CAAC,QAAQ,EAAE,MAAM,GAAG,IAAI,CAGrC;IAED,SAAS,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAGlC;IAED,QAAQ,CAAC,SAAS,EAAE,QAAQ,CAAC,MAAM,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAI5D;IAED,OAAO,CAAC,SAAS;CAMjB;AAED,sEAAsE;AACtE,wBAAgB,OAAO,CAAC,UAAU,EAAE,IAAI,CAAC,cAAc,EAAE,MAAM,CAAC,GAAG,MAAM,CAExE;AAED,kEAAkE;AAClE,qBAAa,kBAAkB;IAC9B,OAAO,CAAC,QAAQ,CAA0D;IAE1E,MAAM,CAAC,GAAG,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,IAAI,CAMzD;IAED,GAAG,CAAC,GAAG,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAA;KAAE,GAAG,SAAS,CAE/E;IAED,sEAAsE;IACtE,kBAAkB,CAAC,GAAG,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,MAAM,CAIpE;CACD;AAED,MAAM,WAAW,SAAS;IACzB,uBAAuB;IACvB,GAAG,EAAE,MAAM,CAAC;IACZ,QAAQ,EAAE,MAAM,CAAC;IACjB,IAAI,EAAE,cAAc,CAAC;IACrB,mFAAmF;IACnF,KAAK,EAAE,MAAM,CAAC;CACd;AAED,MAAM,WAAW,eAAgB,SAAQ,SAAS;IACjD,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,CAAC;IAChB,OAAO,EAAE,MAAM,CAAC;IAChB,QAAQ,EAAE,OAAO,CAAC;CAClB;AAED,MAAM,WAAW,SAAS;IACzB,YAAY,EAAE,cAAc,CAAC;IAC7B,OAAO,EAAE,cAAc,CAAC;IACxB,uBAAuB;IACvB,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,aAAa,CAAC;IACvB,OAAO,EAAE,MAAM,CAAC;IAChB,OAAO,EAAE,kBAAkB,CAAC;IAC5B,KAAK,EAAE,YAAY,CAAC;CACpB;AAED;;;;;GAKG;AACH,wBAAgB,cAAc,CAAC,UAAU,EAAE,SAAS,SAAS,EAAE,EAAE,KAAK,EAAE,SAAS,GAAG,eAAe,EAAE,CAqCpG;AAED,wEAAwE;AACxE,wBAAgB,aAAa,CAAC,GAAG,EAAE,MAAM,GAAG;IAAE,QAAQ,EAAE,MAAM,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAAG,SAAS,CAIvF"}

package/dist/extensions/laya/policy.js

@@ -0,0 +1,222 @@
+import { CAPABILITY_TIERS, VALIDATION_LEVELS, } from "./questions.js";
+/**
+ * Policy engine: turns Laya's recommendation into a model choice the harness accepts.
+ *
+ * Laya recommends; the harness decides. Deterministic rules set floors that Laya cannot lower,
+ * the model registry maps abstract tiers to concrete models, the quota manager and performance
+ * history adjust each candidate, and a profile-weighted utility picks the cheapest model that is
+ * likely enough to succeed.
+ */
+export const POLICY_PROFILES = ["economy", "balanced", "quality", "critical"];
+export const PROFILES = {
+    economy: { quality: 0.3, cost: 0.5, latency: 0.2, risk: 0, minSuccess: 0.75 },
+    balanced: { quality: 0.45, cost: 0.3, latency: 0.15, risk: 0.1, minSuccess: 0.85 },
+    quality: { quality: 0.7, cost: 0.1, latency: 0, risk: 0.2, minSuccess: 0.9 },
+    critical: { quality: 0.7, cost: 0, latency: 0, risk: 0.3, minSuccess: 0.95 },
+};
+/**
+ * Concrete models per tier as `provider/model`, in preference order. Only models whose provider
+ * has credentials are candidates. Override with the `laya.models` setting.
+ */
+export const DEFAULT_MODEL_REGISTRY = {
+    fast: ["anthropic/claude-haiku-4-5", "openai-codex/gpt-6-luna", "google/gemini-3.5-flash"],
+    balanced: ["anthropic/claude-sonnet-5-5", "openai-codex/gpt-5.6-terra", "google/gemini-3.8-flash"],
+    strong: ["anthropic/claude-opus-5-5", "openai-codex/gpt-6-sol", "google/gemini-3.1-pro-preview"],
+    frontier: ["anthropic/claude-fable-5-1", "openai-codex/gpt-6.1-sol"],
+};
+/** Relative quota cost and latency per tier, 0..1. */
+const TIER_PRIORS = {
+    fast: { cost: 0.1, latency: 0.15 },
+    balanced: { cost: 0.35, latency: 0.35 },
+    strong: { cost: 0.65, latency: 0.6 },
+    frontier: { cost: 0.9, latency: 0.85 },
+};
+/**
+ * Prior success probability relative to the tier the task requires: a model of exactly that tier is
+ * sufficient, each tier above adds a little, each tier below loses a lot.
+ */
+const MATCHED_SUCCESS = 0.9;
+const OVERPOWERED_BONUS = 0.04;
+const UNDERPOWERED_PENALTY = 0.18;
+/** Weight of the prior, in runs, when blending it with observed history. */
+const PRIOR_RUNS = 5;
+export const tierIndex = (tier) => CAPABILITY_TIERS.indexOf(tier);
+export const maxTier = (a, b) => (tierIndex(a) >= tierIndex(b) ? a : b);
+export function nextTier(tier) {
+    return CAPABILITY_TIERS[tierIndex(tier) + 1];
+}
+/** Requests that must not run on a weak model, whatever Laya says. */
+export const DETERMINISTIC_RULES = [
+    {
+        id: "production-database-migration",
+        pattern: /(migrat|migra[cç]|\bdrop\b|backfill|alter table).*(produc|produ[cç])|(produc|produ[cç]).*(migrat|migra[cç]|\bdrop\b|backfill|alter table)/i,
+    },
+    { id: "authentication", pattern: /\bauth|login|oauth|\bsso\b|password|senha|session token|jwt/i },
+    { id: "security-vulnerability", pattern: /vulnerab|inje[cç]|\bxss\b|csrf|\bcve-|exploit|privilege escalation/i },
+    { id: "payment-processing", pattern: /payment|pagamento|billing|cobran[cç]a|stripe|checkout|invoice|fatura/i },
+];
+/** The top effort maps to `xhigh`; routing never uses `max`. */
+export const EFFORT_TO_THINKING = {
+    minimal: "minimal",
+    low: "low",
+    medium: "medium",
+    high: "high",
+    extra_high: "xhigh",
+};
+/** Applies the deterministic rules and the confidence threshold to Laya's assessment. */
+export function applyPolicy(assessment, request, options) {
+    const recommendedTier = assessment.recommendation.tier;
+    const rules = DETERMINISTIC_RULES.filter((rule) => rule.pattern.test(request)).map((rule) => rule.id);
+    const escalations = [];
+    let minTier = "fast";
+    if (rules.length > 0 || assessment.securitySensitive || assessment.task.type === "security") {
+        minTier = "strong";
+        escalations.push("security_sensitive");
+    }
+    else if (assessment.task.risk >= 0.66) {
+        minTier = "strong";
+        escalations.push("high_risk");
+    }
+    else if (assessment.task.type === "architecture" && assessment.task.complexity >= 0.75) {
+        minTier = "strong";
+        escalations.push("architecture_complexity");
+    }
+    let requiredTier = maxTier(recommendedTier, minTier);
+    if (assessment.confidence < options.minConfidence) {
+        const raised = nextTier(requiredTier);
+        if (raised) {
+            requiredTier = raised;
+            escalations.push("insufficient_confidence");
+        }
+    }
+    // Floors that do not raise the tier are not reported as escalations.
+    const effectiveEscalations = tierIndex(requiredTier) > tierIndex(recommendedTier) ? escalations : [];
+    const sensitive = minTier === "strong";
+    let effort = assessment.recommendation.effort;
+    if (sensitive && (effort === "minimal" || effort === "low"))
+        effort = "medium";
+    const validationLevel = sensitive && VALIDATION_LEVELS.indexOf(assessment.validation) < VALIDATION_LEVELS.indexOf("review")
+        ? "review"
+        : assessment.validation;
+    return {
+        recommendedTier,
+        requiredTier,
+        minTier,
+        effort,
+        thinkingLevel: EFFORT_TO_THINKING[effort],
+        validation: { required: sensitive || validationLevel !== "none", level: validationLevel },
+        rules,
+        escalations: effectiveEscalations,
+    };
+}
+/**
+ * Relative availability of each provider's subscription quota, 0..1. Configured values come from
+ * the `laya.quota` setting; rate-limit and overload errors lower a provider's availability for
+ * ten minutes, so equivalent models of other providers win in the meantime.
+ */
+export class QuotaManager {
+    configured = {};
+    penalties = new Map();
+    now;
+    constructor(now = Date.now) {
+        this.now = now;
+    }
+    configure(quota) {
+        this.configured = { ...(quota ?? {}) };
+    }
+    /** Records a rate-limit or overload error from `provider`. */
+    recordPressure(provider) {
+        const current = this.penaltyOf(provider);
+        this.penalties.set(provider, { amount: Math.min(0.9, current + 0.5), at: this.now() });
+    }
+    available(provider) {
+        const configured = Math.min(1, Math.max(0, this.configured[provider] ?? 1));
+        return configured * (1 - this.penaltyOf(provider));
+    }
+    snapshot(providers) {
+        return Object.fromEntries([...providers].map((provider) => [provider, Number(this.available(provider).toFixed(2))]));
+    }
+    penaltyOf(provider) {
+        const penalty = this.penalties.get(provider);
+        if (!penalty)
+            return 0;
+        const remaining = 1 - (this.now() - penalty.at) / 600_000;
+        return remaining > 0 ? penalty.amount * remaining : 0;
+    }
+}
+/** History key: task type plus complexity level, e.g. `bug_fix:2`. */
+export function taskKey(assessment) {
+    return `${assessment.task.type}:${Math.round(assessment.task.complexity * 4)}`;
+}
+/** Observed outcomes per task class and model, from telemetry. */
+export class PerformanceHistory {
+    outcomes = new Map();
+    record(key, model, success) {
+        const id = `${key}|${model}`;
+        const entry = this.outcomes.get(id) ?? { runs: 0, successes: 0 };
+        entry.runs++;
+        if (success)
+            entry.successes++;
+        this.outcomes.set(id, entry);
+    }
+    get(key, model) {
+        return this.outcomes.get(`${key}|${model}`);
+    }
+    /** P(success | task, model): the prior blended with observed runs. */
+    successProbability(key, model, prior) {
+        const observed = this.get(key, model);
+        if (!observed)
+            return prior;
+        return (prior * PRIOR_RUNS + observed.successes) / (PRIOR_RUNS + observed.runs);
+    }
+}
+/**
+ * Scores every candidate with the profile's utility
+ * `quality * P(success) - cost * quota_cost - latency * latency - risk * failure_risk`
+ * and sorts the best first. Candidates below the minimum tier are dropped. A candidate is eligible
+ * when its success probability reaches the profile minimum; when none is, the most likely one wins.
+ */
+export function rankCandidates(candidates, input) {
+    const weights = PROFILES[input.profile];
+    const required = tierIndex(input.requiredTier);
+    const scored = candidates
+        .filter((candidate) => tierIndex(candidate.tier) >= tierIndex(input.minTier))
+        .filter((candidate) => input.quota.available(candidate.provider) > 0.05)
+        .map((candidate) => {
+        const priors = TIER_PRIORS[candidate.tier];
+        const gap = tierIndex(candidate.tier) - required;
+        const prior = Math.min(0.99, Math.max(0.05, MATCHED_SUCCESS + (gap > 0 ? OVERPOWERED_BONUS : UNDERPOWERED_PENALTY) * gap));
+        const success = input.history.successProbability(input.taskKey, candidate.ref, prior);
+        // Scarce quota makes a model more expensive to use, not less likely to succeed.
+        const cost = Math.min(1, priors.cost * (2 - input.quota.available(candidate.provider)));
+        const failureRisk = (1 - success) * Math.max(input.risk, 0.2);
+        const utility = weights.quality * success -
+            weights.cost * cost -
+            weights.latency * priors.latency -
+            weights.risk * failureRisk;
+        return {
+            ...candidate,
+            success,
+            cost,
+            latency: priors.latency,
+            utility,
+            eligible: success >= weights.minSuccess,
+        };
+    });
+    const anyEligible = scored.some((candidate) => candidate.eligible);
+    return scored.sort((a, b) => {
+        if (anyEligible && a.eligible !== b.eligible)
+            return a.eligible ? -1 : 1;
+        if (!anyEligible)
+            return b.success - a.success || a.order - b.order;
+        return b.utility - a.utility || a.order - b.order;
+    });
+}
+/** Parses `provider/model`; the model id may itself contain slashes. */
+export function parseModelRef(ref) {
+    const slash = ref.indexOf("/");
+    if (slash <= 0 || slash === ref.length - 1)
+        return undefined;
+    return { provider: ref.slice(0, slash), id: ref.slice(slash + 1) };
+}
+//# sourceMappingURL=policy.js.map

package/dist/extensions/laya/policy.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"policy.js","sourceRoot":"","sources":["../../../src/extensions/laya/policy.ts"],"names":[],"mappings":"AAEA,OAAO,EACN,gBAAgB,EAGhB,iBAAiB,GAEjB,MAAM,gBAAgB,CAAC;AAExB;;;;;;;GAOG;AAEH,MAAM,CAAC,MAAM,eAAe,GAAG,CAAC,SAAS,EAAE,UAAU,EAAE,SAAS,EAAE,UAAU,CAAU,CAAC;AAYvF,MAAM,CAAC,MAAM,QAAQ,GAA0C;IAC9D,OAAO,EAAE,EAAE,OAAO,EAAE,GAAG,EAAE,IAAI,EAAE,GAAG,EAAE,OAAO,EAAE,GAAG,EAAE,IAAI,EAAE,CAAC,EAAE,UAAU,EAAE,IAAI,EAAE;IAC7E,QAAQ,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,GAAG,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,GAAG,EAAE,UAAU,EAAE,IAAI,EAAE;IAClF,OAAO,EAAE,EAAE,OAAO,EAAE,GAAG,EAAE,IAAI,EAAE,GAAG,EAAE,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,GAAG,EAAE,UAAU,EAAE,GAAG,EAAE;IAC5E,QAAQ,EAAE,EAAE,OAAO,EAAE,GAAG,EAAE,IAAI,EAAE,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,GAAG,EAAE,UAAU,EAAE,IAAI,EAAE;CAC5E,CAAC;AAEF;;;GAGG;AACH,MAAM,CAAC,MAAM,sBAAsB,GAAqC;IACvE,IAAI,EAAE,CAAC,4BAA4B,EAAE,yBAAyB,EAAE,yBAAyB,CAAC;IAC1F,QAAQ,EAAE,CAAC,6BAA6B,EAAE,4BAA4B,EAAE,yBAAyB,CAAC;IAClG,MAAM,EAAE,CAAC,2BAA2B,EAAE,wBAAwB,EAAE,+BAA+B,CAAC;IAChG,QAAQ,EAAE,CAAC,4BAA4B,EAAE,0BAA0B,CAAC;CACpE,CAAC;AAEF,sDAAsD;AACtD,MAAM,WAAW,GAA8D;IAC9E,IAAI,EAAE,EAAE,IAAI,EAAE,GAAG,EAAE,OAAO,EAAE,IAAI,EAAE;IAClC,QAAQ,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE;IACvC,MAAM,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,GAAG,EAAE;IACpC,QAAQ,EAAE,EAAE,IAAI,EAAE,GAAG,EAAE,OAAO,EAAE,IAAI,EAAE;CACtC,CAAC;AAEF;;;GAGG;AACH,MAAM,eAAe,GAAG,GAAG,CAAC;AAC5B,MAAM,iBAAiB,GAAG,IAAI,CAAC;AAC/B,MAAM,oBAAoB,GAAG,IAAI,CAAC;AAElC,4EAA4E;AAC5E,MAAM,UAAU,GAAG,CAAC,CAAC;AAErB,MAAM,CAAC,MAAM,SAAS,GAAG,CAAC,IAAoB,EAAE,EAAE,CAAC,gBAAgB,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;AAClF,MAAM,CAAC,MAAM,OAAO,GAAG,CAAC,CAAiB,EAAE,CAAiB,EAAE,EAAE,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AACxG,MAAM,UAAU,QAAQ,CAAC,IAAoB;IAC5C,OAAO,gBAAgB,CAAC,SAAS,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC;AAC9C,CAAC;AAED,sEAAsE;AACtE,MAAM,CAAC,MAAM,mBAAmB,GAA2C;IAC1E;QACC,EAAE,EAAE,+BAA+B;QACnC,OAAO,EACN,4IAA4I;KAC7I;IACD,EAAE,EAAE,EAAE,gBAAgB,EAAE,OAAO,EAAE,8DAA8D,EAAE;IACjG,EAAE,EAAE,EAAE,wBAAwB,EAAE,OAAO,EAAE,qEAAqE,EAAE;IAChH,EAAE,EAAE,EAAE,oBAAoB,EAAE,OAAO,EAAE,uEAAuE,EAAE;CAC9G,CAAC;AAEF,gEAAgE;AAChE,MAAM,CAAC,MAAM,kBAAkB,GAAgD;IAC9E,OAAO,EAAE,SAAS;IAClB,GAAG,EAAE,KAAK;IACV,MAAM,EAAE,QAAQ;IAChB,IAAI,EAAE,MAAM;IACZ,UAAU,EAAE,OAAO;CACnB,CAAC;AAiCF,yFAAyF;AACzF,MAAM,UAAU,WAAW,CAAC,UAA0B,EAAE,OAAe,EAAE,OAAsB;IAC9F,MAAM,eAAe,GAAG,UAAU,CAAC,cAAc,CAAC,IAAI,CAAC;IACvD,MAAM,KAAK,GAAG,mBAAmB,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACtG,MAAM,WAAW,GAAuB,EAAE,CAAC;IAC3C,IAAI,OAAO,GAAmB,MAAM,CAAC;IACrC,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,UAAU,CAAC,iBAAiB,IAAI,UAAU,CAAC,IAAI,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;QAC7F,OAAO,GAAG,QAAQ,CAAC;QACnB,WAAW,CAAC,IAAI,CAAC,oBAAoB,CAAC,CAAC;IACxC,CAAC;SAAM,IAAI,UAAU,CAAC,IAAI,CAAC,IAAI,IAAI,IAAI,EAAE,CAAC;QACzC,OAAO,GAAG,QAAQ,CAAC;QACnB,WAAW,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC;IAC/B,CAAC;SAAM,IAAI,UAAU,CAAC,IAAI,CAAC,IAAI,KAAK,cAAc,IAAI,UAAU,CAAC,IAAI,CAAC,UAAU,IAAI,IAAI,EAAE,CAAC;QAC1F,OAAO,GAAG,QAAQ,CAAC;QACnB,WAAW,CAAC,IAAI,CAAC,yBAAyB,CAAC,CAAC;IAC7C,CAAC;IACD,IAAI,YAAY,GAAG,OAAO,CAAC,eAAe,EAAE,OAAO,CAAC,CAAC;IACrD,IAAI,UAAU,CAAC,UAAU,GAAG,OAAO,CAAC,aAAa,EAAE,CAAC;QACnD,MAAM,MAAM,GAAG,QAAQ,CAAC,YAAY,CAAC,CAAC;QACtC,IAAI,MAAM,EAAE,CAAC;YACZ,YAAY,GAAG,MAAM,CAAC;YACtB,WAAW,CAAC,IAAI,CAAC,yBAAyB,CAAC,CAAC;QAC7C,CAAC;IACF,CAAC;IACD,qEAAqE;IACrE,MAAM,oBAAoB,GAAG,SAAS,CAAC,YAAY,CAAC,GAAG,SAAS,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,EAAE,CAAC;IAErG,MAAM,SAAS,GAAG,OAAO,KAAK,QAAQ,CAAC;IACvC,IAAI,MAAM,GAAG,UAAU,CAAC,cAAc,CAAC,MAAM,CAAC;IAC9C,IAAI,SAAS,IAAI,CAAC,MAAM,KAAK,SAAS,IAAI,MAAM,KAAK,KAAK,CAAC;QAAE,MAAM,GAAG,QAAQ,CAAC;IAC/E,MAAM,eAAe,GACpB,SAAS,IAAI,iBAAiB,CAAC,OAAO,CAAC,UAAU,CAAC,UAAU,CAAC,GAAG,iBAAiB,CAAC,OAAO,CAAC,QAAQ,CAAC;QAClG,CAAC,CAAC,QAAQ;QACV,CAAC,CAAC,UAAU,CAAC,UAAU,CAAC;IAC1B,OAAO;QACN,eAAe;QACf,YAAY;QACZ,OAAO;QACP,MAAM;QACN,aAAa,EAAE,kBAAkB,CAAC,MAAM,CAAC;QACzC,UAAU,EAAE,EAAE,QAAQ,EAAE,SAAS,IAAI,eAAe,KAAK,MAAM,EAAE,KAAK,EAAE,eAAe,EAAE;QACzF,KAAK;QACL,WAAW,EAAE,oBAAoB;KACjC,CAAC;AACH,CAAC;AAED;;;;GAIG;AACH,MAAM,OAAO,YAAY;IAChB,UAAU,GAA2B,EAAE,CAAC;IACxC,SAAS,GAAG,IAAI,GAAG,EAA0C,CAAC;IACrD,GAAG,CAAe;IAEnC,YAAY,GAAG,GAAiB,IAAI,CAAC,GAAG;QACvC,IAAI,CAAC,GAAG,GAAG,GAAG,CAAC;IAChB,CAAC;IAED,SAAS,CAAC,KAAyC;QAClD,IAAI,CAAC,UAAU,GAAG,EAAE,GAAG,CAAC,KAAK,IAAI,EAAE,CAAC,EAAE,CAAC;IACxC,CAAC;IAED,8DAA8D;IAC9D,cAAc,CAAC,QAAgB;QAC9B,MAAM,OAAO,GAAG,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC;QACzC,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,QAAQ,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,OAAO,GAAG,GAAG,CAAC,EAAE,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;IACxF,CAAC;IAED,SAAS,CAAC,QAAgB;QACzB,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,UAAU,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QAC5E,OAAO,UAAU,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,CAAC;IACpD,CAAC;IAED,QAAQ,CAAC,SAA2B;QACnC,OAAO,MAAM,CAAC,WAAW,CACxB,CAAC,GAAG,SAAS,CAAC,CAAC,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,QAAQ,EAAE,MAAM,CAAC,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CACzF,CAAC;IACH,CAAC;IAEO,SAAS,CAAC,QAAgB;QACjC,MAAM,OAAO,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;QAC7C,IAAI,CAAC,OAAO;YAAE,OAAO,CAAC,CAAC;QACvB,MAAM,SAAS,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,EAAE,GAAG,OAAO,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC;QAC1D,OAAO,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC;IACvD,CAAC;CACD;AAED,sEAAsE;AACtE,MAAM,UAAU,OAAO,CAAC,UAAwC;IAC/D,OAAO,GAAG,UAAU,CAAC,IAAI,CAAC,IAAI,IAAI,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,IAAI,CAAC,UAAU,GAAG,CAAC,CAAC,EAAE,CAAC;AAChF,CAAC;AAED,kEAAkE;AAClE,MAAM,OAAO,kBAAkB;IACtB,QAAQ,GAAG,IAAI,GAAG,EAA+C,CAAC;IAE1E,MAAM,CAAC,GAAW,EAAE,KAAa,EAAE,OAAgB;QAClD,MAAM,EAAE,GAAG,GAAG,GAAG,IAAI,KAAK,EAAE,CAAC;QAC7B,MAAM,KAAK,GAAG,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,EAAE,CAAC,IAAI,EAAE,IAAI,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,CAAC;QACjE,KAAK,CAAC,IAAI,EAAE,CAAC;QACb,IAAI,OAAO;YAAE,KAAK,CAAC,SAAS,EAAE,CAAC;QAC/B,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,EAAE,EAAE,KAAK,CAAC,CAAC;IAC9B,CAAC;IAED,GAAG,CAAC,GAAW,EAAE,KAAa;QAC7B,OAAO,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,GAAG,GAAG,IAAI,KAAK,EAAE,CAAC,CAAC;IAC7C,CAAC;IAED,sEAAsE;IACtE,kBAAkB,CAAC,GAAW,EAAE,KAAa,EAAE,KAAa;QAC3D,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;QACtC,IAAI,CAAC,QAAQ;YAAE,OAAO,KAAK,CAAC;QAC5B,OAAO,CAAC,KAAK,GAAG,UAAU,GAAG,QAAQ,CAAC,SAAS,CAAC,GAAG,CAAC,UAAU,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC;IACjF,CAAC;CACD;AA8BD;;;;;GAKG;AACH,MAAM,UAAU,cAAc,CAAC,UAAgC,EAAE,KAAgB;IAChF,MAAM,OAAO,GAAG,QAAQ,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;IACxC,MAAM,QAAQ,GAAG,SAAS,CAAC,KAAK,CAAC,YAAY,CAAC,CAAC;IAC/C,MAAM,MAAM,GAAG,UAAU;SACvB,MAAM,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,SAAS,CAAC,SAAS,CAAC,IAAI,CAAC,IAAI,SAAS,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;SAC5E,MAAM,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,KAAK,CAAC,KAAK,CAAC,SAAS,CAAC,SAAS,CAAC,QAAQ,CAAC,GAAG,IAAI,CAAC;SACvE,GAAG,CAAC,CAAC,SAAS,EAAmB,EAAE;QACnC,MAAM,MAAM,GAAG,WAAW,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC;QAC3C,MAAM,GAAG,GAAG,SAAS,CAAC,SAAS,CAAC,IAAI,CAAC,GAAG,QAAQ,CAAC;QACjD,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CACrB,IAAI,EACJ,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,eAAe,GAAG,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,iBAAiB,CAAC,CAAC,CAAC,oBAAoB,CAAC,GAAG,GAAG,CAAC,CAC5F,CAAC;QACF,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO,CAAC,kBAAkB,CAAC,KAAK,CAAC,OAAO,EAAE,SAAS,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;QACtF,gFAAgF;QAChF,MAAM,IAAI,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,CAAC,IAAI,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,KAAK,CAAC,SAAS,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC;QACxF,MAAM,WAAW,GAAG,CAAC,CAAC,GAAG,OAAO,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC;QAC9D,MAAM,OAAO,GACZ,OAAO,CAAC,OAAO,GAAG,OAAO;YACzB,OAAO,CAAC,IAAI,GAAG,IAAI;YACnB,OAAO,CAAC,OAAO,GAAG,MAAM,CAAC,OAAO;YAChC,OAAO,CAAC,IAAI,GAAG,WAAW,CAAC;QAC5B,OAAO;YACN,GAAG,SAAS;YACZ,OAAO;YACP,IAAI;YACJ,OAAO,EAAE,MAAM,CAAC,OAAO;YACvB,OAAO;YACP,QAAQ,EAAE,OAAO,IAAI,OAAO,CAAC,UAAU;SACvC,CAAC;IACH,CAAC,CAAC,CAAC;IACJ,MAAM,WAAW,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC;IACnE,OAAO,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE;QAC3B,IAAI,WAAW,IAAI,CAAC,CAAC,QAAQ,KAAK,CAAC,CAAC,QAAQ;YAAE,OAAO,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACzE,IAAI,CAAC,WAAW;YAAE,OAAO,CAAC,CAAC,OAAO,GAAG,CAAC,CAAC,OAAO,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC;QACpE,OAAO,CAAC,CAAC,OAAO,GAAG,CAAC,CAAC,OAAO,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC;IACnD,CAAC,CAAC,CAAC;AACJ,CAAC;AAED,wEAAwE;AACxE,MAAM,UAAU,aAAa,CAAC,GAAW;IACxC,MAAM,KAAK,GAAG,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;IAC/B,IAAI,KAAK,IAAI,CAAC,IAAI,KAAK,KAAK,GAAG,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,SAAS,CAAC;IAC7D,OAAO,EAAE,QAAQ,EAAE,GAAG,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,EAAE,EAAE,EAAE,GAAG,CAAC,KAAK,CAAC,KAAK,GAAG,CAAC,CAAC,EAAE,CAAC;AACpE,CAAC","sourcesContent":["import type { ModelThinkingLevel } from \"@relay-harness/ai\";\nimport type { TaskAssessment } from \"./assessment.ts\";\nimport {\n\tCAPABILITY_TIERS,\n\ttype CapabilityTier,\n\ttype ReasoningEffort,\n\tVALIDATION_LEVELS,\n\ttype ValidationLevel,\n} from \"./questions.ts\";\n\n/**\n * Policy engine: turns Laya's recommendation into a model choice the harness accepts.\n *\n * Laya recommends; the harness decides. Deterministic rules set floors that Laya cannot lower,\n * the model registry maps abstract tiers to concrete models, the quota manager and performance\n * history adjust each candidate, and a profile-weighted utility picks the cheapest model that is\n * likely enough to succeed.\n */\n\nexport const POLICY_PROFILES = [\"economy\", \"balanced\", \"quality\", \"critical\"] as const;\nexport type PolicyProfile = (typeof POLICY_PROFILES)[number];\n\nexport interface ProfileWeights {\n\tquality: number;\n\tcost: number;\n\tlatency: number;\n\trisk: number;\n\t/** Minimum estimated success probability a model needs to be chosen. */\n\tminSuccess: number;\n}\n\nexport const PROFILES: Record<PolicyProfile, ProfileWeights> = {\n\teconomy: { quality: 0.3, cost: 0.5, latency: 0.2, risk: 0, minSuccess: 0.75 },\n\tbalanced: { quality: 0.45, cost: 0.3, latency: 0.15, risk: 0.1, minSuccess: 0.85 },\n\tquality: { quality: 0.7, cost: 0.1, latency: 0, risk: 0.2, minSuccess: 0.9 },\n\tcritical: { quality: 0.7, cost: 0, latency: 0, risk: 0.3, minSuccess: 0.95 },\n};\n\n/**\n * Concrete models per tier as `provider/model`, in preference order. Only models whose provider\n * has credentials are candidates. Override with the `laya.models` setting.\n */\nexport const DEFAULT_MODEL_REGISTRY: Record<CapabilityTier, string[]> = {\n\tfast: [\"anthropic/claude-haiku-4-5\", \"openai-codex/gpt-6-luna\", \"google/gemini-3.5-flash\"],\n\tbalanced: [\"anthropic/claude-sonnet-5-5\", \"openai-codex/gpt-5.6-terra\", \"google/gemini-3.8-flash\"],\n\tstrong: [\"anthropic/claude-opus-5-5\", \"openai-codex/gpt-6-sol\", \"google/gemini-3.1-pro-preview\"],\n\tfrontier: [\"anthropic/claude-fable-5-1\", \"openai-codex/gpt-6.1-sol\"],\n};\n\n/** Relative quota cost and latency per tier, 0..1. */\nconst TIER_PRIORS: Record<CapabilityTier, { cost: number; latency: number }> = {\n\tfast: { cost: 0.1, latency: 0.15 },\n\tbalanced: { cost: 0.35, latency: 0.35 },\n\tstrong: { cost: 0.65, latency: 0.6 },\n\tfrontier: { cost: 0.9, latency: 0.85 },\n};\n\n/**\n * Prior success probability relative to the tier the task requires: a model of exactly that tier is\n * sufficient, each tier above adds a little, each tier below loses a lot.\n */\nconst MATCHED_SUCCESS = 0.9;\nconst OVERPOWERED_BONUS = 0.04;\nconst UNDERPOWERED_PENALTY = 0.18;\n\n/** Weight of the prior, in runs, when blending it with observed history. */\nconst PRIOR_RUNS = 5;\n\nexport const tierIndex = (tier: CapabilityTier) => CAPABILITY_TIERS.indexOf(tier);\nexport const maxTier = (a: CapabilityTier, b: CapabilityTier) => (tierIndex(a) >= tierIndex(b) ? a : b);\nexport function nextTier(tier: CapabilityTier): CapabilityTier | undefined {\n\treturn CAPABILITY_TIERS[tierIndex(tier) + 1];\n}\n\n/** Requests that must not run on a weak model, whatever Laya says. */\nexport const DETERMINISTIC_RULES: Array<{ id: string; pattern: RegExp }> = [\n\t{\n\t\tid: \"production-database-migration\",\n\t\tpattern:\n\t\t\t/(migrat|migra[cç]|\\bdrop\\b|backfill|alter table).*(produc|produ[cç])|(produc|produ[cç]).*(migrat|migra[cç]|\\bdrop\\b|backfill|alter table)/i,\n\t},\n\t{ id: \"authentication\", pattern: /\\bauth|login|oauth|\\bsso\\b|password|senha|session token|jwt/i },\n\t{ id: \"security-vulnerability\", pattern: /vulnerab|inje[cç]|\\bxss\\b|csrf|\\bcve-|exploit|privilege escalation/i },\n\t{ id: \"payment-processing\", pattern: /payment|pagamento|billing|cobran[cç]a|stripe|checkout|invoice|fatura/i },\n];\n\n/** The top effort maps to `xhigh`; routing never uses `max`. */\nexport const EFFORT_TO_THINKING: Record<ReasoningEffort, ModelThinkingLevel> = {\n\tminimal: \"minimal\",\n\tlow: \"low\",\n\tmedium: \"medium\",\n\thigh: \"high\",\n\textra_high: \"xhigh\",\n};\n\nexport interface PolicyOptions {\n\t/** Answers below this confidence raise the tier by one (`insufficient_confidence`). */\n\tminConfidence: number;\n}\n\nexport interface PolicyDecision {\n\t/** Tier Laya recommended. */\n\trecommendedTier: CapabilityTier;\n\t/** Tier the policy requires: the recommendation raised by rules and low confidence. */\n\trequiredTier: CapabilityTier;\n\t/** Hard floor: no model below this tier is chosen. */\n\tminTier: CapabilityTier;\n\teffort: ReasoningEffort;\n\tthinkingLevel: ModelThinkingLevel;\n\tvalidation: { required: boolean; level: ValidationLevel };\n\t/** Deterministic rules that fired. */\n\trules: string[];\n\t/** Why the tier differs from Laya's recommendation. */\n\tescalations: EscalationReason[];\n}\n\nexport type EscalationReason =\n\t| \"insufficient_confidence\"\n\t| \"test_failure\"\n\t| \"missing_context\"\n\t| \"high_risk\"\n\t| \"repeated_failure\"\n\t| \"architecture_complexity\"\n\t| \"security_sensitive\"\n\t| \"provider_unavailable\";\n\n/** Applies the deterministic rules and the confidence threshold to Laya's assessment. */\nexport function applyPolicy(assessment: TaskAssessment, request: string, options: PolicyOptions): PolicyDecision {\n\tconst recommendedTier = assessment.recommendation.tier;\n\tconst rules = DETERMINISTIC_RULES.filter((rule) => rule.pattern.test(request)).map((rule) => rule.id);\n\tconst escalations: EscalationReason[] = [];\n\tlet minTier: CapabilityTier = \"fast\";\n\tif (rules.length > 0 || assessment.securitySensitive || assessment.task.type === \"security\") {\n\t\tminTier = \"strong\";\n\t\tescalations.push(\"security_sensitive\");\n\t} else if (assessment.task.risk >= 0.66) {\n\t\tminTier = \"strong\";\n\t\tescalations.push(\"high_risk\");\n\t} else if (assessment.task.type === \"architecture\" && assessment.task.complexity >= 0.75) {\n\t\tminTier = \"strong\";\n\t\tescalations.push(\"architecture_complexity\");\n\t}\n\tlet requiredTier = maxTier(recommendedTier, minTier);\n\tif (assessment.confidence < options.minConfidence) {\n\t\tconst raised = nextTier(requiredTier);\n\t\tif (raised) {\n\t\t\trequiredTier = raised;\n\t\t\tescalations.push(\"insufficient_confidence\");\n\t\t}\n\t}\n\t// Floors that do not raise the tier are not reported as escalations.\n\tconst effectiveEscalations = tierIndex(requiredTier) > tierIndex(recommendedTier) ? escalations : [];\n\n\tconst sensitive = minTier === \"strong\";\n\tlet effort = assessment.recommendation.effort;\n\tif (sensitive && (effort === \"minimal\" || effort === \"low\")) effort = \"medium\";\n\tconst validationLevel: ValidationLevel =\n\t\tsensitive && VALIDATION_LEVELS.indexOf(assessment.validation) < VALIDATION_LEVELS.indexOf(\"review\")\n\t\t\t? \"review\"\n\t\t\t: assessment.validation;\n\treturn {\n\t\trecommendedTier,\n\t\trequiredTier,\n\t\tminTier,\n\t\teffort,\n\t\tthinkingLevel: EFFORT_TO_THINKING[effort],\n\t\tvalidation: { required: sensitive || validationLevel !== \"none\", level: validationLevel },\n\t\trules,\n\t\tescalations: effectiveEscalations,\n\t};\n}\n\n/**\n * Relative availability of each provider's subscription quota, 0..1. Configured values come from\n * the `laya.quota` setting; rate-limit and overload errors lower a provider's availability for\n * ten minutes, so equivalent models of other providers win in the meantime.\n */\nexport class QuotaManager {\n\tprivate configured: Record<string, number> = {};\n\tprivate penalties = new Map<string, { amount: number; at: number }>();\n\tprivate readonly now: () => number;\n\n\tconstructor(now: () => number = Date.now) {\n\t\tthis.now = now;\n\t}\n\n\tconfigure(quota: Record<string, number> | undefined): void {\n\t\tthis.configured = { ...(quota ?? {}) };\n\t}\n\n\t/** Records a rate-limit or overload error from `provider`. */\n\trecordPressure(provider: string): void {\n\t\tconst current = this.penaltyOf(provider);\n\t\tthis.penalties.set(provider, { amount: Math.min(0.9, current + 0.5), at: this.now() });\n\t}\n\n\tavailable(provider: string): number {\n\t\tconst configured = Math.min(1, Math.max(0, this.configured[provider] ?? 1));\n\t\treturn configured * (1 - this.penaltyOf(provider));\n\t}\n\n\tsnapshot(providers: Iterable<string>): Record<string, number> {\n\t\treturn Object.fromEntries(\n\t\t\t[...providers].map((provider) => [provider, Number(this.available(provider).toFixed(2))]),\n\t\t);\n\t}\n\n\tprivate penaltyOf(provider: string): number {\n\t\tconst penalty = this.penalties.get(provider);\n\t\tif (!penalty) return 0;\n\t\tconst remaining = 1 - (this.now() - penalty.at) / 600_000;\n\t\treturn remaining > 0 ? penalty.amount * remaining : 0;\n\t}\n}\n\n/** History key: task type plus complexity level, e.g. `bug_fix:2`. */\nexport function taskKey(assessment: Pick<TaskAssessment, \"task\">): string {\n\treturn `${assessment.task.type}:${Math.round(assessment.task.complexity * 4)}`;\n}\n\n/** Observed outcomes per task class and model, from telemetry. */\nexport class PerformanceHistory {\n\tprivate outcomes = new Map<string, { runs: number; successes: number }>();\n\n\trecord(key: string, model: string, success: boolean): void {\n\t\tconst id = `${key}|${model}`;\n\t\tconst entry = this.outcomes.get(id) ?? { runs: 0, successes: 0 };\n\t\tentry.runs++;\n\t\tif (success) entry.successes++;\n\t\tthis.outcomes.set(id, entry);\n\t}\n\n\tget(key: string, model: string): { runs: number; successes: number } | undefined {\n\t\treturn this.outcomes.get(`${key}|${model}`);\n\t}\n\n\t/** P(success | task, model): the prior blended with observed runs. */\n\tsuccessProbability(key: string, model: string, prior: number): number {\n\t\tconst observed = this.get(key, model);\n\t\tif (!observed) return prior;\n\t\treturn (prior * PRIOR_RUNS + observed.successes) / (PRIOR_RUNS + observed.runs);\n\t}\n}\n\nexport interface Candidate {\n\t/** `provider/model` */\n\tref: string;\n\tprovider: string;\n\ttier: CapabilityTier;\n\t/** Position in the registry; breaks ties in favor of the configured preference. */\n\torder: number;\n}\n\nexport interface ScoredCandidate extends Candidate {\n\tsuccess: number;\n\tcost: number;\n\tlatency: number;\n\tutility: number;\n\teligible: boolean;\n}\n\nexport interface RankInput {\n\trequiredTier: CapabilityTier;\n\tminTier: CapabilityTier;\n\t/** Task risk, 0..1. */\n\trisk: number;\n\tprofile: PolicyProfile;\n\ttaskKey: string;\n\thistory: PerformanceHistory;\n\tquota: QuotaManager;\n}\n\n/**\n * Scores every candidate with the profile's utility\n * `quality * P(success) - cost * quota_cost - latency * latency - risk * failure_risk`\n * and sorts the best first. Candidates below the minimum tier are dropped. A candidate is eligible\n * when its success probability reaches the profile minimum; when none is, the most likely one wins.\n */\nexport function rankCandidates(candidates: readonly Candidate[], input: RankInput): ScoredCandidate[] {\n\tconst weights = PROFILES[input.profile];\n\tconst required = tierIndex(input.requiredTier);\n\tconst scored = candidates\n\t\t.filter((candidate) => tierIndex(candidate.tier) >= tierIndex(input.minTier))\n\t\t.filter((candidate) => input.quota.available(candidate.provider) > 0.05)\n\t\t.map((candidate): ScoredCandidate => {\n\t\t\tconst priors = TIER_PRIORS[candidate.tier];\n\t\t\tconst gap = tierIndex(candidate.tier) - required;\n\t\t\tconst prior = Math.min(\n\t\t\t\t0.99,\n\t\t\t\tMath.max(0.05, MATCHED_SUCCESS + (gap > 0 ? OVERPOWERED_BONUS : UNDERPOWERED_PENALTY) * gap),\n\t\t\t);\n\t\t\tconst success = input.history.successProbability(input.taskKey, candidate.ref, prior);\n\t\t\t// Scarce quota makes a model more expensive to use, not less likely to succeed.\n\t\t\tconst cost = Math.min(1, priors.cost * (2 - input.quota.available(candidate.provider)));\n\t\t\tconst failureRisk = (1 - success) * Math.max(input.risk, 0.2);\n\t\t\tconst utility =\n\t\t\t\tweights.quality * success -\n\t\t\t\tweights.cost * cost -\n\t\t\t\tweights.latency * priors.latency -\n\t\t\t\tweights.risk * failureRisk;\n\t\t\treturn {\n\t\t\t\t...candidate,\n\t\t\t\tsuccess,\n\t\t\t\tcost,\n\t\t\t\tlatency: priors.latency,\n\t\t\t\tutility,\n\t\t\t\teligible: success >= weights.minSuccess,\n\t\t\t};\n\t\t});\n\tconst anyEligible = scored.some((candidate) => candidate.eligible);\n\treturn scored.sort((a, b) => {\n\t\tif (anyEligible && a.eligible !== b.eligible) return a.eligible ? -1 : 1;\n\t\tif (!anyEligible) return b.success - a.success || a.order - b.order;\n\t\treturn b.utility - a.utility || a.order - b.order;\n\t});\n}\n\n/** Parses `provider/model`; the model id may itself contain slashes. */\nexport function parseModelRef(ref: string): { provider: string; id: string } | undefined {\n\tconst slash = ref.indexOf(\"/\");\n\tif (slash <= 0 || slash === ref.length - 1) return undefined;\n\treturn { provider: ref.slice(0, slash), id: ref.slice(slash + 1) };\n}\n"]}