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

package/dist/extensions/laya/runtime.js

@@ -0,0 +1,127 @@
+import { createHash } from "node:crypto";
+import { TRAIN_SCRIPT } from "./train-script.js";
+/** Registry the images are published to. */
+export const LAYA_IMAGE_REPOSITORY = "ghcr.io/eltonssouza/relay-laya";
+/** The server container, kept running between Relay sessions. */
+export const LAYA_CONTAINER = "relay-laya";
+/** The container of a running training; removed when it ends. */
+export const LAYA_TRAIN_CONTAINER = "relay-laya-train";
+/** Volume with the models trained from sessions. */
+export const LAYA_MODELS_VOLUME = "relay-laya-models";
+/** Port the server listens on inside its container. */
+export const LAYA_CONTAINER_PORT = 8000;
+/** Paths inside the image and its containers. */
+export const LAYA_IMAGE_PATHS = {
+    shippedModel: "/opt/laya/model",
+    trainedModels: "/data/models",
+    serveScript: "/opt/laya/serve.py",
+    trainScript: "/opt/laya/train.py",
+    workspace: "/workspace",
+};
+const TORCH_INDEX = {
+    cpu: "https://download.pytorch.org/whl/cpu",
+    cuda: "https://download.pytorch.org/whl/cu128",
+};
+/** Serves the model with the System One protocol Relay's `typesafe-system-one` classifier speaks. */
+export const SERVE_SCRIPT = `import argparse
+
+import laya
+import torch
+import uvicorn
+from fastapi import FastAPI, HTTPException
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--model", required=True)
+parser.add_argument("--host", default="127.0.0.1")
+parser.add_argument("--port", type=int, default=8000)
+args = parser.parse_args()
+
+agent = laya.load(args.model, device="cuda" if torch.cuda.is_available() else "cpu")
+app = FastAPI(title="relay-laya")
+
+
+@app.get("/health")
+def health():
+    return {"ok": True, "model": args.model}
+
+
+@app.post("/v1/systemone")
+def system_one(body: dict):
+    if "state" not in body or not body.get("questions"):
+        raise HTTPException(422, "send 'state' and 'questions'")
+    return agent.predict(body["state"], body["questions"])
+
+
+uvicorn.run(app, host=args.host, port=args.port, log_level="warning")
+`;
+/** Downloads the shipped model while the image is built, checking every file against the manifest. */
+export const FETCH_MODEL_SCRIPT = `import hashlib
+import json
+import pathlib
+import sys
+import urllib.request
+
+manifest = json.loads(pathlib.Path(sys.argv[1]).read_text(encoding="utf-8"))
+target = pathlib.Path(sys.argv[2])
+for file in manifest["files"]:
+    path = target / file["path"]
+    path.parent.mkdir(parents=True, exist_ok=True)
+    digest = hashlib.sha256()
+    size = 0
+    with urllib.request.urlopen(manifest["baseUrl"] + file["asset"]) as response, open(path, "wb") as out:
+        while chunk := response.read(1 << 20):
+            digest.update(chunk)
+            size += len(chunk)
+            out.write(chunk)
+    if size != file["size"] or digest.hexdigest() != file["sha256"]:
+        sys.exit(f"{file['asset']} does not match the manifest (size {size}, sha256 {digest.hexdigest()})")
+    print(f"{file['path']} ok", flush=True)
+`;
+function dockerfile(manifest, variant) {
+    const paths = LAYA_IMAGE_PATHS;
+    return `FROM python:3.11-slim
+LABEL org.opencontainers.image.source="https://github.com/eltonssouza/relay-harness"
+LABEL org.opencontainers.image.description="Laya routing model for Relay (${variant})"
+ENV PYTHONDONTWRITEBYTECODE=1 PYTHONUNBUFFERED=1 PYTHONIOENCODING=utf-8 PIP_NO_CACHE_DIR=1 \\
+    PIP_DISABLE_PIP_VERSION_CHECK=1 USE_TF=0 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 TOKENIZERS_PARALLELISM=false
+RUN pip install --upgrade pip && pip install torch --index-url ${TORCH_INDEX[variant]}
+RUN pip install "${manifest.layaPackage}" fastapi uvicorn
+RUN useradd --create-home laya && mkdir -p ${paths.trainedModels} && chown laya ${paths.trainedModels}
+COPY fetch_model.py model.json /opt/laya/
+RUN python /opt/laya/fetch_model.py /opt/laya/model.json ${paths.shippedModel}
+COPY serve.py train.py /opt/laya/
+USER laya
+EXPOSE ${LAYA_CONTAINER_PORT}
+CMD ["python", "${paths.serveScript}", "--model", "${paths.shippedModel}", "--host", "0.0.0.0", "--port", "${LAYA_CONTAINER_PORT}"]
+`;
+}
+/** Files of the image's build context, by name. */
+export function layaImageContext(manifest, variant) {
+    return {
+        Dockerfile: dockerfile(manifest, variant),
+        "fetch_model.py": FETCH_MODEL_SCRIPT,
+        "model.json": `${JSON.stringify({ baseUrl: manifest.baseUrl, files: manifest.files }, null, 2)}\n`,
+        "serve.py": SERVE_SCRIPT,
+        "train.py": TRAIN_SCRIPT,
+    };
+}
+/**
+ * Image tag: the model version, the variant and a hash of the build context. Changing a script, a
+ * dependency or the model gives a new tag, so a Relay version always runs the image built from its
+ * own sources, and an unchanged context reuses the published image.
+ */
+export function layaImageTag(manifest, variant) {
+    const hash = createHash("sha256");
+    for (const [name, content] of Object.entries(layaImageContext(manifest, variant)).sort(([a], [b]) => a.localeCompare(b))) {
+        hash.update(`${name}\0${content}\0`);
+    }
+    return `${manifest.version}-${variant}-${hash.digest("hex").slice(0, 12)}`;
+}
+export function layaImage(manifest, variant) {
+    return `${LAYA_IMAGE_REPOSITORY}:${layaImageTag(manifest, variant)}`;
+}
+/** Container path of a model trained from sessions. */
+export function trainedModelPath(name) {
+    return `${LAYA_IMAGE_PATHS.trainedModels}/${name}`;
+}
+//# sourceMappingURL=runtime.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"runtime.js","sourceRoot":"","sources":["../../../src/extensions/laya/runtime.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AACzC,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AA2BjD,4CAA4C;AAC5C,MAAM,CAAC,MAAM,qBAAqB,GAAG,gCAAgC,CAAC;AACtE,iEAAiE;AACjE,MAAM,CAAC,MAAM,cAAc,GAAG,YAAY,CAAC;AAC3C,iEAAiE;AACjE,MAAM,CAAC,MAAM,oBAAoB,GAAG,kBAAkB,CAAC;AACvD,oDAAoD;AACpD,MAAM,CAAC,MAAM,kBAAkB,GAAG,mBAAmB,CAAC;AACtD,uDAAuD;AACvD,MAAM,CAAC,MAAM,mBAAmB,GAAG,IAAI,CAAC;AAExC,iDAAiD;AACjD,MAAM,CAAC,MAAM,gBAAgB,GAAG;IAC/B,YAAY,EAAE,iBAAiB;IAC/B,aAAa,EAAE,cAAc;IAC7B,WAAW,EAAE,oBAAoB;IACjC,WAAW,EAAE,oBAAoB;IACjC,SAAS,EAAE,YAAY;CACd,CAAC;AAKX,MAAM,WAAW,GAAqC;IACrD,GAAG,EAAE,sCAAsC;IAC3C,IAAI,EAAE,wCAAwC;CAC9C,CAAC;AAEF,qGAAqG;AACrG,MAAM,CAAC,MAAM,YAAY,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8B3B,CAAC;AAEF,sGAAsG;AACtG,MAAM,CAAC,MAAM,kBAAkB,GAAG;;;;;;;;;;;;;;;;;;;;;CAqBjC,CAAC;AAEF,SAAS,UAAU,CAAC,QAA2B,EAAE,OAAyB;IACzE,MAAM,KAAK,GAAG,gBAAgB,CAAC;IAC/B,OAAO;;4EAEoE,OAAO;;;iEAGlB,WAAW,CAAC,OAAO,CAAC;mBAClE,QAAQ,CAAC,WAAW;6CACM,KAAK,CAAC,aAAa,kBAAkB,KAAK,CAAC,aAAa;;2DAE1C,KAAK,CAAC,YAAY;;;SAGpE,mBAAmB;kBACV,KAAK,CAAC,WAAW,kBAAkB,KAAK,CAAC,YAAY,sCAAsC,mBAAmB;CAC/H,CAAC;AACF,CAAC;AAED,mDAAmD;AACnD,MAAM,UAAU,gBAAgB,CAAC,QAA2B,EAAE,OAAyB;IACtF,OAAO;QACN,UAAU,EAAE,UAAU,CAAC,QAAQ,EAAE,OAAO,CAAC;QACzC,gBAAgB,EAAE,kBAAkB;QACpC,YAAY,EAAE,GAAG,IAAI,CAAC,SAAS,CAAC,EAAE,OAAO,EAAE,QAAQ,CAAC,OAAO,EAAE,KAAK,EAAE,QAAQ,CAAC,KAAK,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI;QAClG,UAAU,EAAE,YAAY;QACxB,UAAU,EAAE,YAAY;KACxB,CAAC;AACH,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,YAAY,CAAC,QAA2B,EAAE,OAAyB;IAClF,MAAM,IAAI,GAAG,UAAU,CAAC,QAAQ,CAAC,CAAC;IAClC,KAAK,MAAM,CAAC,IAAI,EAAE,OAAO,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,gBAAgB,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,EAAE,CACnG,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC,CAClB,EAAE,CAAC;QACH,IAAI,CAAC,MAAM,CAAC,GAAG,IAAI,KAAK,OAAO,IAAI,CAAC,CAAC;IACtC,CAAC;IACD,OAAO,GAAG,QAAQ,CAAC,OAAO,IAAI,OAAO,IAAI,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,CAAC;AAC5E,CAAC;AAED,MAAM,UAAU,SAAS,CAAC,QAA2B,EAAE,OAAyB;IAC/E,OAAO,GAAG,qBAAqB,IAAI,YAAY,CAAC,QAAQ,EAAE,OAAO,CAAC,EAAE,CAAC;AACtE,CAAC;AAED,uDAAuD;AACvD,MAAM,UAAU,gBAAgB,CAAC,IAAY;IAC5C,OAAO,GAAG,gBAAgB,CAAC,aAAa,IAAI,IAAI,EAAE,CAAC;AACpD,CAAC","sourcesContent":["import { createHash } from \"node:crypto\";\nimport { TRAIN_SCRIPT } from \"./train-script.ts\";\n\n/**\n * Laya runs in Docker, never on the host: an image carries Python, torch, the `laya` package, the\n * shipped model and the scripts that serve and train it. Relay pulls the image, keeps one server\n * container running and starts a short-lived container for each training. The host only keeps\n * Relay's own data (training exercises, the model registry, telemetry).\n */\n\nexport interface LayaModelFile {\n\t/** Path inside the model directory, with forward slashes. */\n\tpath: string;\n\t/** File name in the model repository. */\n\tasset: string;\n\tsize: number;\n\tsha256: string;\n}\n\nexport interface LayaModelManifest {\n\tversion: string;\n\t/** pip requirement for the Laya runtime. */\n\tlayaPackage: string;\n\t/** Directory URL the model files are downloaded from when the image is built, ending in a slash. */\n\tbaseUrl: string;\n\tfiles: LayaModelFile[];\n}\n\n/** Registry the images are published to. */\nexport const LAYA_IMAGE_REPOSITORY = \"ghcr.io/eltonssouza/relay-laya\";\n/** The server container, kept running between Relay sessions. */\nexport const LAYA_CONTAINER = \"relay-laya\";\n/** The container of a running training; removed when it ends. */\nexport const LAYA_TRAIN_CONTAINER = \"relay-laya-train\";\n/** Volume with the models trained from sessions. */\nexport const LAYA_MODELS_VOLUME = \"relay-laya-models\";\n/** Port the server listens on inside its container. */\nexport const LAYA_CONTAINER_PORT = 8000;\n\n/** Paths inside the image and its containers. */\nexport const LAYA_IMAGE_PATHS = {\n\tshippedModel: \"/opt/laya/model\",\n\ttrainedModels: \"/data/models\",\n\tserveScript: \"/opt/laya/serve.py\",\n\ttrainScript: \"/opt/laya/train.py\",\n\tworkspace: \"/workspace\",\n} as const;\n\n/** `cpu` runs everywhere; `cuda` carries the CUDA build of torch for machines with an NVIDIA GPU. */\nexport type LayaImageVariant = \"cpu\" | \"cuda\";\n\nconst TORCH_INDEX: Record<LayaImageVariant, string> = {\n\tcpu: \"https://download.pytorch.org/whl/cpu\",\n\tcuda: \"https://download.pytorch.org/whl/cu128\",\n};\n\n/** Serves the model with the System One protocol Relay's `typesafe-system-one` classifier speaks. */\nexport const SERVE_SCRIPT = `import argparse\n\nimport laya\nimport torch\nimport uvicorn\nfrom fastapi import FastAPI, HTTPException\n\nparser = argparse.ArgumentParser()\nparser.add_argument(\"--model\", required=True)\nparser.add_argument(\"--host\", default=\"127.0.0.1\")\nparser.add_argument(\"--port\", type=int, default=8000)\nargs = parser.parse_args()\n\nagent = laya.load(args.model, device=\"cuda\" if torch.cuda.is_available() else \"cpu\")\napp = FastAPI(title=\"relay-laya\")\n\n\n@app.get(\"/health\")\ndef health():\n    return {\"ok\": True, \"model\": args.model}\n\n\n@app.post(\"/v1/systemone\")\ndef system_one(body: dict):\n    if \"state\" not in body or not body.get(\"questions\"):\n        raise HTTPException(422, \"send 'state' and 'questions'\")\n    return agent.predict(body[\"state\"], body[\"questions\"])\n\n\nuvicorn.run(app, host=args.host, port=args.port, log_level=\"warning\")\n`;\n\n/** Downloads the shipped model while the image is built, checking every file against the manifest. */\nexport const FETCH_MODEL_SCRIPT = `import hashlib\nimport json\nimport pathlib\nimport sys\nimport urllib.request\n\nmanifest = json.loads(pathlib.Path(sys.argv[1]).read_text(encoding=\"utf-8\"))\ntarget = pathlib.Path(sys.argv[2])\nfor file in manifest[\"files\"]:\n    path = target / file[\"path\"]\n    path.parent.mkdir(parents=True, exist_ok=True)\n    digest = hashlib.sha256()\n    size = 0\n    with urllib.request.urlopen(manifest[\"baseUrl\"] + file[\"asset\"]) as response, open(path, \"wb\") as out:\n        while chunk := response.read(1 << 20):\n            digest.update(chunk)\n            size += len(chunk)\n            out.write(chunk)\n    if size != file[\"size\"] or digest.hexdigest() != file[\"sha256\"]:\n        sys.exit(f\"{file['asset']} does not match the manifest (size {size}, sha256 {digest.hexdigest()})\")\n    print(f\"{file['path']} ok\", flush=True)\n`;\n\nfunction dockerfile(manifest: LayaModelManifest, variant: LayaImageVariant): string {\n\tconst paths = LAYA_IMAGE_PATHS;\n\treturn `FROM python:3.11-slim\nLABEL org.opencontainers.image.source=\"https://github.com/eltonssouza/relay-harness\"\nLABEL org.opencontainers.image.description=\"Laya routing model for Relay (${variant})\"\nENV PYTHONDONTWRITEBYTECODE=1 PYTHONUNBUFFERED=1 PYTHONIOENCODING=utf-8 PIP_NO_CACHE_DIR=1 \\\\\n    PIP_DISABLE_PIP_VERSION_CHECK=1 USE_TF=0 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 TOKENIZERS_PARALLELISM=false\nRUN pip install --upgrade pip && pip install torch --index-url ${TORCH_INDEX[variant]}\nRUN pip install \"${manifest.layaPackage}\" fastapi uvicorn\nRUN useradd --create-home laya && mkdir -p ${paths.trainedModels} && chown laya ${paths.trainedModels}\nCOPY fetch_model.py model.json /opt/laya/\nRUN python /opt/laya/fetch_model.py /opt/laya/model.json ${paths.shippedModel}\nCOPY serve.py train.py /opt/laya/\nUSER laya\nEXPOSE ${LAYA_CONTAINER_PORT}\nCMD [\"python\", \"${paths.serveScript}\", \"--model\", \"${paths.shippedModel}\", \"--host\", \"0.0.0.0\", \"--port\", \"${LAYA_CONTAINER_PORT}\"]\n`;\n}\n\n/** Files of the image's build context, by name. */\nexport function layaImageContext(manifest: LayaModelManifest, variant: LayaImageVariant): Record<string, string> {\n\treturn {\n\t\tDockerfile: dockerfile(manifest, variant),\n\t\t\"fetch_model.py\": FETCH_MODEL_SCRIPT,\n\t\t\"model.json\": `${JSON.stringify({ baseUrl: manifest.baseUrl, files: manifest.files }, null, 2)}\\n`,\n\t\t\"serve.py\": SERVE_SCRIPT,\n\t\t\"train.py\": TRAIN_SCRIPT,\n\t};\n}\n\n/**\n * Image tag: the model version, the variant and a hash of the build context. Changing a script, a\n * dependency or the model gives a new tag, so a Relay version always runs the image built from its\n * own sources, and an unchanged context reuses the published image.\n */\nexport function layaImageTag(manifest: LayaModelManifest, variant: LayaImageVariant): string {\n\tconst hash = createHash(\"sha256\");\n\tfor (const [name, content] of Object.entries(layaImageContext(manifest, variant)).sort(([a], [b]) =>\n\t\ta.localeCompare(b),\n\t)) {\n\t\thash.update(`${name}\\0${content}\\0`);\n\t}\n\treturn `${manifest.version}-${variant}-${hash.digest(\"hex\").slice(0, 12)}`;\n}\n\nexport function layaImage(manifest: LayaModelManifest, variant: LayaImageVariant): string {\n\treturn `${LAYA_IMAGE_REPOSITORY}:${layaImageTag(manifest, variant)}`;\n}\n\n/** Container path of a model trained from sessions. */\nexport function trainedModelPath(name: string): string {\n\treturn `${LAYA_IMAGE_PATHS.trainedModels}/${name}`;\n}\n"]}

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

@@ -0,0 +1,39 @@
+import { type AgentId, type CapabilityTier, type TaskScope, type TaskType, type ValidationLevel } from "./questions.ts";
+/**
+ * Synthetic bootstrap exercises for training Laya before the harness has telemetry.
+ *
+ * Each template fixes the task profile; the tier and effort follow from it by the same rule for
+ * every template, so the labels are consistent. These are opinion-based labels. Replace them over
+ * time with evidence-based labels exported from telemetry (`/laya export`).
+ */
+interface Template {
+    type: TaskType;
+    /** Indexes into the score levels of `questions.ts`. */
+    complexity: 0 | 1 | 2 | 3 | 4;
+    scope: TaskScope;
+    risk: 0 | 1 | 2 | 3;
+    ambiguity: 0 | 1 | 2;
+    reasoning: 0 | 1 | 2 | 3 | 4;
+    agent: AgentId;
+    validation: ValidationLevel;
+    /** w=write s=shell t=tests W=web b=browser d=database g=git */
+    tools: string;
+    security?: boolean;
+    texts: string[];
+}
+/** Tier from the task profile: the cheapest tier for its difficulty, raised for risk and security. */
+export declare function seedTier(template: Pick<Template, "complexity" | "reasoning" | "risk" | "security">): CapabilityTier;
+export interface SeedRow {
+    state: {
+        request: string;
+    };
+    expected: Record<string, string | number | boolean>;
+    source: "synthetic";
+}
+/**
+ * Generates up to `count` distinct exercises in laya-trainer's dataset format, the same number per
+ * template, so templates with many slot combinations do not crowd out the others.
+ */
+export declare function generateSeed(count: number, seed?: number): SeedRow[];
+export {};
+//# sourceMappingURL=seed.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"seed.d.ts","sourceRoot":"","sources":["../../../src/extensions/laya/seed.ts"],"names":[],"mappings":"AAAA,OAAO,EACN,KAAK,OAAO,EAEZ,KAAK,cAAc,EAEnB,KAAK,SAAS,EACd,KAAK,QAAQ,EAEb,KAAK,eAAe,EACpB,MAAM,gBAAgB,CAAC;AAExB;;;;;;GAMG;AAEH,UAAU,QAAQ;IACjB,IAAI,EAAE,QAAQ,CAAC;IACf,uDAAuD;IACvD,UAAU,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IAC9B,KAAK,EAAE,SAAS,CAAC;IACjB,IAAI,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IACpB,SAAS,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IACrB,SAAS,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IAC7B,KAAK,EAAE,OAAO,CAAC;IACf,UAAU,EAAE,eAAe,CAAC;IAC5B,+DAA+D;IAC/D,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,KAAK,EAAE,MAAM,EAAE,CAAC;CAChB;AA8VD,sGAAsG;AACtG,wBAAgB,QAAQ,CAAC,QAAQ,EAAE,IAAI,CAAC,QAAQ,EAAE,YAAY,GAAG,WAAW,GAAG,MAAM,GAAG,UAAU,CAAC,GAAG,cAAc,CAKnH;AAaD,MAAM,WAAW,OAAO;IACvB,KAAK,EAAE;QAAE,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC;IAC3B,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,GAAG,OAAO,CAAC,CAAC;IACpD,MAAM,EAAE,WAAW,CAAC;CACpB;AAED;;;GAGG;AACH,wBAAgB,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,SAAO,GAAG,OAAO,EAAE,CA+ClE"}

package/dist/extensions/laya/seed.js

@@ -0,0 +1,358 @@
+import { CAPABILITY_TIERS, REASONING_EFFORTS, } from "./questions.js";
+const TOOL_FLAGS = {
+    w: "requires_write",
+    s: "requires_shell",
+    t: "requires_tests",
+    W: "requires_web",
+    b: "requires_browser",
+    d: "requires_database",
+    g: "requires_git",
+};
+const t = (type, complexity, scope, risk, ambiguity, reasoning, agent, validation, tools, texts, security = false) => ({ type, complexity, scope, risk, ambiguity, reasoning, agent, validation, tools, texts, security });
+const TEMPLATES = [
+    t("question", 0, "single_file", 0, 0, 0, "researcher", "none", "", [
+        "What does the function {fn} in {file} return?",
+        "O que a função {fn} em {file} retorna?",
+        "Where is {fn} defined?",
+        "Onde fica definida a função {fn}?",
+    ]),
+    t("question", 1, "module", 0, 0, 1, "researcher", "none", "", [
+        "How does the {module} module load its configuration?",
+        "Como o módulo {module} carrega a configuração?",
+        "Explain how {module} handles errors.",
+        "Explique como o {module} trata erros.",
+    ]),
+    t("question", 2, "repository", 0, 1, 2, "software-architect", "none", "", [
+        "Why does the {module} package depend on {module2}? Is that coupling necessary?",
+        "Por que o pacote {module} depende do {module2}? Esse acoplamento é necessário?",
+    ]),
+    t("question", 1, "single_file", 0, 2, 1, "researcher", "none", "", [
+        "Is this approach ok?",
+        "Isso está certo?",
+        "Can you take a look at {file}?",
+        "Dá uma olhada no {file}?",
+    ]),
+    t("code_edit", 0, "single_line", 0, 0, 0, "frontend-engineer", "syntax", "w", [
+        "Change the text of the submit button to 'Save'.",
+        "Corrija o texto deste botão para 'Salvar'.",
+        "Fix the typo in the error message of {file}.",
+        "Corrija o erro de digitação na mensagem de erro de {file}.",
+    ]),
+    t("code_edit", 0, "single_line", 0, 0, 0, "software-engineer", "compile", "wt", [
+        "Rename the variable {var} to {var2} in {file}.",
+        "Renomeie a variável {var} para {var2} em {file}.",
+        "Change the default timeout in {file} from 30 to 60 seconds.",
+        "Mude o timeout padrão em {file} de 30 para 60 segundos.",
+    ]),
+    t("code_edit", 1, "single_file", 0, 0, 1, "software-engineer", "compile", "wt", [
+        "Remove the unused imports from {file}.",
+        "Remova os imports não usados de {file}.",
+        "Make {fn} in {file} async.",
+        "Transforme {fn} em {file} em uma função assíncrona.",
+    ]),
+    t("code_edit", 0, "single_file", 0, 0, 0, "software-engineer", "none", "g", [
+        "Commit the staged changes with a descriptive message.",
+        "Faça o commit das mudanças com uma mensagem descritiva.",
+        "Create a branch for this fix.",
+        "Crie uma branch para essa correção.",
+    ]),
+    t("code_edit", 1, "single_file", 1, 0, 1, "devops-engineer", "compile", "ws", [
+        "Bump Node to 22 in the Dockerfile.",
+        "Atualize a versão do Node para 22 no Dockerfile.",
+        "Add a cache step for npm in the CI workflow.",
+        "Adicione cache do npm no workflow do CI.",
+    ]),
+    t("bug_fix", 1, "single_file", 1, 0, 1, "software-engineer", "unit_test", "wt", [
+        "{fn} in {file} returns undefined when the list is empty; it should return an empty array.",
+        "A função {fn} em {file} retorna undefined quando a lista está vazia; deveria retornar um array vazio.",
+        "Off-by-one error in the pagination of {module}: the last page is skipped.",
+        "Erro de off-by-one na paginação do {module}: a última página é pulada.",
+    ]),
+    t("bug_fix", 1, "single_file", 0, 0, 1, "frontend-engineer", "syntax", "w", [
+        "The header is misaligned on small screens.",
+        "O cabeçalho está desalinhado em telas pequenas.",
+        "Fix the CSS so the footer sticks to the bottom.",
+        "Arrume o CSS para o rodapé ficar no final da página.",
+    ]),
+    t("bug_fix", 2, "module", 1, 1, 2, "backend-engineer", "unit_test", "wt", [
+        "The {module} API returns 500 when the request body has no 'email' field. It should return 400.",
+        "A API do {module} retorna 500 quando o corpo não tem o campo 'email'. Deveria retornar 400.",
+        "Dates are saved in local time instead of UTC in the {module} service.",
+        "As datas são salvas no horário local em vez de UTC no serviço {module}.",
+    ]),
+    t("bug_fix", 2, "single_file", 1, 1, 2, "frontend-engineer", "integration_test", "wtb", [
+        "The dropdown in the settings page closes immediately on mobile.",
+        "O dropdown da página de configurações fecha sozinho no celular.",
+        "The modal does not trap focus; keyboard users can tab behind it.",
+        "O modal não prende o foco; quem usa teclado consegue navegar por trás dele.",
+    ]),
+    t("bug_fix", 3, "multi_module", 2, 1, 3, "backend-engineer", "integration_test", "wts", [
+        "Race condition: two workers in {module} sometimes process the same job twice.",
+        "Condição de corrida: dois workers do {module} às vezes processam o mesmo job duas vezes.",
+        "Memory leak in the {module} cache under load; memory grows until the process is killed.",
+        "Vazamento de memória no cache do {module} sob carga; a memória cresce até o processo cair.",
+    ]),
+    t("bug_fix", 3, "module", 2, 1, 3, "mobile-engineer", "integration_test", "wt", [
+        "The Android app crashes on startup on Android 14 only.",
+        "O app Android fecha ao abrir apenas no Android 14.",
+    ]),
+    t("debugging", 2, "single_file", 1, 1, 2, "software-engineer", "unit_test", "wts", [
+        "I get 'TypeError: cannot read properties of undefined' when running the app. Help.",
+        "Estou recebendo 'TypeError: cannot read properties of undefined' ao rodar o app. Ajuda.",
+    ]),
+    t("debugging", 3, "module", 1, 2, 3, "software-engineer", "unit_test", "wts", [
+        "The test '{test}' fails only on CI with a timeout and passes locally. Find out why.",
+        "O teste '{test}' falha só no CI com timeout e passa localmente. Descubra o motivo.",
+        "Something broke after the last merge: {module} hangs on startup. Investigate.",
+        "Alguma coisa quebrou depois do último merge: o {module} trava na inicialização. Investigue.",
+    ]),
+    t("debugging", 4, "multi_module", 2, 2, 4, "software-architect", "integration_test", "wts", [
+        "Intermittent data corruption between the {module} and {module2} services, no clear reproduction. Find the root cause.",
+        "Corrupção de dados intermitente entre os serviços {module} e {module2}, sem reprodução clara. Ache a causa raiz.",
+        "Deadlock under high concurrency in the scheduler; it happens about once a day in production.",
+        "Deadlock com alta concorrência no scheduler; acontece mais ou menos uma vez por dia em produção.",
+    ]),
+    t("refactor", 1, "single_file", 0, 0, 1, "software-engineer", "compile", "wt", [
+        "Convert {file} from JavaScript to TypeScript.",
+        "Converta {file} de JavaScript para TypeScript.",
+        "Replace var with const and let in {file}.",
+        "Troque var por const e let em {file}.",
+    ]),
+    t("refactor", 2, "single_file", 1, 0, 2, "software-engineer", "unit_test", "wt", [
+        "Extract the validation logic of {file} into its own function.",
+        "Extraia a lógica de validação de {file} para uma função separada.",
+        "Split {file}; it has 1500 lines.",
+        "Divida {file}, que tem 1500 linhas.",
+    ]),
+    t("refactor", 3, "multi_module", 2, 1, 3, "software-architect", "unit_test", "wt", [
+        "Replace the global singleton config in {module} with dependency injection across the packages.",
+        "Substitua o singleton global de configuração do {module} por injeção de dependência em todos os pacotes.",
+        "Migrate all callbacks in {module} to async/await.",
+        "Migre todos os callbacks do {module} para async/await.",
+    ]),
+    t("refactor", 4, "repository", 2, 2, 4, "software-architect", "integration_test", "wtg", [
+        "Restructure the monorepo so {module} becomes an independent package without breaking consumers.",
+        "Reestruture o monorepo para o {module} virar um pacote independente sem quebrar quem usa.",
+    ]),
+    t("feature", 1, "single_file", 0, 0, 1, "software-engineer", "unit_test", "wt", [
+        "Add a --verbose flag to the CLI that prints debug logs.",
+        "Adicione uma flag --verbose na CLI que imprima logs de debug.",
+        "Add a helper in {file} that formats bytes as KB or MB.",
+        "Crie em {file} uma função que formata bytes como KB ou MB.",
+    ]),
+    t("feature", 2, "module", 1, 1, 2, "backend-engineer", "unit_test", "wt", [
+        "Add a GET /users/:id/orders endpoint with pagination.",
+        "Crie um endpoint GET /users/:id/orders com paginação.",
+        "Add retry with exponential backoff to the HTTP client in {module}.",
+        "Adicione retry com backoff exponencial no cliente HTTP do {module}.",
+    ]),
+    t("feature", 2, "module", 1, 1, 2, "frontend-engineer", "integration_test", "wtb", [
+        "Add a dark mode toggle to the settings page.",
+        "Adicione um botão de modo escuro na página de configurações.",
+        "Create a reusable date picker component.",
+        "Crie um componente reutilizável de seleção de data.",
+    ]),
+    t("feature", 2, "module", 1, 1, 2, "mobile-engineer", "integration_test", "wt", [
+        "Add push notifications to the Android app.",
+        "Adicione notificações push no app Android.",
+        "The iOS app needs offline caching of the feed.",
+        "O app iOS precisa de cache offline do feed.",
+    ]),
+    t("feature", 2, "module", 1, 1, 2, "software-engineer", "unit_test", "wtW", [
+        "Integrate the {tech} SDK following its official docs.",
+        "Integre o SDK do {tech} seguindo a documentação oficial.",
+    ]),
+    t("feature", 2, "single_file", 1, 0, 2, "database-engineer", "unit_test", "wtd", [
+        "Add an index to speed up the orders-by-customer query.",
+        "Adicione um índice para acelerar a consulta de pedidos por cliente.",
+        "Write a SQL query that returns monthly revenue per product.",
+        "Escreva uma consulta SQL que retorne o faturamento mensal por produto.",
+    ]),
+    t("feature", 2, "module", 3, 1, 3, "database-engineer", "review", "wd", [
+        "Write a migration that drops the legacy column from the production orders table.",
+        "Escreva uma migration que remove a coluna legada da tabela de pedidos de produção.",
+        "Run a data migration on the production database to backfill user emails.",
+        "Rode uma migração de dados no banco de produção para preencher os e-mails dos usuários.",
+    ], true),
+    t("feature", 3, "multi_module", 2, 1, 3, "backend-engineer", "integration_test", "wtsd", [
+        "Implement multi-tenant support in {module}: every query must be scoped by tenant.",
+        "Implemente suporte multi-tenant no {module}: toda consulta tem que filtrar pelo tenant.",
+    ]),
+    t("feature", 3, "repository", 2, 1, 3, "devops-engineer", "integration_test", "wsg", [
+        "Set up a release pipeline that builds, tests and publishes the packages on tag.",
+        "Monte um pipeline de release que compila, testa e publica os pacotes quando criar uma tag.",
+        "Deploy the {module} service to Kubernetes with zero downtime.",
+        "Faça o deploy do serviço {module} no Kubernetes sem downtime.",
+    ]),
+    t("feature", 4, "multi_module", 2, 2, 4, "software-architect", "integration_test", "wts", [
+        "Build a plugin system so third parties can add commands at runtime.",
+        "Crie um sistema de plugins para terceiros adicionarem comandos em tempo de execução.",
+        "Implement a CRDT-based collaborative editing engine.",
+        "Implemente um motor de edição colaborativa baseado em CRDT.",
+    ]),
+    t("test", 0, "single_file", 0, 0, 0, "qa-engineer", "unit_test", "t", [
+        "Run the tests and tell me what fails.",
+        "Rode os testes e me diga o que falha.",
+        "Run the type checker.",
+        "Rode a checagem de tipos.",
+    ]),
+    t("test", 1, "single_file", 0, 0, 1, "qa-engineer", "unit_test", "wt", [
+        "Write unit tests for {fn} in {file}.",
+        "Escreva testes unitários para {fn} em {file}.",
+        "Add a regression test for the empty input case of {fn}.",
+        "Adicione um teste de regressão para a entrada vazia de {fn}.",
+    ]),
+    t("test", 2, "module", 1, 1, 2, "qa-engineer", "integration_test", "wts", [
+        "The test suite of {module} is flaky; stabilize the flaky tests.",
+        "A suíte de testes do {module} está instável; estabilize os testes flaky.",
+        "Add integration tests for the checkout flow.",
+        "Adicione testes de integração para o fluxo de checkout.",
+    ]),
+    t("architecture", 3, "multi_module", 1, 2, 3, "software-architect", "review", "", [
+        "Propose an event-driven architecture for the notification system of our services.",
+        "Proponha uma arquitetura orientada a eventos para o sistema de notificações dos nossos serviços.",
+        "Evaluate whether we should split {module} into microservices.",
+        "Avalie se devemos dividir o {module} em microsserviços.",
+    ]),
+    t("architecture", 4, "multi_repository", 2, 2, 4, "software-architect", "review", "W", [
+        "Design a distributed cache invalidation strategy across our three regions with strong consistency.",
+        "Desenhe uma estratégia de invalidação de cache distribuído entre nossas três regiões com consistência forte.",
+        "Analyze the concurrency model of the system and propose how to remove the global lock.",
+        "Analise o modelo de concorrência do sistema e proponha como remover o lock global.",
+    ]),
+    t("security", 3, "module", 3, 1, 3, "security-engineer", "review", "wt", [
+        "Fix the SQL injection in the search endpoint of {module}.",
+        "Corrija o SQL injection no endpoint de busca do {module}.",
+        "The password reset token never expires; fix it.",
+        "O token de redefinição de senha nunca expira; corrija.",
+    ], true),
+    t("security", 3, "module", 3, 1, 3, "security-engineer", "review", "", [
+        "Review the authentication middleware for vulnerabilities.",
+        "Revise o middleware de autenticação procurando vulnerabilidades.",
+        "Audit how API keys are stored in {module}.",
+        "Audite como as chaves de API são guardadas no {module}.",
+    ], true),
+    t("security", 4, "multi_module", 3, 1, 4, "security-engineer", "review", "wtd", [
+        "Implement OAuth2 login with refresh token rotation.",
+        "Implemente login OAuth2 com rotação de refresh token.",
+        "Add payment processing with Stripe webhooks and idempotency.",
+        "Adicione processamento de pagamentos com webhooks do Stripe e idempotência.",
+    ], true),
+    t("documentation", 1, "single_file", 0, 0, 1, "software-engineer", "syntax", "w", [
+        "Update the README with the new installation steps.",
+        "Atualize o README com os novos passos de instalação.",
+        "Add doc comments to the public functions of {file}.",
+        "Escreva a documentação da função {fn}.",
+    ]),
+    t("documentation", 2, "module", 0, 1, 2, "software-architect", "syntax", "w", [
+        "Write an ADR documenting why we chose {tech} for the job queue.",
+        "Escreva um ADR documentando por que escolhemos {tech} para a fila de jobs.",
+    ]),
+    t("research", 1, "single_file", 0, 1, 2, "researcher", "none", "W", [
+        "Compare {tech} and {tech2} for our use case.",
+        "Compare {tech} e {tech2} para o nosso caso.",
+        "What changed in the latest version of {tech}?",
+        "Pesquise as novidades da última versão do {tech}.",
+    ]),
+    t("research", 2, "repository", 0, 1, 2, "researcher", "none", "", [
+        "Map every place where {module} writes to disk.",
+        "Mapeie todos os lugares onde o {module} grava em disco.",
+        "List the modules that would be affected by changing the signature of {fn}.",
+        "Liste os módulos afetados se mudarmos a assinatura de {fn}.",
+    ]),
+];
+const SLOTS = {
+    file: [
+        "src/utils/date.ts",
+        "app/models/user.py",
+        "src/main/java/OrderService.java",
+        "lib/cache.go",
+        "src/components/Header.tsx",
+        "internal/auth/token.go",
+        "src/api/routes.ts",
+        "services/billing.rb",
+    ],
+    fn: ["parseDate", "calculateTotal", "getUser", "formatPrice", "loadConfig", "mergeOptions", "validateInput"],
+    module: ["billing", "auth", "orders", "notifications", "search", "sync", "scheduler", "gateway"],
+    var: ["data", "tmp", "res", "cfg"],
+    var2: ["payload", "buffer", "response", "config"],
+    test: ["should refresh token", "renders header", "handles retries"],
+    tech: ["Redis", "Kafka", "RabbitMQ", "PostgreSQL", "Stripe", "Sentry"],
+};
+/** Optional stack mention appended to a request; it varies the text without changing the labels. */
+const STACKS = ["TypeScript", "Python", "Go", "Java", "Kotlin", "Swift", "Ruby", "Rust"];
+const STACK_FORMATS = ["", " ({stack})", " [{stack}]", " - {stack}"];
+/** Tier from the task profile: the cheapest tier for its difficulty, raised for risk and security. */
+export function seedTier(template) {
+    const difficulty = Math.max(template.complexity, template.reasoning);
+    let index = difficulty <= 1 ? 0 : difficulty - 1;
+    if (template.risk >= 2 || template.security)
+        index = Math.max(index, 2);
+    return CAPABILITY_TIERS[Math.min(index, CAPABILITY_TIERS.length - 1)];
+}
+/** Deterministic PRNG (mulberry32), so the same count yields the same exercises. */
+function random(seed) {
+    let state = seed;
+    return () => {
+        state = (state + 0x6d2b79f5) | 0;
+        let value = Math.imul(state ^ (state >>> 15), 1 | state);
+        value = (value + Math.imul(value ^ (value >>> 7), 61 | value)) ^ value;
+        return ((value ^ (value >>> 14)) >>> 0) / 4294967296;
+    };
+}
+/**
+ * Generates up to `count` distinct exercises in laya-trainer's dataset format, the same number per
+ * template, so templates with many slot combinations do not crowd out the others.
+ */
+export function generateSeed(count, seed = 2026) {
+    const next = random(seed);
+    const pick = (values) => values[Math.floor(next() * values.length)];
+    const rows = [];
+    const seen = new Set();
+    const quota = Math.ceil(count / TEMPLATES.length);
+    const produced = new Map();
+    for (let attempt = 0; rows.length < count && attempt < count * 50; attempt++) {
+        const template = TEMPLATES[attempt % TEMPLATES.length];
+        if ((produced.get(template) ?? 0) >= quota)
+            continue;
+        const module = pick(SLOTS.module);
+        const tech = pick(SLOTS.tech);
+        const stack = pick(STACK_FORMATS).replace("{stack}", pick(STACKS));
+        const request = pick(template.texts).replace(/\{(\w+)\}/g, (_, slot) => {
+            if (slot === "module")
+                return module;
+            if (slot === "module2")
+                return pick(SLOTS.module.filter((value) => value !== module));
+            if (slot === "tech")
+                return tech;
+            if (slot === "tech2")
+                return pick(SLOTS.tech.filter((value) => value !== tech));
+            return pick(SLOTS[slot]);
+        }) + stack;
+        if (seen.has(request))
+            continue;
+        seen.add(request);
+        produced.set(template, (produced.get(template) ?? 0) + 1);
+        const tools = Object.fromEntries(Object.entries(TOOL_FLAGS).map(([flag, id]) => [id, template.tools.includes(flag)]));
+        rows.push({
+            state: { request },
+            expected: {
+                task_type: template.type,
+                complexity: template.complexity,
+                scope: template.scope,
+                risk: template.risk,
+                ambiguity: template.ambiguity,
+                reasoning_requirement: template.reasoning,
+                capability_tier: seedTier(template),
+                reasoning_effort: REASONING_EFFORTS[template.reasoning],
+                agent: template.agent,
+                validation_level: template.validation,
+                ...tools,
+                security_sensitive: template.security ?? false,
+            },
+            source: "synthetic",
+        });
+    }
+    return rows;
+}
+//# sourceMappingURL=seed.js.map