@tangle-network/agent-eval 0.123.8 → 0.125.0

package/dist/chunk-MAX3TN3C.js

@@ -0,0 +1,249 @@
+// src/rollout/schema.ts
+var ROLLOUT_SCHEMA = "tangle.rollout.v1";
+var ROLLOUT_FORMAT = ROLLOUT_SCHEMA;
+var ROLLOUT_ROLES = [
+  "agent",
+  "supervisor",
+  "worker",
+  "proposer",
+  "judge",
+  "analyst"
+];
+var ROLLOUT_SPLITS = [
+  "search",
+  "dev",
+  "holdout",
+  "canary",
+  "train"
+];
+var TRAINABLE_SPLITS = ["search", "train"];
+function isTrainableSplit(split) {
+  return TRAINABLE_SPLITS.includes(split);
+}
+var ROLLOUT_CAPTURES = ["mint", "settle-time", "backfill"];
+var CHAT_ROLES = ["system", "user", "assistant", "tool"];
+var isRecord = (v) => typeof v === "object" && v !== null && !Array.isArray(v);
+var isNumberOrNull = (v) => v === null || typeof v === "number";
+var isStringOrNull = (v) => v === null || typeof v === "string";
+var isIntegerOrNull = (v) => v === null || Number.isInteger(v);
+function validateChatMessage(value, path, errors) {
+  if (!isRecord(value)) {
+    errors.push(`${path}: not an object`);
+    return;
+  }
+  if (!CHAT_ROLES.includes(value.role))
+    errors.push(`${path}.role: invalid role ${String(value.role)}`);
+  if (!isStringOrNull(value.content)) errors.push(`${path}.content: must be string|null`);
+  if (value.reasoning_content !== void 0 && typeof value.reasoning_content !== "string") {
+    errors.push(`${path}.reasoning_content: must be string when present`);
+  }
+  if (value.tool_call_id !== void 0 && typeof value.tool_call_id !== "string") {
+    errors.push(`${path}.tool_call_id: must be string when present`);
+  }
+  if (value.role === "tool" && typeof value.tool_call_id !== "string") {
+    errors.push(`${path}.tool_call_id: required on role:"tool"`);
+  }
+  if (value.tool_calls !== void 0) {
+    if (!Array.isArray(value.tool_calls)) {
+      errors.push(`${path}.tool_calls: must be an array when present`);
+    } else {
+      value.tool_calls.forEach((call, i) => {
+        if (!isRecord(call) || typeof call.id !== "string" || call.type !== "function") {
+          errors.push(`${path}.tool_calls[${i}]: must be {id, type:"function", function}`);
+          return;
+        }
+        const fn = call.function;
+        if (!isRecord(fn) || typeof fn.name !== "string" || typeof fn.arguments !== "string") {
+          errors.push(
+            `${path}.tool_calls[${i}].function: must be {name: string, arguments: string}`
+          );
+        }
+      });
+    }
+  }
+}
+function validateSection(value, path, fields, errors) {
+  if (!isRecord(value)) {
+    errors.push(`${path}: not an object`);
+    return;
+  }
+  for (const [name, check, expect] of fields) {
+    if (!check(value[name])) errors.push(`${path}.${name}: expected ${expect}`);
+  }
+}
+function validateRolloutLine(value) {
+  const errors = [];
+  if (!isRecord(value)) return ["line: not an object"];
+  if (value.schema !== ROLLOUT_SCHEMA) errors.push(`schema: expected "${ROLLOUT_SCHEMA}"`);
+  if (typeof value.rollout_id !== "string" || value.rollout_id.length === 0)
+    errors.push("rollout_id: expected non-empty string");
+  if (!isStringOrNull(value.parent_rollout_id))
+    errors.push("parent_rollout_id: expected string|null");
+  if (typeof value.run_id !== "string" || value.run_id.length === 0)
+    errors.push("run_id: expected non-empty string");
+  if (value.experiment_id !== void 0 && !isStringOrNull(value.experiment_id)) {
+    errors.push("experiment_id: expected string|null when present");
+  }
+  if (value.candidate_id !== void 0 && !isStringOrNull(value.candidate_id)) {
+    errors.push("candidate_id: expected string|null when present");
+  }
+  if (!isIntegerOrNull(value.generation)) errors.push("generation: expected integer|null");
+  if (!isIntegerOrNull(value.candidate_index)) errors.push("candidate_index: expected integer|null");
+  if (!ROLLOUT_ROLES.includes(value.role))
+    errors.push(`role: invalid role ${String(value.role)}`);
+  validateSection(
+    value.task,
+    "task",
+    [
+      ["suite", (v) => typeof v === "string" && v.length > 0, "non-empty string"],
+      ["instance_id", (v) => typeof v === "string" && v.length > 0, "non-empty string"],
+      [
+        "split",
+        (v) => ROLLOUT_SPLITS.includes(v),
+        `one of ${ROLLOUT_SPLITS.join("|")}`
+      ],
+      ["seed", isNumberOrNull, "number|null"],
+      ["rep", (v) => Number.isInteger(v), "integer"]
+    ],
+    errors
+  );
+  validateSection(
+    value.policy,
+    "policy",
+    [
+      ["harness", isStringOrNull, "string|null"],
+      ["harness_version", isStringOrNull, "string|null"],
+      ["model", isStringOrNull, "string|null"],
+      ["provider", isStringOrNull, "string|null"],
+      ["profile_commit", isStringOrNull, "string|null"],
+      ["sampling", (v) => v === null || isRecord(v), "object|null"]
+    ],
+    errors
+  );
+  if (isRecord(value.policy)) {
+    for (const key of ["prompt_hash", "config_hash", "agent_profile_cell_id"]) {
+      if (value.policy[key] !== void 0 && !isStringOrNull(value.policy[key])) {
+        errors.push(`policy.${key}: expected string|null when present`);
+      }
+    }
+  }
+  if (!Array.isArray(value.messages)) {
+    errors.push("messages: expected array");
+  } else {
+    for (const [i, m] of value.messages.entries()) validateChatMessage(m, `messages[${i}]`, errors);
+  }
+  if (!Array.isArray(value.tool_defs)) {
+    errors.push("tool_defs: expected array");
+  } else {
+    value.tool_defs.forEach((d, i) => {
+      if (!isRecord(d) || d.type !== "function" || !isRecord(d.function) || typeof d.function.name !== "string") {
+        errors.push(`tool_defs[${i}]: must be {type:"function", function:{name}}`);
+      }
+    });
+  }
+  if (value.steps !== void 0) {
+    if (!Array.isArray(value.steps)) {
+      errors.push("steps: expected array when present");
+    } else {
+      value.steps.forEach((s, i) => {
+        if (!isRecord(s) || typeof s.kind !== "string" || typeof s.name !== "string") {
+          errors.push(`steps[${i}]: must be {kind: string, name: string, \u2026}`);
+        }
+      });
+    }
+  }
+  validateSection(
+    value.outcome,
+    "outcome",
+    [
+      ["reward", isNumberOrNull, "number|null"],
+      ["reward_source", isStringOrNull, "string|null"],
+      ["metrics", isRecord, "object"],
+      ["is_completed", (v) => typeof v === "boolean", "boolean"],
+      ["is_truncated", (v) => typeof v === "boolean", "boolean"],
+      ["error", isStringOrNull, "string|null"]
+    ],
+    errors
+  );
+  if (isRecord(value.outcome)) {
+    if (!("verdict" in value.outcome)) errors.push("outcome.verdict: field required (may be null)");
+    if (value.outcome.realness_gated !== void 0 && typeof value.outcome.realness_gated !== "boolean") {
+      errors.push("outcome.realness_gated: expected boolean when present");
+    }
+  }
+  validateSection(
+    value.cost,
+    "cost",
+    [
+      ["usd", isNumberOrNull, "number|null"],
+      ["tokens_in", isNumberOrNull, "number|null"],
+      ["tokens_out", isNumberOrNull, "number|null"],
+      ["tokens_reasoning", isNumberOrNull, "number|null"],
+      ["cache_read", isNumberOrNull, "number|null"],
+      ["cache_write", isNumberOrNull, "number|null"],
+      ["wall_s", isNumberOrNull, "number|null"]
+    ],
+    errors
+  );
+  validateSection(
+    value.artifacts,
+    "artifacts",
+    [
+      ["patch_path", isStringOrNull, "string|null"],
+      ["run_dir", isStringOrNull, "string|null"],
+      ["transcript_ref", isStringOrNull, "string|null"]
+    ],
+    errors
+  );
+  validateSection(
+    value.provenance,
+    "provenance",
+    [
+      [
+        "captured_at",
+        (v) => typeof v === "string" && !Number.isNaN(Date.parse(v)),
+        "ISO-8601 timestamp"
+      ],
+      [
+        "capture",
+        (v) => ROLLOUT_CAPTURES.includes(v),
+        `one of ${ROLLOUT_CAPTURES.join("|")}`
+      ]
+    ],
+    errors
+  );
+  if (isRecord(value.provenance) && value.provenance.gap !== void 0 && typeof value.provenance.gap !== "string") {
+    errors.push("provenance.gap: must be string when present");
+  }
+  if (Array.isArray(value.messages) && isRecord(value.provenance)) {
+    if (value.messages.length === 0 && typeof value.provenance.gap !== "string") {
+      errors.push("provenance.gap: required when messages is empty");
+    }
+  }
+  return errors;
+}
+function assertRolloutLine(value, context = "rollout line") {
+  const errors = validateRolloutLine(value);
+  if (errors.length > 0) {
+    throw new Error(`invalid ${context}:
+  ${errors.join("\n  ")}`);
+  }
+}
+function isRolloutLine(value) {
+  return validateRolloutLine(value).length === 0;
+}
+
+export {
+  ROLLOUT_SCHEMA,
+  ROLLOUT_FORMAT,
+  ROLLOUT_ROLES,
+  ROLLOUT_SPLITS,
+  TRAINABLE_SPLITS,
+  isTrainableSplit,
+  ROLLOUT_CAPTURES,
+  CHAT_ROLES,
+  validateRolloutLine,
+  assertRolloutLine,
+  isRolloutLine
+};
+//# sourceMappingURL=chunk-MAX3TN3C.js.map

package/dist/chunk-MAX3TN3C.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/rollout/schema.ts"],"sourcesContent":["/**\n * `tangle.rollout.v1` — THE canonical rollout serialization, owned by\n * agent-eval. One JSONL line per agent invocation (a solo eval run, a\n * supervisor episode, a worker session, a proposer shot, a judge call, an\n * analyst pass), labeled with its task/split coordinates and a single\n * scalar reward, carrying the FULL message transcript inline.\n *\n * This schema is the reconciliation of two prior producers:\n *   - agent-eval's RunRecord-joined rollout rows (PR #410): identity,\n *     provenance hashes, the realness gate travelling into the reward,\n *     trace-derived steps.\n *   - the bench rollout-ledger (agent-runtime PR #591): the wire shape —\n *     role, task.split/rep, parent_rollout_id, policy provenance, capture\n *     provenance, inline canonical chat-with-tools messages.\n * Where the two conflicted, RunRecord-derived semantics won; the wire\n * field names follow the ledger (snake_case). See `docs/rollout.md` for\n * the field-by-field decision table.\n *\n * Messages are inlined — never referenced — because every harness store a\n * rollout can be recovered from is mutable or garbage-collected. A line\n * must stay a complete training/eval example on its own.\n *\n * `outcome.reward` is THE single scalar (null = no verdict exists — a\n * labeled gap, never 0). `outcome.realness_gated` is the anti-Goodhart\n * flag: a gated line must never export as a positive training example.\n */\n\nexport const ROLLOUT_SCHEMA = 'tangle.rollout.v1'\n/** @deprecated alias kept for consumers of the pre-unification constant name. */\nexport const ROLLOUT_FORMAT = ROLLOUT_SCHEMA\n\n/** `agent` = a solo evaluation run (no multi-agent topology). */\nexport type RolloutRole = 'agent' | 'supervisor' | 'worker' | 'proposer' | 'judge' | 'analyst'\nexport const ROLLOUT_ROLES: readonly RolloutRole[] = [\n  'agent',\n  'supervisor',\n  'worker',\n  'proposer',\n  'judge',\n  'analyst',\n]\n\n/**\n * Split vocabulary follows `RunRecord.splitTag` ('search' is the pool the\n * optimizer may read — the trainable split), extended with the ledger's\n * 'canary'. 'train' is a legacy alias for 'search' emitted by\n * pre-unification ledgers; it validates and counts as trainable, but new\n * producers must emit 'search'.\n */\nexport type RolloutSplit = 'search' | 'dev' | 'holdout' | 'canary' | 'train'\nexport const ROLLOUT_SPLITS: readonly RolloutSplit[] = [\n  'search',\n  'dev',\n  'holdout',\n  'canary',\n  'train',\n]\n/** Splits that may ship in training exports. Everything else is fail-closed excluded. */\nexport const TRAINABLE_SPLITS: readonly RolloutSplit[] = ['search', 'train']\n\nexport function isTrainableSplit(split: RolloutSplit): boolean {\n  return TRAINABLE_SPLITS.includes(split)\n}\n\n/** 'mint' = joined live from RunRecord + trace by `mintRolloutRows`. */\nexport type RolloutCapture = 'mint' | 'settle-time' | 'backfill'\nexport const ROLLOUT_CAPTURES: readonly RolloutCapture[] = ['mint', 'settle-time', 'backfill']\n\n// ---------------------------------------------------------------------------\n// Canonical message format — OpenAI chat-with-tools, full fidelity.\n// ---------------------------------------------------------------------------\n\nexport type ChatRole = 'system' | 'user' | 'assistant' | 'tool'\nexport const CHAT_ROLES: readonly ChatRole[] = ['system', 'user', 'assistant', 'tool']\n\nexport interface ChatToolCall {\n  id: string\n  type: 'function'\n  function: {\n    name: string\n    /** JSON-encoded argument object, exactly as the model emitted it. */\n    arguments: string\n  }\n}\n\nexport interface ChatMessage {\n  role: ChatRole\n  content: string | null\n  /** Reasoning/thinking channel where the harness captured it (full fidelity). */\n  reasoning_content?: string\n  tool_calls?: ChatToolCall[]\n  /** Required on role:\"tool\" — the ChatToolCall this result answers. */\n  tool_call_id?: string\n  name?: string\n}\n\nexport interface ToolDef {\n  type: 'function'\n  function: {\n    name: string\n    description?: string\n    parameters?: Record<string, unknown>\n  }\n}\n\n/**\n * Compact trace-span projection (llm/tool step) carried alongside the\n * conversation when the line was minted from a trace. Optional: lines\n * recovered from harness stores have no span structure.\n */\nexport interface RolloutStep {\n  kind: string\n  name: string\n  /** llm: last-message summary · tool: stringified args. Scrubbed. */\n  input?: string\n  /** llm: output text · tool: stringified result. Scrubbed. */\n  output?: string\n  status?: 'ok' | 'error'\n  durationMs?: number\n}\n\n// ---------------------------------------------------------------------------\n// Ledger line sections.\n// ---------------------------------------------------------------------------\n\nexport interface RolloutTask {\n  /** Benchmark/suite id (e.g. \"swe-bench-verified\") or the experiment id. */\n  suite: string\n  instance_id: string\n  split: RolloutSplit\n  /** Sampling seed the campaign pinned; null = not recorded. */\n  seed: number | null\n  /** Replicate index (0-based). */\n  rep: number\n}\n\nexport interface RolloutPolicy {\n  /** Harness that drove the invocation (e.g. \"opencode\", \"claude\", \"pi-loops\"). */\n  harness: string | null\n  harness_version: string | null\n  model: string | null\n  provider: string | null\n  /** Commit of the agent profile / candidate under evaluation. */\n  profile_commit: string | null\n  /** sha256 of the effective prompt (post-steering), when recorded. */\n  prompt_hash?: string | null\n  /** sha256 of the effective run config, when recorded. */\n  config_hash?: string | null\n  /** Canonical agent-profile cell identity, when the run carries one. */\n  agent_profile_cell_id?: string | null\n  /** Sampling params (temperature, top_p, max_tokens…); null = not recorded. */\n  sampling: Record<string, unknown> | null\n}\n\nexport interface RolloutOutcome {\n  /**\n   * THE single scalar training signal — the official verdict.\n   * null = no verdict exists for this invocation (a labeled gap, never 0).\n   */\n  reward: number | null\n  /** Where the reward came from (judge id; \"/inherited\" = parent episode's). */\n  reward_source: string | null\n  /** Raw judge verdict record, verbatim. */\n  verdict: unknown\n  /** Everything that is NOT the scalar reward. */\n  metrics: Record<string, unknown>\n  is_completed: boolean\n  is_truncated: boolean\n  error: string | null\n  /**\n   * Anti-Goodhart flag from `RunRecord.outcome.realness.gated`: the run\n   * faked its success signal. Reward is forced to 0 at mint time and the\n   * line never qualifies for SFT. Optional on the wire (absent = false)\n   * so pre-unification ledgers stay readable.\n   */\n  realness_gated?: boolean\n}\n\nexport interface RolloutCostBlock {\n  usd: number | null\n  tokens_in: number | null\n  tokens_out: number | null\n  tokens_reasoning: number | null\n  cache_read: number | null\n  cache_write: number | null\n  wall_s: number | null\n}\n\nexport interface RolloutArtifacts {\n  patch_path: string | null\n  run_dir: string | null\n  /** Source-of-truth transcript pointer (session id / jsonl path) for audit. */\n  transcript_ref: string | null\n}\n\nexport interface RolloutProvenance {\n  captured_at: string\n  capture: RolloutCapture\n  /** Present on gap lines: why `messages` could not be recovered. */\n  gap?: string\n}\n\nexport interface RolloutLine {\n  schema: typeof ROLLOUT_SCHEMA\n  rollout_id: string\n  /** Spawning invocation within the same episode (worker → supervisor). */\n  parent_rollout_id: string | null\n  run_id: string\n  /** Logical experiment grouping from `RunRecord.experimentId`. Optional on\n   *  the wire (pre-unification ledgers lack it); null = not recorded. */\n  experiment_id?: string | null\n  /** Stable candidate identity from `RunRecord.candidateId`; null = not recorded. */\n  candidate_id?: string | null\n  /** Improvement-loop generation (-1 = baseline); null = not an improvement loop. */\n  generation: number | null\n  /** Improvement-loop candidate index (-1 = baseline); null = not an improvement loop. */\n  candidate_index: number | null\n  role: RolloutRole\n  task: RolloutTask\n  policy: RolloutPolicy\n  /** Full transcript, inline. [] = gap line (see provenance.gap). */\n  messages: ChatMessage[]\n  tool_defs: ToolDef[]\n  /** Trace-span projections, when minted from a trace. */\n  steps?: RolloutStep[]\n  outcome: RolloutOutcome\n  cost: RolloutCostBlock\n  artifacts: RolloutArtifacts\n  provenance: RolloutProvenance\n}\n\n// ---------------------------------------------------------------------------\n// Validation — pure TS, no runtime schema dependency, mirroring the\n// run-record validator's fail-loud discipline. Returns [] when the value\n// is a valid RolloutLine; otherwise one dotted-path error per defect.\n// ---------------------------------------------------------------------------\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n  typeof v === 'object' && v !== null && !Array.isArray(v)\n\nconst isNumberOrNull = (v: unknown): boolean => v === null || typeof v === 'number'\nconst isStringOrNull = (v: unknown): boolean => v === null || typeof v === 'string'\nconst isIntegerOrNull = (v: unknown): boolean => v === null || Number.isInteger(v)\n\nfunction validateChatMessage(value: unknown, path: string, errors: string[]): void {\n  if (!isRecord(value)) {\n    errors.push(`${path}: not an object`)\n    return\n  }\n  if (!CHAT_ROLES.includes(value.role as ChatRole))\n    errors.push(`${path}.role: invalid role ${String(value.role)}`)\n  if (!isStringOrNull(value.content)) errors.push(`${path}.content: must be string|null`)\n  if (value.reasoning_content !== undefined && typeof value.reasoning_content !== 'string') {\n    errors.push(`${path}.reasoning_content: must be string when present`)\n  }\n  if (value.tool_call_id !== undefined && typeof value.tool_call_id !== 'string') {\n    errors.push(`${path}.tool_call_id: must be string when present`)\n  }\n  if (value.role === 'tool' && typeof value.tool_call_id !== 'string') {\n    errors.push(`${path}.tool_call_id: required on role:\"tool\"`)\n  }\n  if (value.tool_calls !== undefined) {\n    if (!Array.isArray(value.tool_calls)) {\n      errors.push(`${path}.tool_calls: must be an array when present`)\n    } else {\n      value.tool_calls.forEach((call, i) => {\n        if (!isRecord(call) || typeof call.id !== 'string' || call.type !== 'function') {\n          errors.push(`${path}.tool_calls[${i}]: must be {id, type:\"function\", function}`)\n          return\n        }\n        const fn = call.function\n        if (!isRecord(fn) || typeof fn.name !== 'string' || typeof fn.arguments !== 'string') {\n          errors.push(\n            `${path}.tool_calls[${i}].function: must be {name: string, arguments: string}`,\n          )\n        }\n      })\n    }\n  }\n}\n\nfunction validateSection(\n  value: unknown,\n  path: string,\n  fields: Array<[name: string, check: (v: unknown) => boolean, expect: string]>,\n  errors: string[],\n): void {\n  if (!isRecord(value)) {\n    errors.push(`${path}: not an object`)\n    return\n  }\n  for (const [name, check, expect] of fields) {\n    if (!check(value[name])) errors.push(`${path}.${name}: expected ${expect}`)\n  }\n}\n\nexport function validateRolloutLine(value: unknown): string[] {\n  const errors: string[] = []\n  if (!isRecord(value)) return ['line: not an object']\n\n  if (value.schema !== ROLLOUT_SCHEMA) errors.push(`schema: expected \"${ROLLOUT_SCHEMA}\"`)\n  if (typeof value.rollout_id !== 'string' || value.rollout_id.length === 0)\n    errors.push('rollout_id: expected non-empty string')\n  if (!isStringOrNull(value.parent_rollout_id))\n    errors.push('parent_rollout_id: expected string|null')\n  if (typeof value.run_id !== 'string' || value.run_id.length === 0)\n    errors.push('run_id: expected non-empty string')\n  if (value.experiment_id !== undefined && !isStringOrNull(value.experiment_id)) {\n    errors.push('experiment_id: expected string|null when present')\n  }\n  if (value.candidate_id !== undefined && !isStringOrNull(value.candidate_id)) {\n    errors.push('candidate_id: expected string|null when present')\n  }\n  if (!isIntegerOrNull(value.generation)) errors.push('generation: expected integer|null')\n  if (!isIntegerOrNull(value.candidate_index)) errors.push('candidate_index: expected integer|null')\n  if (!ROLLOUT_ROLES.includes(value.role as RolloutRole))\n    errors.push(`role: invalid role ${String(value.role)}`)\n\n  validateSection(\n    value.task,\n    'task',\n    [\n      ['suite', (v) => typeof v === 'string' && v.length > 0, 'non-empty string'],\n      ['instance_id', (v) => typeof v === 'string' && v.length > 0, 'non-empty string'],\n      [\n        'split',\n        (v) => ROLLOUT_SPLITS.includes(v as RolloutSplit),\n        `one of ${ROLLOUT_SPLITS.join('|')}`,\n      ],\n      ['seed', isNumberOrNull, 'number|null'],\n      ['rep', (v) => Number.isInteger(v), 'integer'],\n    ],\n    errors,\n  )\n\n  validateSection(\n    value.policy,\n    'policy',\n    [\n      ['harness', isStringOrNull, 'string|null'],\n      ['harness_version', isStringOrNull, 'string|null'],\n      ['model', isStringOrNull, 'string|null'],\n      ['provider', isStringOrNull, 'string|null'],\n      ['profile_commit', isStringOrNull, 'string|null'],\n      ['sampling', (v) => v === null || isRecord(v), 'object|null'],\n    ],\n    errors,\n  )\n  if (isRecord(value.policy)) {\n    for (const key of ['prompt_hash', 'config_hash', 'agent_profile_cell_id'] as const) {\n      if (value.policy[key] !== undefined && !isStringOrNull(value.policy[key])) {\n        errors.push(`policy.${key}: expected string|null when present`)\n      }\n    }\n  }\n\n  if (!Array.isArray(value.messages)) {\n    errors.push('messages: expected array')\n  } else {\n    for (const [i, m] of value.messages.entries()) validateChatMessage(m, `messages[${i}]`, errors)\n  }\n\n  if (!Array.isArray(value.tool_defs)) {\n    errors.push('tool_defs: expected array')\n  } else {\n    value.tool_defs.forEach((d, i) => {\n      if (\n        !isRecord(d) ||\n        d.type !== 'function' ||\n        !isRecord(d.function) ||\n        typeof d.function.name !== 'string'\n      ) {\n        errors.push(`tool_defs[${i}]: must be {type:\"function\", function:{name}}`)\n      }\n    })\n  }\n\n  if (value.steps !== undefined) {\n    if (!Array.isArray(value.steps)) {\n      errors.push('steps: expected array when present')\n    } else {\n      value.steps.forEach((s, i) => {\n        if (!isRecord(s) || typeof s.kind !== 'string' || typeof s.name !== 'string') {\n          errors.push(`steps[${i}]: must be {kind: string, name: string, …}`)\n        }\n      })\n    }\n  }\n\n  validateSection(\n    value.outcome,\n    'outcome',\n    [\n      ['reward', isNumberOrNull, 'number|null'],\n      ['reward_source', isStringOrNull, 'string|null'],\n      ['metrics', isRecord, 'object'],\n      ['is_completed', (v) => typeof v === 'boolean', 'boolean'],\n      ['is_truncated', (v) => typeof v === 'boolean', 'boolean'],\n      ['error', isStringOrNull, 'string|null'],\n    ],\n    errors,\n  )\n  if (isRecord(value.outcome)) {\n    if (!('verdict' in value.outcome)) errors.push('outcome.verdict: field required (may be null)')\n    if (\n      value.outcome.realness_gated !== undefined &&\n      typeof value.outcome.realness_gated !== 'boolean'\n    ) {\n      errors.push('outcome.realness_gated: expected boolean when present')\n    }\n  }\n\n  validateSection(\n    value.cost,\n    'cost',\n    [\n      ['usd', isNumberOrNull, 'number|null'],\n      ['tokens_in', isNumberOrNull, 'number|null'],\n      ['tokens_out', isNumberOrNull, 'number|null'],\n      ['tokens_reasoning', isNumberOrNull, 'number|null'],\n      ['cache_read', isNumberOrNull, 'number|null'],\n      ['cache_write', isNumberOrNull, 'number|null'],\n      ['wall_s', isNumberOrNull, 'number|null'],\n    ],\n    errors,\n  )\n\n  validateSection(\n    value.artifacts,\n    'artifacts',\n    [\n      ['patch_path', isStringOrNull, 'string|null'],\n      ['run_dir', isStringOrNull, 'string|null'],\n      ['transcript_ref', isStringOrNull, 'string|null'],\n    ],\n    errors,\n  )\n\n  validateSection(\n    value.provenance,\n    'provenance',\n    [\n      [\n        'captured_at',\n        (v) => typeof v === 'string' && !Number.isNaN(Date.parse(v)),\n        'ISO-8601 timestamp',\n      ],\n      [\n        'capture',\n        (v) => ROLLOUT_CAPTURES.includes(v as RolloutCapture),\n        `one of ${ROLLOUT_CAPTURES.join('|')}`,\n      ],\n    ],\n    errors,\n  )\n  if (\n    isRecord(value.provenance) &&\n    value.provenance.gap !== undefined &&\n    typeof value.provenance.gap !== 'string'\n  ) {\n    errors.push('provenance.gap: must be string when present')\n  }\n\n  // A gap line must say WHY it is a gap; a full line must not carry a gap note.\n  if (Array.isArray(value.messages) && isRecord(value.provenance)) {\n    if (value.messages.length === 0 && typeof value.provenance.gap !== 'string') {\n      errors.push('provenance.gap: required when messages is empty')\n    }\n  }\n\n  return errors\n}\n\nexport function assertRolloutLine(\n  value: unknown,\n  context = 'rollout line',\n): asserts value is RolloutLine {\n  const errors = validateRolloutLine(value)\n  if (errors.length > 0) {\n    throw new Error(`invalid ${context}:\\n  ${errors.join('\\n  ')}`)\n  }\n}\n\nexport function isRolloutLine(value: unknown): value is RolloutLine {\n  return validateRolloutLine(value).length === 0\n}\n"],"mappings":";AA2BO,IAAM,iBAAiB;AAEvB,IAAM,iBAAiB;AAIvB,IAAM,gBAAwC;AAAA,EACnD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAUO,IAAM,iBAA0C;AAAA,EACrD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEO,IAAM,mBAA4C,CAAC,UAAU,OAAO;AAEpE,SAAS,iBAAiB,OAA8B;AAC7D,SAAO,iBAAiB,SAAS,KAAK;AACxC;AAIO,IAAM,mBAA8C,CAAC,QAAQ,eAAe,UAAU;AAOtF,IAAM,aAAkC,CAAC,UAAU,QAAQ,aAAa,MAAM;AAoKrF,IAAM,WAAW,CAAC,MAChB,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;AAEzD,IAAM,iBAAiB,CAAC,MAAwB,MAAM,QAAQ,OAAO,MAAM;AAC3E,IAAM,iBAAiB,CAAC,MAAwB,MAAM,QAAQ,OAAO,MAAM;AAC3E,IAAM,kBAAkB,CAAC,MAAwB,MAAM,QAAQ,OAAO,UAAU,CAAC;AAEjF,SAAS,oBAAoB,OAAgB,MAAc,QAAwB;AACjF,MAAI,CAAC,SAAS,KAAK,GAAG;AACpB,WAAO,KAAK,GAAG,IAAI,iBAAiB;AACpC;AAAA,EACF;AACA,MAAI,CAAC,WAAW,SAAS,MAAM,IAAgB;AAC7C,WAAO,KAAK,GAAG,IAAI,uBAAuB,OAAO,MAAM,IAAI,CAAC,EAAE;AAChE,MAAI,CAAC,eAAe,MAAM,OAAO,EAAG,QAAO,KAAK,GAAG,IAAI,+BAA+B;AACtF,MAAI,MAAM,sBAAsB,UAAa,OAAO,MAAM,sBAAsB,UAAU;AACxF,WAAO,KAAK,GAAG,IAAI,iDAAiD;AAAA,EACtE;AACA,MAAI,MAAM,iBAAiB,UAAa,OAAO,MAAM,iBAAiB,UAAU;AAC9E,WAAO,KAAK,GAAG,IAAI,4CAA4C;AAAA,EACjE;AACA,MAAI,MAAM,SAAS,UAAU,OAAO,MAAM,iBAAiB,UAAU;AACnE,WAAO,KAAK,GAAG,IAAI,wCAAwC;AAAA,EAC7D;AACA,MAAI,MAAM,eAAe,QAAW;AAClC,QAAI,CAAC,MAAM,QAAQ,MAAM,UAAU,GAAG;AACpC,aAAO,KAAK,GAAG,IAAI,4CAA4C;AAAA,IACjE,OAAO;AACL,YAAM,WAAW,QAAQ,CAAC,MAAM,MAAM;AACpC,YAAI,CAAC,SAAS,IAAI,KAAK,OAAO,KAAK,OAAO,YAAY,KAAK,SAAS,YAAY;AAC9E,iBAAO,KAAK,GAAG,IAAI,eAAe,CAAC,4CAA4C;AAC/E;AAAA,QACF;AACA,cAAM,KAAK,KAAK;AAChB,YAAI,CAAC,SAAS,EAAE,KAAK,OAAO,GAAG,SAAS,YAAY,OAAO,GAAG,cAAc,UAAU;AACpF,iBAAO;AAAA,YACL,GAAG,IAAI,eAAe,CAAC;AAAA,UACzB;AAAA,QACF;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAEA,SAAS,gBACP,OACA,MACA,QACA,QACM;AACN,MAAI,CAAC,SAAS,KAAK,GAAG;AACpB,WAAO,KAAK,GAAG,IAAI,iBAAiB;AACpC;AAAA,EACF;AACA,aAAW,CAAC,MAAM,OAAO,MAAM,KAAK,QAAQ;AAC1C,QAAI,CAAC,MAAM,MAAM,IAAI,CAAC,EAAG,QAAO,KAAK,GAAG,IAAI,IAAI,IAAI,cAAc,MAAM,EAAE;AAAA,EAC5E;AACF;AAEO,SAAS,oBAAoB,OAA0B;AAC5D,QAAM,SAAmB,CAAC;AAC1B,MAAI,CAAC,SAAS,KAAK,EAAG,QAAO,CAAC,qBAAqB;AAEnD,MAAI,MAAM,WAAW,eAAgB,QAAO,KAAK,qBAAqB,cAAc,GAAG;AACvF,MAAI,OAAO,MAAM,eAAe,YAAY,MAAM,WAAW,WAAW;AACtE,WAAO,KAAK,uCAAuC;AACrD,MAAI,CAAC,eAAe,MAAM,iBAAiB;AACzC,WAAO,KAAK,yCAAyC;AACvD,MAAI,OAAO,MAAM,WAAW,YAAY,MAAM,OAAO,WAAW;AAC9D,WAAO,KAAK,mCAAmC;AACjD,MAAI,MAAM,kBAAkB,UAAa,CAAC,eAAe,MAAM,aAAa,GAAG;AAC7E,WAAO,KAAK,kDAAkD;AAAA,EAChE;AACA,MAAI,MAAM,iBAAiB,UAAa,CAAC,eAAe,MAAM,YAAY,GAAG;AAC3E,WAAO,KAAK,iDAAiD;AAAA,EAC/D;AACA,MAAI,CAAC,gBAAgB,MAAM,UAAU,EAAG,QAAO,KAAK,mCAAmC;AACvF,MAAI,CAAC,gBAAgB,MAAM,eAAe,EAAG,QAAO,KAAK,wCAAwC;AACjG,MAAI,CAAC,cAAc,SAAS,MAAM,IAAmB;AACnD,WAAO,KAAK,sBAAsB,OAAO,MAAM,IAAI,CAAC,EAAE;AAExD;AAAA,IACE,MAAM;AAAA,IACN;AAAA,IACA;AAAA,MACE,CAAC,SAAS,CAAC,MAAM,OAAO,MAAM,YAAY,EAAE,SAAS,GAAG,kBAAkB;AAAA,MAC1E,CAAC,eAAe,CAAC,MAAM,OAAO,MAAM,YAAY,EAAE,SAAS,GAAG,kBAAkB;AAAA,MAChF;AAAA,QACE;AAAA,QACA,CAAC,MAAM,eAAe,SAAS,CAAiB;AAAA,QAChD,UAAU,eAAe,KAAK,GAAG,CAAC;AAAA,MACpC;AAAA,MACA,CAAC,QAAQ,gBAAgB,aAAa;AAAA,MACtC,CAAC,OAAO,CAAC,MAAM,OAAO,UAAU,CAAC,GAAG,SAAS;AAAA,IAC/C;AAAA,IACA;AAAA,EACF;AAEA;AAAA,IACE,MAAM;AAAA,IACN;AAAA,IACA;AAAA,MACE,CAAC,WAAW,gBAAgB,aAAa;AAAA,MACzC,CAAC,mBAAmB,gBAAgB,aAAa;AAAA,MACjD,CAAC,SAAS,gBAAgB,aAAa;AAAA,MACvC,CAAC,YAAY,gBAAgB,aAAa;AAAA,MAC1C,CAAC,kBAAkB,gBAAgB,aAAa;AAAA,MAChD,CAAC,YAAY,CAAC,MAAM,MAAM,QAAQ,SAAS,CAAC,GAAG,aAAa;AAAA,IAC9D;AAAA,IACA;AAAA,EACF;AACA,MAAI,SAAS,MAAM,MAAM,GAAG;AAC1B,eAAW,OAAO,CAAC,eAAe,eAAe,uBAAuB,GAAY;AAClF,UAAI,MAAM,OAAO,GAAG,MAAM,UAAa,CAAC,eAAe,MAAM,OAAO,GAAG,CAAC,GAAG;AACzE,eAAO,KAAK,UAAU,GAAG,qCAAqC;AAAA,MAChE;AAAA,IACF;AAAA,EACF;AAEA,MAAI,CAAC,MAAM,QAAQ,MAAM,QAAQ,GAAG;AAClC,WAAO,KAAK,0BAA0B;AAAA,EACxC,OAAO;AACL,eAAW,CAAC,GAAG,CAAC,KAAK,MAAM,SAAS,QAAQ,EAAG,qBAAoB,GAAG,YAAY,CAAC,KAAK,MAAM;AAAA,EAChG;AAEA,MAAI,CAAC,MAAM,QAAQ,MAAM,SAAS,GAAG;AACnC,WAAO,KAAK,2BAA2B;AAAA,EACzC,OAAO;AACL,UAAM,UAAU,QAAQ,CAAC,GAAG,MAAM;AAChC,UACE,CAAC,SAAS,CAAC,KACX,EAAE,SAAS,cACX,CAAC,SAAS,EAAE,QAAQ,KACpB,OAAO,EAAE,SAAS,SAAS,UAC3B;AACA,eAAO,KAAK,aAAa,CAAC,+CAA+C;AAAA,MAC3E;AAAA,IACF,CAAC;AAAA,EACH;AAEA,MAAI,MAAM,UAAU,QAAW;AAC7B,QAAI,CAAC,MAAM,QAAQ,MAAM,KAAK,GAAG;AAC/B,aAAO,KAAK,oCAAoC;AAAA,IAClD,OAAO;AACL,YAAM,MAAM,QAAQ,CAAC,GAAG,MAAM;AAC5B,YAAI,CAAC,SAAS,CAAC,KAAK,OAAO,EAAE,SAAS,YAAY,OAAO,EAAE,SAAS,UAAU;AAC5E,iBAAO,KAAK,SAAS,CAAC,iDAA4C;AAAA,QACpE;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAEA;AAAA,IACE,MAAM;AAAA,IACN;AAAA,IACA;AAAA,MACE,CAAC,UAAU,gBAAgB,aAAa;AAAA,MACxC,CAAC,iBAAiB,gBAAgB,aAAa;AAAA,MAC/C,CAAC,WAAW,UAAU,QAAQ;AAAA,MAC9B,CAAC,gBAAgB,CAAC,MAAM,OAAO,MAAM,WAAW,SAAS;AAAA,MACzD,CAAC,gBAAgB,CAAC,MAAM,OAAO,MAAM,WAAW,SAAS;AAAA,MACzD,CAAC,SAAS,gBAAgB,aAAa;AAAA,IACzC;AAAA,IACA;AAAA,EACF;AACA,MAAI,SAAS,MAAM,OAAO,GAAG;AAC3B,QAAI,EAAE,aAAa,MAAM,SAAU,QAAO,KAAK,+CAA+C;AAC9F,QACE,MAAM,QAAQ,mBAAmB,UACjC,OAAO,MAAM,QAAQ,mBAAmB,WACxC;AACA,aAAO,KAAK,uDAAuD;AAAA,IACrE;AAAA,EACF;AAEA;AAAA,IACE,MAAM;AAAA,IACN;AAAA,IACA;AAAA,MACE,CAAC,OAAO,gBAAgB,aAAa;AAAA,MACrC,CAAC,aAAa,gBAAgB,aAAa;AAAA,MAC3C,CAAC,cAAc,gBAAgB,aAAa;AAAA,MAC5C,CAAC,oBAAoB,gBAAgB,aAAa;AAAA,MAClD,CAAC,cAAc,gBAAgB,aAAa;AAAA,MAC5C,CAAC,eAAe,gBAAgB,aAAa;AAAA,MAC7C,CAAC,UAAU,gBAAgB,aAAa;AAAA,IAC1C;AAAA,IACA;AAAA,EACF;AAEA;AAAA,IACE,MAAM;AAAA,IACN;AAAA,IACA;AAAA,MACE,CAAC,cAAc,gBAAgB,aAAa;AAAA,MAC5C,CAAC,WAAW,gBAAgB,aAAa;AAAA,MACzC,CAAC,kBAAkB,gBAAgB,aAAa;AAAA,IAClD;AAAA,IACA;AAAA,EACF;AAEA;AAAA,IACE,MAAM;AAAA,IACN;AAAA,IACA;AAAA,MACE;AAAA,QACE;AAAA,QACA,CAAC,MAAM,OAAO,MAAM,YAAY,CAAC,OAAO,MAAM,KAAK,MAAM,CAAC,CAAC;AAAA,QAC3D;AAAA,MACF;AAAA,MACA;AAAA,QACE;AAAA,QACA,CAAC,MAAM,iBAAiB,SAAS,CAAmB;AAAA,QACpD,UAAU,iBAAiB,KAAK,GAAG,CAAC;AAAA,MACtC;AAAA,IACF;AAAA,IACA;AAAA,EACF;AACA,MACE,SAAS,MAAM,UAAU,KACzB,MAAM,WAAW,QAAQ,UACzB,OAAO,MAAM,WAAW,QAAQ,UAChC;AACA,WAAO,KAAK,6CAA6C;AAAA,EAC3D;AAGA,MAAI,MAAM,QAAQ,MAAM,QAAQ,KAAK,SAAS,MAAM,UAAU,GAAG;AAC/D,QAAI,MAAM,SAAS,WAAW,KAAK,OAAO,MAAM,WAAW,QAAQ,UAAU;AAC3E,aAAO,KAAK,iDAAiD;AAAA,IAC/D;AAAA,EACF;AAEA,SAAO;AACT;AAEO,SAAS,kBACd,OACA,UAAU,gBACoB;AAC9B,QAAM,SAAS,oBAAoB,KAAK;AACxC,MAAI,OAAO,SAAS,GAAG;AACrB,UAAM,IAAI,MAAM,WAAW,OAAO;AAAA,IAAQ,OAAO,KAAK,MAAM,CAAC,EAAE;AAAA,EACjE;AACF;AAEO,SAAS,cAAc,OAAsC;AAClE,SAAO,oBAAoB,KAAK,EAAE,WAAW;AAC/C;","names":[]}

package/dist/{chunk-QVGVJQMR.js → chunk-PMITBABE.js}

@@ -15,7 +15,7 @@ import {
   clamp01,
   mapConcurrent,
   validatePolicyEditCandidateRecord
-} from "./chunk-EEHHBAWH.js";
+} from "./chunk-DT7OXY3C.js";
 import {
   detectRewardHacking
 } from "./chunk-ARU2PZFM.js";
@@ -25,7 +25,7 @@ import {
   costReceiptFromLlmError,
   maximumChargeForLlmRequest,
   stripFencedJson
-} from "./chunk-PDHIOKRE.js";
+} from "./chunk-EQUK3RFS.js";
 import {
   pairedBootstrap,
   weightedComposite
@@ -2494,7 +2494,8 @@ function gepaProposer(opts) {
           ],
           jsonMode: true,
           temperature: opts.temperature ?? 0.7,
-          maxTokens
+          maxTokens,
+          thinking: opts.thinking
         };
         const paid = await costLedger.runPaidCall({
           channel: "driver",
@@ -2542,7 +2543,8 @@ ${analyst}` : userPrompt;
           ],
           jsonMode: true,
           temperature: opts.temperature ?? 0.7,
-          maxTokens
+          maxTokens,
+          thinking: opts.thinking
         };
         const paid = await costLedger.runPaidCall({
           channel: "driver",
@@ -3836,4 +3838,4 @@ export {
   provenanceSpansPath,
   emitLoopProvenance
 };
-//# sourceMappingURL=chunk-QVGVJQMR.js.map
+//# sourceMappingURL=chunk-PMITBABE.js.map