@elizaos/plugin-training 2.0.3-beta.6 → 2.0.3-beta.7

package/dist/optimizers/scoring.js

@@ -0,0 +1,299 @@
+function createPromptScorer(adapter, options = {}) {
+  const temperature = options.temperature ?? 0;
+  const maxTokens = options.maxTokens ?? 512;
+  const compare = options.compare ?? scoreAgreement;
+  return async (prompt, examples) => {
+    if (examples.length === 0) return 0;
+    const cap = options.maxExamples ?? examples.length;
+    const limited = examples.slice(0, Math.max(1, cap));
+    let total = 0;
+    for (const example of limited) {
+      const completion = await adapter.complete({
+        system: prompt,
+        user: example.input.user,
+        temperature,
+        maxTokens
+      });
+      total += compare(completion, example.expectedOutput);
+    }
+    return total / limited.length;
+  };
+}
+function stripOutputFences(text) {
+  return text.trim().replace(/^```[a-z0-9_-]*\s*/i, "").replace(/\s*```$/i, "").trim();
+}
+function parseJsonObject(text) {
+  const trimmed = stripOutputFences(text);
+  if (!trimmed.startsWith("{")) {
+    return null;
+  }
+  try {
+    const parsed = JSON.parse(trimmed);
+    return parsed && typeof parsed === "object" && !Array.isArray(parsed) ? parsed : null;
+  } catch {
+    return null;
+  }
+}
+function readLegacyField(text, fieldName) {
+  const lineMatch = new RegExp(`(^|\\n)${fieldName}:\\s*([^\\n]+)`, "i").exec(
+    text
+  );
+  const value = lineMatch?.[2]?.trim();
+  return value ? value : void 0;
+}
+function parsePlannerObject(text) {
+  const parsed = parseJsonObject(text);
+  if (parsed) {
+    return parsed;
+  }
+  const legacyFields = {};
+  for (const fieldName of ["action", "actionName", "name", "type", "actions"]) {
+    const value = readLegacyField(text, fieldName);
+    if (value) {
+      legacyFields[fieldName] = value;
+    }
+  }
+  return legacyFields;
+}
+function extractPlannerAction(text) {
+  if (!text) return null;
+  const parsed = parsePlannerObject(text);
+  if (parsed && Array.isArray(parsed.toolCalls)) {
+    const first = parsed.toolCalls[0];
+    if (first && typeof first === "object") {
+      const record = first;
+      const name = record.name ?? record.action ?? record.actionName;
+      if (typeof name === "string" && name.trim().length > 0) {
+        return name.trim().toUpperCase();
+      }
+    }
+  }
+  const raw = parsed.action ?? parsed.actionName ?? parsed.name ?? parsed.type ?? parsed.actions;
+  if (typeof raw === "string" && raw.trim()) {
+    return raw.split(",")[0]?.trim().toUpperCase() ?? null;
+  }
+  if (Array.isArray(raw)) {
+    const first = raw[0];
+    if (typeof first === "string") return first.trim().toUpperCase();
+    if (first && typeof first === "object") {
+      const record = first;
+      const name = record.name ?? record.action ?? record.actionName ?? record.type;
+      if (typeof name === "string") return name.trim().toUpperCase();
+    }
+  }
+  const nameMatch = text.match(/\b([A-Z][A-Z0-9_]{2,})\b/);
+  return nameMatch?.[1] ?? null;
+}
+function readViewFromArgs(obj) {
+  if (!obj || typeof obj !== "object") return null;
+  const r = obj;
+  const v = r.view ?? r.viewId ?? r.id ?? r.target;
+  return typeof v === "string" && v.trim() ? v.trim().toLowerCase() : null;
+}
+function extractPlannerView(text) {
+  if (!text) return null;
+  const parsed = parsePlannerObject(text);
+  if (!parsed) return null;
+  if (Array.isArray(parsed.toolCalls)) {
+    const first = parsed.toolCalls[0];
+    if (first && typeof first === "object") {
+      const record = first;
+      const fromCall = readViewFromArgs(
+        record.args ?? record.arguments ?? record.parameters
+      );
+      if (fromCall) return fromCall;
+    }
+  }
+  const fromParams = readViewFromArgs(
+    parsed.parameters ?? parsed.args ?? parsed.arguments
+  );
+  if (fromParams) return fromParams;
+  return readViewFromArgs(parsed);
+}
+function scorePlannerAction(actual, expected) {
+  const actualAction = extractPlannerAction(actual);
+  const expectedAction = extractPlannerAction(expected);
+  if (!expectedAction) return 0;
+  if (!actualAction) return 0;
+  if (actualAction !== expectedAction) return 0;
+  const expectedView = extractPlannerView(expected);
+  if (!expectedView) return 1;
+  const actualView = extractPlannerView(actual);
+  if (!actualView) return 0.5;
+  return actualView === expectedView ? 1 : 0.5;
+}
+function scoreViewSelection(actual, expected) {
+  const expectedView = extractPlannerView(expected);
+  const actualView = extractPlannerView(actual);
+  if (expectedView === null && actualView === null) return 1;
+  return expectedView === actualView ? 1 : 0;
+}
+function scoreAgreement(actual, expected) {
+  const actualTokens = tokenize(actual);
+  const expectedTokens = tokenize(expected);
+  if (expectedTokens.size === 0 && actualTokens.size === 0) return 1;
+  if (expectedTokens.size === 0 || actualTokens.size === 0) return 0;
+  let intersection = 0;
+  for (const token of actualTokens) {
+    if (expectedTokens.has(token)) intersection += 1;
+  }
+  const union = actualTokens.size + expectedTokens.size - intersection;
+  if (union === 0) return 0;
+  return intersection / union;
+}
+function tokenize(text) {
+  const tokens = text.toLowerCase().replace(/[^a-z0-9\s_-]+/g, " ").split(/\s+/).filter((token) => token.length > 0);
+  return new Set(tokens);
+}
+const LIFEOPS_SCORER_TASKS = [
+  "calendar_extract",
+  "schedule_plan",
+  "reminder_dispatch",
+  "inbox_triage",
+  "meeting_prep",
+  "morning_brief",
+  "health_checkin",
+  "screentime_recap"
+];
+const LIFEOPS_STRUCTURED_SCORER_TASKS = [
+  "calendar_extract",
+  "schedule_plan",
+  "reminder_dispatch",
+  "inbox_triage",
+  "meeting_prep",
+  "health_checkin",
+  "screentime_recap"
+];
+const LIFEOPS_EXTRACTION_TASKS = new Set(
+  LIFEOPS_STRUCTURED_SCORER_TASKS
+);
+function stripFence(text) {
+  return text.trim().replace(/^```(?:json)?\s*/i, "").replace(/\s*```$/i, "");
+}
+function parseJsonLoose(text) {
+  const trimmed = stripFence(text);
+  const start = trimmed.indexOf("{");
+  const end = trimmed.lastIndexOf("}");
+  if (start === -1 || end === -1 || end <= start) return null;
+  try {
+    const parsed = JSON.parse(trimmed.slice(start, end + 1));
+    return parsed && typeof parsed === "object" && !Array.isArray(parsed) ? parsed : null;
+  } catch {
+    return null;
+  }
+}
+function parseLineFieldsLoose(text) {
+  const fields = {};
+  let parsedLines = 0;
+  for (const line of stripFence(text).split(/\r?\n/u)) {
+    const match = /^([a-zA-Z_][a-zA-Z0-9_-]*)\s*:\s*(.*)$/u.exec(line.trim());
+    if (!match) continue;
+    const key = match[1]?.trim();
+    if (!key) continue;
+    fields[key] = match[2]?.trim() ?? "";
+    parsedLines += 1;
+  }
+  return parsedLines > 0 ? fields : null;
+}
+function parseStructuredFieldsLoose(text) {
+  return parseJsonLoose(text) ?? parseLineFieldsLoose(text);
+}
+function normalizeScalar(value) {
+  if (value === null || value === void 0) return "";
+  if (typeof value === "object") return JSON.stringify(value);
+  return String(value).trim().toLowerCase();
+}
+function scoreStructuredFields(actual, expected, fields) {
+  const expectedObj = parseStructuredFieldsLoose(expected);
+  if (!expectedObj) return 0;
+  const actualObj = parseStructuredFieldsLoose(actual) ?? {};
+  const keys = fields && fields.length > 0 ? [...fields] : Object.keys(expectedObj);
+  if (keys.length === 0) {
+    return Object.keys(actualObj).length === 0 ? 1 : 0;
+  }
+  let matched = 0;
+  for (const key of keys) {
+    if (normalizeScalar(actualObj[key]) === normalizeScalar(expectedObj[key])) {
+      matched += 1;
+    }
+  }
+  return matched / keys.length;
+}
+function actionTokens(text) {
+  const obj = parseStructuredFieldsLoose(text);
+  const source = obj ? [
+    obj.action,
+    obj.subaction,
+    obj.category,
+    obj.priority,
+    obj.channel,
+    obj.suggestion
+  ].map(normalizeScalar).filter(Boolean).join(" ") : text;
+  return new Set(
+    source.toLowerCase().split(/[\s,|]+/).map((token) => token.trim()).filter(Boolean)
+  );
+}
+function scoreActionSet(actual, expected) {
+  const actualSet = actionTokens(actual);
+  const expectedSet = actionTokens(expected);
+  if (actualSet.size === 0 && expectedSet.size === 0) return 1;
+  if (actualSet.size === 0 || expectedSet.size === 0) return 0;
+  let intersection = 0;
+  for (const token of actualSet) {
+    if (expectedSet.has(token)) intersection += 1;
+  }
+  const union = actualSet.size + expectedSet.size - intersection;
+  return union === 0 ? 0 : intersection / union;
+}
+function scoreLifeOpsTask(task, actual, expected) {
+  if (LIFEOPS_EXTRACTION_TASKS.has(task)) {
+    return scoreStructuredFields(actual, expected);
+  }
+  return scoreAgreement(actual, expected);
+}
+function subsample(items, count, rng = Math.random) {
+  if (count >= items.length) return [...items];
+  const indices = /* @__PURE__ */ new Set();
+  const out = [];
+  while (out.length < count) {
+    const idx = Math.floor(rng() * items.length);
+    if (indices.has(idx)) continue;
+    indices.add(idx);
+    const item = items[idx];
+    if (item !== void 0) out.push(item);
+  }
+  return out;
+}
+function createRuntimeAdapter(useModel) {
+  return {
+    async complete(input) {
+      const composed = input.system ? `${input.system}
+
+${input.user}` : input.user;
+      const response = await useModel({
+        prompt: composed,
+        temperature: input.temperature,
+        maxTokens: input.maxTokens
+      });
+      if (typeof response === "string") return response;
+      if (response === void 0 || response === null) return "";
+      return JSON.stringify(response);
+    }
+  };
+}
+export {
+  LIFEOPS_SCORER_TASKS,
+  LIFEOPS_STRUCTURED_SCORER_TASKS,
+  createPromptScorer,
+  createRuntimeAdapter,
+  extractPlannerAction,
+  extractPlannerView,
+  scoreActionSet,
+  scoreAgreement,
+  scoreLifeOpsTask,
+  scorePlannerAction,
+  scoreStructuredFields,
+  scoreViewSelection,
+  subsample
+};
+//# sourceMappingURL=scoring.js.map

package/dist/optimizers/scoring.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/optimizers/scoring.ts"],"sourcesContent":["/**\n * Scoring utilities for native optimizers.\n *\n * The default scorer measures token-overlap agreement between the model's\n * actual output and the expected output recorded in the trajectory dataset.\n * It is deliberately cheap and deterministic — the optimizers run hundreds\n * of completions per round, so we cannot afford a model-graded scorer.\n *\n * Token-overlap agreement (Jaccard over normalized tokens) is the same primitive\n * that `replay-validator.ts` uses for `scoreSkill`-style success measurement,\n * just lifted to the (output vs reference) comparison instead of (skill vs\n * trajectory). When a richer signal becomes available, the scorer factory can\n * be swapped without changing any optimizer code.\n */\n\nimport type { LlmAdapter, PromptScorer } from \"./types.js\";\n\ninterface ScorerOptions {\n  /** Cap on examples scored per call. Defaults to all examples. */\n  maxExamples?: number;\n  /** Temperature passed to the adapter. Defaults to 0 for determinism. */\n  temperature?: number;\n  /** Max tokens for each completion. Defaults to 512. */\n  maxTokens?: number;\n  /**\n   * Per-example comparator. Defaults to Jaccard token overlap.\n   * Returning 1.0 means a perfect match, 0.0 means no credit.\n   */\n  compare?: (actual: string, expected: string) => number;\n}\n\n/**\n * Build a `PromptScorer` backed by a real LLM adapter.\n *\n * For each example:\n *   1. Run `prompt` (as system) + `example.input.user` through the adapter.\n *   2. Compare the completion against `example.expectedOutput` via Jaccard\n *      similarity over normalized tokens.\n *   3. Return the mean score.\n *\n * Reuses the same normalization heuristic as the trajectory-task-datasets\n * exporter (lower-case, strip punctuation, drop empty tokens).\n */\nexport function createPromptScorer(\n  adapter: LlmAdapter,\n  options: ScorerOptions = {},\n): PromptScorer {\n  const temperature = options.temperature ?? 0;\n  const maxTokens = options.maxTokens ?? 512;\n  const compare = options.compare ?? scoreAgreement;\n  return async (prompt, examples) => {\n    if (examples.length === 0) return 0;\n    const cap = options.maxExamples ?? examples.length;\n    const limited = examples.slice(0, Math.max(1, cap));\n    let total = 0;\n    for (const example of limited) {\n      const completion = await adapter.complete({\n        system: prompt,\n        user: example.input.user,\n        temperature,\n        maxTokens,\n      });\n      total += compare(completion, example.expectedOutput);\n    }\n    return total / limited.length;\n  };\n}\n\nfunction stripOutputFences(text: string): string {\n  return text\n    .trim()\n    .replace(/^```[a-z0-9_-]*\\s*/i, \"\")\n    .replace(/\\s*```$/i, \"\")\n    .trim();\n}\n\nfunction parseJsonObject(text: string): Record<string, unknown> | null {\n  const trimmed = stripOutputFences(text);\n  if (!trimmed.startsWith(\"{\")) {\n    return null;\n  }\n  try {\n    const parsed = JSON.parse(trimmed) as unknown;\n    return parsed && typeof parsed === \"object\" && !Array.isArray(parsed)\n      ? (parsed as Record<string, unknown>)\n      : null;\n  } catch {\n    return null;\n  }\n}\n\nfunction readLegacyField(text: string, fieldName: string): string | undefined {\n  const lineMatch = new RegExp(`(^|\\\\n)${fieldName}:\\\\s*([^\\\\n]+)`, \"i\").exec(\n    text,\n  );\n  const value = lineMatch?.[2]?.trim();\n  return value ? value : undefined;\n}\n\nfunction parsePlannerObject(text: string): Record<string, unknown> {\n  const parsed = parseJsonObject(text);\n  if (parsed) {\n    return parsed;\n  }\n\n  const legacyFields: Record<string, unknown> = {};\n  for (const fieldName of [\"action\", \"actionName\", \"name\", \"type\", \"actions\"]) {\n    const value = readLegacyField(text, fieldName);\n    if (value) {\n      legacyFields[fieldName] = value;\n    }\n  }\n  return legacyFields;\n}\n\n/**\n * Extract the first action name from planner output. JSON is preferred; a\n * small line-based reader keeps older key/value rows comparable.\n *\n * Schemas understood (in priority order):\n *   1. v5 planner: `{toolCalls:[{name:\"OWNER_TODOS\",\"args\":{...}}]}` — handled directly.\n *   2. Legacy structured: top-level `action`/`actionName`/`name`/`type`/`actions` field.\n *   3. Legacy line-based: `action: OWNER_TODOS` or similar key:value rows.\n *   4. Last-resort: any uppercase identifier in the text.\n *\n * The regex fallback is intentionally last — it matches identifiers like\n * `OWNER`, `OPTIONAL`, `JSON`, etc. that show up in field names, so it can\n * mislabel non-action text. Prefer the JSON paths when the runtime emits\n * structured output (which is the common case post-v5).\n */\nexport function extractPlannerAction(text: string): string | null {\n  if (!text) return null;\n  const parsed = parsePlannerObject(text);\n  // v5 toolCalls shape — most common in current trajectories\n  if (parsed && Array.isArray(parsed.toolCalls)) {\n    const first = parsed.toolCalls[0];\n    if (first && typeof first === \"object\") {\n      const record = first as Record<string, unknown>;\n      const name = record.name ?? record.action ?? record.actionName;\n      if (typeof name === \"string\" && name.trim().length > 0) {\n        return name.trim().toUpperCase();\n      }\n    }\n  }\n  const raw =\n    parsed.action ??\n    parsed.actionName ??\n    parsed.name ??\n    parsed.type ??\n    parsed.actions;\n  if (typeof raw === \"string\" && raw.trim()) {\n    return raw.split(\",\")[0]?.trim().toUpperCase() ?? null;\n  }\n  if (Array.isArray(raw)) {\n    const first = raw[0];\n    if (typeof first === \"string\") return first.trim().toUpperCase();\n    if (first && typeof first === \"object\") {\n      const record = first as Record<string, unknown>;\n      const name =\n        record.name ?? record.action ?? record.actionName ?? record.type;\n      if (typeof name === \"string\") return name.trim().toUpperCase();\n    }\n  }\n  const nameMatch = text.match(/\\b([A-Z][A-Z0-9_]{2,})\\b/);\n  return nameMatch?.[1] ?? null;\n}\n\n/**\n * Pull a target-view id out of a planner argument object. View navigation\n * carries the surface in one of a few alias keys (`view`/`viewId`/`id`/`target`)\n * — the VIEWS action declares all of them. `name` is intentionally excluded: at\n * tool-call top level `name` is the ACTION name, so reading it as a view would\n * mislabel every call.\n */\nfunction readViewFromArgs(obj: unknown): string | null {\n  if (!obj || typeof obj !== \"object\") return null;\n  const r = obj as Record<string, unknown>;\n  const v = r.view ?? r.viewId ?? r.id ?? r.target;\n  return typeof v === \"string\" && v.trim() ? v.trim().toLowerCase() : null;\n}\n\n/**\n * Extract the target view id from planner output when the chosen action is a\n * view navigation. Understands the same shapes as {@link extractPlannerAction}:\n *   1. tool-call: `{toolCalls:[{name:\"VIEWS\", args/arguments/parameters:{view}}]}`\n *   2. bare action: `{action:\"VIEWS\", parameters/args:{view}}`\n *   3. top-level alias: `{view}` / `{viewId}`\n * Returns the lower-cased view id, or `null` when none is present.\n */\nexport function extractPlannerView(text: string): string | null {\n  if (!text) return null;\n  const parsed = parsePlannerObject(text);\n  if (!parsed) return null;\n  if (Array.isArray(parsed.toolCalls)) {\n    const first = parsed.toolCalls[0];\n    if (first && typeof first === \"object\") {\n      const record = first as Record<string, unknown>;\n      const fromCall = readViewFromArgs(\n        record.args ?? record.arguments ?? record.parameters,\n      );\n      if (fromCall) return fromCall;\n    }\n  }\n  const fromParams = readViewFromArgs(\n    parsed.parameters ?? parsed.args ?? parsed.arguments,\n  );\n  if (fromParams) return fromParams;\n  return readViewFromArgs(parsed);\n}\n\n/**\n * Action-name comparator: returns 1.0 when both outputs resolve to the same\n * planner action name, 0.0 otherwise. This is the right primitive for\n * optimizing the `action_planner` task because token overlap under-credits\n * correct choices when surrounding rationale varies stochastically.\n *\n * View-aware refinement: when the expected output pins a specific view (a VIEWS\n * navigation target), a matching action alone is NOT full credit — the view has\n * to match too. Without this the optimizer can never learn correct view\n * selection, because every `VIEWS/<anything>` would score 1.0 against a\n * `VIEWS/calendar` reference (the exact gap that made entry-tier wrong-view\n * outputs look perfect). Partial credit (right action, wrong/missing view =\n * 0.5) keeps a usable gradient for the optimizer. Expected outputs without a\n * view (every non-navigation action) are scored action-only, unchanged.\n */\nexport function scorePlannerAction(actual: string, expected: string): number {\n  const actualAction = extractPlannerAction(actual);\n  const expectedAction = extractPlannerAction(expected);\n  if (!expectedAction) return 0;\n  if (!actualAction) return 0;\n  if (actualAction !== expectedAction) return 0;\n  const expectedView = extractPlannerView(expected);\n  if (!expectedView) return 1;\n  const actualView = extractPlannerView(actual);\n  if (!actualView) return 0.5;\n  return actualView === expectedView ? 1 : 0.5;\n}\n\n/**\n * View-selection comparator for the contextual view evaluator (`view_context`\n * task). Both outputs are `{viewId, reason}` (or \"none\"); credit is 1.0 when the\n * chosen view id matches the reference, 0.0 otherwise. Case-insensitive, and a\n * matching \"none\" (correctly declining to navigate) scores 1.0 — so it rewards\n * both opening the right surface AND staying put on non-navigational turns.\n */\nexport function scoreViewSelection(actual: string, expected: string): number {\n  const expectedView = extractPlannerView(expected);\n  const actualView = extractPlannerView(actual);\n  if (expectedView === null && actualView === null) return 1;\n  return expectedView === actualView ? 1 : 0;\n}\n\n/**\n * Jaccard similarity over normalized token sets, in `[0, 1]`. Empty inputs\n * collapse to 0 (no overlap to measure).\n */\nexport function scoreAgreement(actual: string, expected: string): number {\n  const actualTokens = tokenize(actual);\n  const expectedTokens = tokenize(expected);\n  if (expectedTokens.size === 0 && actualTokens.size === 0) return 1;\n  if (expectedTokens.size === 0 || actualTokens.size === 0) return 0;\n  let intersection = 0;\n  for (const token of actualTokens) {\n    if (expectedTokens.has(token)) intersection += 1;\n  }\n  const union = actualTokens.size + expectedTokens.size - intersection;\n  if (union === 0) return 0;\n  return intersection / union;\n}\n\nfunction tokenize(text: string): Set<string> {\n  const tokens = text\n    .toLowerCase()\n    .replace(/[^a-z0-9\\s_-]+/g, \" \")\n    .split(/\\s+/)\n    .filter((token) => token.length > 0);\n  return new Set(tokens);\n}\n\n// -----------------------------------------------------------------------------\n// LifeOps per-capability scorers (#8795 item 4).\n//\n// The LifeOps optimization tasks split into two scoring shapes:\n//   - Extraction tasks emit structured fields (JSON objects for most tasks,\n//     line-based `key: value` fields for a few legacy planners). These are\n//     graded on structured-field exact-match — the fraction of expected fields\n//     the model reproduced. Date/time/recurrence/recipient are exactly the\n//     fields that must be right, so partial-credit-by-field is the right signal.\n//   - Chat-shaped tasks emit free text (the morning brief). These fall back to\n//     token agreement here; the real optimization loop gates them on the\n//     `responseJudge` rubric instead of this cheap proxy.\n// Both shapes are deterministic and allocation-light, matching the optimizer's\n// hundreds-of-completions-per-round budget.\n// -----------------------------------------------------------------------------\n\n/** LifeOps tasks with per-capability scorers (#8795). */\nexport const LIFEOPS_SCORER_TASKS = [\n  \"calendar_extract\",\n  \"schedule_plan\",\n  \"reminder_dispatch\",\n  \"inbox_triage\",\n  \"meeting_prep\",\n  \"morning_brief\",\n  \"health_checkin\",\n  \"screentime_recap\",\n] as const;\n\n/** LifeOps tasks whose output is a structured JSON object (exact-field match). */\nexport const LIFEOPS_STRUCTURED_SCORER_TASKS = [\n  \"calendar_extract\",\n  \"schedule_plan\",\n  \"reminder_dispatch\",\n  \"inbox_triage\",\n  \"meeting_prep\",\n  \"health_checkin\",\n  \"screentime_recap\",\n] as const;\n\nconst LIFEOPS_EXTRACTION_TASKS: ReadonlySet<string> = new Set(\n  LIFEOPS_STRUCTURED_SCORER_TASKS,\n);\n\nfunction stripFence(text: string): string {\n  return text\n    .trim()\n    .replace(/^```(?:json)?\\s*/i, \"\")\n    .replace(/\\s*```$/i, \"\");\n}\n\n/** Parse a JSON object out of a completion, tolerating ```json fences/prose. */\nfunction parseJsonLoose(text: string): Record<string, unknown> | null {\n  const trimmed = stripFence(text);\n  const start = trimmed.indexOf(\"{\");\n  const end = trimmed.lastIndexOf(\"}\");\n  if (start === -1 || end === -1 || end <= start) return null;\n  try {\n    const parsed = JSON.parse(trimmed.slice(start, end + 1)) as unknown;\n    return parsed && typeof parsed === \"object\" && !Array.isArray(parsed)\n      ? (parsed as Record<string, unknown>)\n      : null;\n  } catch {\n    return null;\n  }\n}\n\n/** Parse legacy line-based planner output: `field: value` per line. */\nfunction parseLineFieldsLoose(text: string): Record<string, unknown> | null {\n  const fields: Record<string, string> = {};\n  let parsedLines = 0;\n  for (const line of stripFence(text).split(/\\r?\\n/u)) {\n    const match = /^([a-zA-Z_][a-zA-Z0-9_-]*)\\s*:\\s*(.*)$/u.exec(line.trim());\n    if (!match) continue;\n    const key = match[1]?.trim();\n    if (!key) continue;\n    fields[key] = match[2]?.trim() ?? \"\";\n    parsedLines += 1;\n  }\n  return parsedLines > 0 ? fields : null;\n}\n\nfunction parseStructuredFieldsLoose(\n  text: string,\n): Record<string, unknown> | null {\n  return parseJsonLoose(text) ?? parseLineFieldsLoose(text);\n}\n\nfunction normalizeScalar(value: unknown): string {\n  if (value === null || value === undefined) return \"\";\n  if (typeof value === \"object\") return JSON.stringify(value);\n  return String(value).trim().toLowerCase();\n}\n\n/**\n * Structured-field exact-match score in `[0, 1]`: the fraction of expected\n * fields whose value the actual output reproduced. Both inputs are parsed as\n * JSON or line-based fields (tolerating fences/prose). When `fields` is supplied only those keys are\n * scored; otherwise every key in `expected` is scored. Returns 0 when expected\n * is unparseable (nothing to credit) and 1 when both parse to empty objects.\n */\nexport function scoreStructuredFields(\n  actual: string,\n  expected: string,\n  fields?: readonly string[],\n): number {\n  const expectedObj = parseStructuredFieldsLoose(expected);\n  if (!expectedObj) return 0;\n  const actualObj = parseStructuredFieldsLoose(actual) ?? {};\n  const keys =\n    fields && fields.length > 0 ? [...fields] : Object.keys(expectedObj);\n  if (keys.length === 0) {\n    return Object.keys(actualObj).length === 0 ? 1 : 0;\n  }\n  let matched = 0;\n  for (const key of keys) {\n    if (normalizeScalar(actualObj[key]) === normalizeScalar(expectedObj[key])) {\n      matched += 1;\n    }\n  }\n  return matched / keys.length;\n}\n\n/** Tokenize an output into an action/label set (JSON fields or raw words). */\nfunction actionTokens(text: string): Set<string> {\n  const obj = parseStructuredFieldsLoose(text);\n  const source = obj\n    ? [\n        obj.action,\n        obj.subaction,\n        obj.category,\n        obj.priority,\n        obj.channel,\n        obj.suggestion,\n      ]\n        .map(normalizeScalar)\n        .filter(Boolean)\n        .join(\" \")\n    : text;\n  return new Set(\n    source\n      .toLowerCase()\n      .split(/[\\s,|]+/)\n      .map((token) => token.trim())\n      .filter(Boolean),\n  );\n}\n\n/**\n * Action/label set-overlap (Jaccard) in `[0, 1]`. For tasks whose target is\n * \"did the agent pick the right action/category set\" rather than exact text.\n * Two empty sets score 1.0 (both correctly produced nothing actionable).\n */\nexport function scoreActionSet(actual: string, expected: string): number {\n  const actualSet = actionTokens(actual);\n  const expectedSet = actionTokens(expected);\n  if (actualSet.size === 0 && expectedSet.size === 0) return 1;\n  if (actualSet.size === 0 || expectedSet.size === 0) return 0;\n  let intersection = 0;\n  for (const token of actualSet) {\n    if (expectedSet.has(token)) intersection += 1;\n  }\n  const union = actualSet.size + expectedSet.size - intersection;\n  return union === 0 ? 0 : intersection / union;\n}\n\n/**\n * Authoritative per-task comparator for the LifeOps optimization tasks (the\n * GEPA Pareto scorer dispatches through here). Extraction tasks →\n * structured-field exact-match; the chat-shaped morning brief → token\n * agreement (proxy for the judge rubric); anything else → token agreement.\n */\nexport function scoreLifeOpsTask(\n  task: string,\n  actual: string,\n  expected: string,\n): number {\n  if (LIFEOPS_EXTRACTION_TASKS.has(task)) {\n    return scoreStructuredFields(actual, expected);\n  }\n  return scoreAgreement(actual, expected);\n}\n\n/**\n * Random-without-replacement subsample, used by optimizer rounds to keep\n * scoring cheap on large datasets without sacrificing comparability across\n * rounds (deterministic when `rng` is supplied).\n */\nexport function subsample<T>(\n  items: T[],\n  count: number,\n  rng: () => number = Math.random,\n): T[] {\n  if (count >= items.length) return [...items];\n  const indices = new Set<number>();\n  const out: T[] = [];\n  while (out.length < count) {\n    const idx = Math.floor(rng() * items.length);\n    if (indices.has(idx)) continue;\n    indices.add(idx);\n    const item = items[idx];\n    if (item !== undefined) out.push(item);\n  }\n  return out;\n}\n\n/**\n * Wraps `IAgentRuntime.useModel` into the `LlmAdapter` shape. We accept a\n * loose runtime type so this module stays free of `@elizaos/core` import\n * cycles — the native backend supplies the bound `useModel` directly.\n */\nexport type UseModelHandler = (input: {\n  prompt: string;\n  temperature?: number;\n  maxTokens?: number;\n}) => Promise<string | object | undefined>;\n\nexport function createRuntimeAdapter(useModel: UseModelHandler): LlmAdapter {\n  return {\n    async complete(input) {\n      const composed = input.system\n        ? `${input.system}\\n\\n${input.user}`\n        : input.user;\n      const response = await useModel({\n        prompt: composed,\n        temperature: input.temperature,\n        maxTokens: input.maxTokens,\n      });\n      if (typeof response === \"string\") return response;\n      if (response === undefined || response === null) return \"\";\n      return JSON.stringify(response);\n    },\n  };\n}\n"],"mappings":"AA2CO,SAAS,mBACd,SACA,UAAyB,CAAC,GACZ;AACd,QAAM,cAAc,QAAQ,eAAe;AAC3C,QAAM,YAAY,QAAQ,aAAa;AACvC,QAAM,UAAU,QAAQ,WAAW;AACnC,SAAO,OAAO,QAAQ,aAAa;AACjC,QAAI,SAAS,WAAW,EAAG,QAAO;AAClC,UAAM,MAAM,QAAQ,eAAe,SAAS;AAC5C,UAAM,UAAU,SAAS,MAAM,GAAG,KAAK,IAAI,GAAG,GAAG,CAAC;AAClD,QAAI,QAAQ;AACZ,eAAW,WAAW,SAAS;AAC7B,YAAM,aAAa,MAAM,QAAQ,SAAS;AAAA,QACxC,QAAQ;AAAA,QACR,MAAM,QAAQ,MAAM;AAAA,QACpB;AAAA,QACA;AAAA,MACF,CAAC;AACD,eAAS,QAAQ,YAAY,QAAQ,cAAc;AAAA,IACrD;AACA,WAAO,QAAQ,QAAQ;AAAA,EACzB;AACF;AAEA,SAAS,kBAAkB,MAAsB;AAC/C,SAAO,KACJ,KAAK,EACL,QAAQ,uBAAuB,EAAE,EACjC,QAAQ,YAAY,EAAE,EACtB,KAAK;AACV;AAEA,SAAS,gBAAgB,MAA8C;AACrE,QAAM,UAAU,kBAAkB,IAAI;AACtC,MAAI,CAAC,QAAQ,WAAW,GAAG,GAAG;AAC5B,WAAO;AAAA,EACT;AACA,MAAI;AACF,UAAM,SAAS,KAAK,MAAM,OAAO;AACjC,WAAO,UAAU,OAAO,WAAW,YAAY,CAAC,MAAM,QAAQ,MAAM,IAC/D,SACD;AAAA,EACN,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAEA,SAAS,gBAAgB,MAAc,WAAuC;AAC5E,QAAM,YAAY,IAAI,OAAO,UAAU,SAAS,kBAAkB,GAAG,EAAE;AAAA,IACrE;AAAA,EACF;AACA,QAAM,QAAQ,YAAY,CAAC,GAAG,KAAK;AACnC,SAAO,QAAQ,QAAQ;AACzB;AAEA,SAAS,mBAAmB,MAAuC;AACjE,QAAM,SAAS,gBAAgB,IAAI;AACnC,MAAI,QAAQ;AACV,WAAO;AAAA,EACT;AAEA,QAAM,eAAwC,CAAC;AAC/C,aAAW,aAAa,CAAC,UAAU,cAAc,QAAQ,QAAQ,SAAS,GAAG;AAC3E,UAAM,QAAQ,gBAAgB,MAAM,SAAS;AAC7C,QAAI,OAAO;AACT,mBAAa,SAAS,IAAI;AAAA,IAC5B;AAAA,EACF;AACA,SAAO;AACT;AAiBO,SAAS,qBAAqB,MAA6B;AAChE,MAAI,CAAC,KAAM,QAAO;AAClB,QAAM,SAAS,mBAAmB,IAAI;AAEtC,MAAI,UAAU,MAAM,QAAQ,OAAO,SAAS,GAAG;AAC7C,UAAM,QAAQ,OAAO,UAAU,CAAC;AAChC,QAAI,SAAS,OAAO,UAAU,UAAU;AACtC,YAAM,SAAS;AACf,YAAM,OAAO,OAAO,QAAQ,OAAO,UAAU,OAAO;AACpD,UAAI,OAAO,SAAS,YAAY,KAAK,KAAK,EAAE,SAAS,GAAG;AACtD,eAAO,KAAK,KAAK,EAAE,YAAY;AAAA,MACjC;AAAA,IACF;AAAA,EACF;AACA,QAAM,MACJ,OAAO,UACP,OAAO,cACP,OAAO,QACP,OAAO,QACP,OAAO;AACT,MAAI,OAAO,QAAQ,YAAY,IAAI,KAAK,GAAG;AACzC,WAAO,IAAI,MAAM,GAAG,EAAE,CAAC,GAAG,KAAK,EAAE,YAAY,KAAK;AAAA,EACpD;AACA,MAAI,MAAM,QAAQ,GAAG,GAAG;AACtB,UAAM,QAAQ,IAAI,CAAC;AACnB,QAAI,OAAO,UAAU,SAAU,QAAO,MAAM,KAAK,EAAE,YAAY;AAC/D,QAAI,SAAS,OAAO,UAAU,UAAU;AACtC,YAAM,SAAS;AACf,YAAM,OACJ,OAAO,QAAQ,OAAO,UAAU,OAAO,cAAc,OAAO;AAC9D,UAAI,OAAO,SAAS,SAAU,QAAO,KAAK,KAAK,EAAE,YAAY;AAAA,IAC/D;AAAA,EACF;AACA,QAAM,YAAY,KAAK,MAAM,0BAA0B;AACvD,SAAO,YAAY,CAAC,KAAK;AAC3B;AASA,SAAS,iBAAiB,KAA6B;AACrD,MAAI,CAAC,OAAO,OAAO,QAAQ,SAAU,QAAO;AAC5C,QAAM,IAAI;AACV,QAAM,IAAI,EAAE,QAAQ,EAAE,UAAU,EAAE,MAAM,EAAE;AAC1C,SAAO,OAAO,MAAM,YAAY,EAAE,KAAK,IAAI,EAAE,KAAK,EAAE,YAAY,IAAI;AACtE;AAUO,SAAS,mBAAmB,MAA6B;AAC9D,MAAI,CAAC,KAAM,QAAO;AAClB,QAAM,SAAS,mBAAmB,IAAI;AACtC,MAAI,CAAC,OAAQ,QAAO;AACpB,MAAI,MAAM,QAAQ,OAAO,SAAS,GAAG;AACnC,UAAM,QAAQ,OAAO,UAAU,CAAC;AAChC,QAAI,SAAS,OAAO,UAAU,UAAU;AACtC,YAAM,SAAS;AACf,YAAM,WAAW;AAAA,QACf,OAAO,QAAQ,OAAO,aAAa,OAAO;AAAA,MAC5C;AACA,UAAI,SAAU,QAAO;AAAA,IACvB;AAAA,EACF;AACA,QAAM,aAAa;AAAA,IACjB,OAAO,cAAc,OAAO,QAAQ,OAAO;AAAA,EAC7C;AACA,MAAI,WAAY,QAAO;AACvB,SAAO,iBAAiB,MAAM;AAChC;AAiBO,SAAS,mBAAmB,QAAgB,UAA0B;AAC3E,QAAM,eAAe,qBAAqB,MAAM;AAChD,QAAM,iBAAiB,qBAAqB,QAAQ;AACpD,MAAI,CAAC,eAAgB,QAAO;AAC5B,MAAI,CAAC,aAAc,QAAO;AAC1B,MAAI,iBAAiB,eAAgB,QAAO;AAC5C,QAAM,eAAe,mBAAmB,QAAQ;AAChD,MAAI,CAAC,aAAc,QAAO;AAC1B,QAAM,aAAa,mBAAmB,MAAM;AAC5C,MAAI,CAAC,WAAY,QAAO;AACxB,SAAO,eAAe,eAAe,IAAI;AAC3C;AASO,SAAS,mBAAmB,QAAgB,UAA0B;AAC3E,QAAM,eAAe,mBAAmB,QAAQ;AAChD,QAAM,aAAa,mBAAmB,MAAM;AAC5C,MAAI,iBAAiB,QAAQ,eAAe,KAAM,QAAO;AACzD,SAAO,iBAAiB,aAAa,IAAI;AAC3C;AAMO,SAAS,eAAe,QAAgB,UAA0B;AACvE,QAAM,eAAe,SAAS,MAAM;AACpC,QAAM,iBAAiB,SAAS,QAAQ;AACxC,MAAI,eAAe,SAAS,KAAK,aAAa,SAAS,EAAG,QAAO;AACjE,MAAI,eAAe,SAAS,KAAK,aAAa,SAAS,EAAG,QAAO;AACjE,MAAI,eAAe;AACnB,aAAW,SAAS,cAAc;AAChC,QAAI,eAAe,IAAI,KAAK,EAAG,iBAAgB;AAAA,EACjD;AACA,QAAM,QAAQ,aAAa,OAAO,eAAe,OAAO;AACxD,MAAI,UAAU,EAAG,QAAO;AACxB,SAAO,eAAe;AACxB;AAEA,SAAS,SAAS,MAA2B;AAC3C,QAAM,SAAS,KACZ,YAAY,EACZ,QAAQ,mBAAmB,GAAG,EAC9B,MAAM,KAAK,EACX,OAAO,CAAC,UAAU,MAAM,SAAS,CAAC;AACrC,SAAO,IAAI,IAAI,MAAM;AACvB;AAmBO,MAAM,uBAAuB;AAAA,EAClC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAGO,MAAM,kCAAkC;AAAA,EAC7C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,MAAM,2BAAgD,IAAI;AAAA,EACxD;AACF;AAEA,SAAS,WAAW,MAAsB;AACxC,SAAO,KACJ,KAAK,EACL,QAAQ,qBAAqB,EAAE,EAC/B,QAAQ,YAAY,EAAE;AAC3B;AAGA,SAAS,eAAe,MAA8C;AACpE,QAAM,UAAU,WAAW,IAAI;AAC/B,QAAM,QAAQ,QAAQ,QAAQ,GAAG;AACjC,QAAM,MAAM,QAAQ,YAAY,GAAG;AACnC,MAAI,UAAU,MAAM,QAAQ,MAAM,OAAO,MAAO,QAAO;AACvD,MAAI;AACF,UAAM,SAAS,KAAK,MAAM,QAAQ,MAAM,OAAO,MAAM,CAAC,CAAC;AACvD,WAAO,UAAU,OAAO,WAAW,YAAY,CAAC,MAAM,QAAQ,MAAM,IAC/D,SACD;AAAA,EACN,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAGA,SAAS,qBAAqB,MAA8C;AAC1E,QAAM,SAAiC,CAAC;AACxC,MAAI,cAAc;AAClB,aAAW,QAAQ,WAAW,IAAI,EAAE,MAAM,QAAQ,GAAG;AACnD,UAAM,QAAQ,0CAA0C,KAAK,KAAK,KAAK,CAAC;AACxE,QAAI,CAAC,MAAO;AACZ,UAAM,MAAM,MAAM,CAAC,GAAG,KAAK;AAC3B,QAAI,CAAC,IAAK;AACV,WAAO,GAAG,IAAI,MAAM,CAAC,GAAG,KAAK,KAAK;AAClC,mBAAe;AAAA,EACjB;AACA,SAAO,cAAc,IAAI,SAAS;AACpC;AAEA,SAAS,2BACP,MACgC;AAChC,SAAO,eAAe,IAAI,KAAK,qBAAqB,IAAI;AAC1D;AAEA,SAAS,gBAAgB,OAAwB;AAC/C,MAAI,UAAU,QAAQ,UAAU,OAAW,QAAO;AAClD,MAAI,OAAO,UAAU,SAAU,QAAO,KAAK,UAAU,KAAK;AAC1D,SAAO,OAAO,KAAK,EAAE,KAAK,EAAE,YAAY;AAC1C;AASO,SAAS,sBACd,QACA,UACA,QACQ;AACR,QAAM,cAAc,2BAA2B,QAAQ;AACvD,MAAI,CAAC,YAAa,QAAO;AACzB,QAAM,YAAY,2BAA2B,MAAM,KAAK,CAAC;AACzD,QAAM,OACJ,UAAU,OAAO,SAAS,IAAI,CAAC,GAAG,MAAM,IAAI,OAAO,KAAK,WAAW;AACrE,MAAI,KAAK,WAAW,GAAG;AACrB,WAAO,OAAO,KAAK,SAAS,EAAE,WAAW,IAAI,IAAI;AAAA,EACnD;AACA,MAAI,UAAU;AACd,aAAW,OAAO,MAAM;AACtB,QAAI,gBAAgB,UAAU,GAAG,CAAC,MAAM,gBAAgB,YAAY,GAAG,CAAC,GAAG;AACzE,iBAAW;AAAA,IACb;AAAA,EACF;AACA,SAAO,UAAU,KAAK;AACxB;AAGA,SAAS,aAAa,MAA2B;AAC/C,QAAM,MAAM,2BAA2B,IAAI;AAC3C,QAAM,SAAS,MACX;AAAA,IACE,IAAI;AAAA,IACJ,IAAI;AAAA,IACJ,IAAI;AAAA,IACJ,IAAI;AAAA,IACJ,IAAI;AAAA,IACJ,IAAI;AAAA,EACN,EACG,IAAI,eAAe,EACnB,OAAO,OAAO,EACd,KAAK,GAAG,IACX;AACJ,SAAO,IAAI;AAAA,IACT,OACG,YAAY,EACZ,MAAM,SAAS,EACf,IAAI,CAAC,UAAU,MAAM,KAAK,CAAC,EAC3B,OAAO,OAAO;AAAA,EACnB;AACF;AAOO,SAAS,eAAe,QAAgB,UAA0B;AACvE,QAAM,YAAY,aAAa,MAAM;AACrC,QAAM,cAAc,aAAa,QAAQ;AACzC,MAAI,UAAU,SAAS,KAAK,YAAY,SAAS,EAAG,QAAO;AAC3D,MAAI,UAAU,SAAS,KAAK,YAAY,SAAS,EAAG,QAAO;AAC3D,MAAI,eAAe;AACnB,aAAW,SAAS,WAAW;AAC7B,QAAI,YAAY,IAAI,KAAK,EAAG,iBAAgB;AAAA,EAC9C;AACA,QAAM,QAAQ,UAAU,OAAO,YAAY,OAAO;AAClD,SAAO,UAAU,IAAI,IAAI,eAAe;AAC1C;AAQO,SAAS,iBACd,MACA,QACA,UACQ;AACR,MAAI,yBAAyB,IAAI,IAAI,GAAG;AACtC,WAAO,sBAAsB,QAAQ,QAAQ;AAAA,EAC/C;AACA,SAAO,eAAe,QAAQ,QAAQ;AACxC;AAOO,SAAS,UACd,OACA,OACA,MAAoB,KAAK,QACpB;AACL,MAAI,SAAS,MAAM,OAAQ,QAAO,CAAC,GAAG,KAAK;AAC3C,QAAM,UAAU,oBAAI,IAAY;AAChC,QAAM,MAAW,CAAC;AAClB,SAAO,IAAI,SAAS,OAAO;AACzB,UAAM,MAAM,KAAK,MAAM,IAAI,IAAI,MAAM,MAAM;AAC3C,QAAI,QAAQ,IAAI,GAAG,EAAG;AACtB,YAAQ,IAAI,GAAG;AACf,UAAM,OAAO,MAAM,GAAG;AACtB,QAAI,SAAS,OAAW,KAAI,KAAK,IAAI;AAAA,EACvC;AACA,SAAO;AACT;AAaO,SAAS,qBAAqB,UAAuC;AAC1E,SAAO;AAAA,IACL,MAAM,SAAS,OAAO;AACpB,YAAM,WAAW,MAAM,SACnB,GAAG,MAAM,MAAM;AAAA;AAAA,EAAO,MAAM,IAAI,KAChC,MAAM;AACV,YAAM,WAAW,MAAM,SAAS;AAAA,QAC9B,QAAQ;AAAA,QACR,aAAa,MAAM;AAAA,QACnB,WAAW,MAAM;AAAA,MACnB,CAAC;AACD,UAAI,OAAO,aAAa,SAAU,QAAO;AACzC,UAAI,aAAa,UAAa,aAAa,KAAM,QAAO;AACxD,aAAO,KAAK,UAAU,QAAQ;AAAA,IAChC;AAAA,EACF;AACF;","names":[]}

package/dist/optimizers/types.d.ts

@@ -0,0 +1,105 @@
+/**
+ * Native prompt-optimization primitives.
+ *
+ * The optimizers under `plugins/plugin-training/src/optimizers/` (instruction-search,
+ * prompt-evolution, bootstrap-fewshot) all operate on the same triple:
+ *
+ *   - `OptimizationExample`: a labeled (input -> expected output) row, mirroring
+ *     the JSONL shape produced by `trajectory-task-datasets.ts`.
+ *   - `PromptScorer`: pluggable evaluator that returns a score in `[0, 1]` for
+ *     a candidate prompt against a held-out subset of examples.
+ *   - `LlmAdapter`: thin wrapper over `runtime.useModel()` so optimizer code
+ *     never depends on a specific provider.
+ *
+ * Decoupling the LLM behind `LlmAdapter` makes the optimizer modules unit
+ * testable with a deterministic in-memory adapter — no HTTP, no fixtures.
+ */
+import type { TrajectoryTrainingTask } from "../core/trajectory-task-datasets.js";
+/**
+ * One row in the optimization dataset.
+ *
+ * `input.system`/`input.user` mirror the `messages` arrays produced by the
+ * trajectory dataset exporter (`should_respond`, `response`, etc.). The
+ * exporter always wires those into the `system`/`user` slots, so this is the
+ * canonical shape for all native optimizer code.
+ */
+export interface OptimizationExample {
+    /** Stable identifier for traceability. Defaults to the row index. */
+    id?: string;
+    input: {
+        system?: string;
+        user: string;
+    };
+    /** Reference output the model should produce. Compared by the scorer. */
+    expectedOutput: string;
+    /**
+     * Optional reward signal recorded with the trajectory (e.g. successful
+     * completion = 1). Bootstrap-fewshot uses this to pick top-K demonstrations.
+     */
+    reward?: number;
+    /** Optional per-row metadata (task name, source trajectory id, ...). */
+    metadata?: Record<string, unknown>;
+}
+/**
+ * Scorer signature.
+ *
+ * Returns the mean score in `[0, 1]` across the supplied examples. A scorer
+ * MUST be deterministic given the same prompt + dataset + adapter, otherwise
+ * the optimizer cannot tell signal from noise across rounds.
+ */
+export type PromptScorer = (prompt: string, examples: OptimizationExample[]) => Promise<number>;
+/**
+ * Minimal LLM interface the optimizers depend on. Implementations route to
+ * `runtime.useModel(ModelType.TEXT_LARGE, …)` in production and to a
+ * deterministic adapter in tests.
+ */
+export interface LlmAdapter {
+    /**
+     * Run a single completion. Returns plain text (no parsing).
+     *
+     * `temperature` is optional because some adapters (e.g. tests) ignore it,
+     * but optimizers should pass it explicitly when they want diverse samples.
+     */
+    complete(input: {
+        system?: string;
+        user: string;
+        temperature?: number;
+        maxTokens?: number;
+    }): Promise<string>;
+}
+/** Per-round bookkeeping returned by every optimizer. */
+export interface OptimizerLineageEntry {
+    round: number;
+    variant: number;
+    score: number;
+    notes?: string;
+}
+/** Common shape returned by all native optimizers. */
+export interface OptimizerResult {
+    optimizedPrompt: string;
+    score: number;
+    baseline: number;
+    lineage: OptimizerLineageEntry[];
+    /** Demonstrations injected into the prompt (bootstrap-fewshot only). */
+    fewShotExamples?: OptimizationExample[];
+}
+export type OptimizerName = "instruction-search" | "prompt-evolution" | "gepa" | "bootstrap-fewshot" | "dspy-bootstrap-fewshot" | "dspy-copro" | "dspy-mipro";
+/**
+ * Persisted artifact written by the native backend and consumed by
+ * `OptimizedPromptService`. The persisted-on-disk schema is checked field by
+ * field on read; required fields stay required (no `?? null` fallbacks).
+ */
+export interface OptimizedPromptArtifact {
+    task: TrajectoryTrainingTask;
+    optimizer: OptimizerName;
+    baseline: string;
+    prompt: string;
+    score: number;
+    baselineScore: number;
+    datasetId: string;
+    datasetSize: number;
+    generatedAt: string;
+    fewShotExamples?: OptimizationExample[];
+    lineage: OptimizerLineageEntry[];
+}
+//# sourceMappingURL=types.d.ts.map

package/dist/optimizers/types.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/optimizers/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,KAAK,EAAE,sBAAsB,EAAE,MAAM,qCAAqC,CAAC;AAElF;;;;;;;GAOG;AACH,MAAM,WAAW,mBAAmB;IAClC,qEAAqE;IACrE,EAAE,CAAC,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE;QACL,MAAM,CAAC,EAAE,MAAM,CAAC;QAChB,IAAI,EAAE,MAAM,CAAC;KACd,CAAC;IACF,yEAAyE;IACzE,cAAc,EAAE,MAAM,CAAC;IACvB;;;OAGG;IACH,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,wEAAwE;IACxE,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACpC;AAED;;;;;;GAMG;AACH,MAAM,MAAM,YAAY,GAAG,CACzB,MAAM,EAAE,MAAM,EACd,QAAQ,EAAE,mBAAmB,EAAE,KAC5B,OAAO,CAAC,MAAM,CAAC,CAAC;AAErB;;;;GAIG;AACH,MAAM,WAAW,UAAU;IACzB;;;;;OAKG;IACH,QAAQ,CAAC,KAAK,EAAE;QACd,MAAM,CAAC,EAAE,MAAM,CAAC;QAChB,IAAI,EAAE,MAAM,CAAC;QACb,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,SAAS,CAAC,EAAE,MAAM,CAAC;KACpB,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;CACrB;AAED,yDAAyD;AACzD,MAAM,WAAW,qBAAqB;IACpC,KAAK,EAAE,MAAM,CAAC;IACd,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,sDAAsD;AACtD,MAAM,WAAW,eAAe;IAC9B,eAAe,EAAE,MAAM,CAAC;IACxB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;IACjB,OAAO,EAAE,qBAAqB,EAAE,CAAC;IACjC,wEAAwE;IACxE,eAAe,CAAC,EAAE,mBAAmB,EAAE,CAAC;CACzC;AAED,MAAM,MAAM,aAAa,GACrB,oBAAoB,GACpB,kBAAkB,GAClB,MAAM,GACN,mBAAmB,GACnB,wBAAwB,GACxB,YAAY,GACZ,YAAY,CAAC;AAEjB;;;;GAIG;AACH,MAAM,WAAW,uBAAuB;IACtC,IAAI,EAAE,sBAAsB,CAAC;IAC7B,SAAS,EAAE,aAAa,CAAC;IACzB,QAAQ,EAAE,MAAM,CAAC;IACjB,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,aAAa,EAAE,MAAM,CAAC;IACtB,SAAS,EAAE,MAAM,CAAC;IAClB,WAAW,EAAE,MAAM,CAAC;IACpB,WAAW,EAAE,MAAM,CAAC;IACpB,eAAe,CAAC,EAAE,mBAAmB,EAAE,CAAC;IACxC,OAAO,EAAE,qBAAqB,EAAE,CAAC;CAClC"}

package/dist/optimizers/types.js

@@ -0,0 +1 @@
+//# sourceMappingURL=types.js.map

package/dist/optimizers/types.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}

package/dist/register-runtime.d.ts

@@ -0,0 +1,3 @@
+import type { AgentRuntime } from "@elizaos/core";
+export declare function registerTrainingRuntimeHooks(runtime: AgentRuntime): Promise<void>;
+//# sourceMappingURL=register-runtime.d.ts.map

package/dist/register-runtime.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"register-runtime.d.ts","sourceRoot":"","sources":["../src/register-runtime.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,YAAY,EAAW,MAAM,eAAe,CAAC;AAiB3D,wBAAsB,4BAA4B,CAChD,OAAO,EAAE,YAAY,GACpB,OAAO,CAAC,IAAI,CAAC,CAwDf"}

package/dist/register-runtime.js

@@ -0,0 +1,60 @@
+import { logger, OptimizedPromptService } from "@elizaos/core";
+import { registerSkillScoringCron } from "./core/skill-scoring-cron.js";
+import { registerTrajectoryExportCron } from "./core/trajectory-export-cron.js";
+import {
+  bootstrapOptimizationFromAccumulatedTrajectories,
+  registerTrainingTriggerService
+} from "./services/training-trigger.js";
+function trainingCronRegistrationDisabled() {
+  const raw = process.env.ELIZA_DISABLE_TRAINING_CRONS;
+  if (!raw) {
+    return false;
+  }
+  return ["1", "true", "yes"].includes(raw.trim().toLowerCase());
+}
+async function registerTrainingRuntimeHooks(runtime) {
+  try {
+    let optimizedPromptService = await runtime.getServiceLoadPromise(OptimizedPromptService.serviceType).catch(() => null);
+    if (!optimizedPromptService) {
+      await runtime.registerService(
+        OptimizedPromptService
+      );
+      optimizedPromptService = await runtime.getServiceLoadPromise(
+        OptimizedPromptService.serviceType
+      );
+    }
+    logger.info(
+      "[eliza] Registered OptimizedPromptService (action_planner / media_description / etc. will pick up artifacts from <stateDir>/optimized-prompts/)"
+    );
+  } catch (err) {
+    logger.warn(
+      `[eliza] OptimizedPromptService registration failed: ${err instanceof Error ? err.message : String(err)}`
+    );
+  }
+  const skipCronRegistration = trainingCronRegistrationDisabled();
+  if (skipCronRegistration) {
+    logger.info("[eliza] Training cron registration skipped");
+  } else {
+    await registerTrajectoryExportCron(runtime);
+    await registerSkillScoringCron(runtime);
+  }
+  const triggerService = registerTrainingTriggerService(runtime);
+  logger.info(
+    skipCronRegistration ? "[eliza] Registered Track C auto-train trigger service" : "[eliza] Registered Track C training crons + auto-train trigger service"
+  );
+  void bootstrapOptimizationFromAccumulatedTrajectories(runtime, triggerService).then((fired) => {
+    if (fired.length > 0) {
+      logger.info(
+        `[eliza] Bootstrapped prompt optimization for ${fired.join(", ")}`
+      );
+    }
+  }).catch((err) => {
+    logger.error(
+      `[eliza] bootstrapOptimizationFromAccumulatedTrajectories failed: ${err instanceof Error ? err.stack ?? err.message : String(err)}`
+    );
+  });
+}
+export {
+  registerTrainingRuntimeHooks
+};
+//# sourceMappingURL=register-runtime.js.map

package/dist/register-runtime.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/register-runtime.ts"],"sourcesContent":["import type { AgentRuntime, Service } from \"@elizaos/core\";\nimport { logger, OptimizedPromptService } from \"@elizaos/core\";\nimport { registerSkillScoringCron } from \"./core/skill-scoring-cron.js\";\nimport { registerTrajectoryExportCron } from \"./core/trajectory-export-cron.js\";\nimport {\n  bootstrapOptimizationFromAccumulatedTrajectories,\n  registerTrainingTriggerService,\n} from \"./services/training-trigger.js\";\n\nfunction trainingCronRegistrationDisabled(): boolean {\n  const raw = process.env.ELIZA_DISABLE_TRAINING_CRONS;\n  if (!raw) {\n    return false;\n  }\n  return [\"1\", \"true\", \"yes\"].includes(raw.trim().toLowerCase());\n}\n\nexport async function registerTrainingRuntimeHooks(\n  runtime: AgentRuntime,\n): Promise<void> {\n  // Register the OptimizedPromptService so the planner-loop + media handler\n  // can pick up artifacts written by `bun run train -- --backend native`\n  // (or by the in-runtime trigger service) without operator intervention.\n  // Without this, runtime.getService(OPTIMIZED_PROMPT_SERVICE) always\n  // returns null and the optimized prompt is never substituted in.\n  try {\n    let optimizedPromptService: Service | null = await runtime\n      .getServiceLoadPromise(OptimizedPromptService.serviceType)\n      .catch(() => null);\n    if (!optimizedPromptService) {\n      await runtime.registerService(\n        OptimizedPromptService as Parameters<\n          typeof runtime.registerService\n        >[0],\n      );\n      optimizedPromptService = await runtime.getServiceLoadPromise(\n        OptimizedPromptService.serviceType,\n      );\n    }\n    logger.info(\n      \"[eliza] Registered OptimizedPromptService (action_planner / media_description / etc. will pick up artifacts from <stateDir>/optimized-prompts/)\",\n    );\n  } catch (err) {\n    logger.warn(\n      `[eliza] OptimizedPromptService registration failed: ${err instanceof Error ? err.message : String(err)}`,\n    );\n  }\n\n  const skipCronRegistration = trainingCronRegistrationDisabled();\n  if (skipCronRegistration) {\n    logger.info(\"[eliza] Training cron registration skipped\");\n  } else {\n    await registerTrajectoryExportCron(runtime);\n    await registerSkillScoringCron(runtime);\n  }\n  const triggerService = registerTrainingTriggerService(runtime);\n  logger.info(\n    skipCronRegistration\n      ? \"[eliza] Registered Track C auto-train trigger service\"\n      : \"[eliza] Registered Track C training crons + auto-train trigger service\",\n  );\n\n  void bootstrapOptimizationFromAccumulatedTrajectories(runtime, triggerService)\n    .then((fired) => {\n      if (fired.length > 0) {\n        logger.info(\n          `[eliza] Bootstrapped prompt optimization for ${fired.join(\", \")}`,\n        );\n      }\n    })\n    .catch((err) => {\n      logger.error(\n        `[eliza] bootstrapOptimizationFromAccumulatedTrajectories failed: ${err instanceof Error ? (err.stack ?? err.message) : String(err)}`,\n      );\n    });\n}\n"],"mappings":"AACA,SAAS,QAAQ,8BAA8B;AAC/C,SAAS,gCAAgC;AACzC,SAAS,oCAAoC;AAC7C;AAAA,EACE;AAAA,EACA;AAAA,OACK;AAEP,SAAS,mCAA4C;AACnD,QAAM,MAAM,QAAQ,IAAI;AACxB,MAAI,CAAC,KAAK;AACR,WAAO;AAAA,EACT;AACA,SAAO,CAAC,KAAK,QAAQ,KAAK,EAAE,SAAS,IAAI,KAAK,EAAE,YAAY,CAAC;AAC/D;AAEA,eAAsB,6BACpB,SACe;AAMf,MAAI;AACF,QAAI,yBAAyC,MAAM,QAChD,sBAAsB,uBAAuB,WAAW,EACxD,MAAM,MAAM,IAAI;AACnB,QAAI,CAAC,wBAAwB;AAC3B,YAAM,QAAQ;AAAA,QACZ;AAAA,MAGF;AACA,+BAAyB,MAAM,QAAQ;AAAA,QACrC,uBAAuB;AAAA,MACzB;AAAA,IACF;AACA,WAAO;AAAA,MACL;AAAA,IACF;AAAA,EACF,SAAS,KAAK;AACZ,WAAO;AAAA,MACL,uDAAuD,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,IACzG;AAAA,EACF;AAEA,QAAM,uBAAuB,iCAAiC;AAC9D,MAAI,sBAAsB;AACxB,WAAO,KAAK,4CAA4C;AAAA,EAC1D,OAAO;AACL,UAAM,6BAA6B,OAAO;AAC1C,UAAM,yBAAyB,OAAO;AAAA,EACxC;AACA,QAAM,iBAAiB,+BAA+B,OAAO;AAC7D,SAAO;AAAA,IACL,uBACI,0DACA;AAAA,EACN;AAEA,OAAK,iDAAiD,SAAS,cAAc,EAC1E,KAAK,CAAC,UAAU;AACf,QAAI,MAAM,SAAS,GAAG;AACpB,aAAO;AAAA,QACL,gDAAgD,MAAM,KAAK,IAAI,CAAC;AAAA,MAClE;AAAA,IACF;AAAA,EACF,CAAC,EACA,MAAM,CAAC,QAAQ;AACd,WAAO;AAAA,MACL,oEAAoE,eAAe,QAAS,IAAI,SAAS,IAAI,UAAW,OAAO,GAAG,CAAC;AAAA,IACrI;AAAA,EACF,CAAC;AACL;","names":[]}

package/dist/register-terminal-view.d.ts

@@ -0,0 +1,15 @@
+/**
+ * Register the training view for terminal rendering.
+ *
+ * The agent terminal mounts plugin views by id from the `@elizaos/tui` terminal
+ * registry. This makes the `training` view's `tui` modality render for real in
+ * the terminal (the unified {@link FineTuningSpatialView}) rather than only
+ * navigating a GUI shell. A module-level snapshot lets a host push the live
+ * training state; with no host it defaults to an empty snapshot.
+ */
+import { type FineTuningSnapshot } from "./ui/FineTuningSpatialView.tsx";
+/** Update the snapshot the registered terminal view renders from. */
+export declare function setFineTuningTerminalSnapshot(next: FineTuningSnapshot): void;
+/** Register the training terminal view; returns an unregister function. */
+export declare function registerFineTuningTerminalView(): () => void;
+//# sourceMappingURL=register-terminal-view.d.ts.map

package/dist/register-terminal-view.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"register-terminal-view.d.ts","sourceRoot":"","sources":["../src/register-terminal-view.tsx"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAIH,OAAO,EACL,KAAK,kBAAkB,EAExB,MAAM,gCAAgC,CAAC;AAexC,qEAAqE;AACrE,wBAAgB,6BAA6B,CAAC,IAAI,EAAE,kBAAkB,GAAG,IAAI,CAE5E;AAED,2EAA2E;AAC3E,wBAAgB,8BAA8B,IAAI,MAAM,IAAI,CAI3D"}

package/dist/register-terminal-view.js

@@ -0,0 +1,31 @@
+import { registerSpatialTerminalView } from "@elizaos/ui/spatial/tui";
+import { createElement } from "react";
+import {
+  FineTuningSpatialView
+} from "./ui/FineTuningSpatialView.js";
+const EMPTY = {
+  runtimeAvailable: false,
+  runningJobs: 0,
+  queuedJobs: 0,
+  completedJobs: 0,
+  failedJobs: 0,
+  jobs: [],
+  models: 0,
+  datasets: 0,
+  trajectoryCount: 0
+};
+let current = EMPTY;
+function setFineTuningTerminalSnapshot(next) {
+  current = next;
+}
+function registerFineTuningTerminalView() {
+  return registerSpatialTerminalView(
+    "training",
+    () => createElement(FineTuningSpatialView, { snapshot: current })
+  );
+}
+export {
+  registerFineTuningTerminalView,
+  setFineTuningTerminalSnapshot
+};
+//# sourceMappingURL=register-terminal-view.js.map

package/dist/register-terminal-view.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/register-terminal-view.tsx"],"sourcesContent":["/**\n * Register the training view for terminal rendering.\n *\n * The agent terminal mounts plugin views by id from the `@elizaos/tui` terminal\n * registry. This makes the `training` view's `tui` modality render for real in\n * the terminal (the unified {@link FineTuningSpatialView}) rather than only\n * navigating a GUI shell. A module-level snapshot lets a host push the live\n * training state; with no host it defaults to an empty snapshot.\n */\n\nimport { registerSpatialTerminalView } from \"@elizaos/ui/spatial/tui\";\nimport { createElement } from \"react\";\nimport {\n  type FineTuningSnapshot,\n  FineTuningSpatialView,\n} from \"./ui/FineTuningSpatialView.js\";\n\nconst EMPTY: FineTuningSnapshot = {\n  runtimeAvailable: false,\n  runningJobs: 0,\n  queuedJobs: 0,\n  completedJobs: 0,\n  failedJobs: 0,\n  jobs: [],\n  models: 0,\n  datasets: 0,\n  trajectoryCount: 0,\n};\nlet current: FineTuningSnapshot = EMPTY;\n\n/** Update the snapshot the registered terminal view renders from. */\nexport function setFineTuningTerminalSnapshot(next: FineTuningSnapshot): void {\n  current = next;\n}\n\n/** Register the training terminal view; returns an unregister function. */\nexport function registerFineTuningTerminalView(): () => void {\n  return registerSpatialTerminalView(\"training\", () =>\n    createElement(FineTuningSpatialView, { snapshot: current }),\n  );\n}\n"],"mappings":"AAUA,SAAS,mCAAmC;AAC5C,SAAS,qBAAqB;AAC9B;AAAA,EAEE;AAAA,OACK;AAEP,MAAM,QAA4B;AAAA,EAChC,kBAAkB;AAAA,EAClB,aAAa;AAAA,EACb,YAAY;AAAA,EACZ,eAAe;AAAA,EACf,YAAY;AAAA,EACZ,MAAM,CAAC;AAAA,EACP,QAAQ;AAAA,EACR,UAAU;AAAA,EACV,iBAAiB;AACnB;AACA,IAAI,UAA8B;AAG3B,SAAS,8BAA8B,MAAgC;AAC5E,YAAU;AACZ;AAGO,SAAS,iCAA6C;AAC3D,SAAO;AAAA,IAA4B;AAAA,IAAY,MAC7C,cAAc,uBAAuB,EAAE,UAAU,QAAQ,CAAC;AAAA,EAC5D;AACF;","names":[]}

package/dist/routes/experience-routes.d.ts

@@ -0,0 +1,21 @@
+/**
+ * Experience HTTP routes.
+ *
+ * Migrated from `packages/agent/src/api/experience-routes.ts` so the experience
+ * service surface lives next to the rest of the training/trajectory plumbing.
+ *
+ * The runtime mounts these routes through the plugin route registry; see
+ * `setup-routes.ts` for the registered Plugin and the `rawPath: true` paths.
+ */
+import type { AgentRuntime } from "@elizaos/core";
+import type { RouteRequestContext } from "@elizaos/shared";
+export interface ExperienceRouteContext extends RouteRequestContext {
+    runtime: AgentRuntime | null;
+    url: URL;
+}
+export declare function handleExperienceRoutes(ctx: ExperienceRouteContext): Promise<boolean>;
+export declare const EXPERIENCE_ROUTE_PATHS: Array<{
+    type: string;
+    path: string;
+}>;
+//# sourceMappingURL=experience-routes.d.ts.map

package/dist/routes/experience-routes.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"experience-routes.d.ts","sourceRoot":"","sources":["../../src/routes/experience-routes.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAQ,MAAM,eAAe,CAAC;AACxD,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,iBAAiB,CAAC;AA0G3D,MAAM,WAAW,sBAAuB,SAAQ,mBAAmB;IACjE,OAAO,EAAE,YAAY,GAAG,IAAI,CAAC;IAC7B,GAAG,EAAE,GAAG,CAAC;CACV;AA8ZD,wBAAsB,sBAAsB,CAC1C,GAAG,EAAE,sBAAsB,GAC1B,OAAO,CAAC,OAAO,CAAC,CAoLlB;AAED,eAAO,MAAM,sBAAsB,EAAE,KAAK,CAAC;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,CAiBxE,CAAC"}