@warlock.js/ai 4.4.0 → 4.6.0

package/esm/vcr/vcr.mjs

@@ -0,0 +1,248 @@
+import { redact } from "../security/redact.mjs";
+import { emptyCassette, loadCassette, saveCassette } from "./cassette-io.mjs";
+import { VcrCassetteMissError } from "./errors.mjs";
+import { DEFAULT_HASH_OPTIONS, hashRequest } from "./hash-request.mjs";
+
+//#region ../@warlock.js/ai/src/vcr/vcr.ts
+/**
+* Internal decorator that wraps an inner `ModelContract`, intercepting only
+* `complete()`/`stream()` and delegating every identity getter to the inner
+* model. Drives the record/replay state machine over a single in-memory
+* {@link Cassette}.
+*
+* **Why a class.** It holds mutable per-instance state (the loaded cassette,
+* the dirty flag, the load promise) behind a stable `ModelContract` surface;
+* the public API is the `vcr()` factory, never `new`.
+*/
+var Vcr = class {
+	constructor(inner, options) {
+		this.inner = inner;
+		this.dirty = false;
+		this.warnedVerbatim = false;
+		this.path = options.path;
+		this.mode = options.mode ?? "auto";
+		this.hashOptions = options.hashOptions ?? DEFAULT_HASH_OPTIONS;
+		this.recordRequest = options.recordRequest ?? "verbatim";
+		this.redactRequestHook = options.redactRequest;
+		this.redactResponseHook = options.redactResponse;
+		this.redactErrorHook = options.redactError;
+		this.loadedCassette = emptyCassette(inner.name, inner.provider);
+	}
+	/** Inner model identifier — delegated verbatim. */
+	get name() {
+		return this.inner.name;
+	}
+	/** Inner provider — delegated verbatim. */
+	get provider() {
+		return this.inner.provider;
+	}
+	/** Inner capability flags — delegated verbatim. */
+	get capabilities() {
+		return this.inner.capabilities;
+	}
+	/** Inner pricing — delegated verbatim so cost accounting is unchanged. */
+	get pricing() {
+		return this.inner.pricing;
+	}
+	/** Loaded/recorded cassette, exposed for assertions. */
+	get cassette() {
+		return this.loadedCassette;
+	}
+	/**
+	* Load the on-disk cassette exactly once. Pure `record` mode skips the
+	* read — it always writes fresh — but the in-memory cassette still starts
+	* empty so a record run never accidentally replays a stale entry.
+	*/
+	async ensureLoaded() {
+		if (this.loadPromise) return this.loadPromise;
+		this.loadPromise = this.mode === "record" ? Promise.resolve() : (async () => {
+			this.loadedCassette = await loadCassette(this.path, this.inner.name, this.inner.provider);
+		})();
+		return this.loadPromise;
+	}
+	/** Find a recorded entry whose hash matches the current request. */
+	findEntry(hash) {
+		return this.loadedCassette.entries.find((entry) => entry.requestHash === hash);
+	}
+	/** Re-throw a recorded error by reconstructing a plain `Error`. */
+	throwRecordedError(entry) {
+		const error = new Error(entry.error?.message ?? "Recorded error");
+		error.name = entry.error?.name ?? "Error";
+		throw error;
+	}
+	/**
+	* Non-streaming call. In `replay` a miss throws; in `auto`/`record` a miss
+	* calls the inner model and records the outcome (response or error).
+	*/
+	async complete(messages, options) {
+		await this.ensureLoaded();
+		const hash = hashRequest(messages, options, this.hashOptions);
+		if (this.mode !== "record") {
+			const entry = this.findEntry(hash);
+			if (entry) {
+				if (entry.error) this.throwRecordedError(entry);
+				if (entry.response) return entry.response;
+			}
+			if (this.mode === "replay") throw new VcrCassetteMissError(`No cassette entry for this request (model "${this.inner.name}", hash ${hash}).`, {
+				requestHash: hash,
+				path: this.path
+			});
+		}
+		try {
+			const response = await this.inner.complete(messages, options);
+			this.record({
+				requestHash: hash,
+				request: {
+					messages,
+					options
+				},
+				response
+			});
+			return response;
+		} catch (error) {
+			this.record({
+				requestHash: hash,
+				request: {
+					messages,
+					options
+				},
+				error: {
+					name: error.name,
+					message: error.message
+				}
+			});
+			throw error;
+		}
+	}
+	/**
+	* Streaming call. On replay the stored `chunks` are re-yielded in order
+	* (reproducing the `delta`/`tool-call`/`done` sequence) or the stored
+	* error is re-thrown. On record the inner stream is buffered into
+	* `chunks[]` while being re-emitted, then recorded once exhausted.
+	*/
+	async *stream(messages, options) {
+		await this.ensureLoaded();
+		const hash = hashRequest(messages, options, this.hashOptions);
+		if (this.mode !== "record") {
+			const entry = this.findEntry(hash);
+			if (entry) {
+				if (entry.error) this.throwRecordedError(entry);
+				if (entry.chunks) {
+					for (const chunk of entry.chunks) yield chunk;
+					return;
+				}
+			}
+			if (this.mode === "replay") throw new VcrCassetteMissError(`No cassette entry for this request (model "${this.inner.name}", hash ${hash}).`, {
+				requestHash: hash,
+				path: this.path
+			});
+		}
+		const chunks = [];
+		try {
+			for await (const chunk of this.inner.stream(messages, options)) {
+				chunks.push(chunk);
+				yield chunk;
+			}
+		} catch (error) {
+			this.record({
+				requestHash: hash,
+				request: {
+					messages,
+					options
+				},
+				error: {
+					name: error.name,
+					message: error.message
+				}
+			});
+			throw error;
+		}
+		this.record({
+			requestHash: hash,
+			request: {
+				messages,
+				options
+			},
+			chunks
+		});
+	}
+	/**
+	* Append an entry to the in-memory cassette and mark it dirty, applying
+	* the configured request/response/error redaction first (S2). Pure
+	* `replay` never reaches this path, so no replay run is ever dirtied.
+	*/
+	record(entry) {
+		this.loadedCassette.entries.push(this.applyRedaction(entry));
+		this.dirty = true;
+		this.maybeWarnVerbatim();
+	}
+	/**
+	* Apply the persisted-body privacy controls to an entry before it is
+	* stored. The request body follows `recordRequest`; response/error
+	* redactors are applied only when supplied. Replay matching is by the
+	* recomputed hash (kept verbatim), so none of this affects replay.
+	*/
+	applyRedaction(entry) {
+		const out = {
+			requestHash: entry.requestHash,
+			request: entry.request
+		};
+		if (this.recordRequest === "hash-only") out.request = { messages: [] };
+		else if (this.recordRequest === "redacted") out.request = this.redactRequestHook ? this.redactRequestHook(entry.request) : redact(entry.request);
+		if (entry.response) out.response = this.redactResponseHook ? this.redactResponseHook(entry.response) : entry.response;
+		if (entry.chunks) out.chunks = entry.chunks;
+		if (entry.error) out.error = this.redactErrorHook ? this.redactErrorHook(entry.error) : entry.error;
+		return out;
+	}
+	/**
+	* Warn once (outside tests) when the cassette is recording verbatim
+	* request bodies — they may carry prompts, tool args, and PII, so the
+	* file is not safe to commit until sanitized.
+	*/
+	maybeWarnVerbatim() {
+		if (this.warnedVerbatim || this.recordRequest !== "verbatim") return;
+		if (process.env.VITEST || process.env.NODE_ENV === "test") return;
+		this.warnedVerbatim = true;
+		console.warn(`[warlock-ai] VCR is recording verbatim request bodies to "${this.path}" — prompts, tool args, and any PII are stored unredacted. Sanitize before committing, or set recordRequest: "redacted" | "hash-only".`);
+	}
+	/**
+	* Flush newly recorded entries to `path`. No-op when nothing was recorded
+	* (pure replay, or a record/auto run that only ever hit cached entries).
+	*/
+	async save() {
+		if (!this.dirty) return;
+		await saveCassette(this.path, this.loadedCassette);
+		this.dirty = false;
+	}
+};
+/**
+* Wrap any `ModelContract` in a record/replay decorator backed by a JSON
+* cassette on disk.
+*
+* **What it does.** Intercepts only `complete()`/`stream()` — the single
+* seam every agent trip funnels through — and delegates `name`, `provider`,
+* `capabilities`, and `pricing` to the inner model untouched. On a call it
+* computes a stable hash over `{ messages, picked options }` and, depending
+* on `mode`:
+*
+* - **`record`** — always calls the inner model and appends a cassette entry.
+* - **`replay`** — returns the matching entry (or re-yields its chunks /
+*   re-throws its error); a miss throws `VcrCassetteMissError`, never a live
+*   call.
+* - **`auto`** (default) — replays a hit, records a miss.
+*
+* Composes *below* `fallbackModel` and works with any adapter because it
+* depends only on `ModelContract`. Call `save()` to flush new entries.
+*
+* @example
+* const model = vcr(liveModel, { path: "./cassettes/support.json" });
+* const response = await model.complete(messages);
+* await model.save(); // first run records; later runs replay deterministically.
+*/
+function vcr(model, options) {
+	return new Vcr(model, options);
+}
+
+//#endregion
+export { vcr };
+//# sourceMappingURL=vcr.mjs.map

package/esm/vcr/vcr.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"vcr.mjs","names":[],"sources":["../../../../../../../@warlock.js/ai/src/vcr/vcr.ts"],"sourcesContent":["import type { Message } from \"../contracts/conversation-message.type\";\nimport type {\n  ModelCallOptions,\n  ModelCapabilities,\n  ModelContract,\n  ModelResponse,\n  ModelStreamChunk,\n} from \"../contracts/model.contract\";\nimport type { ModelPricing } from \"../contracts/result/model-pricing.type\";\nimport { redact } from \"../security/redact\";\nimport { emptyCassette, loadCassette, saveCassette } from \"./cassette-io\";\nimport { VcrCassetteMissError } from \"./errors\";\nimport { DEFAULT_HASH_OPTIONS, hashRequest } from \"./hash-request\";\nimport type { Cassette, CassetteEntry, VcrMode, VcrModel, VcrOptions } from \"./vcr.type\";\n\n/**\n * Internal decorator that wraps an inner `ModelContract`, intercepting only\n * `complete()`/`stream()` and delegating every identity getter to the inner\n * model. Drives the record/replay state machine over a single in-memory\n * {@link Cassette}.\n *\n * **Why a class.** It holds mutable per-instance state (the loaded cassette,\n * the dirty flag, the load promise) behind a stable `ModelContract` surface;\n * the public API is the `vcr()` factory, never `new`.\n */\nclass Vcr implements VcrModel {\n  private readonly mode: VcrMode;\n  private readonly path: string;\n  private readonly hashOptions: readonly string[];\n\n  /** Loaded + newly recorded entries. Mutated in place as we record. */\n  private loadedCassette: Cassette;\n\n  /** Set when an entry is recorded so `save()` knows there's work to flush. */\n  private dirty = false;\n\n  /** One-shot lazy load of the on-disk cassette, shared across calls. */\n  private loadPromise: Promise<void> | undefined;\n\n  /** Persisted-body privacy controls (S2). */\n  private readonly recordRequest: NonNullable<VcrOptions[\"recordRequest\"]>;\n  private readonly redactRequestHook: VcrOptions[\"redactRequest\"];\n  private readonly redactResponseHook: VcrOptions[\"redactResponse\"];\n  private readonly redactErrorHook: VcrOptions[\"redactError\"];\n\n  /** Verbatim-recording warning fires at most once per instance. */\n  private warnedVerbatim = false;\n\n  public constructor(\n    private readonly inner: ModelContract,\n    options: VcrOptions,\n  ) {\n    this.path = options.path;\n    this.mode = options.mode ?? \"auto\";\n    this.hashOptions = options.hashOptions ?? DEFAULT_HASH_OPTIONS;\n    this.recordRequest = options.recordRequest ?? \"verbatim\";\n    this.redactRequestHook = options.redactRequest;\n    this.redactResponseHook = options.redactResponse;\n    this.redactErrorHook = options.redactError;\n    this.loadedCassette = emptyCassette(inner.name, inner.provider);\n  }\n\n  /** Inner model identifier — delegated verbatim. */\n  public get name(): string {\n    return this.inner.name;\n  }\n\n  /** Inner provider — delegated verbatim. */\n  public get provider(): string {\n    return this.inner.provider;\n  }\n\n  /** Inner capability flags — delegated verbatim. */\n  public get capabilities(): ModelCapabilities | undefined {\n    return this.inner.capabilities;\n  }\n\n  /** Inner pricing — delegated verbatim so cost accounting is unchanged. */\n  public get pricing(): ModelPricing | undefined {\n    return this.inner.pricing;\n  }\n\n  /** Loaded/recorded cassette, exposed for assertions. */\n  public get cassette(): Cassette {\n    return this.loadedCassette;\n  }\n\n  /**\n   * Load the on-disk cassette exactly once. Pure `record` mode skips the\n   * read — it always writes fresh — but the in-memory cassette still starts\n   * empty so a record run never accidentally replays a stale entry.\n   */\n  private async ensureLoaded(): Promise<void> {\n    if (this.loadPromise) {\n      return this.loadPromise;\n    }\n\n    this.loadPromise =\n      this.mode === \"record\"\n        ? Promise.resolve()\n        : (async () => {\n            this.loadedCassette = await loadCassette(\n              this.path,\n              this.inner.name,\n              this.inner.provider,\n            );\n          })();\n\n    return this.loadPromise;\n  }\n\n  /** Find a recorded entry whose hash matches the current request. */\n  private findEntry(hash: string): CassetteEntry | undefined {\n    return this.loadedCassette.entries.find((entry) => entry.requestHash === hash);\n  }\n\n  /** Re-throw a recorded error by reconstructing a plain `Error`. */\n  private throwRecordedError(entry: CassetteEntry): never {\n    const error = new Error(entry.error?.message ?? \"Recorded error\");\n\n    error.name = entry.error?.name ?? \"Error\";\n\n    throw error;\n  }\n\n  /**\n   * Non-streaming call. In `replay` a miss throws; in `auto`/`record` a miss\n   * calls the inner model and records the outcome (response or error).\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    await this.ensureLoaded();\n\n    const hash = hashRequest(messages, options, this.hashOptions);\n\n    if (this.mode !== \"record\") {\n      const entry = this.findEntry(hash);\n\n      if (entry) {\n        if (entry.error) {\n          this.throwRecordedError(entry);\n        }\n\n        if (entry.response) {\n          return entry.response;\n        }\n      }\n\n      if (this.mode === \"replay\") {\n        throw new VcrCassetteMissError(\n          `No cassette entry for this request (model \"${this.inner.name}\", hash ${hash}).`,\n          { requestHash: hash, path: this.path },\n        );\n      }\n    }\n\n    try {\n      const response = await this.inner.complete(messages, options);\n\n      this.record({ requestHash: hash, request: { messages, options }, response });\n\n      return response;\n    } catch (error) {\n      this.record({\n        requestHash: hash,\n        request: { messages, options },\n        error: { name: (error as Error).name, message: (error as Error).message },\n      });\n\n      throw error;\n    }\n  }\n\n  /**\n   * Streaming call. On replay the stored `chunks` are re-yielded in order\n   * (reproducing the `delta`/`tool-call`/`done` sequence) or the stored\n   * error is re-thrown. On record the inner stream is buffered into\n   * `chunks[]` while being re-emitted, then recorded once exhausted.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    await this.ensureLoaded();\n\n    const hash = hashRequest(messages, options, this.hashOptions);\n\n    if (this.mode !== \"record\") {\n      const entry = this.findEntry(hash);\n\n      if (entry) {\n        if (entry.error) {\n          this.throwRecordedError(entry);\n        }\n\n        if (entry.chunks) {\n          for (const chunk of entry.chunks) {\n            yield chunk;\n          }\n\n          return;\n        }\n      }\n\n      if (this.mode === \"replay\") {\n        throw new VcrCassetteMissError(\n          `No cassette entry for this request (model \"${this.inner.name}\", hash ${hash}).`,\n          { requestHash: hash, path: this.path },\n        );\n      }\n    }\n\n    const chunks: ModelStreamChunk[] = [];\n\n    try {\n      for await (const chunk of this.inner.stream(messages, options)) {\n        chunks.push(chunk);\n\n        yield chunk;\n      }\n    } catch (error) {\n      this.record({\n        requestHash: hash,\n        request: { messages, options },\n        error: { name: (error as Error).name, message: (error as Error).message },\n      });\n\n      throw error;\n    }\n\n    this.record({ requestHash: hash, request: { messages, options }, chunks });\n  }\n\n  /**\n   * Append an entry to the in-memory cassette and mark it dirty, applying\n   * the configured request/response/error redaction first (S2). Pure\n   * `replay` never reaches this path, so no replay run is ever dirtied.\n   */\n  private record(entry: CassetteEntry): void {\n    this.loadedCassette.entries.push(this.applyRedaction(entry));\n    this.dirty = true;\n    this.maybeWarnVerbatim();\n  }\n\n  /**\n   * Apply the persisted-body privacy controls to an entry before it is\n   * stored. The request body follows `recordRequest`; response/error\n   * redactors are applied only when supplied. Replay matching is by the\n   * recomputed hash (kept verbatim), so none of this affects replay.\n   */\n  private applyRedaction(entry: CassetteEntry): CassetteEntry {\n    const out: CassetteEntry = {\n      requestHash: entry.requestHash,\n      request: entry.request,\n    };\n\n    if (this.recordRequest === \"hash-only\") {\n      out.request = { messages: [] };\n    } else if (this.recordRequest === \"redacted\") {\n      out.request = this.redactRequestHook\n        ? this.redactRequestHook(entry.request)\n        : redact(entry.request);\n    }\n\n    if (entry.response) {\n      out.response = this.redactResponseHook\n        ? this.redactResponseHook(entry.response)\n        : entry.response;\n    }\n    if (entry.chunks) {\n      out.chunks = entry.chunks;\n    }\n    if (entry.error) {\n      out.error = this.redactErrorHook\n        ? this.redactErrorHook(entry.error)\n        : entry.error;\n    }\n\n    return out;\n  }\n\n  /**\n   * Warn once (outside tests) when the cassette is recording verbatim\n   * request bodies — they may carry prompts, tool args, and PII, so the\n   * file is not safe to commit until sanitized.\n   */\n  private maybeWarnVerbatim(): void {\n    if (this.warnedVerbatim || this.recordRequest !== \"verbatim\") return;\n    if (process.env.VITEST || process.env.NODE_ENV === \"test\") return;\n\n    this.warnedVerbatim = true;\n    console.warn(\n      `[warlock-ai] VCR is recording verbatim request bodies to \"${this.path}\" — prompts, tool args, and any PII are stored unredacted. ` +\n        'Sanitize before committing, or set recordRequest: \"redacted\" | \"hash-only\".',\n    );\n  }\n\n  /**\n   * Flush newly recorded entries to `path`. No-op when nothing was recorded\n   * (pure replay, or a record/auto run that only ever hit cached entries).\n   */\n  public async save(): Promise<void> {\n    if (!this.dirty) {\n      return;\n    }\n\n    await saveCassette(this.path, this.loadedCassette);\n    this.dirty = false;\n  }\n}\n\n/**\n * Wrap any `ModelContract` in a record/replay decorator backed by a JSON\n * cassette on disk.\n *\n * **What it does.** Intercepts only `complete()`/`stream()` — the single\n * seam every agent trip funnels through — and delegates `name`, `provider`,\n * `capabilities`, and `pricing` to the inner model untouched. On a call it\n * computes a stable hash over `{ messages, picked options }` and, depending\n * on `mode`:\n *\n * - **`record`** — always calls the inner model and appends a cassette entry.\n * - **`replay`** — returns the matching entry (or re-yields its chunks /\n *   re-throws its error); a miss throws `VcrCassetteMissError`, never a live\n *   call.\n * - **`auto`** (default) — replays a hit, records a miss.\n *\n * Composes *below* `fallbackModel` and works with any adapter because it\n * depends only on `ModelContract`. Call `save()` to flush new entries.\n *\n * @example\n * const model = vcr(liveModel, { path: \"./cassettes/support.json\" });\n * const response = await model.complete(messages);\n * await model.save(); // first run records; later runs replay deterministically.\n */\nexport function vcr(model: ModelContract, options: VcrOptions): VcrModel {\n  return new Vcr(model, options);\n}\n"],"mappings":";;;;;;;;;;;;;;;;AAyBA,IAAM,MAAN,MAA8B;CAuB5B,AAAO,YACL,AAAiB,OACjB,SACA;EAFiB;eAfH;wBAYS;EAMvB,KAAK,OAAO,QAAQ;EACpB,KAAK,OAAO,QAAQ,QAAQ;EAC5B,KAAK,cAAc,QAAQ,eAAe;EAC1C,KAAK,gBAAgB,QAAQ,iBAAiB;EAC9C,KAAK,oBAAoB,QAAQ;EACjC,KAAK,qBAAqB,QAAQ;EAClC,KAAK,kBAAkB,QAAQ;EAC/B,KAAK,iBAAiB,cAAc,MAAM,MAAM,MAAM,QAAQ;CAChE;;CAGA,IAAW,OAAe;EACxB,OAAO,KAAK,MAAM;CACpB;;CAGA,IAAW,WAAmB;EAC5B,OAAO,KAAK,MAAM;CACpB;;CAGA,IAAW,eAA8C;EACvD,OAAO,KAAK,MAAM;CACpB;;CAGA,IAAW,UAAoC;EAC7C,OAAO,KAAK,MAAM;CACpB;;CAGA,IAAW,WAAqB;EAC9B,OAAO,KAAK;CACd;;;;;;CAOA,MAAc,eAA8B;EAC1C,IAAI,KAAK,aACP,OAAO,KAAK;EAGd,KAAK,cACH,KAAK,SAAS,WACV,QAAQ,QAAQ,KACf,YAAY;GACX,KAAK,iBAAiB,MAAM,aAC1B,KAAK,MACL,KAAK,MAAM,MACX,KAAK,MAAM,QACb;EACF,EAAC,CAAE;EAET,OAAO,KAAK;CACd;;CAGA,AAAQ,UAAU,MAAyC;EACzD,OAAO,KAAK,eAAe,QAAQ,MAAM,UAAU,MAAM,gBAAgB,IAAI;CAC/E;;CAGA,AAAQ,mBAAmB,OAA6B;EACtD,MAAM,QAAQ,IAAI,MAAM,MAAM,OAAO,WAAW,gBAAgB;EAEhE,MAAM,OAAO,MAAM,OAAO,QAAQ;EAElC,MAAM;CACR;;;;;CAMA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,MAAM,KAAK,aAAa;EAExB,MAAM,OAAO,YAAY,UAAU,SAAS,KAAK,WAAW;EAE5D,IAAI,KAAK,SAAS,UAAU;GAC1B,MAAM,QAAQ,KAAK,UAAU,IAAI;GAEjC,IAAI,OAAO;IACT,IAAI,MAAM,OACR,KAAK,mBAAmB,KAAK;IAG/B,IAAI,MAAM,UACR,OAAO,MAAM;GAEjB;GAEA,IAAI,KAAK,SAAS,UAChB,MAAM,IAAI,qBACR,8CAA8C,KAAK,MAAM,KAAK,UAAU,KAAK,KAC7E;IAAE,aAAa;IAAM,MAAM,KAAK;GAAK,CACvC;EAEJ;EAEA,IAAI;GACF,MAAM,WAAW,MAAM,KAAK,MAAM,SAAS,UAAU,OAAO;GAE5D,KAAK,OAAO;IAAE,aAAa;IAAM,SAAS;KAAE;KAAU;IAAQ;IAAG;GAAS,CAAC;GAE3E,OAAO;EACT,SAAS,OAAO;GACd,KAAK,OAAO;IACV,aAAa;IACb,SAAS;KAAE;KAAU;IAAQ;IAC7B,OAAO;KAAE,MAAO,MAAgB;KAAM,SAAU,MAAgB;IAAQ;GAC1E,CAAC;GAED,MAAM;EACR;CACF;;;;;;;CAQA,OAAc,OACZ,UACA,SACiC;EACjC,MAAM,KAAK,aAAa;EAExB,MAAM,OAAO,YAAY,UAAU,SAAS,KAAK,WAAW;EAE5D,IAAI,KAAK,SAAS,UAAU;GAC1B,MAAM,QAAQ,KAAK,UAAU,IAAI;GAEjC,IAAI,OAAO;IACT,IAAI,MAAM,OACR,KAAK,mBAAmB,KAAK;IAG/B,IAAI,MAAM,QAAQ;KAChB,KAAK,MAAM,SAAS,MAAM,QACxB,MAAM;KAGR;IACF;GACF;GAEA,IAAI,KAAK,SAAS,UAChB,MAAM,IAAI,qBACR,8CAA8C,KAAK,MAAM,KAAK,UAAU,KAAK,KAC7E;IAAE,aAAa;IAAM,MAAM,KAAK;GAAK,CACvC;EAEJ;EAEA,MAAM,SAA6B,CAAC;EAEpC,IAAI;GACF,WAAW,MAAM,SAAS,KAAK,MAAM,OAAO,UAAU,OAAO,GAAG;IAC9D,OAAO,KAAK,KAAK;IAEjB,MAAM;GACR;EACF,SAAS,OAAO;GACd,KAAK,OAAO;IACV,aAAa;IACb,SAAS;KAAE;KAAU;IAAQ;IAC7B,OAAO;KAAE,MAAO,MAAgB;KAAM,SAAU,MAAgB;IAAQ;GAC1E,CAAC;GAED,MAAM;EACR;EAEA,KAAK,OAAO;GAAE,aAAa;GAAM,SAAS;IAAE;IAAU;GAAQ;GAAG;EAAO,CAAC;CAC3E;;;;;;CAOA,AAAQ,OAAO,OAA4B;EACzC,KAAK,eAAe,QAAQ,KAAK,KAAK,eAAe,KAAK,CAAC;EAC3D,KAAK,QAAQ;EACb,KAAK,kBAAkB;CACzB;;;;;;;CAQA,AAAQ,eAAe,OAAqC;EAC1D,MAAM,MAAqB;GACzB,aAAa,MAAM;GACnB,SAAS,MAAM;EACjB;EAEA,IAAI,KAAK,kBAAkB,aACzB,IAAI,UAAU,EAAE,UAAU,CAAC,EAAE;OACxB,IAAI,KAAK,kBAAkB,YAChC,IAAI,UAAU,KAAK,oBACf,KAAK,kBAAkB,MAAM,OAAO,IACpC,OAAO,MAAM,OAAO;EAG1B,IAAI,MAAM,UACR,IAAI,WAAW,KAAK,qBAChB,KAAK,mBAAmB,MAAM,QAAQ,IACtC,MAAM;EAEZ,IAAI,MAAM,QACR,IAAI,SAAS,MAAM;EAErB,IAAI,MAAM,OACR,IAAI,QAAQ,KAAK,kBACb,KAAK,gBAAgB,MAAM,KAAK,IAChC,MAAM;EAGZ,OAAO;CACT;;;;;;CAOA,AAAQ,oBAA0B;EAChC,IAAI,KAAK,kBAAkB,KAAK,kBAAkB,YAAY;EAC9D,IAAI,QAAQ,IAAI,UAAU,QAAQ,IAAI,aAAa,QAAQ;EAE3D,KAAK,iBAAiB;EACtB,QAAQ,KACN,6DAA6D,KAAK,KAAK,uIAEzE;CACF;;;;;CAMA,MAAa,OAAsB;EACjC,IAAI,CAAC,KAAK,OACR;EAGF,MAAM,aAAa,KAAK,MAAM,KAAK,cAAc;EACjD,KAAK,QAAQ;CACf;AACF;;;;;;;;;;;;;;;;;;;;;;;;;AA0BA,SAAgB,IAAI,OAAsB,SAA+B;CACvE,OAAO,IAAI,IAAI,OAAO,OAAO;AAC/B"}

package/esm/vcr/vcr.type.d.mts

@@ -0,0 +1,118 @@
+import { Message } from "../contracts/conversation-message.type.mjs";
+import { ModelCallOptions, ModelContract, ModelResponse, ModelStreamChunk } from "../contracts/model.contract.mjs";
+
+//#region ../@warlock.js/ai/src/vcr/vcr.type.d.ts
+/**
+ * Record/replay mode for a {@link VcrModel}.
+ *
+ * - **`record`** — always call the inner model and append the result to
+ *   the cassette. Never replays. Use to (re)capture a fresh cassette.
+ * - **`replay`** — never call the inner model. A cassette hit returns the
+ *   stored response/chunks/error; a miss throws `VcrCassetteMissError`
+ *   (never a silent live call). Use in CI for deterministic, offline tests.
+ * - **`auto`** (default) — replay on a cassette hit, record on a miss.
+ *   The friendliest mode for local dev: records once, replays thereafter.
+ */
+type VcrMode = "record" | "replay" | "auto";
+/**
+ * A single recorded request → response pair on a {@link Cassette}.
+ *
+ * Exactly one of `response` / `chunks` / `error` is populated, mirroring
+ * the three outcomes of a model call: a non-streaming reply, a streamed
+ * chunk sequence, or a thrown provider error.
+ */
+type CassetteEntry = {
+  /** Stable hash of the normalized request (messages + hashed options). */requestHash: string;
+  /**
+   * What was sent — stored verbatim for human readability and so the
+   * cassette can be re-hashed if the hashing format ever changes.
+   */
+  request: {
+    messages: Message[];
+    options?: ModelCallOptions;
+  }; /** A non-streaming reply, recorded from `complete()`. */
+  response?: ModelResponse; /** A streaming reply, captured from `stream()` as the ordered chunk list. */
+  chunks?: ModelStreamChunk[]; /** A thrown provider error, replayed by re-throwing a reconstructed `Error`. */
+  error?: {
+    name: string;
+    message: string;
+  };
+};
+/**
+ * On-disk cassette format: the inner model identity plus every recorded
+ * {@link CassetteEntry}. Serialized as JSON to {@link VcrOptions.path}.
+ */
+type Cassette = {
+  /** Cassette schema version. Bump only on a breaking format change. */version: 1; /** Inner `model.name` at record time — informational. */
+  model: string; /** Inner `model.provider` at record time — informational. */
+  provider: string; /** Recorded request → response pairs, in capture order. */
+  entries: CassetteEntry[];
+};
+/**
+ * Options for {@link vcr}.
+ */
+type VcrOptions = {
+  /** Cassette file path (JSON). Read on construct, written on `save()`. */path: string; /** Record/replay behavior. Default `"auto"`. */
+  mode?: VcrMode;
+  /**
+   * Fields of `ModelCallOptions` to include in the request hash. Default
+   * `["temperature","maxTokens","responseSchema","tools","reasoning"]`.
+   * `signal` and unknown provider keys are always excluded so an
+   * otherwise-identical logical call still matches. `tools` are hashed by
+   * name + description + input-schema shape, not object identity.
+   */
+  hashOptions?: string[];
+  /**
+   * How the request BODY is persisted (S2). Replay matching is by the
+   * recomputed hash, never the stored body — so privacy modes don't break
+   * replay. Default `"verbatim"`.
+   *
+   * - `"verbatim"` — store the full request (prompts, tool args, options)
+   *   for human-readable diffs and re-hashing. Records a loud one-time
+   *   warning (outside tests) since the body may carry PII / secrets.
+   * - `"redacted"` — run the request through {@link redactRequest}, or —
+   *   without one — the key-based `redact()` (strips structured secrets
+   *   like API keys in options; free-text prompts need a custom hook).
+   * - `"hash-only"` — drop the request body entirely; keep only the hash
+   *   and the outcome. Maximum privacy; loses human readability and the
+   *   ability to re-hash.
+   */
+  recordRequest?: "verbatim" | "redacted" | "hash-only";
+  /**
+   * Custom request redactor for `recordRequest: "redacted"`. Receives the
+   * `{ messages, options }` about to be stored; return a sanitized copy.
+   * Overrides the default key-based `redact()`.
+   */
+  redactRequest?: (request: {
+    messages: Message[];
+    options?: ModelCallOptions;
+  }) => {
+    messages: Message[];
+    options?: ModelCallOptions;
+  };
+  /**
+   * Optional redactor for the recorded RESPONSE. Off by default — redacting
+   * the response changes what replay returns, so opt in only when a secret
+   * leaks into model output (e.g. an echoed token).
+   */
+  redactResponse?: (response: ModelResponse) => ModelResponse; /** Optional redactor for a recorded error's `{ name, message }`. */
+  redactError?: (error: {
+    name: string;
+    message: string;
+  }) => {
+    name: string;
+    message: string;
+  };
+};
+/**
+ * A `ModelContract` decorator returned by {@link vcr}. Delegates
+ * `name`/`provider`/`capabilities`/`pricing` to the inner model and
+ * intercepts only `complete()`/`stream()`.
+ */
+type VcrModel = ModelContract & {
+  /** Flush newly recorded entries to `path`. No-op in pure replay. */save(): Promise<void>; /** Recorded/loaded entries, exposed for assertions. */
+  readonly cassette: Cassette;
+};
+//#endregion
+export { Cassette, CassetteEntry, VcrMode, VcrModel, VcrOptions };
+//# sourceMappingURL=vcr.type.d.mts.map

package/esm/vcr/vcr.type.d.mts.map

@@ -0,0 +1 @@
+{"version":3,"file":"vcr.type.d.mts","names":[],"sources":["../../../../../../../@warlock.js/ai/src/vcr/vcr.type.ts"],"mappings":";;;;;;AAmBA;;;;AAAmB;AASnB;;;;KATY,OAAA;;;;;;;;KASA,aAAA;EASR,yEAPF,WAAA;EAUA;;;;EALA,OAAA;IACE,QAAA,EAAU,OAAA;IACV,OAAA,GAAU,gBAAA;EAAA,GASH;EANT,QAAA,GAAW,aAAA,EAcO;EAZlB,MAAA,GAAS,gBAAA,IAoBa;EAlBtB,KAAA;IACE,IAAA;IACA,OAAA;EAAA;AAAA;;AAgBoB;AAMxB;;KAdY,QAAA;EAkBH,sEAhBP,OAAA,KAgDY;EA9CZ,KAAA,UA+CuC;EA7CvC,QAAA,UAmD8C;EAjD9C,OAAA,EAAS,aAAa;AAAA;;;;KAMZ,UAAA;EA4BV,yEA1BA,IAAA,UAiCE;EA/BF,IAAA,GAAO,OAAA;EAgCL;;;;;;;EAxBF,WAAA;EA+B4B;;;;;;;;;;AAInB;AASX;;;;EA5BE,aAAA;EAgCmB;;;;;EA1BnB,aAAA,IAAiB,OAAA;IACf,QAAA,EAAU,OAAA;IACV,OAAA,GAAU,gBAAA;EAAA;IACJ,QAAA,EAAU,OAAA;IAAW,OAAA,GAAU,gBAAA;EAAA;;;;;;EAMvC,cAAA,IAAkB,QAAA,EAAU,aAAA,KAAkB,aAAA;EAE9C,WAAA,IAAe,KAAA;IAAS,IAAA;IAAc,OAAA;EAAA;IACpC,IAAA;IACA,OAAA;EAAA;AAAA;;;;;;KASQ,QAAA,GAAW,aAAA;sEAErB,IAAA,IAAQ,OAAA;WAEC,QAAA,EAAU,QAAA;AAAA"}

package/esm/workflow/as-tool.mjs

@@ -32,8 +32,8 @@ function asTool(workflowInstance, options) {
 		name: workflowInstance.name,
 		description: options.description ?? `Invoke workflow "${workflowInstance.name}" as a tool.`,
 		input: options.inputSchema,
-		execute: async (input) => {
-			const result = await workflowInstance.execute(input);
+		execute: async (input, ctx) => {
+			const result = await workflowInstance.execute(input, ctx?.signal ? { signal: ctx.signal } : void 0);
 			if (result.error) throw result.error;
 			return {
 				data: result.data,

package/esm/workflow/as-tool.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"as-tool.mjs","names":[],"sources":["../../../../../../../@warlock.js/ai/src/workflow/as-tool.ts"],"sourcesContent":["import type { StandardSchemaV1 } from \"@standard-schema/spec\";\nimport type { WorkflowInstance } from \"../contracts/workflow/workflow.contract\";\nimport { WorkflowError } from \"../errors\";\nimport { compositeAsTool, type ToolContract } from \"../tool/tool\";\n\n/**\n * Wrap a `WorkflowInstance` as a `ToolContract` so an agent can invoke\n * it inside its tool-call loop. Closes the agent-calls-workflow\n * composition gap.\n *\n * Behavior:\n * - Tool `name` mirrors `workflow.name` — workflows without a name throw\n *   `WorkflowError` at wrap time (the agent surface needs a stable id).\n * - Tool `input` is the supplied `inputSchema`; the validated value is\n *   forwarded straight to `workflow.execute(input)`.\n * - On `result.error`, the workflow error is wrapped in\n *   `ToolExecutionError` with `cause` set to the original\n *   `WorkflowError` subclass — the agent's tool-call loop sees a\n *   uniform error class regardless of which primitive failed.\n *\n * @example\n * const wf = workflow({ name: \"triage\", steps: [...] });\n * const triageTool = asTool(wf, {\n *   description: \"Run the support-ticket triage flow\",\n *   inputSchema: ticketSchema,\n * });\n * const a = ai.agent({ model, tools: [triageTool] });\n */\nexport function asTool<TInput, TOutput, TToolInput = TInput>(\n  workflowInstance: WorkflowInstance<TInput, TOutput>,\n  options: {\n    description?: string;\n    inputSchema: StandardSchemaV1<TToolInput>;\n  },\n): ToolContract<TToolInput, TOutput> {\n  if (!workflowInstance.name || typeof workflowInstance.name !== \"string\") {\n    throw new WorkflowError(\n      \"workflow.asTool(): workflow must have a `name` to be wrapped as a tool\",\n    );\n  }\n\n  return compositeAsTool<TToolInput, TOutput>({\n    name: workflowInstance.name,\n    description: options.description ?? `Invoke workflow \"${workflowInstance.name}\" as a tool.`,\n    input: options.inputSchema,\n    execute: async (input) => {\n      const result = await workflowInstance.execute(input as unknown as TInput);\n\n      if (result.error) {\n        // Throw the workflow error so the surrounding wrapper catches\n        // it and produces a `ToolExecutionError` with `cause` pointing\n        // back at the original `WorkflowError` subclass — keeps the\n        // agent's tool-call loop seeing one uniform error class\n        // regardless of which primitive failed.\n        throw result.error;\n      }\n\n      return {\n        data: result.data as TOutput,\n        usage: result.usage,\n        report: result.report,\n      };\n    },\n  });\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4BA,SAAgB,OACd,kBACA,SAImC;CACnC,IAAI,CAAC,iBAAiB,QAAQ,OAAO,iBAAiB,SAAS,UAC7D,MAAM,IAAI,cACR,wEACF;CAGF,OAAO,gBAAqC;EAC1C,MAAM,iBAAiB;EACvB,aAAa,QAAQ,eAAe,oBAAoB,iBAAiB,KAAK;EAC9E,OAAO,QAAQ;EACf,SAAS,OAAO,UAAU;GACxB,MAAM,SAAS,MAAM,iBAAiB,QAAQ,KAA0B;GAExE,IAAI,OAAO,OAMT,MAAM,OAAO;GAGf,OAAO;IACL,MAAM,OAAO;IACb,OAAO,OAAO;IACd,QAAQ,OAAO;GACjB;EACF;CACF,CAAC;AACH"}
+{"version":3,"file":"as-tool.mjs","names":[],"sources":["../../../../../../../@warlock.js/ai/src/workflow/as-tool.ts"],"sourcesContent":["import type { StandardSchemaV1 } from \"@standard-schema/spec\";\nimport type { WorkflowInstance } from \"../contracts/workflow/workflow.contract\";\nimport { WorkflowError } from \"../errors\";\nimport { compositeAsTool, type ToolContract } from \"../tool/tool\";\n\n/**\n * Wrap a `WorkflowInstance` as a `ToolContract` so an agent can invoke\n * it inside its tool-call loop. Closes the agent-calls-workflow\n * composition gap.\n *\n * Behavior:\n * - Tool `name` mirrors `workflow.name` — workflows without a name throw\n *   `WorkflowError` at wrap time (the agent surface needs a stable id).\n * - Tool `input` is the supplied `inputSchema`; the validated value is\n *   forwarded straight to `workflow.execute(input)`.\n * - On `result.error`, the workflow error is wrapped in\n *   `ToolExecutionError` with `cause` set to the original\n *   `WorkflowError` subclass — the agent's tool-call loop sees a\n *   uniform error class regardless of which primitive failed.\n *\n * @example\n * const wf = workflow({ name: \"triage\", steps: [...] });\n * const triageTool = asTool(wf, {\n *   description: \"Run the support-ticket triage flow\",\n *   inputSchema: ticketSchema,\n * });\n * const a = ai.agent({ model, tools: [triageTool] });\n */\nexport function asTool<TInput, TOutput, TToolInput = TInput>(\n  workflowInstance: WorkflowInstance<TInput, TOutput>,\n  options: {\n    description?: string;\n    inputSchema: StandardSchemaV1<TToolInput>;\n  },\n): ToolContract<TToolInput, TOutput> {\n  if (!workflowInstance.name || typeof workflowInstance.name !== \"string\") {\n    throw new WorkflowError(\n      \"workflow.asTool(): workflow must have a `name` to be wrapped as a tool\",\n    );\n  }\n\n  return compositeAsTool<TToolInput, TOutput>({\n    name: workflowInstance.name,\n    description: options.description ?? `Invoke workflow \"${workflowInstance.name}\" as a tool.`,\n    input: options.inputSchema,\n    execute: async (input, ctx) => {\n      // Relay the outer agent's cancellation signal so cancelling the\n      // parent aborts this nested workflow run (C2).\n      const result = await workflowInstance.execute(\n        input as unknown as TInput,\n        ctx?.signal ? { signal: ctx.signal } : undefined,\n      );\n\n      if (result.error) {\n        // Throw the workflow error so the surrounding wrapper catches\n        // it and produces a `ToolExecutionError` with `cause` pointing\n        // back at the original `WorkflowError` subclass — keeps the\n        // agent's tool-call loop seeing one uniform error class\n        // regardless of which primitive failed.\n        throw result.error;\n      }\n\n      return {\n        data: result.data as TOutput,\n        usage: result.usage,\n        report: result.report,\n      };\n    },\n  });\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4BA,SAAgB,OACd,kBACA,SAImC;CACnC,IAAI,CAAC,iBAAiB,QAAQ,OAAO,iBAAiB,SAAS,UAC7D,MAAM,IAAI,cACR,wEACF;CAGF,OAAO,gBAAqC;EAC1C,MAAM,iBAAiB;EACvB,aAAa,QAAQ,eAAe,oBAAoB,iBAAiB,KAAK;EAC9E,OAAO,QAAQ;EACf,SAAS,OAAO,OAAO,QAAQ;GAG7B,MAAM,SAAS,MAAM,iBAAiB,QACpC,OACA,KAAK,SAAS,EAAE,QAAQ,IAAI,OAAO,IAAI,MACzC;GAEA,IAAI,OAAO,OAMT,MAAM,OAAO;GAGf,OAAO;IACL,MAAM,OAAO;IACb,OAAO,OAAO;IACd,QAAQ,OAAO;GACjB;EACF;CACF,CAAC;AACH"}

package/esm/workflow/engine.mjs

@@ -258,6 +258,7 @@ async function runWorkflow(params) {
 		workflowName: definition.name,
 		signature,
 		status,
+		...error ? { error } : {},
 		startedAt,
 		endedAt,
 		duration,

package/esm/workflow/engine.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"engine.mjs","names":[],"sources":["../../../../../../../@warlock.js/ai/src/workflow/engine.ts"],"sourcesContent":["import { log } from \"@warlock.js/logger\";\nimport type { BaseReport } from \"../contracts/result/base-report.type\";\nimport type { StepSnapshot } from \"../contracts/result/step-result.type\";\nimport type { Usage } from \"../contracts/result/usage.type\";\nimport type {\n  WorkflowReport,\n  WorkflowResult,\n} from \"../contracts/result/workflow-result.type\";\nimport type { StepDefinition } from \"../contracts/workflow/step.contract\";\nimport type { WorkflowContext } from \"../contracts/workflow/workflow-context.type\";\nimport type { WorkflowSnapshot } from \"../contracts/workflow/workflow-snapshot.type\";\nimport type {\n  WorkflowDefinition,\n  WorkflowEventHandlers,\n} from \"../contracts/workflow/workflow.contract\";\nimport {\n  AIError,\n  MaxStepsExceededError,\n  RoutingError,\n  SchemaValidationError,\n  WorkflowCancelledError,\n  WorkflowError,\n} from \"../errors\";\nimport { stampReportLineage } from \"../utils\";\nimport { createCancelledError } from \"./cancellation\";\nimport type { WorkflowEmitter } from \"./emitter\";\nimport { mapNextStep, nextDeclaredStep, resolveNextStep } from \"./router\";\nimport { runScopedEmitter } from \"./run-scoped-emitter\";\nimport { persistSnapshot } from \"./snapshot\";\nimport { cloneState, deepFreeze } from \"./state\";\nimport { executeStep, finalizeSnapshot, toAIError } from \"./step-runner\";\n\nexport { loadSnapshotForResume } from \"./snapshot\";\n\nconst DEFAULT_MAX_STEPS = 100;\nconst DEFAULT_LOOP_WARN = 5;\nconst LOG_MODULE_BASE = \"ai.workflow\";\n\ntype EngineParams<TOutput> = {\n  definition: WorkflowDefinition<any, TOutput, any, any>;\n  signature: string;\n  emitter: WorkflowEmitter;\n  input: unknown;\n  /**\n   * Request-scoped envelope, frozen and exposed as `ctx.context` to\n   * every step. Never persisted in snapshots; resume callers supply\n   * it fresh via `WorkflowResumeOptions.context`. Defaults to a\n   * frozen empty object when caller omits it.\n   */\n  context?: unknown;\n  runId: string;\n  signal?: AbortSignal;\n  executionHandlers?: WorkflowEventHandlers;\n  resumeFrom?: WorkflowSnapshot;\n  /**\n   * Opaque session identifier propagated onto every report node this\n   * run produces — including agent reports from child steps. Threaded\n   * from `WorkflowRunOptions.sessionId`. Omitted leaves the field\n   * undefined throughout the tree.\n   */\n  sessionId?: string;\n};\n\n/**\n * Main workflow driver. Walks the declared steps, handling routing,\n * cancellation, retries, parallel execution, and snapshot\n * persistence. Delegates the step lifecycle to `step-runner.ts`,\n * routing to `router.ts`, persistence to `snapshot.ts`. Never throws\n * — every failure funnels into `result.error`.\n */\nexport async function runWorkflow<TOutput>(\n  params: EngineParams<TOutput>,\n): Promise<WorkflowResult<TOutput>> {\n  const { definition, signature, input, runId, signal } = params;\n  // Bind the factory-scoped emitter to THIS run's identity. Every\n  // `emitter.emit(...)` below — and the one threaded into\n  // `executeStep` — now stamps `runId` / `rootRunId` automatically.\n  const emitter = runScopedEmitter(params.emitter, {\n    runId,\n    rootRunId: runId,\n  });\n\n  // Freeze the envelope once at run start. Default to `{}` so step\n  // code can always read `ctx.context` without an undefined guard.\n  const context = Object.freeze(params.context ?? {});\n  const maxSteps = definition.maxSteps ?? DEFAULT_MAX_STEPS;\n  const loopWarnAfter = definition.loopWarnAfter ?? DEFAULT_LOOP_WARN;\n\n  const logger = log;\n  const logModule = `${LOG_MODULE_BASE}.${definition.name}`;\n\n  const stepByName = new Map<string, StepDefinition>();\n  for (const s of definition.steps) stepByName.set(s.name, s);\n\n  const state: Record<string, unknown> = params.resumeFrom\n    ? { ...params.resumeFrom.state }\n    : {};\n  const steps: Record<string, StepSnapshot> = params.resumeFrom\n    ? { ...params.resumeFrom.steps }\n    : {};\n  const enteredCount = new Map<string, number>();\n  const usage: Usage = { input: 0, output: 0, total: 0 };\n\n  const startedAt = params.resumeFrom?.startedAt ?? new Date().toISOString();\n  const startedAtDate = new Date(startedAt);\n  const runStartPerf = performance.now();\n\n  let error: AIError | undefined;\n  let status: \"completed\" | \"failed\" | \"cancelled\" = \"completed\";\n  let cancelledAt: string | undefined;\n  let lastGoto: string | null = null;\n  // Captured when a step throws after retries exhaust (and `onFailure`\n  // didn't recover). Used to point the final snapshot's `next` at the\n  // failed step so `resume()` re-runs it after the cause is fixed.\n  let failedStepName: string | undefined;\n\n  const buildContext = (current?: {\n    state: Record<string, unknown>;\n    agentResult?: unknown;\n  }): WorkflowContext => ({\n    input,\n    context,\n    steps: steps as Readonly<Record<string, StepSnapshot>>,\n    state: current?.state ?? state,\n    agentResult: current?.agentResult as WorkflowContext[\"agentResult\"],\n    runId,\n    signal,\n    startedAt: startedAtDate,\n  });\n\n  emitter.emit(\n    \"workflow.starting\",\n    { workflowName: definition.name, input },\n    params.executionHandlers,\n  );\n  logger.info(logModule, \"starting\", \"workflow starting\", { runId });\n\n  let currentName: string | null = resolveInitialStep(\n    definition,\n    params.resumeFrom,\n  );\n  let stepCount = 0;\n\n  try {\n    while (currentName !== null) {\n      if (signal?.aborted) throw createCancelledError(signal);\n\n      stepCount += 1;\n      if (stepCount > maxSteps) {\n        throw new MaxStepsExceededError(\n          `workflow \"${definition.name}\" exceeded maxSteps=${maxSteps}`,\n          { maxSteps },\n        );\n      }\n\n      const entered = (enteredCount.get(currentName) ?? 0) + 1;\n      enteredCount.set(currentName, entered);\n      if (entered === loopWarnAfter) {\n        emitter.emit(\n          \"workflow.loop.warning\",\n          { step: currentName, enteredCount: entered, lastGoto },\n          params.executionHandlers,\n        );\n        logger.warn(logModule, \"loop.warning\", \"loop warning\", {\n          step: currentName,\n          enteredCount: entered,\n        });\n      }\n\n      const step = stepByName.get(currentName);\n      if (!step) {\n        throw new RoutingError(\n          `workflow \"${definition.name}\": unknown step \"${currentName}\"`,\n          { stepName: currentName },\n        );\n      }\n\n      const snapshot = await executeStep({\n        step,\n        state,\n        emitter,\n        executionHandlers: params.executionHandlers,\n        logger,\n        logModule,\n        signal,\n        buildContext,\n        usage,\n        workflowDefaultRetry: definition.defaultRetry,\n      });\n\n      Object.assign(state, snapshot.state);\n      steps[step.name] = finalizeSnapshot(snapshot);\n      // Parallel children — flat-path addressing alongside nested.\n      if (snapshot.steps) {\n        for (const [childName, childSnap] of Object.entries(snapshot.steps)) {\n          steps[childName] = childSnap;\n        }\n      }\n\n      // Failure path: retries exhausted. Give `onFailure` a chance to\n      // recover; otherwise checkpoint at the failed step (so resume\n      // re-runs it) and throw — workflow halts.\n      if (snapshot.status === \"failed\" && snapshot.error) {\n        const failureRoute = await resolveFailureRoute({\n          step,\n          definition,\n          error: snapshot.error,\n          ctx: buildContext({ state, agentResult: snapshot.executionResult }),\n        });\n\n        if (failureRoute === undefined) {\n          // No recovery — persist with `next: step.name` so resume\n          // re-runs this step after the user fixes the cause.\n          const persistOutcome = await persistSnapshot({\n            definition,\n            signature,\n            runId,\n            startedAt,\n            input,\n            state,\n            steps,\n            next: step.name,\n            status: \"running\",\n          });\n          if (!persistOutcome.ok) {\n            const persistErr = toAIError(persistOutcome.error);\n            emitter.emit(\n              \"workflow.error\",\n              { error: persistErr },\n              params.executionHandlers,\n            );\n            logger.error(\n              logModule,\n              \"persist.failed\",\n              \"snapshot persist failed\",\n              {\n                step: step.name,\n                code: persistErr.code,\n                message: persistErr.message,\n              },\n            );\n          }\n          failedStepName = step.name;\n          throw snapshot.error;\n        }\n\n        // onFailure routed — workflow continues. Checkpoint at the\n        // routed target (or `null` for `end`) so resume picks up there.\n        const failureNext = failureRoute === \"end\" ? null : failureRoute;\n        if (failureNext !== null && !stepByName.has(failureNext)) {\n          throw new RoutingError(\n            `workflow \"${definition.name}\": step \"${step.name}\" onFailure routed to unknown target \"${failureNext}\"`,\n            { stepName: step.name, targetName: failureNext },\n          );\n        }\n\n        const failurePersist = await persistSnapshot({\n          definition,\n          signature,\n          runId,\n          startedAt,\n          input,\n          state,\n          steps,\n          next: failureNext,\n          status: \"running\",\n        });\n        if (!failurePersist.ok) {\n          const persistErr = toAIError(failurePersist.error);\n          emitter.emit(\n            \"workflow.error\",\n            { error: persistErr },\n            params.executionHandlers,\n          );\n          logger.error(\n            logModule,\n            \"persist.failed\",\n            \"snapshot persist failed\",\n            {\n              step: step.name,\n              code: persistErr.code,\n              message: persistErr.message,\n            },\n          );\n        }\n\n        if (signal?.aborted) throw createCancelledError(signal);\n\n        if (failureRoute === \"end\") {\n          currentName = null;\n          break;\n        }\n\n        lastGoto = failureRoute;\n        currentName = failureRoute;\n        continue;\n      }\n\n      // Resolve next step for checkpoint accuracy BEFORE routing errors\n      // bubble — so the snapshot records where resume should resume from.\n      const resolved = await resolveNextStep({\n        step,\n        definition,\n        ctx: buildContext({ state, agentResult: snapshot.executionResult }),\n      });\n\n      const nextName =\n        resolved === \"end\"\n          ? null\n          : typeof resolved === \"string\"\n            ? resolved\n            : nextDeclaredStep(definition, step.name);\n\n      // Checkpoint after every step with the resolved `next`.\n      const outcome = await persistSnapshot({\n        definition,\n        signature,\n        runId,\n        startedAt,\n        input,\n        state,\n        steps,\n        next: nextName,\n        status: \"running\",\n      });\n      if (!outcome.ok) {\n        const persistErr = toAIError(outcome.error);\n        emitter.emit(\n          \"workflow.error\",\n          { error: persistErr },\n          params.executionHandlers,\n        );\n        logger.error(logModule, \"persist.failed\", \"snapshot persist failed\", {\n          step: step.name,\n          code: persistErr.code,\n          message: persistErr.message,\n        });\n      }\n\n      if (signal?.aborted) throw createCancelledError(signal);\n\n      if (resolved === \"end\") {\n        currentName = null;\n        break;\n      }\n\n      if (typeof resolved === \"string\") {\n        if (!stepByName.has(resolved)) {\n          throw new RoutingError(\n            `workflow \"${definition.name}\": step \"${step.name}\" goto unknown target \"${resolved}\"`,\n            { stepName: step.name, targetName: resolved },\n          );\n        }\n        lastGoto = resolved;\n        currentName = resolved;\n        continue;\n      }\n\n      currentName = nextName;\n      lastGoto = currentName;\n    }\n  } catch (err) {\n    if (err instanceof WorkflowCancelledError) {\n      status = \"cancelled\";\n      cancelledAt = err.cancelledAt;\n      error = err;\n    } else if (err instanceof AIError) {\n      status = \"failed\";\n      error = err;\n    } else {\n      status = \"failed\";\n      error = new WorkflowError(\n        err instanceof Error ? err.message : String(err),\n        { cause: err },\n      );\n    }\n  }\n\n  // Note: a `failed` step always throws (caught above) unless its\n  // `onFailure` recovered the run. A `completed` workflow may still\n  // contain `failed` step snapshots — those are the recovered cases\n  // and are intentionally preserved for forensic trace.\n\n  const endedAt = new Date().toISOString();\n  const duration = performance.now() - runStartPerf;\n\n  let data: TOutput | undefined;\n  if (status === \"completed\" && definition.output) {\n    try {\n      const extracted = await definition.output.extract(\n        buildContext({ state }),\n      );\n      data = (await validateWorkflowOutput(\n        definition.output.schema,\n        extracted,\n      )) as TOutput;\n    } catch (err) {\n      status = \"failed\";\n      error =\n        err instanceof AIError\n          ? err\n          : new WorkflowError(\n              err instanceof Error ? err.message : String(err),\n              { cause: err },\n            );\n    }\n  }\n\n  // Collect child executable reports from every step that ran one.\n  // Today the step runner surfaces agent reports (the only executable\n  // kind steps can invoke natively via `step.agent`); custom `run`\n  // callbacks that call tools/workflows/supervisors can't be observed\n  // here without a richer step-runner API — that's a v1.x follow-up\n  // (see backlog: \"step.run executables surface in workflow tree\").\n  const children: BaseReport[] = [];\n  for (const stepName in steps) {\n    const snap = steps[stepName];\n    if (snap.agentReport) {\n      children.push(snap.agentReport);\n    }\n  }\n\n  const report: WorkflowReport = {\n    runId,\n    rootRunId: runId,\n    name: definition.name,\n    version: definition.version,\n    type: \"workflow\",\n    workflowName: definition.name,\n    signature,\n    status,\n    startedAt,\n    endedAt,\n    duration,\n    cancelledAt,\n    usage,\n    children,\n    steps,\n    state: deepFreeze(cloneState(state)),\n  };\n\n  // Stamp lineage on the assembled tree exactly once. Walker rewrites\n  // any inner self-roots that nested agent reports brought in (each\n  // agent's `buildResult` set its own runId as root), propagates\n  // sessionId, and writes `reportSchemaVersion` on the root.\n  stampReportLineage(report, {\n    rootRunId: runId,\n    sessionId: params.sessionId,\n  });\n\n  // On a failed run with a captured `failedStepName`, point `next` at\n  // the failed step so `resume()` re-runs it. The pre-throw checkpoint\n  // already wrote this value, but the final snapshot would otherwise\n  // overwrite it with `null` and force resume to fall back to the\n  // first non-completed step (which is the same step in practice, but\n  // less informative for tooling reading the snapshot).\n  const finalNext = status === \"failed\" ? failedStepName ?? null : null;\n\n  const finalOutcome = await persistSnapshot({\n    definition,\n    signature,\n    runId,\n    startedAt,\n    input,\n    state,\n    steps,\n    next: finalNext,\n    status,\n  });\n  if (!finalOutcome.ok) {\n    const persistErr = toAIError(finalOutcome.error);\n    emitter.emit(\n      \"workflow.error\",\n      { error: persistErr },\n      params.executionHandlers,\n    );\n    logger.error(logModule, \"persist.failed\", \"final snapshot persist failed\", {\n      code: persistErr.code,\n      message: persistErr.message,\n    });\n  }\n\n  const result: WorkflowResult<TOutput> = {\n    type: \"workflow\",\n    data,\n    report,\n    usage,\n    error,\n  };\n\n  if (status === \"cancelled\") {\n    emitter.emit(\n      \"workflow.cancelled\",\n      {\n        cancelledAt: cancelledAt ?? endedAt,\n        reason: (error as WorkflowCancelledError | undefined)?.reason ?? \"\",\n      },\n      params.executionHandlers,\n    );\n    logger.warn(logModule, \"cancelled\", \"workflow cancelled\", { runId });\n  }\n\n  if (status === \"failed\" && error) {\n    emitter.emit(\"workflow.error\", { error }, params.executionHandlers);\n    logger.error(logModule, \"error\", \"workflow failed\", {\n      runId,\n      code: error.code,\n      message: error.message,\n    });\n  }\n\n  emitter.emit(\n    \"workflow.completed\",\n    { result: result as WorkflowResult<unknown> },\n    params.executionHandlers,\n  );\n  logger.info(logModule, \"completed\", \"workflow completed\", {\n    runId,\n    status,\n    duration,\n  });\n\n  return result;\n}\n\n/**\n * Run a failed step's `onFailure` hook (if present) and translate its\n * result into a route. Returns `undefined` when the workflow should\n * halt with the original error; `\"end\"` for clean termination; or a\n * step name to redirect to. A throw inside `onFailure` is wrapped in\n * `RoutingError` — routing is authoritative, never retried.\n */\nasync function resolveFailureRoute<T>(params: {\n  step: StepDefinition;\n  definition: WorkflowDefinition<any, T, any, any>;\n  error: AIError;\n  ctx: WorkflowContext;\n}): Promise<\"end\" | string | undefined> {\n  const { step, definition, error, ctx } = params;\n  if (!step.onFailure) return undefined;\n\n  let outcome;\n  try {\n    outcome = await step.onFailure(ctx, error);\n  } catch (err) {\n    throw new RoutingError(\n      `workflow \"${definition.name}\": step \"${step.name}\" onFailure threw`,\n      { stepName: step.name, cause: err },\n    );\n  }\n  return mapNextStep(outcome);\n}\n\nfunction resolveInitialStep<T>(\n  definition: WorkflowDefinition<any, T, any, any>,\n  resumeFrom: WorkflowSnapshot | undefined,\n): string | null {\n  if (!resumeFrom) return definition.steps[0]?.name ?? null;\n\n  // Prefer the explicitly-recorded `next` (now populated on every\n  // checkpoint). Falls back to first step whose snapshot is missing\n  // or not in a terminal-success state — covers older snapshots\n  // written before `next` was wired.\n  if (\n    resumeFrom.next &&\n    definition.steps.some(s => s.name === resumeFrom.next)\n  ) {\n    return resumeFrom.next;\n  }\n\n  for (const step of definition.steps) {\n    const snap = resumeFrom.steps[step.name];\n    if (!snap || (snap.status !== \"completed\" && snap.status !== \"skipped\")) {\n      return step.name;\n    }\n  }\n\n  return null;\n}\n\nasync function validateWorkflowOutput(\n  schema: unknown,\n  value: unknown,\n): Promise<unknown> {\n  if (!schema) return value;\n\n  const result = await (\n    schema as {\n      \"~standard\": { validate: (v: unknown) => Promise<unknown> | unknown };\n    }\n  )[\"~standard\"].validate(value);\n\n  if (\n    result &&\n    typeof result === \"object\" &&\n    \"issues\" in result &&\n    (result as { issues: unknown }).issues\n  ) {\n    throw new SchemaValidationError(\n      \"workflow output failed schema validation\",\n      {\n        issues: (result as { issues: any }).issues,\n      },\n    );\n  }\n\n  return (result as { value: unknown }).value;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAkCA,MAAM,oBAAoB;AAC1B,MAAM,oBAAoB;AAC1B,MAAM,kBAAkB;;;;;;;;AAkCxB,eAAsB,YACpB,QACkC;CAClC,MAAM,EAAE,YAAY,WAAW,OAAO,OAAO,WAAW;CAIxD,MAAM,UAAU,iBAAiB,OAAO,SAAS;EAC/C;EACA,WAAW;CACb,CAAC;CAID,MAAM,UAAU,OAAO,OAAO,OAAO,WAAW,CAAC,CAAC;CAClD,MAAM,WAAW,WAAW,YAAY;CACxC,MAAM,gBAAgB,WAAW,iBAAiB;CAElD,MAAM,SAAS;CACf,MAAM,YAAY,GAAG,gBAAgB,GAAG,WAAW;CAEnD,MAAM,6BAAa,IAAI,IAA4B;CACnD,KAAK,MAAM,KAAK,WAAW,OAAO,WAAW,IAAI,EAAE,MAAM,CAAC;CAE1D,MAAM,QAAiC,OAAO,aAC1C,EAAE,GAAG,OAAO,WAAW,MAAM,IAC7B,CAAC;CACL,MAAM,QAAsC,OAAO,aAC/C,EAAE,GAAG,OAAO,WAAW,MAAM,IAC7B,CAAC;CACL,MAAM,+BAAe,IAAI,IAAoB;CAC7C,MAAM,QAAe;EAAE,OAAO;EAAG,QAAQ;EAAG,OAAO;CAAE;CAErD,MAAM,YAAY,OAAO,YAAY,8BAAa,IAAI,KAAK,EAAC,CAAC,YAAY;CACzE,MAAM,gBAAgB,IAAI,KAAK,SAAS;CACxC,MAAM,eAAe,YAAY,IAAI;CAErC,IAAI;CACJ,IAAI,SAA+C;CACnD,IAAI;CACJ,IAAI,WAA0B;CAI9B,IAAI;CAEJ,MAAM,gBAAgB,aAGE;EACtB;EACA;EACO;EACP,OAAO,SAAS,SAAS;EACzB,aAAa,SAAS;EACtB;EACA;EACA,WAAW;CACb;CAEA,QAAQ,KACN,qBACA;EAAE,cAAc,WAAW;EAAM;CAAM,GACvC,OAAO,iBACT;CACA,OAAO,KAAK,WAAW,YAAY,qBAAqB,EAAE,MAAM,CAAC;CAEjE,IAAI,cAA6B,mBAC/B,YACA,OAAO,UACT;CACA,IAAI,YAAY;CAEhB,IAAI;EACF,OAAO,gBAAgB,MAAM;GAC3B,IAAI,QAAQ,SAAS,MAAM,qBAAqB,MAAM;GAEtD,aAAa;GACb,IAAI,YAAY,UACd,MAAM,IAAI,sBACR,aAAa,WAAW,KAAK,sBAAsB,YACnD,EAAE,SAAS,CACb;GAGF,MAAM,WAAW,aAAa,IAAI,WAAW,KAAK,KAAK;GACvD,aAAa,IAAI,aAAa,OAAO;GACrC,IAAI,YAAY,eAAe;IAC7B,QAAQ,KACN,yBACA;KAAE,MAAM;KAAa,cAAc;KAAS;IAAS,GACrD,OAAO,iBACT;IACA,OAAO,KAAK,WAAW,gBAAgB,gBAAgB;KACrD,MAAM;KACN,cAAc;IAChB,CAAC;GACH;GAEA,MAAM,OAAO,WAAW,IAAI,WAAW;GACvC,IAAI,CAAC,MACH,MAAM,IAAI,aACR,aAAa,WAAW,KAAK,mBAAmB,YAAY,IAC5D,EAAE,UAAU,YAAY,CAC1B;GAGF,MAAM,WAAW,MAAM,YAAY;IACjC;IACA;IACA;IACA,mBAAmB,OAAO;IAC1B;IACA;IACA;IACA;IACA;IACA,sBAAsB,WAAW;GACnC,CAAC;GAED,OAAO,OAAO,OAAO,SAAS,KAAK;GACnC,MAAM,KAAK,QAAQ,iBAAiB,QAAQ;GAE5C,IAAI,SAAS,OACX,KAAK,MAAM,CAAC,WAAW,cAAc,OAAO,QAAQ,SAAS,KAAK,GAChE,MAAM,aAAa;GAOvB,IAAI,SAAS,WAAW,YAAY,SAAS,OAAO;IAClD,MAAM,eAAe,MAAM,oBAAoB;KAC7C;KACA;KACA,OAAO,SAAS;KAChB,KAAK,aAAa;MAAE;MAAO,aAAa,SAAS;KAAgB,CAAC;IACpE,CAAC;IAED,IAAI,iBAAiB,QAAW;KAG9B,MAAM,iBAAiB,MAAM,gBAAgB;MAC3C;MACA;MACA;MACA;MACA;MACA;MACA;MACA,MAAM,KAAK;MACX,QAAQ;KACV,CAAC;KACD,IAAI,CAAC,eAAe,IAAI;MACtB,MAAM,aAAa,UAAU,eAAe,KAAK;MACjD,QAAQ,KACN,kBACA,EAAE,OAAO,WAAW,GACpB,OAAO,iBACT;MACA,OAAO,MACL,WACA,kBACA,2BACA;OACE,MAAM,KAAK;OACX,MAAM,WAAW;OACjB,SAAS,WAAW;MACtB,CACF;KACF;KACA,iBAAiB,KAAK;KACtB,MAAM,SAAS;IACjB;IAIA,MAAM,cAAc,iBAAiB,QAAQ,OAAO;IACpD,IAAI,gBAAgB,QAAQ,CAAC,WAAW,IAAI,WAAW,GACrD,MAAM,IAAI,aACR,aAAa,WAAW,KAAK,WAAW,KAAK,KAAK,wCAAwC,YAAY,IACtG;KAAE,UAAU,KAAK;KAAM,YAAY;IAAY,CACjD;IAGF,MAAM,iBAAiB,MAAM,gBAAgB;KAC3C;KACA;KACA;KACA;KACA;KACA;KACA;KACA,MAAM;KACN,QAAQ;IACV,CAAC;IACD,IAAI,CAAC,eAAe,IAAI;KACtB,MAAM,aAAa,UAAU,eAAe,KAAK;KACjD,QAAQ,KACN,kBACA,EAAE,OAAO,WAAW,GACpB,OAAO,iBACT;KACA,OAAO,MACL,WACA,kBACA,2BACA;MACE,MAAM,KAAK;MACX,MAAM,WAAW;MACjB,SAAS,WAAW;KACtB,CACF;IACF;IAEA,IAAI,QAAQ,SAAS,MAAM,qBAAqB,MAAM;IAEtD,IAAI,iBAAiB,OAAO;KAC1B,cAAc;KACd;IACF;IAEA,WAAW;IACX,cAAc;IACd;GACF;GAIA,MAAM,WAAW,MAAM,gBAAgB;IACrC;IACA;IACA,KAAK,aAAa;KAAE;KAAO,aAAa,SAAS;IAAgB,CAAC;GACpE,CAAC;GAED,MAAM,WACJ,aAAa,QACT,OACA,OAAO,aAAa,WAClB,WACA,iBAAiB,YAAY,KAAK,IAAI;GAG9C,MAAM,UAAU,MAAM,gBAAgB;IACpC;IACA;IACA;IACA;IACA;IACA;IACA;IACA,MAAM;IACN,QAAQ;GACV,CAAC;GACD,IAAI,CAAC,QAAQ,IAAI;IACf,MAAM,aAAa,UAAU,QAAQ,KAAK;IAC1C,QAAQ,KACN,kBACA,EAAE,OAAO,WAAW,GACpB,OAAO,iBACT;IACA,OAAO,MAAM,WAAW,kBAAkB,2BAA2B;KACnE,MAAM,KAAK;KACX,MAAM,WAAW;KACjB,SAAS,WAAW;IACtB,CAAC;GACH;GAEA,IAAI,QAAQ,SAAS,MAAM,qBAAqB,MAAM;GAEtD,IAAI,aAAa,OAAO;IACtB,cAAc;IACd;GACF;GAEA,IAAI,OAAO,aAAa,UAAU;IAChC,IAAI,CAAC,WAAW,IAAI,QAAQ,GAC1B,MAAM,IAAI,aACR,aAAa,WAAW,KAAK,WAAW,KAAK,KAAK,yBAAyB,SAAS,IACpF;KAAE,UAAU,KAAK;KAAM,YAAY;IAAS,CAC9C;IAEF,WAAW;IACX,cAAc;IACd;GACF;GAEA,cAAc;GACd,WAAW;EACb;CACF,SAAS,KAAK;EACZ,IAAI,eAAe,wBAAwB;GACzC,SAAS;GACT,cAAc,IAAI;GAClB,QAAQ;EACV,OAAO,IAAI,eAAe,SAAS;GACjC,SAAS;GACT,QAAQ;EACV,OAAO;GACL,SAAS;GACT,QAAQ,IAAI,cACV,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GAC/C,EAAE,OAAO,IAAI,CACf;EACF;CACF;CAOA,MAAM,2BAAU,IAAI,KAAK,EAAC,CAAC,YAAY;CACvC,MAAM,WAAW,YAAY,IAAI,IAAI;CAErC,IAAI;CACJ,IAAI,WAAW,eAAe,WAAW,QACvC,IAAI;EACF,MAAM,YAAY,MAAM,WAAW,OAAO,QACxC,aAAa,EAAE,MAAM,CAAC,CACxB;EACA,OAAQ,MAAM,uBACZ,WAAW,OAAO,QAClB,SACF;CACF,SAAS,KAAK;EACZ,SAAS;EACT,QACE,eAAe,UACX,MACA,IAAI,cACF,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GAC/C,EAAE,OAAO,IAAI,CACf;CACR;CASF,MAAM,WAAyB,CAAC;CAChC,KAAK,MAAM,YAAY,OAAO;EAC5B,MAAM,OAAO,MAAM;EACnB,IAAI,KAAK,aACP,SAAS,KAAK,KAAK,WAAW;CAElC;CAEA,MAAM,SAAyB;EAC7B;EACA,WAAW;EACX,MAAM,WAAW;EACjB,SAAS,WAAW;EACpB,MAAM;EACN,cAAc,WAAW;EACzB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,OAAO,WAAW,WAAW,KAAK,CAAC;CACrC;CAMA,mBAAmB,QAAQ;EACzB,WAAW;EACX,WAAW,OAAO;CACpB,CAAC;CAUD,MAAM,eAAe,MAAM,gBAAgB;EACzC;EACA;EACA;EACA;EACA;EACA;EACA;EACA,MAVgB,WAAW,WAAW,kBAAkB,OAAO;EAW/D;CACF,CAAC;CACD,IAAI,CAAC,aAAa,IAAI;EACpB,MAAM,aAAa,UAAU,aAAa,KAAK;EAC/C,QAAQ,KACN,kBACA,EAAE,OAAO,WAAW,GACpB,OAAO,iBACT;EACA,OAAO,MAAM,WAAW,kBAAkB,iCAAiC;GACzE,MAAM,WAAW;GACjB,SAAS,WAAW;EACtB,CAAC;CACH;CAEA,MAAM,SAAkC;EACtC,MAAM;EACN;EACA;EACA;EACA;CACF;CAEA,IAAI,WAAW,aAAa;EAC1B,QAAQ,KACN,sBACA;GACE,aAAa,eAAe;GAC5B,QAAS,OAA8C,UAAU;EACnE,GACA,OAAO,iBACT;EACA,OAAO,KAAK,WAAW,aAAa,sBAAsB,EAAE,MAAM,CAAC;CACrE;CAEA,IAAI,WAAW,YAAY,OAAO;EAChC,QAAQ,KAAK,kBAAkB,EAAE,MAAM,GAAG,OAAO,iBAAiB;EAClE,OAAO,MAAM,WAAW,SAAS,mBAAmB;GAClD;GACA,MAAM,MAAM;GACZ,SAAS,MAAM;EACjB,CAAC;CACH;CAEA,QAAQ,KACN,sBACA,EAAU,OAAkC,GAC5C,OAAO,iBACT;CACA,OAAO,KAAK,WAAW,aAAa,sBAAsB;EACxD;EACA;EACA;CACF,CAAC;CAED,OAAO;AACT;;;;;;;;AASA,eAAe,oBAAuB,QAKE;CACtC,MAAM,EAAE,MAAM,YAAY,OAAO,QAAQ;CACzC,IAAI,CAAC,KAAK,WAAW,OAAO;CAE5B,IAAI;CACJ,IAAI;EACF,UAAU,MAAM,KAAK,UAAU,KAAK,KAAK;CAC3C,SAAS,KAAK;EACZ,MAAM,IAAI,aACR,aAAa,WAAW,KAAK,WAAW,KAAK,KAAK,oBAClD;GAAE,UAAU,KAAK;GAAM,OAAO;EAAI,CACpC;CACF;CACA,OAAO,YAAY,OAAO;AAC5B;AAEA,SAAS,mBACP,YACA,YACe;CACf,IAAI,CAAC,YAAY,OAAO,WAAW,MAAM,EAAE,EAAE,QAAQ;CAMrD,IACE,WAAW,QACX,WAAW,MAAM,MAAK,MAAK,EAAE,SAAS,WAAW,IAAI,GAErD,OAAO,WAAW;CAGpB,KAAK,MAAM,QAAQ,WAAW,OAAO;EACnC,MAAM,OAAO,WAAW,MAAM,KAAK;EACnC,IAAI,CAAC,QAAS,KAAK,WAAW,eAAe,KAAK,WAAW,WAC3D,OAAO,KAAK;CAEhB;CAEA,OAAO;AACT;AAEA,eAAe,uBACb,QACA,OACkB;CAClB,IAAI,CAAC,QAAQ,OAAO;CAEpB,MAAM,SAAS,MACb,OAGA,YAAY,CAAC,SAAS,KAAK;CAE7B,IACE,UACA,OAAO,WAAW,YAClB,YAAY,UACX,OAA+B,QAEhC,MAAM,IAAI,sBACR,4CACA,EACE,QAAS,OAA2B,OACtC,CACF;CAGF,OAAQ,OAA8B;AACxC"}
+{"version":3,"file":"engine.mjs","names":[],"sources":["../../../../../../../@warlock.js/ai/src/workflow/engine.ts"],"sourcesContent":["import { log } from \"@warlock.js/logger\";\nimport type { BaseReport } from \"../contracts/result/base-report.type\";\nimport type { StepSnapshot } from \"../contracts/result/step-result.type\";\nimport type { Usage } from \"../contracts/result/usage.type\";\nimport type {\n  WorkflowReport,\n  WorkflowResult,\n} from \"../contracts/result/workflow-result.type\";\nimport type { StepDefinition } from \"../contracts/workflow/step.contract\";\nimport type { WorkflowContext } from \"../contracts/workflow/workflow-context.type\";\nimport type { WorkflowSnapshot } from \"../contracts/workflow/workflow-snapshot.type\";\nimport type {\n  WorkflowDefinition,\n  WorkflowEventHandlers,\n} from \"../contracts/workflow/workflow.contract\";\nimport {\n  AIError,\n  MaxStepsExceededError,\n  RoutingError,\n  SchemaValidationError,\n  WorkflowCancelledError,\n  WorkflowError,\n} from \"../errors\";\nimport { stampReportLineage } from \"../utils\";\nimport { createCancelledError } from \"./cancellation\";\nimport type { WorkflowEmitter } from \"./emitter\";\nimport { mapNextStep, nextDeclaredStep, resolveNextStep } from \"./router\";\nimport { runScopedEmitter } from \"./run-scoped-emitter\";\nimport { persistSnapshot } from \"./snapshot\";\nimport { cloneState, deepFreeze } from \"./state\";\nimport { executeStep, finalizeSnapshot, toAIError } from \"./step-runner\";\n\nexport { loadSnapshotForResume } from \"./snapshot\";\n\nconst DEFAULT_MAX_STEPS = 100;\nconst DEFAULT_LOOP_WARN = 5;\nconst LOG_MODULE_BASE = \"ai.workflow\";\n\ntype EngineParams<TOutput> = {\n  definition: WorkflowDefinition<any, TOutput, any, any>;\n  signature: string;\n  emitter: WorkflowEmitter;\n  input: unknown;\n  /**\n   * Request-scoped envelope, frozen and exposed as `ctx.context` to\n   * every step. Never persisted in snapshots; resume callers supply\n   * it fresh via `WorkflowResumeOptions.context`. Defaults to a\n   * frozen empty object when caller omits it.\n   */\n  context?: unknown;\n  runId: string;\n  signal?: AbortSignal;\n  executionHandlers?: WorkflowEventHandlers;\n  resumeFrom?: WorkflowSnapshot;\n  /**\n   * Opaque session identifier propagated onto every report node this\n   * run produces — including agent reports from child steps. Threaded\n   * from `WorkflowRunOptions.sessionId`. Omitted leaves the field\n   * undefined throughout the tree.\n   */\n  sessionId?: string;\n};\n\n/**\n * Main workflow driver. Walks the declared steps, handling routing,\n * cancellation, retries, parallel execution, and snapshot\n * persistence. Delegates the step lifecycle to `step-runner.ts`,\n * routing to `router.ts`, persistence to `snapshot.ts`. Never throws\n * — every failure funnels into `result.error`.\n */\nexport async function runWorkflow<TOutput>(\n  params: EngineParams<TOutput>,\n): Promise<WorkflowResult<TOutput>> {\n  const { definition, signature, input, runId, signal } = params;\n  // Bind the factory-scoped emitter to THIS run's identity. Every\n  // `emitter.emit(...)` below — and the one threaded into\n  // `executeStep` — now stamps `runId` / `rootRunId` automatically.\n  const emitter = runScopedEmitter(params.emitter, {\n    runId,\n    rootRunId: runId,\n  });\n\n  // Freeze the envelope once at run start. Default to `{}` so step\n  // code can always read `ctx.context` without an undefined guard.\n  const context = Object.freeze(params.context ?? {});\n  const maxSteps = definition.maxSteps ?? DEFAULT_MAX_STEPS;\n  const loopWarnAfter = definition.loopWarnAfter ?? DEFAULT_LOOP_WARN;\n\n  const logger = log;\n  const logModule = `${LOG_MODULE_BASE}.${definition.name}`;\n\n  const stepByName = new Map<string, StepDefinition>();\n  for (const s of definition.steps) stepByName.set(s.name, s);\n\n  const state: Record<string, unknown> = params.resumeFrom\n    ? { ...params.resumeFrom.state }\n    : {};\n  const steps: Record<string, StepSnapshot> = params.resumeFrom\n    ? { ...params.resumeFrom.steps }\n    : {};\n  const enteredCount = new Map<string, number>();\n  const usage: Usage = { input: 0, output: 0, total: 0 };\n\n  const startedAt = params.resumeFrom?.startedAt ?? new Date().toISOString();\n  const startedAtDate = new Date(startedAt);\n  const runStartPerf = performance.now();\n\n  let error: AIError | undefined;\n  let status: \"completed\" | \"failed\" | \"cancelled\" = \"completed\";\n  let cancelledAt: string | undefined;\n  let lastGoto: string | null = null;\n  // Captured when a step throws after retries exhaust (and `onFailure`\n  // didn't recover). Used to point the final snapshot's `next` at the\n  // failed step so `resume()` re-runs it after the cause is fixed.\n  let failedStepName: string | undefined;\n\n  const buildContext = (current?: {\n    state: Record<string, unknown>;\n    agentResult?: unknown;\n  }): WorkflowContext => ({\n    input,\n    context,\n    steps: steps as Readonly<Record<string, StepSnapshot>>,\n    state: current?.state ?? state,\n    agentResult: current?.agentResult as WorkflowContext[\"agentResult\"],\n    runId,\n    signal,\n    startedAt: startedAtDate,\n  });\n\n  emitter.emit(\n    \"workflow.starting\",\n    { workflowName: definition.name, input },\n    params.executionHandlers,\n  );\n  logger.info(logModule, \"starting\", \"workflow starting\", { runId });\n\n  let currentName: string | null = resolveInitialStep(\n    definition,\n    params.resumeFrom,\n  );\n  let stepCount = 0;\n\n  try {\n    while (currentName !== null) {\n      if (signal?.aborted) throw createCancelledError(signal);\n\n      stepCount += 1;\n      if (stepCount > maxSteps) {\n        throw new MaxStepsExceededError(\n          `workflow \"${definition.name}\" exceeded maxSteps=${maxSteps}`,\n          { maxSteps },\n        );\n      }\n\n      const entered = (enteredCount.get(currentName) ?? 0) + 1;\n      enteredCount.set(currentName, entered);\n      if (entered === loopWarnAfter) {\n        emitter.emit(\n          \"workflow.loop.warning\",\n          { step: currentName, enteredCount: entered, lastGoto },\n          params.executionHandlers,\n        );\n        logger.warn(logModule, \"loop.warning\", \"loop warning\", {\n          step: currentName,\n          enteredCount: entered,\n        });\n      }\n\n      const step = stepByName.get(currentName);\n      if (!step) {\n        throw new RoutingError(\n          `workflow \"${definition.name}\": unknown step \"${currentName}\"`,\n          { stepName: currentName },\n        );\n      }\n\n      const snapshot = await executeStep({\n        step,\n        state,\n        emitter,\n        executionHandlers: params.executionHandlers,\n        logger,\n        logModule,\n        signal,\n        buildContext,\n        usage,\n        workflowDefaultRetry: definition.defaultRetry,\n      });\n\n      Object.assign(state, snapshot.state);\n      steps[step.name] = finalizeSnapshot(snapshot);\n      // Parallel children — flat-path addressing alongside nested.\n      if (snapshot.steps) {\n        for (const [childName, childSnap] of Object.entries(snapshot.steps)) {\n          steps[childName] = childSnap;\n        }\n      }\n\n      // Failure path: retries exhausted. Give `onFailure` a chance to\n      // recover; otherwise checkpoint at the failed step (so resume\n      // re-runs it) and throw — workflow halts.\n      if (snapshot.status === \"failed\" && snapshot.error) {\n        const failureRoute = await resolveFailureRoute({\n          step,\n          definition,\n          error: snapshot.error,\n          ctx: buildContext({ state, agentResult: snapshot.executionResult }),\n        });\n\n        if (failureRoute === undefined) {\n          // No recovery — persist with `next: step.name` so resume\n          // re-runs this step after the user fixes the cause.\n          const persistOutcome = await persistSnapshot({\n            definition,\n            signature,\n            runId,\n            startedAt,\n            input,\n            state,\n            steps,\n            next: step.name,\n            status: \"running\",\n          });\n          if (!persistOutcome.ok) {\n            const persistErr = toAIError(persistOutcome.error);\n            emitter.emit(\n              \"workflow.error\",\n              { error: persistErr },\n              params.executionHandlers,\n            );\n            logger.error(\n              logModule,\n              \"persist.failed\",\n              \"snapshot persist failed\",\n              {\n                step: step.name,\n                code: persistErr.code,\n                message: persistErr.message,\n              },\n            );\n          }\n          failedStepName = step.name;\n          throw snapshot.error;\n        }\n\n        // onFailure routed — workflow continues. Checkpoint at the\n        // routed target (or `null` for `end`) so resume picks up there.\n        const failureNext = failureRoute === \"end\" ? null : failureRoute;\n        if (failureNext !== null && !stepByName.has(failureNext)) {\n          throw new RoutingError(\n            `workflow \"${definition.name}\": step \"${step.name}\" onFailure routed to unknown target \"${failureNext}\"`,\n            { stepName: step.name, targetName: failureNext },\n          );\n        }\n\n        const failurePersist = await persistSnapshot({\n          definition,\n          signature,\n          runId,\n          startedAt,\n          input,\n          state,\n          steps,\n          next: failureNext,\n          status: \"running\",\n        });\n        if (!failurePersist.ok) {\n          const persistErr = toAIError(failurePersist.error);\n          emitter.emit(\n            \"workflow.error\",\n            { error: persistErr },\n            params.executionHandlers,\n          );\n          logger.error(\n            logModule,\n            \"persist.failed\",\n            \"snapshot persist failed\",\n            {\n              step: step.name,\n              code: persistErr.code,\n              message: persistErr.message,\n            },\n          );\n        }\n\n        if (signal?.aborted) throw createCancelledError(signal);\n\n        if (failureRoute === \"end\") {\n          currentName = null;\n          break;\n        }\n\n        lastGoto = failureRoute;\n        currentName = failureRoute;\n        continue;\n      }\n\n      // Resolve next step for checkpoint accuracy BEFORE routing errors\n      // bubble — so the snapshot records where resume should resume from.\n      const resolved = await resolveNextStep({\n        step,\n        definition,\n        ctx: buildContext({ state, agentResult: snapshot.executionResult }),\n      });\n\n      const nextName =\n        resolved === \"end\"\n          ? null\n          : typeof resolved === \"string\"\n            ? resolved\n            : nextDeclaredStep(definition, step.name);\n\n      // Checkpoint after every step with the resolved `next`.\n      const outcome = await persistSnapshot({\n        definition,\n        signature,\n        runId,\n        startedAt,\n        input,\n        state,\n        steps,\n        next: nextName,\n        status: \"running\",\n      });\n      if (!outcome.ok) {\n        const persistErr = toAIError(outcome.error);\n        emitter.emit(\n          \"workflow.error\",\n          { error: persistErr },\n          params.executionHandlers,\n        );\n        logger.error(logModule, \"persist.failed\", \"snapshot persist failed\", {\n          step: step.name,\n          code: persistErr.code,\n          message: persistErr.message,\n        });\n      }\n\n      if (signal?.aborted) throw createCancelledError(signal);\n\n      if (resolved === \"end\") {\n        currentName = null;\n        break;\n      }\n\n      if (typeof resolved === \"string\") {\n        if (!stepByName.has(resolved)) {\n          throw new RoutingError(\n            `workflow \"${definition.name}\": step \"${step.name}\" goto unknown target \"${resolved}\"`,\n            { stepName: step.name, targetName: resolved },\n          );\n        }\n        lastGoto = resolved;\n        currentName = resolved;\n        continue;\n      }\n\n      currentName = nextName;\n      lastGoto = currentName;\n    }\n  } catch (err) {\n    if (err instanceof WorkflowCancelledError) {\n      status = \"cancelled\";\n      cancelledAt = err.cancelledAt;\n      error = err;\n    } else if (err instanceof AIError) {\n      status = \"failed\";\n      error = err;\n    } else {\n      status = \"failed\";\n      error = new WorkflowError(\n        err instanceof Error ? err.message : String(err),\n        { cause: err },\n      );\n    }\n  }\n\n  // Note: a `failed` step always throws (caught above) unless its\n  // `onFailure` recovered the run. A `completed` workflow may still\n  // contain `failed` step snapshots — those are the recovered cases\n  // and are intentionally preserved for forensic trace.\n\n  const endedAt = new Date().toISOString();\n  const duration = performance.now() - runStartPerf;\n\n  let data: TOutput | undefined;\n  if (status === \"completed\" && definition.output) {\n    try {\n      const extracted = await definition.output.extract(\n        buildContext({ state }),\n      );\n      data = (await validateWorkflowOutput(\n        definition.output.schema,\n        extracted,\n      )) as TOutput;\n    } catch (err) {\n      status = \"failed\";\n      error =\n        err instanceof AIError\n          ? err\n          : new WorkflowError(\n              err instanceof Error ? err.message : String(err),\n              { cause: err },\n            );\n    }\n  }\n\n  // Collect child executable reports from every step that ran one.\n  // Today the step runner surfaces agent reports (the only executable\n  // kind steps can invoke natively via `step.agent`); custom `run`\n  // callbacks that call tools/workflows/supervisors can't be observed\n  // here without a richer step-runner API — that's a v1.x follow-up\n  // (see backlog: \"step.run executables surface in workflow tree\").\n  const children: BaseReport[] = [];\n  for (const stepName in steps) {\n    const snap = steps[stepName];\n    if (snap.agentReport) {\n      children.push(snap.agentReport);\n    }\n  }\n\n  const report: WorkflowReport = {\n    runId,\n    rootRunId: runId,\n    name: definition.name,\n    version: definition.version,\n    type: \"workflow\",\n    workflowName: definition.name,\n    signature,\n    status,\n    // Stamp the terminal error onto the report so it travels with the tree\n    // (observe path has no result envelope to fall back on). A failed `run`\n    // step's cause lives in `steps[name].error`, but the workflow-level\n    // error is what a consumer reads off the root span. Absent on success.\n    ...(error ? { error } : {}),\n    startedAt,\n    endedAt,\n    duration,\n    cancelledAt,\n    usage,\n    children,\n    steps,\n    state: deepFreeze(cloneState(state)),\n  };\n\n  // Stamp lineage on the assembled tree exactly once. Walker rewrites\n  // any inner self-roots that nested agent reports brought in (each\n  // agent's `buildResult` set its own runId as root), propagates\n  // sessionId, and writes `reportSchemaVersion` on the root.\n  stampReportLineage(report, {\n    rootRunId: runId,\n    sessionId: params.sessionId,\n  });\n\n  // On a failed run with a captured `failedStepName`, point `next` at\n  // the failed step so `resume()` re-runs it. The pre-throw checkpoint\n  // already wrote this value, but the final snapshot would otherwise\n  // overwrite it with `null` and force resume to fall back to the\n  // first non-completed step (which is the same step in practice, but\n  // less informative for tooling reading the snapshot).\n  const finalNext = status === \"failed\" ? failedStepName ?? null : null;\n\n  const finalOutcome = await persistSnapshot({\n    definition,\n    signature,\n    runId,\n    startedAt,\n    input,\n    state,\n    steps,\n    next: finalNext,\n    status,\n  });\n  if (!finalOutcome.ok) {\n    const persistErr = toAIError(finalOutcome.error);\n    emitter.emit(\n      \"workflow.error\",\n      { error: persistErr },\n      params.executionHandlers,\n    );\n    logger.error(logModule, \"persist.failed\", \"final snapshot persist failed\", {\n      code: persistErr.code,\n      message: persistErr.message,\n    });\n  }\n\n  const result: WorkflowResult<TOutput> = {\n    type: \"workflow\",\n    data,\n    report,\n    usage,\n    error,\n  };\n\n  if (status === \"cancelled\") {\n    emitter.emit(\n      \"workflow.cancelled\",\n      {\n        cancelledAt: cancelledAt ?? endedAt,\n        reason: (error as WorkflowCancelledError | undefined)?.reason ?? \"\",\n      },\n      params.executionHandlers,\n    );\n    logger.warn(logModule, \"cancelled\", \"workflow cancelled\", { runId });\n  }\n\n  if (status === \"failed\" && error) {\n    emitter.emit(\"workflow.error\", { error }, params.executionHandlers);\n    logger.error(logModule, \"error\", \"workflow failed\", {\n      runId,\n      code: error.code,\n      message: error.message,\n    });\n  }\n\n  emitter.emit(\n    \"workflow.completed\",\n    { result: result as WorkflowResult<unknown> },\n    params.executionHandlers,\n  );\n  logger.info(logModule, \"completed\", \"workflow completed\", {\n    runId,\n    status,\n    duration,\n  });\n\n  return result;\n}\n\n/**\n * Run a failed step's `onFailure` hook (if present) and translate its\n * result into a route. Returns `undefined` when the workflow should\n * halt with the original error; `\"end\"` for clean termination; or a\n * step name to redirect to. A throw inside `onFailure` is wrapped in\n * `RoutingError` — routing is authoritative, never retried.\n */\nasync function resolveFailureRoute<T>(params: {\n  step: StepDefinition;\n  definition: WorkflowDefinition<any, T, any, any>;\n  error: AIError;\n  ctx: WorkflowContext;\n}): Promise<\"end\" | string | undefined> {\n  const { step, definition, error, ctx } = params;\n  if (!step.onFailure) return undefined;\n\n  let outcome;\n  try {\n    outcome = await step.onFailure(ctx, error);\n  } catch (err) {\n    throw new RoutingError(\n      `workflow \"${definition.name}\": step \"${step.name}\" onFailure threw`,\n      { stepName: step.name, cause: err },\n    );\n  }\n  return mapNextStep(outcome);\n}\n\nfunction resolveInitialStep<T>(\n  definition: WorkflowDefinition<any, T, any, any>,\n  resumeFrom: WorkflowSnapshot | undefined,\n): string | null {\n  if (!resumeFrom) return definition.steps[0]?.name ?? null;\n\n  // Prefer the explicitly-recorded `next` (now populated on every\n  // checkpoint). Falls back to first step whose snapshot is missing\n  // or not in a terminal-success state — covers older snapshots\n  // written before `next` was wired.\n  if (\n    resumeFrom.next &&\n    definition.steps.some(s => s.name === resumeFrom.next)\n  ) {\n    return resumeFrom.next;\n  }\n\n  for (const step of definition.steps) {\n    const snap = resumeFrom.steps[step.name];\n    if (!snap || (snap.status !== \"completed\" && snap.status !== \"skipped\")) {\n      return step.name;\n    }\n  }\n\n  return null;\n}\n\nasync function validateWorkflowOutput(\n  schema: unknown,\n  value: unknown,\n): Promise<unknown> {\n  if (!schema) return value;\n\n  const result = await (\n    schema as {\n      \"~standard\": { validate: (v: unknown) => Promise<unknown> | unknown };\n    }\n  )[\"~standard\"].validate(value);\n\n  if (\n    result &&\n    typeof result === \"object\" &&\n    \"issues\" in result &&\n    (result as { issues: unknown }).issues\n  ) {\n    throw new SchemaValidationError(\n      \"workflow output failed schema validation\",\n      {\n        issues: (result as { issues: any }).issues,\n      },\n    );\n  }\n\n  return (result as { value: unknown }).value;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAkCA,MAAM,oBAAoB;AAC1B,MAAM,oBAAoB;AAC1B,MAAM,kBAAkB;;;;;;;;AAkCxB,eAAsB,YACpB,QACkC;CAClC,MAAM,EAAE,YAAY,WAAW,OAAO,OAAO,WAAW;CAIxD,MAAM,UAAU,iBAAiB,OAAO,SAAS;EAC/C;EACA,WAAW;CACb,CAAC;CAID,MAAM,UAAU,OAAO,OAAO,OAAO,WAAW,CAAC,CAAC;CAClD,MAAM,WAAW,WAAW,YAAY;CACxC,MAAM,gBAAgB,WAAW,iBAAiB;CAElD,MAAM,SAAS;CACf,MAAM,YAAY,GAAG,gBAAgB,GAAG,WAAW;CAEnD,MAAM,6BAAa,IAAI,IAA4B;CACnD,KAAK,MAAM,KAAK,WAAW,OAAO,WAAW,IAAI,EAAE,MAAM,CAAC;CAE1D,MAAM,QAAiC,OAAO,aAC1C,EAAE,GAAG,OAAO,WAAW,MAAM,IAC7B,CAAC;CACL,MAAM,QAAsC,OAAO,aAC/C,EAAE,GAAG,OAAO,WAAW,MAAM,IAC7B,CAAC;CACL,MAAM,+BAAe,IAAI,IAAoB;CAC7C,MAAM,QAAe;EAAE,OAAO;EAAG,QAAQ;EAAG,OAAO;CAAE;CAErD,MAAM,YAAY,OAAO,YAAY,8BAAa,IAAI,KAAK,EAAC,CAAC,YAAY;CACzE,MAAM,gBAAgB,IAAI,KAAK,SAAS;CACxC,MAAM,eAAe,YAAY,IAAI;CAErC,IAAI;CACJ,IAAI,SAA+C;CACnD,IAAI;CACJ,IAAI,WAA0B;CAI9B,IAAI;CAEJ,MAAM,gBAAgB,aAGE;EACtB;EACA;EACO;EACP,OAAO,SAAS,SAAS;EACzB,aAAa,SAAS;EACtB;EACA;EACA,WAAW;CACb;CAEA,QAAQ,KACN,qBACA;EAAE,cAAc,WAAW;EAAM;CAAM,GACvC,OAAO,iBACT;CACA,OAAO,KAAK,WAAW,YAAY,qBAAqB,EAAE,MAAM,CAAC;CAEjE,IAAI,cAA6B,mBAC/B,YACA,OAAO,UACT;CACA,IAAI,YAAY;CAEhB,IAAI;EACF,OAAO,gBAAgB,MAAM;GAC3B,IAAI,QAAQ,SAAS,MAAM,qBAAqB,MAAM;GAEtD,aAAa;GACb,IAAI,YAAY,UACd,MAAM,IAAI,sBACR,aAAa,WAAW,KAAK,sBAAsB,YACnD,EAAE,SAAS,CACb;GAGF,MAAM,WAAW,aAAa,IAAI,WAAW,KAAK,KAAK;GACvD,aAAa,IAAI,aAAa,OAAO;GACrC,IAAI,YAAY,eAAe;IAC7B,QAAQ,KACN,yBACA;KAAE,MAAM;KAAa,cAAc;KAAS;IAAS,GACrD,OAAO,iBACT;IACA,OAAO,KAAK,WAAW,gBAAgB,gBAAgB;KACrD,MAAM;KACN,cAAc;IAChB,CAAC;GACH;GAEA,MAAM,OAAO,WAAW,IAAI,WAAW;GACvC,IAAI,CAAC,MACH,MAAM,IAAI,aACR,aAAa,WAAW,KAAK,mBAAmB,YAAY,IAC5D,EAAE,UAAU,YAAY,CAC1B;GAGF,MAAM,WAAW,MAAM,YAAY;IACjC;IACA;IACA;IACA,mBAAmB,OAAO;IAC1B;IACA;IACA;IACA;IACA;IACA,sBAAsB,WAAW;GACnC,CAAC;GAED,OAAO,OAAO,OAAO,SAAS,KAAK;GACnC,MAAM,KAAK,QAAQ,iBAAiB,QAAQ;GAE5C,IAAI,SAAS,OACX,KAAK,MAAM,CAAC,WAAW,cAAc,OAAO,QAAQ,SAAS,KAAK,GAChE,MAAM,aAAa;GAOvB,IAAI,SAAS,WAAW,YAAY,SAAS,OAAO;IAClD,MAAM,eAAe,MAAM,oBAAoB;KAC7C;KACA;KACA,OAAO,SAAS;KAChB,KAAK,aAAa;MAAE;MAAO,aAAa,SAAS;KAAgB,CAAC;IACpE,CAAC;IAED,IAAI,iBAAiB,QAAW;KAG9B,MAAM,iBAAiB,MAAM,gBAAgB;MAC3C;MACA;MACA;MACA;MACA;MACA;MACA;MACA,MAAM,KAAK;MACX,QAAQ;KACV,CAAC;KACD,IAAI,CAAC,eAAe,IAAI;MACtB,MAAM,aAAa,UAAU,eAAe,KAAK;MACjD,QAAQ,KACN,kBACA,EAAE,OAAO,WAAW,GACpB,OAAO,iBACT;MACA,OAAO,MACL,WACA,kBACA,2BACA;OACE,MAAM,KAAK;OACX,MAAM,WAAW;OACjB,SAAS,WAAW;MACtB,CACF;KACF;KACA,iBAAiB,KAAK;KACtB,MAAM,SAAS;IACjB;IAIA,MAAM,cAAc,iBAAiB,QAAQ,OAAO;IACpD,IAAI,gBAAgB,QAAQ,CAAC,WAAW,IAAI,WAAW,GACrD,MAAM,IAAI,aACR,aAAa,WAAW,KAAK,WAAW,KAAK,KAAK,wCAAwC,YAAY,IACtG;KAAE,UAAU,KAAK;KAAM,YAAY;IAAY,CACjD;IAGF,MAAM,iBAAiB,MAAM,gBAAgB;KAC3C;KACA;KACA;KACA;KACA;KACA;KACA;KACA,MAAM;KACN,QAAQ;IACV,CAAC;IACD,IAAI,CAAC,eAAe,IAAI;KACtB,MAAM,aAAa,UAAU,eAAe,KAAK;KACjD,QAAQ,KACN,kBACA,EAAE,OAAO,WAAW,GACpB,OAAO,iBACT;KACA,OAAO,MACL,WACA,kBACA,2BACA;MACE,MAAM,KAAK;MACX,MAAM,WAAW;MACjB,SAAS,WAAW;KACtB,CACF;IACF;IAEA,IAAI,QAAQ,SAAS,MAAM,qBAAqB,MAAM;IAEtD,IAAI,iBAAiB,OAAO;KAC1B,cAAc;KACd;IACF;IAEA,WAAW;IACX,cAAc;IACd;GACF;GAIA,MAAM,WAAW,MAAM,gBAAgB;IACrC;IACA;IACA,KAAK,aAAa;KAAE;KAAO,aAAa,SAAS;IAAgB,CAAC;GACpE,CAAC;GAED,MAAM,WACJ,aAAa,QACT,OACA,OAAO,aAAa,WAClB,WACA,iBAAiB,YAAY,KAAK,IAAI;GAG9C,MAAM,UAAU,MAAM,gBAAgB;IACpC;IACA;IACA;IACA;IACA;IACA;IACA;IACA,MAAM;IACN,QAAQ;GACV,CAAC;GACD,IAAI,CAAC,QAAQ,IAAI;IACf,MAAM,aAAa,UAAU,QAAQ,KAAK;IAC1C,QAAQ,KACN,kBACA,EAAE,OAAO,WAAW,GACpB,OAAO,iBACT;IACA,OAAO,MAAM,WAAW,kBAAkB,2BAA2B;KACnE,MAAM,KAAK;KACX,MAAM,WAAW;KACjB,SAAS,WAAW;IACtB,CAAC;GACH;GAEA,IAAI,QAAQ,SAAS,MAAM,qBAAqB,MAAM;GAEtD,IAAI,aAAa,OAAO;IACtB,cAAc;IACd;GACF;GAEA,IAAI,OAAO,aAAa,UAAU;IAChC,IAAI,CAAC,WAAW,IAAI,QAAQ,GAC1B,MAAM,IAAI,aACR,aAAa,WAAW,KAAK,WAAW,KAAK,KAAK,yBAAyB,SAAS,IACpF;KAAE,UAAU,KAAK;KAAM,YAAY;IAAS,CAC9C;IAEF,WAAW;IACX,cAAc;IACd;GACF;GAEA,cAAc;GACd,WAAW;EACb;CACF,SAAS,KAAK;EACZ,IAAI,eAAe,wBAAwB;GACzC,SAAS;GACT,cAAc,IAAI;GAClB,QAAQ;EACV,OAAO,IAAI,eAAe,SAAS;GACjC,SAAS;GACT,QAAQ;EACV,OAAO;GACL,SAAS;GACT,QAAQ,IAAI,cACV,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GAC/C,EAAE,OAAO,IAAI,CACf;EACF;CACF;CAOA,MAAM,2BAAU,IAAI,KAAK,EAAC,CAAC,YAAY;CACvC,MAAM,WAAW,YAAY,IAAI,IAAI;CAErC,IAAI;CACJ,IAAI,WAAW,eAAe,WAAW,QACvC,IAAI;EACF,MAAM,YAAY,MAAM,WAAW,OAAO,QACxC,aAAa,EAAE,MAAM,CAAC,CACxB;EACA,OAAQ,MAAM,uBACZ,WAAW,OAAO,QAClB,SACF;CACF,SAAS,KAAK;EACZ,SAAS;EACT,QACE,eAAe,UACX,MACA,IAAI,cACF,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GAC/C,EAAE,OAAO,IAAI,CACf;CACR;CASF,MAAM,WAAyB,CAAC;CAChC,KAAK,MAAM,YAAY,OAAO;EAC5B,MAAM,OAAO,MAAM;EACnB,IAAI,KAAK,aACP,SAAS,KAAK,KAAK,WAAW;CAElC;CAEA,MAAM,SAAyB;EAC7B;EACA,WAAW;EACX,MAAM,WAAW;EACjB,SAAS,WAAW;EACpB,MAAM;EACN,cAAc,WAAW;EACzB;EACA;EAKA,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC;EACzB;EACA;EACA;EACA;EACA;EACA;EACA;EACA,OAAO,WAAW,WAAW,KAAK,CAAC;CACrC;CAMA,mBAAmB,QAAQ;EACzB,WAAW;EACX,WAAW,OAAO;CACpB,CAAC;CAUD,MAAM,eAAe,MAAM,gBAAgB;EACzC;EACA;EACA;EACA;EACA;EACA;EACA;EACA,MAVgB,WAAW,WAAW,kBAAkB,OAAO;EAW/D;CACF,CAAC;CACD,IAAI,CAAC,aAAa,IAAI;EACpB,MAAM,aAAa,UAAU,aAAa,KAAK;EAC/C,QAAQ,KACN,kBACA,EAAE,OAAO,WAAW,GACpB,OAAO,iBACT;EACA,OAAO,MAAM,WAAW,kBAAkB,iCAAiC;GACzE,MAAM,WAAW;GACjB,SAAS,WAAW;EACtB,CAAC;CACH;CAEA,MAAM,SAAkC;EACtC,MAAM;EACN;EACA;EACA;EACA;CACF;CAEA,IAAI,WAAW,aAAa;EAC1B,QAAQ,KACN,sBACA;GACE,aAAa,eAAe;GAC5B,QAAS,OAA8C,UAAU;EACnE,GACA,OAAO,iBACT;EACA,OAAO,KAAK,WAAW,aAAa,sBAAsB,EAAE,MAAM,CAAC;CACrE;CAEA,IAAI,WAAW,YAAY,OAAO;EAChC,QAAQ,KAAK,kBAAkB,EAAE,MAAM,GAAG,OAAO,iBAAiB;EAClE,OAAO,MAAM,WAAW,SAAS,mBAAmB;GAClD;GACA,MAAM,MAAM;GACZ,SAAS,MAAM;EACjB,CAAC;CACH;CAEA,QAAQ,KACN,sBACA,EAAU,OAAkC,GAC5C,OAAO,iBACT;CACA,OAAO,KAAK,WAAW,aAAa,sBAAsB;EACxD;EACA;EACA;CACF,CAAC;CAED,OAAO;AACT;;;;;;;;AASA,eAAe,oBAAuB,QAKE;CACtC,MAAM,EAAE,MAAM,YAAY,OAAO,QAAQ;CACzC,IAAI,CAAC,KAAK,WAAW,OAAO;CAE5B,IAAI;CACJ,IAAI;EACF,UAAU,MAAM,KAAK,UAAU,KAAK,KAAK;CAC3C,SAAS,KAAK;EACZ,MAAM,IAAI,aACR,aAAa,WAAW,KAAK,WAAW,KAAK,KAAK,oBAClD;GAAE,UAAU,KAAK;GAAM,OAAO;EAAI,CACpC;CACF;CACA,OAAO,YAAY,OAAO;AAC5B;AAEA,SAAS,mBACP,YACA,YACe;CACf,IAAI,CAAC,YAAY,OAAO,WAAW,MAAM,EAAE,EAAE,QAAQ;CAMrD,IACE,WAAW,QACX,WAAW,MAAM,MAAK,MAAK,EAAE,SAAS,WAAW,IAAI,GAErD,OAAO,WAAW;CAGpB,KAAK,MAAM,QAAQ,WAAW,OAAO;EACnC,MAAM,OAAO,WAAW,MAAM,KAAK;EACnC,IAAI,CAAC,QAAS,KAAK,WAAW,eAAe,KAAK,WAAW,WAC3D,OAAO,KAAK;CAEhB;CAEA,OAAO;AACT;AAEA,eAAe,uBACb,QACA,OACkB;CAClB,IAAI,CAAC,QAAQ,OAAO;CAEpB,MAAM,SAAS,MACb,OAGA,YAAY,CAAC,SAAS,KAAK;CAE7B,IACE,UACA,OAAO,WAAW,YAClB,YAAY,UACX,OAA+B,QAEhC,MAAM,IAAI,sBACR,4CACA,EACE,QAAS,OAA2B,OACtC,CACF;CAGF,OAAQ,OAA8B;AACxC"}

package/esm/workflow/step-runner.mjs

@@ -4,6 +4,8 @@ import { SchemaValidationError } from "../errors/schema-validation-error.mjs";
 import { StepFailedError } from "../errors/step-failed-error.mjs";
 import { WorkflowCancelledError } from "../errors/workflow-cancelled-error.mjs";
 import "../errors/index.mjs";
+import { mergeUsage } from "../utils/compute-cost.mjs";
+import { withoutRunFrame } from "../utils/run-context.mjs";
 import { isAbortError, resolveBackoff, resolveRetryConfig } from "./retry.mjs";
 import { createCancelledError, sleep } from "./cancellation.mjs";
 import { cloneState, deepFreeze } from "./state.mjs";
@@ -78,17 +80,14 @@ async function executeStep(params) {
 			const attemptState = cloneState(params.state);
 			if (step.before) await step.before(params.buildContext({ state: attemptState }));
 			if (step.agent) {
+				const agent = step.agent;
 				const { prompt, ...agentOpts } = step.input ? await step.input(params.buildContext({ state: attemptState })) : { prompt: "" };
-				const result = await step.agent.execute(prompt, {
+				const result = await withoutRunFrame(() => agent.execute(prompt, {
 					...agentOpts,
 					signal
-				});
+				}));
 				executionResult = result;
-				if (result.usage) {
-					params.usage.input += result.usage.input ?? 0;
-					params.usage.output += result.usage.output ?? 0;
-					params.usage.total += result.usage.total ?? 0;
-				}
+				if (result.usage) mergeUsage(params.usage, result.usage);
 				if (result.error) throw result.error;
 			} else if (step.run) executionResult = await step.run(params.buildContext({ state: attemptState }));
 			if (step.output) {
@@ -222,25 +221,25 @@ async function runParallelStep(params) {
 	const childSnapshots = {};
 	let firstError;
 	const results = await Promise.all((step.parallel ?? []).map(async (child) => {
-		const snap = await executeStep({
-			step: child,
-			state: sharedState,
-			emitter,
-			executionHandlers,
-			logger,
-			logModule,
-			signal,
-			buildContext: params.buildContext,
-			usage: params.usage
-		});
-		Object.assign(sharedState, snap.state);
 		return {
 			child,
-			snap
+			snap: await executeStep({
+				step: child,
+				state: sharedState,
+				emitter,
+				executionHandlers,
+				logger,
+				logModule,
+				signal,
+				buildContext: params.buildContext,
+				usage: params.usage
+			})
 		};
 	}));
 	for (const { child, snap } of results) {
 		childSnapshots[child.name] = finalizeSnapshot(snap);
+		if (step.mergeState) step.mergeState(sharedState, snap.state, child.name);
+		else Object.assign(sharedState, snap.state);
 		if (snap.status === "failed" && !firstError && snap.error) firstError = snap.error;
 	}
 	const endedAt = (/* @__PURE__ */ new Date()).toISOString();