@giveitsmaller/sdk 0.8.0 → 0.10.0

package/dist/file-first.js

@@ -9,12 +9,21 @@
  *
  * Mirrors `packages/php/src/FileFirst/*`.
  */
-import { GislApiError, GislConfigError, GislNoSuchKeyError, GislSinkError, GislTimeoutError } from './errors.js';
+import { GislConfigError, GislNetworkError, GislNoSuchKeyError, GislSinkError, GislTimeoutError, SseEndedWithoutTerminal } from './errors.js';
 import { _detectCompressMedia, _consumeSseToTerminal, _pollToTerminal, _parseMaxWait, _checkAborted, } from './builder.js';
-import { HttpDownloader } from './http-downloader.js';
+import { LazyHttpDownloader } from './lazy-downloader.js';
 import { resolveCompressOptions, } from './ergonomic/preset_resolver.js';
 import { OptimizeFor } from './generated/sdk_spec/enums.js';
-import { uploadSource } from './types.js';
+import { uploadSource, jobOutputSource } from './types.js';
+// Value import used only at call-time (inside MergedRecipe.toWorkflowPayload),
+// never at module-eval, so the file-first <-> merge <-> handle import cycle
+// resolves cleanly under ESM (same deferred-usage discipline as `Handle`).
+import { wireMergeOptions } from './merge.js';
+// Deferred-usage-only import: `Handle` is constructed inside `submit()` at call
+// time, never at module-eval, so the handle.ts <-> file-first.ts back-edge
+// (handle.ts imports RunResult/projectDownloadsToRunResult from here) resolves
+// cleanly under ESM. Mirrors builder.ts/merge.ts importing Handle the same way.
+import { Handle } from './handle.js';
 /**
  * Result of a file-first run. Coexists with the operation-first `Result`
  * (in `builder.ts`) until FF6.
@@ -163,6 +172,190 @@ export class RunResult {
         return this.downloader;
     }
 }
+/**
+ * Flatten the terminal workflow status + its downloads into a {@link RunResult}.
+ *
+ * Shared by {@link Recipe.run} (passes its recipe key) and the file-first
+ * {@link Handle} reattach surface (`Handle.wait()`/`Handle.result()`, FF5a —
+ * passes `null` because a reattached handle carries no recipe key).
+ *
+ * **Partition invariant (carries a prior codex-review fix — do NOT let it
+ * drift):** success is ONLY `state === 'completed'`. Every other terminal
+ * state — `failed`, `partially_failed`, `cancelled`, `expired`,
+ * `paused_insufficient_credits` — partitions into `failed[]` so a caller's
+ * `ok`/`succeeded` check can never treat a cancelled/expired/paused run as a
+ * clean result.
+ *
+ * @internal Exported for reuse by the file-first `Handle`; not part of the
+ *   caller-facing fluent surface.
+ */
+export function projectDownloadsToRunResult(workflowId, finalStatus, jobDownloads, key, downloader) {
+    // Flatten to the lean OutputFile[] (the four file-first fields only).
+    const artifacts = [];
+    for (const job of jobDownloads) {
+        for (const f of job.files) {
+            artifacts.push({
+                url: f.downloadUrl,
+                filename: f.filename,
+                sizeBytes: f.sizeBytes,
+                operation: f.operation,
+            });
+        }
+    }
+    const state = finalStatus.status;
+    let succeeded;
+    let failed;
+    if (state === 'completed') {
+        succeeded = [{ key, outputs: artifacts }];
+        failed = [];
+    }
+    else {
+        const firstError = (finalStatus.jobs ?? [])
+            .flatMap((j) => j.operations ?? [])
+            .map((op) => op.errorMessage)
+            .find((m) => m !== undefined);
+        succeeded = [];
+        failed = [
+            { key, error: new Error(firstError !== undefined ? `${state}: ${firstError}` : state) },
+        ];
+    }
+    return new RunResult(workflowId, state, artifacts, succeeded, failed, downloader);
+}
+/**
+ * Flatten a terminal multi-job workflow (the `client.files([...])` fan-out)
+ * into a partitioned {@link RunResult}. One job per input file, keyed by the
+ * `file-{i}` job ref the {@link FilesRecipe} lowering assigns; the result's
+ * `succeeded` / `failed` partition is PER JOB, so one bad input does not sink
+ * the rest.
+ *
+ * Join model: `finalStatus.jobs[]` carries the per-job {@link JobStatus} +
+ * `operations[]` (for the error message); `jobDownloads[]` carries the per-job
+ * output files. Both are joined on the job `ref` ("file-{i}"); the partition
+ * key is the index `"{i}"` parsed out of that ref. The flat `artifacts[]` is
+ * every job's outputs in job order (the order `finalStatus.jobs[]` lists them).
+ *
+ * **Partition invariant (mirrors {@link projectDownloadsToRunResult} PER JOB —
+ * do NOT let it drift):** a job is a SUCCESS only when its
+ * {@link JobResponse.status} `=== 'completed'`. Any other per-job status —
+ * `failed`, `pending`, `waiting`, `blocked_insufficient_credits`,
+ * `in_progress` — partitions that job into `failed[]` (with that job's first
+ * operation error message, scoped to THAT job only).
+ *
+ * @internal Exported for the file-first `client.files([...]).run()` producer;
+ *   not part of the caller-facing fluent surface.
+ */
+export function projectMultiJobToRunResult(workflowId, finalStatus, jobDownloads, keyByRef, downloader) {
+    // Group downloads by job ref so a job's outputs can be flattened AFTER the
+    // per-job partition is decided (grouping is unrecoverable post-flatten).
+    const filesByRef = new Map();
+    for (const job of jobDownloads) {
+        filesByRef.set(job.ref, job.files);
+    }
+    const artifacts = [];
+    const succeeded = [];
+    const failed = [];
+    const jobs = finalStatus.jobs ?? [];
+    for (const job of jobs) {
+        const key = keyByRef.get(job.ref) ?? jobIndexFromRef(job.ref);
+        const outputs = (filesByRef.get(job.ref) ?? []).map((f) => ({
+            url: f.downloadUrl,
+            filename: f.filename,
+            sizeBytes: f.sizeBytes,
+            operation: f.operation,
+        }));
+        // The flat artifacts[] keeps every job's outputs in job order.
+        artifacts.push(...outputs);
+        if (job.status === 'completed') {
+            succeeded.push({ key, outputs });
+        }
+        else {
+            const firstError = (job.operations ?? [])
+                .map((op) => op.errorMessage)
+                .find((m) => m !== undefined);
+            failed.push({
+                key,
+                error: new Error(firstError !== undefined ? `${job.status}: ${firstError}` : String(job.status)),
+            });
+        }
+    }
+    return new RunResult(workflowId, finalStatus.status, artifacts, succeeded, failed, downloader);
+}
+/** Derive the partition key `"{i}"` from a `file-{i}` job ref; the ref verbatim otherwise. */
+function jobIndexFromRef(ref) {
+    return ref.startsWith('file-') ? ref.slice('file-'.length) : ref;
+}
+const _FANOUT_REF = /^file-\d+$/;
+/**
+ * True when a terminal status describes a homogeneous `files([...])` fan-out —
+ * i.e. it has at least one job and EVERY job ref is `file-{i}` (the ids the
+ * {@link FilesRecipe} lowering assigns). A single-file {@link Recipe} omits the
+ * job id, so its job carries a non-`file-N` ref (e.g. `op`) and this is false.
+ *
+ * This is the data-driven seam that lets {@link Handle.wait}/{@link Handle.result}
+ * pick the per-job producer ({@link projectMultiJobToRunResult}) over the
+ * single-output one for a fan-out — WITHOUT a construction-time marker, so a
+ * fan-out **reattached** via `client.workflow(id)` (which carries no marker)
+ * still partitions per job. Keys are recovered from the `file-{i}` refs.
+ *
+ * @internal Exported for the file-first `Handle`; not part of the public API.
+ */
+export function isFanoutStatus(finalStatus) {
+    const jobs = finalStatus.jobs ?? [];
+    return jobs.length > 0 && jobs.every((job) => _FANOUT_REF.test(job.ref));
+}
+const _MERGE_SRC_REF = /^src_\d+$/;
+/**
+ * True when a terminal status describes a fluent `files([...]).merge(...)`
+ * combine — at least one job ref `merge` and every OTHER job ref is `src_{i}`
+ * (the ids the {@link MergedRecipe} lowering assigns). The data-driven seam that
+ * lets {@link Handle.wait}/{@link Handle.result} project ONLY the merged output
+ * — filtering the `src_*` passthrough plumbing — even after a
+ * `client.workflow(id)` reattach (no construction-time marker), matching
+ * {@link MergedRecipe.run}'s `ref === 'merge'` filter. Mutually exclusive with
+ * {@link isFanoutStatus} (a fan-out's refs are all `file-{i}`).
+ *
+ * @internal Exported for the file-first `Handle`; not part of the public API.
+ */
+export function isMergeStatus(finalStatus) {
+    const jobs = finalStatus.jobs ?? [];
+    if (jobs.length === 0)
+        return false;
+    let hasMerge = false;
+    for (const job of jobs) {
+        if (job.ref === 'merge') {
+            hasMerge = true;
+            continue;
+        }
+        if (!_MERGE_SRC_REF.test(job.ref))
+            return false;
+    }
+    return hasMerge;
+}
+/**
+ * True when a terminal status describes a fluent `files([...]).archive(...)`
+ * bundle — at least one job ref `archive` and every OTHER job ref is `src_{i}`
+ * (the ids the {@link ArchivedRecipe} lowering assigns). Lets
+ * {@link Handle.wait}/{@link Handle.result} project ONLY the archive output —
+ * filtering the `src_*` passthrough plumbing — even after a `client.workflow(id)`
+ * reattach. Mutually exclusive with {@link isFanoutStatus} / {@link isMergeStatus}.
+ *
+ * @internal Exported for the file-first `Handle`; not part of the public API.
+ */
+export function isArchiveStatus(finalStatus) {
+    const jobs = finalStatus.jobs ?? [];
+    if (jobs.length === 0)
+        return false;
+    let hasArchive = false;
+    for (const job of jobs) {
+        if (job.ref === 'archive') {
+            hasArchive = true;
+            continue;
+        }
+        if (!_MERGE_SRC_REF.test(job.ref))
+            return false;
+    }
+    return hasArchive;
+}
 /** Named constructors for {@link FileInput} — mirror the PHP static factories. */
 export const fileInput = {
     path(path) {
@@ -171,6 +364,19 @@ export const fileInput = {
     blob(blob) {
         return { kind: 'blob', blob };
     },
+    /**
+     * Reference an already-uploaded file by its upload id, instead of
+     * re-uploading bytes.
+     *
+     * Auth-ownership: an upload created by an **authenticated** caller is owned
+     * by that caller. If you reuse the id from a client configured with a
+     * *different* auth context (a different `apiKey` / session), workflow-create
+     * returns `404 upload_not_found` — the server enforces ownership (api
+     * PqpD9ySv). Reference an upload id only under the SAME auth that created it.
+     * The normal upload-then-create-in-one-client flow is consistent by
+     * construction (the same `Authorization` rides every request). Ownerless
+     * (anonymous-intake) uploads are unaffected.
+     */
     uploadId(fileId) {
         return { kind: 'uploadId', fileId };
     },
@@ -252,16 +458,20 @@ export class Recipe {
      * `operations[]`; the job `id` is omitted (a single job referenced by
      * nothing — the server auto-assigns `job_N`).
      *
-     * @internal Consumed by FF2b's `run()` (after a real upload) and by the
-     *   cross-language parity harness (with a fixed id). Not part of the
-     *   caller-facing fluent surface.
+     * When `callbackUrl` is given (the file-first `submit()` path), it is built
+     * INTO the payload at construction (`callback_url`) rather than spread onto an
+     * already-built readonly payload. `run()` passes no `callbackUrl`.
+     *
+     * @internal Consumed by FF2b's `run()` (after a real upload), FF5b's
+     *   `submit()` (with a webhook), and the cross-language parity harness (with a
+     *   fixed id). Not part of the caller-facing fluent surface.
      */
-    toWorkflowPayload(fileId) {
+    toWorkflowPayload(fileId, callbackUrl) {
         const operations = this.steps.map((step) => this.lowerStep(step));
         // Key order (source, operations) matches the PHP `toWire()` so the
         // JSON-string serialisation is byte-identical across languages.
         const job = { source: uploadSource(fileId), operations };
-        return { jobs: [job] };
+        return callbackUrl === undefined ? { jobs: [job] } : { jobs: [job], callback_url: callbackUrl };
     }
     /** The result-addressing key passed to `file()`, or undefined. */
     key() {
@@ -271,6 +481,14 @@ export class Recipe {
     get stepCount() {
         return this.steps.length;
     }
+    /**
+     * The captured op chain. Read by {@link FilesRecipe} to compose a shared
+     * chain across many inputs without duplicating the chain-method validation.
+     * @internal
+     */
+    get recipeSteps() {
+        return this.steps;
+    }
     /**
      * Execute the recipe end-to-end: upload the input (when required), create
      * the workflow, await a terminal state (SSE with poll fallback), then
@@ -289,34 +507,9 @@ export class Recipe {
             throw new GislConfigError('Recipe.run() requires a client; build the recipe via gisl().file(...) rather than constructing Recipe directly.', { reason: 'no_client' });
         }
         const deadline = Date.now() + _parseMaxWait(options.maxWait ?? 300_000);
-        // 1. Resolve the upload id. A pre-uploaded id skips the upload entirely;
-        // a path / blob is uploaded now, emitting {phase:'upload'} progress.
-        let fileId;
-        if (this.input.kind === 'uploadId') {
-            fileId = this.input.fileId;
-        }
-        else {
-            const source = this.input.kind === 'path' ? this.input.path : this.input.blob;
-            const up = await this.client.uploadFile(source, {
-                signal,
-                ...(onProgress !== undefined
-                    ? {
-                        onProgress: (uploadedBytes, totalBytes) => {
-                            onProgress({ phase: 'upload', uploadedBytes, totalBytes });
-                        },
-                    }
-                    : {}),
-            });
-            fileId = up.fileId;
-        }
-        _checkAborted(signal);
-        if (Date.now() >= deadline) {
-            throw new GislTimeoutError('Upload completed but maxWait elapsed before workflow could be created');
-        }
-        // 2. Create the workflow from the lowered payload.
-        const payload = this.toWorkflowPayload(fileId);
-        const created = await this.client.createWorkflow(payload);
-        _checkAborted(signal);
+        // 1+2. Upload (when required) + create the workflow. Shared with submit()
+        // (which passes a webhook → callback_url). run() passes no webhook.
+        const created = await this._uploadAndCreate(undefined, deadline, onProgress, signal);
         // 3. Wait to terminal status — SSE first, poll on a genuine SSE error.
         // Caller-aborted + deadline-elapsed errors MUST propagate (not transient).
         let finalStatus;
@@ -329,18 +522,16 @@ export class Recipe {
             });
         }
         catch (err) {
-            // Only genuine SSE transport / clean-stream-end failures fall through to
-            // poll. Caller-deadline, abort, and API errors (a 401/402/etc. from
-            // /events, or an onProgress callback throw surfacing as GislApiError)
-            // MUST propagate — re-issuing the same doomed request via poll would mask
-            // them. Mirrors the PHP BuilderInternals::awaitTerminal sealed-marker
-            // discipline (codex review medium).
-            if (err instanceof GislTimeoutError)
-                throw err;
-            if (err instanceof DOMException && err.name === 'AbortError')
-                throw err;
-            if (err instanceof GislApiError)
+            // TDqmkWpX: poll-fallback ONLY on a clean SSE stream-end
+            // (SseEndedWithoutTerminal) or a typed transport error (GislNetworkError).
+            // Everything else — caller-deadline, abort, an API error from /events, an
+            // onProgress callback throw (propagates as-is, NOT wrapped), anything
+            // unexpected — MUST propagate; re-issuing the same doomed request via poll
+            // would mask it. Mirrors the PHP BuilderInternals::awaitTerminal sealed-
+            // marker discipline.
+            if (!(err instanceof SseEndedWithoutTerminal || err instanceof GislNetworkError)) {
                 throw err;
+            }
             finalStatus = await _pollToTerminal(this.client, {
                 workflowId: created.workflowId,
                 deadline,
@@ -354,46 +545,85 @@ export class Recipe {
             throw new GislTimeoutError(`Workflow ${created.workflowId} reached terminal status but maxWait elapsed before downloads could be fetched`);
         }
         const downloads = await this.client.getWorkflowDownloads(created.workflowId);
-        // Flatten to the lean OutputFile[] (the four file-first fields only).
-        const artifacts = [];
-        for (const job of downloads.downloads) {
-            for (const f of job.files) {
-                artifacts.push({
-                    url: f.downloadUrl,
-                    filename: f.filename,
-                    sizeBytes: f.sizeBytes,
-                    operation: f.operation,
-                });
-            }
+        // TDqmkWpX: re-check AFTER the downloads fetch so a slow getWorkflowDownloads
+        // cannot return a success past the advertised maxWait deadline.
+        if (Date.now() >= deadline) {
+            throw new GislTimeoutError(`Workflow ${created.workflowId} downloads fetch completed after maxWait elapsed`);
         }
-        // Partition the single input into succeeded / failed by terminal state.
-        // Success is ONLY `completed`: every other terminal state — `failed`,
-        // `partially_failed`, `cancelled`, `expired`,
-        // `paused_insufficient_credits` — is a non-success and partitions into
-        // `failed` so a caller's `ok`/`succeeded` check can never treat a
-        // cancelled/expired/paused run as a clean result (codex review high).
-        const state = finalStatus.status;
-        const key = this.recipeKey ?? null;
-        let succeeded;
-        let failed;
-        if (state === 'completed') {
-            succeeded = [{ key, outputs: artifacts }];
-            failed = [];
+        // Download URLs from getWorkflowDownloads are pre-signed and require no SDK
+        // auth, so the downloader issues a plain unauthenticated fetch.
+        const downloader = new LazyHttpDownloader();
+        return projectDownloadsToRunResult(created.workflowId, finalStatus, downloads.downloads, this.recipeKey ?? null, downloader);
+    }
+    /**
+     * Fire-and-forget the recipe: upload the input (when required), create the
+     * workflow (wiring `webhook` into `callback_url` when given), and return a
+     * client-bound {@link Handle} carrying the workflow id + webhook secret + the
+     * recipe key. Does NOT wait for terminal status — call `handle.wait()` /
+     * `handle.result()` later to collect the {@link RunResult}.
+     *
+     * Requires a client bound at construction time (same `no_client` guard as
+     * {@link run}). `webhook` is OPTIONAL: when omitted, no `callback_url` is
+     * sent. Mirrors the PHP `Recipe.submit()`.
+     *
+     * @param webhook Absolute callback URL the server POSTs lifecycle events to.
+     */
+    async submit(webhook) {
+        if (this.client === undefined) {
+            throw new GislConfigError('Recipe.submit() requires a client; build the recipe via gisl().file(...) rather than constructing Recipe directly.', { reason: 'no_client' });
+        }
+        // submit() is fire-and-forget — NO whole-run deadline. The upload may be
+        // large (a multi-GB master, example 12) and is bounded by the HTTP client's
+        // own request timeout, not an arbitrary submit-side cap. Pass `undefined`
+        // so the post-upload deadline check is skipped: a 300s cap here would throw
+        // on a slow-but-successful big upload before createWorkflow (codex).
+        const created = await this._uploadAndCreate(webhook, undefined);
+        return new Handle(created.workflowId, created.webhookSecret != null ? created.webhookSecret : undefined, this.client, this.recipeKey ?? null);
+    }
+    /**
+     * Resolve the upload id (verbatim for a pre-uploaded id; uploading a path /
+     * blob otherwise, emitting `{phase:'upload'}` progress), check the post-upload
+     * deadline, lower to the workflow-create payload (wiring `webhook` into
+     * `callback_url`), and create the workflow. Shared first half of
+     * {@link run} + {@link submit}.
+     *
+     * The post-upload deadline check carries a prior codex fix (9a117f04eb59): a
+     * slow upload must not proceed to createWorkflow past the deadline.
+     */
+    async _uploadAndCreate(webhook, deadline, onProgress, signal) {
+        // 1. Resolve the upload id. A pre-uploaded id skips the upload entirely;
+        // a path / blob is uploaded now, emitting {phase:'upload'} progress.
+        let fileId;
+        if (this.input.kind === 'uploadId') {
+            fileId = this.input.fileId;
         }
         else {
-            const firstError = (finalStatus.jobs ?? [])
-                .flatMap((j) => j.operations ?? [])
-                .map((op) => op.errorMessage)
-                .find((m) => m !== undefined);
-            succeeded = [];
-            failed = [
-                { key, error: new Error(firstError !== undefined ? `${state}: ${firstError}` : state) },
-            ];
+            const source = this.input.kind === 'path' ? this.input.path : this.input.blob;
+            const up = await this.client.uploadFile(source, {
+                signal,
+                ...(onProgress !== undefined
+                    ? {
+                        onProgress: (uploadedBytes, totalBytes) => {
+                            onProgress({ phase: 'upload', uploadedBytes, totalBytes });
+                        },
+                    }
+                    : {}),
+            });
+            fileId = up.fileId;
         }
-        // Download URLs from getWorkflowDownloads are pre-signed and require no SDK
-        // auth, so the downloader issues a plain unauthenticated fetch.
-        const downloader = new HttpDownloader();
-        return new RunResult(created.workflowId, state, artifacts, succeeded, failed, downloader);
+        _checkAborted(signal);
+        // run() passes a whole-run deadline (the codex 9a117f04eb59 fix: a slow
+        // upload must not proceed to createWorkflow past maxWait); submit() passes
+        // `undefined` (fire-and-forget, no upload cap), so the check is skipped.
+        if (deadline !== undefined && Date.now() >= deadline) {
+            throw new GislTimeoutError('Upload completed but maxWait elapsed before workflow could be created');
+        }
+        // 2. Create the workflow from the lowered payload (callback_url built into
+        // the payload at construction when a webhook is given).
+        const payload = this.toWorkflowPayload(fileId, webhook);
+        const created = await this.client.createWorkflow(payload);
+        _checkAborted(signal);
+        return created;
     }
     withStep(step) {
         return new Recipe(this.input, this.recipeKey, [...this.steps, step], this.presetDefaults, this.scopedPresetDefaults, this.client);
@@ -443,3 +673,732 @@ export class Recipe {
         return undefined;
     }
 }
+/**
+ * The homogeneous fan-out builder value (FF3a). `client.files([a, b, c])`
+ * returns a `FilesRecipe`; the op-chain methods (`compress`, `convert`,
+ * `thumbnail`, `textWatermark`) build ONE shared recipe (chain) that is applied
+ * to EVERY input file in ONE workflow. `run()` returns a partitioned
+ * {@link RunResult} — one `succeeded`/`failed` entry per input, keyed by its
+ * 0-based index ("0", "1", …) so one bad input does not sink the rest.
+ *
+ * **Immutable / clone-on-write**, exactly like {@link Recipe}: every op returns
+ * a NEW `FilesRecipe` carrying the appended step. The inputs are held as an
+ * ORDERED list (NOT a map) so the per-file index is the partition key.
+ *
+ * **Lowering composes {@link Recipe} per file** rather than duplicating
+ * `lowerStep`/`lowerCompressOptions`: for each input `i` it builds an internal
+ * single-file `Recipe(input_i, …, steps)`, calls its `toWorkflowPayload` to get
+ * that file's one-job payload, then merges all jobs into ONE
+ * {@link WorkflowCreatePayload} with `jobs[i].id = "file-{i}"`. This preserves
+ * each file's media-hint (different extensions per input resolve compress
+ * presets independently).
+ *
+ * Exposes both `run()` (blocking, returns a partitioned {@link RunResult}) and
+ * `submit(webhook?)` (fire-and-forget, returns a {@link Handle}). Mirrors the
+ * PHP `FilesRecipe`.
+ */
+export class FilesRecipe {
+    inputs;
+    steps;
+    presetDefaults;
+    scopedPresetDefaults;
+    client;
+    constructor(inputs, steps = [], presetDefaults, scopedPresetDefaults, client) {
+        this.inputs = inputs;
+        this.steps = steps;
+        this.presetDefaults = presetDefaults;
+        this.scopedPresetDefaults = scopedPresetDefaults;
+        this.client = client;
+    }
+    /**
+     * Reduce file size on every input. `optimize` selects a per-media preset
+     * (resolved per file at lower-time, so each input's extension picks its own
+     * preset). Reuses {@link Recipe}'s validation — a directly-constructed
+     * lowering builds an internal Recipe that throws the same `GislConfigError`.
+     */
+    compress(optimize) {
+        return this.withStep(this.baseRecipe().compress(optimize));
+    }
+    /** Change every input's format. `format` lowers verbatim to the `format` option. */
+    convert(format) {
+        return this.withStep(this.baseRecipe().convert(format));
+    }
+    /** Generate a preview of every input. Omitted dimensions are dropped from the wire options. */
+    thumbnail(options = {}) {
+        return this.withStep(this.baseRecipe().thumbnail(options));
+    }
+    /** Apply the same text watermark to every input. */
+    textWatermark(text) {
+        return this.withStep(this.baseRecipe().textWatermark(text));
+    }
+    /**
+     * Combine the inputs into ONE output (N→1), in array order (FF3b). Returns a
+     * single-output {@link MergedRecipe} you chain further ops on
+     * (`files([...]).merge().compress()`). Reuses the operation-first
+     * {@link MergeOptions} for the merge-level options, so the wire shape matches
+     * `client.merge([...], options)`.
+     *
+     * `merge()` must be the FIRST op on `files([...])` — per-file ops before a
+     * combine (compress-each-then-merge) are a separate follow-up, rejected here
+     * with `GislConfigError` reason `pre_merge_ops_unsupported`.
+     */
+    merge(options = {}) {
+        if (this.steps.length !== 0) {
+            throw new GislConfigError('merge() must be the first operation on files([...]); applying per-file ops before a combine ' +
+                '(compress-each-then-merge) is not yet supported — call merge() directly, then chain ops on the merged output.', { reason: 'pre_merge_ops_unsupported' });
+        }
+        return new MergedRecipe(this.inputs, options, [], this.presetDefaults, this.scopedPresetDefaults, this.client);
+    }
+    /**
+     * Bundle the inputs into ONE archive (N→1, zip / tar.gz) — media-agnostic,
+     * inputs may mix types. Returns a terminal {@link ArchivedRecipe} (a zip is
+     * the final artefact — no post-bundle chain). `format` / `folderStructure` are
+     * optional; the server defaults to zip + flat.
+     *
+     * `archive()` must be the FIRST op on `files([...])` → `GislConfigError` reason
+     * `pre_archive_ops_unsupported` otherwise.
+     */
+    archive(options = {}) {
+        if (this.steps.length !== 0) {
+            throw new GislConfigError('archive() must be the first operation on files([...]); applying per-file ops before a bundle ' +
+                'is not yet supported — call archive() directly on the files you want to bundle.', { reason: 'pre_archive_ops_unsupported' });
+        }
+        return new ArchivedRecipe(this.inputs, options, this.client);
+    }
+    /** The number of inputs in this fan-out (introspection / tests). */
+    get inputCount() {
+        return this.inputs.length;
+    }
+    /** The number of operations chained so far (introspection / tests). */
+    get stepCount() {
+        return this.steps.length;
+    }
+    /**
+     * Lower this fan-out to a single multi-job workflow-create payload against a
+     * list of resolved upload ids (one per input, in input order). Each input `i`
+     * becomes ONE job with `id = "file-{i}"`, its `source: upload(fileIds[i])`,
+     * and the SHARED lowered `operations[]`. Composes the single-file
+     * {@link Recipe.toWorkflowPayload} per file so per-file media-hints resolve
+     * independently and lowering logic is not duplicated.
+     *
+     * @internal Consumed by {@link run} (after uploading all inputs) and the
+     *   cross-language parity harness (with fixed ids). Not caller-facing.
+     */
+    toWorkflowPayload(fileIds, callbackUrl) {
+        const jobs = this.inputs.map((input, i) => {
+            const single = new Recipe(input, undefined, this.steps, this.presetDefaults, this.scopedPresetDefaults);
+            const oneJob = single.toWorkflowPayload(fileIds[i]).jobs[0];
+            // Key order (id, source, operations) matches the PHP `toWire()` so the
+            // JSON-string serialisation is byte-identical across languages.
+            return { id: `file-${i}`, source: oneJob.source, operations: oneJob.operations };
+        });
+        // When `callbackUrl` is given (the file-first `submit()` path) it is built
+        // INTO the payload (`callback_url`) — mirrors Recipe.toWorkflowPayload.
+        // `run()` passes no callbackUrl.
+        return callbackUrl === undefined ? { jobs } : { jobs, callback_url: callbackUrl };
+    }
+    /**
+     * Execute the fan-out end-to-end: upload EVERY input, create ONE workflow
+     * with one job per input, await a terminal state (SSE with poll fallback),
+     * then resolve the per-job downloads into a partitioned {@link RunResult}.
+     * `partially_failed` is a NORMAL terminal state here — its successful jobs
+     * land in `succeeded`, its failed jobs in `failed`.
+     *
+     * Requires a client bound at construction time — `gisl().files(...)` wires
+     * it; a directly-constructed `FilesRecipe` throws {@link GislConfigError}.
+     * Mirrors the single-file {@link Recipe.run}; see {@link submit} for the
+     * fire-and-forget arm.
+     */
+    async run(options = {}) {
+        const signal = options.signal;
+        const onProgress = options.onProgress;
+        if (this.client === undefined) {
+            throw new GislConfigError('FilesRecipe.run() requires a client; build the fan-out via gisl().files(...) rather than constructing FilesRecipe directly.', { reason: 'no_client' });
+        }
+        const deadline = Date.now() + _parseMaxWait(options.maxWait ?? 300_000);
+        // 1+2. Upload EVERY input + create ONE multi-job workflow. Shared with
+        // submit() (which passes a webhook → callback_url and no deadline).
+        const created = await this._uploadAllAndCreate(undefined, deadline, onProgress, signal);
+        // 3. Wait to terminal status — SSE first, poll on a genuine SSE error.
+        // `partially_failed` is a normal terminal state here (the helper treats it
+        // as terminal); only caller-aborted / deadline / API errors propagate.
+        let finalStatus;
+        try {
+            finalStatus = await _consumeSseToTerminal(this.client, {
+                workflowId: created.workflowId,
+                deadline,
+                signal,
+                onProgress,
+            });
+        }
+        catch (err) {
+            if (!(err instanceof SseEndedWithoutTerminal || err instanceof GislNetworkError)) {
+                throw err;
+            }
+            finalStatus = await _pollToTerminal(this.client, {
+                workflowId: created.workflowId,
+                deadline,
+                signal,
+                pollIntervalMs: options.pollIntervalMs,
+            });
+        }
+        // 4. Fetch downloads + project per-job into the partitioned RunResult.
+        if (Date.now() >= deadline) {
+            throw new GislTimeoutError(`Workflow ${created.workflowId} reached terminal status but maxWait elapsed before downloads could be fetched`);
+        }
+        const downloads = await this.client.getWorkflowDownloads(created.workflowId);
+        // TDqmkWpX: re-check AFTER the downloads fetch so a slow getWorkflowDownloads
+        // cannot return a success past the advertised maxWait deadline.
+        if (Date.now() >= deadline) {
+            throw new GislTimeoutError(`Workflow ${created.workflowId} downloads fetch completed after maxWait elapsed`);
+        }
+        // keyByRef maps each job ref ("file-{i}") to the partition key. Today the
+        // key is just the index string; the Map seam leaves room for the FF3b
+        // keyed-fan-out card to map refs to caller-supplied keys without changing
+        // the producer's signature.
+        const keyByRef = new Map(this.inputs.map((_, i) => [`file-${i}`, String(i)]));
+        const downloader = new LazyHttpDownloader();
+        return projectMultiJobToRunResult(created.workflowId, finalStatus, downloads.downloads, keyByRef, downloader);
+    }
+    /**
+     * Fire-and-forget the fan-out: upload every input, create ONE multi-job
+     * workflow (wiring `webhook` into `callback_url` when given), and return a
+     * client-bound {@link Handle}. Does NOT wait for terminal status — call
+     * `handle.wait()` / `handle.result()` later to collect the partitioned
+     * {@link RunResult}. The Handle detects the fan-out from the wire `file-{i}`
+     * job refs, so per-file `byKey()` works even after a `client.workflow(id)`
+     * reattach (the keys are the input indices `"0"`, `"1"`, …).
+     *
+     * Requires a client bound at construction time (same `no_client` guard as
+     * {@link run}). `webhook` is OPTIONAL. Fire-and-forget, so NO whole-run
+     * deadline (a multi-GB upload is bounded by the HTTP client's own timeout).
+     * Mirrors the single-file {@link Recipe.submit}.
+     *
+     * @param webhook Absolute callback URL the server POSTs lifecycle events to.
+     */
+    async submit(webhook) {
+        if (this.client === undefined) {
+            throw new GislConfigError('FilesRecipe.submit() requires a client; build the fan-out via gisl().files(...) rather than constructing FilesRecipe directly.', { reason: 'no_client' });
+        }
+        const created = await this._uploadAllAndCreate(webhook, undefined);
+        return new Handle(created.workflowId, created.webhookSecret != null ? created.webhookSecret : undefined, this.client, null);
+    }
+    /**
+     * Upload every input (verbatim for a pre-uploaded id; uploading a path /
+     * blob otherwise, emitting `{phase:'upload'}` progress) then create ONE
+     * multi-job workflow (one job per input, `callback_url` built in when
+     * `webhook` is given). Shared first half of {@link run} + {@link submit}.
+     *
+     * Uploads are sequential so progress events stay ordered and the abort
+     * signal is honoured promptly; a resource arm is impossible in TS (Blob).
+     * `run()` passes a whole-run deadline (a slow upload must not proceed to
+     * createWorkflow past maxWait); `submit()` passes `undefined`, so the
+     * deadline checks are skipped.
+     */
+    async _uploadAllAndCreate(webhook, deadline, onProgress, signal) {
+        const fileIds = [];
+        for (const input of this.inputs) {
+            // Fail fast between uploads — a deadline that elapses mid-batch should
+            // not force every remaining input to upload before throwing.
+            _checkAborted(signal);
+            if (deadline !== undefined && Date.now() >= deadline) {
+                throw new GislTimeoutError('maxWait elapsed during fan-out uploads before all inputs were uploaded');
+            }
+            if (input.kind === 'uploadId') {
+                fileIds.push(input.fileId);
+            }
+            else {
+                const source = input.kind === 'path' ? input.path : input.blob;
+                const up = await this.client.uploadFile(source, {
+                    signal,
+                    ...(onProgress !== undefined
+                        ? {
+                            onProgress: (uploadedBytes, totalBytes) => {
+                                onProgress({ phase: 'upload', uploadedBytes, totalBytes });
+                            },
+                        }
+                        : {}),
+                });
+                fileIds.push(up.fileId);
+            }
+        }
+        _checkAborted(signal);
+        if (deadline !== undefined && Date.now() >= deadline) {
+            throw new GislTimeoutError('Uploads completed but maxWait elapsed before workflow could be created');
+        }
+        const created = await this.client.createWorkflow(this.toWorkflowPayload(fileIds, webhook));
+        _checkAborted(signal);
+        return created;
+    }
+    /**
+     * The shared single-file {@link Recipe} that captures the op chain (input is
+     * a placeholder — only the steps are read). Reuses Recipe's op-chain
+     * validation + coercion so a `FilesRecipe.compress(bad)` throws the identical
+     * `GislConfigError` as `Recipe.compress(bad)`.
+     */
+    baseRecipe() {
+        // The placeholder input never reaches the wire (only `steps` are read off
+        // the returned Recipe). A path placeholder gives compress() a media hint so
+        // optimize validation matches the single-file path; per-file lowering in
+        // toWorkflowPayload() rebuilds a Recipe with the REAL input.
+        return new Recipe(this.inputs[0] ?? fileInput.path('placeholder'), undefined, this.steps, this.presetDefaults, this.scopedPresetDefaults);
+    }
+    withStep(recipeWithStep) {
+        return new FilesRecipe(this.inputs, recipeWithStep.recipeSteps, this.presetDefaults, this.scopedPresetDefaults, this.client);
+    }
+}
+/**
+ * The single-output recipe you're in AFTER a fluent `files([...]).merge(...)`
+ * (FF3b). Merge collapses the N inputs into ONE output, so the per-file ops
+ * ({@link FilesRecipe.compress} etc.) no longer apply — instead this exposes the
+ * SAME chain ops as the single-file {@link Recipe}, applied to the merged
+ * result. `files([...]).merge().compress()` is the flagship case (example 14).
+ *
+ * **Lowering (one workflow):** each input is uploaded once and wrapped in its
+ * own single-input `passthrough` source job (`src_N`); the `merge` job consumes
+ * those via `job_output` inputs (array order = play order) and carries the merge
+ * op FIRST in its `operations[]`, followed by any post-combine ops (compress /
+ * convert / thumbnail) so they run on the merged output in the same job. The
+ * merge-level wire options reuse {@link wireMergeOptions} so a fluent merge
+ * lowers identically to the operation-first `client.merge()`.
+ *
+ * Immutable / clone-on-write like {@link Recipe} / {@link FilesRecipe}. Mirrors
+ * the PHP `MergedRecipe` in `packages/php/src/FileFirst/MergedRecipe.php`.
+ */
+export class MergedRecipe {
+    inputs;
+    mergeOptions;
+    postSteps;
+    presetDefaults;
+    scopedPresetDefaults;
+    client;
+    constructor(inputs, mergeOptions, postSteps = [], presetDefaults, scopedPresetDefaults, client) {
+        this.inputs = inputs;
+        this.mergeOptions = mergeOptions;
+        this.postSteps = postSteps;
+        this.presetDefaults = presetDefaults;
+        this.scopedPresetDefaults = scopedPresetDefaults;
+        this.client = client;
+    }
+    /** Reduce the merged output's size. See {@link Recipe.compress}. */
+    compress(optimize) {
+        if (optimize !== undefined && !Object.values(OptimizeFor).includes(optimize)) {
+            const allowed = Object.values(OptimizeFor).join(', ');
+            throw new GislConfigError(`compress 'optimize' must be one of ${allowed}; got '${String(optimize)}'.`, { reason: 'invalid_optimize', conflictingFields: ['optimize'] });
+        }
+        return this.withStep({ opType: 'compress', options: optimize === undefined ? {} : { optimize } });
+    }
+    /** Change the merged output's format. See {@link Recipe.convert}. */
+    convert(format) {
+        return this.withStep({ opType: 'convert', options: { format } });
+    }
+    /** Thumbnail the merged output. Omitted dimensions are dropped from the wire options. */
+    thumbnail(options = {}) {
+        const wire = {};
+        if (options.width !== undefined)
+            wire.width = options.width;
+        if (options.height !== undefined)
+            wire.height = options.height;
+        return this.withStep({ opType: 'thumbnail', options: wire });
+    }
+    /**
+     * Lower to the merge DAG: one `passthrough` source job per input + one
+     * `merge` job whose `operations[]` is `[merge, ...post-combine ops]`. The
+     * merge job's `inputs[]` consume the source jobs via `job_output` in input
+     * (play) order.
+     *
+     * @internal Consumed by {@link run} (after uploading all inputs), {@link submit}
+     *   (with a webhook), and the cross-language parity harness (with fixed ids).
+     */
+    toWorkflowPayload(fileIds, callbackUrl) {
+        const mediaKind = this.inferMediaKind();
+        const sourceJobs = [];
+        const inputs = [];
+        fileIds.forEach((fileId, i) => {
+            const srcId = `src_${i}`;
+            // Key order (id, source, operations) matches the PHP `toWire()` so the
+            // JSON-string serialisation is byte-identical across languages.
+            sourceJobs.push({ id: srcId, source: uploadSource(fileId), operations: [{ type: 'passthrough' }] });
+            inputs.push({ source: jobOutputSource(srcId) });
+        });
+        const operations = [
+            { type: 'merge', options: wireMergeOptions(this.mergeOptions, mediaKind) },
+            ...this.lowerPostSteps(mediaKind),
+        ];
+        const mergeJob = { id: 'merge', inputs, operations };
+        const jobs = [...sourceJobs, mergeJob];
+        return callbackUrl === undefined ? { jobs } : { jobs, callback_url: callbackUrl };
+    }
+    /** The number of inputs being combined (introspection / tests). */
+    get inputCount() {
+        return this.inputs.length;
+    }
+    /** The number of post-combine ops chained so far (introspection / tests). */
+    get stepCount() {
+        return this.postSteps.length;
+    }
+    /**
+     * Execute end-to-end: upload every input, create the merge workflow, await a
+     * terminal state (SSE with poll fallback), then resolve ONLY the merged output
+     * into a {@link RunResult}. Throws {@link GislTimeoutError} on `maxWait`.
+     *
+     * Requires a client bound at construction time — `gisl().files(...).merge(...)`
+     * wires it; a directly-constructed `MergedRecipe` throws {@link GislConfigError}.
+     * Mirrors the single-file {@link Recipe.run}.
+     */
+    async run(options = {}) {
+        const signal = options.signal;
+        const onProgress = options.onProgress;
+        if (this.client === undefined) {
+            throw new GislConfigError('MergedRecipe.run() requires a client; build the merge via gisl().files(...).merge(...) rather than constructing MergedRecipe directly.', { reason: 'no_client' });
+        }
+        const deadline = Date.now() + _parseMaxWait(options.maxWait ?? 300_000);
+        const created = await this._uploadAllAndCreate(undefined, deadline, onProgress, signal);
+        let finalStatus;
+        try {
+            finalStatus = await _consumeSseToTerminal(this.client, {
+                workflowId: created.workflowId,
+                deadline,
+                signal,
+                onProgress,
+            });
+        }
+        catch (err) {
+            if (!(err instanceof SseEndedWithoutTerminal || err instanceof GislNetworkError)) {
+                throw err;
+            }
+            finalStatus = await _pollToTerminal(this.client, {
+                workflowId: created.workflowId,
+                deadline,
+                signal,
+                pollIntervalMs: options.pollIntervalMs,
+            });
+        }
+        if (Date.now() >= deadline) {
+            throw new GislTimeoutError(`Workflow ${created.workflowId} reached terminal status but maxWait elapsed before downloads could be fetched`);
+        }
+        const downloads = await this.client.getWorkflowDownloads(created.workflowId);
+        // TDqmkWpX: re-check AFTER the downloads fetch so a slow getWorkflowDownloads
+        // cannot return a success past the advertised maxWait deadline.
+        if (Date.now() >= deadline) {
+            throw new GislTimeoutError(`Workflow ${created.workflowId} downloads fetch completed after maxWait elapsed`);
+        }
+        // Project ONLY the merge job's output — the `src_*` passthrough jobs
+        // re-expose the raw uploads, which are plumbing, not the deliverable
+        // (mirrors the operation-first merge.ts `ref === 'merge'` filter + PHP).
+        const mergeDownloads = downloads.downloads.filter((d) => d.ref === 'merge');
+        const downloader = new LazyHttpDownloader();
+        return projectDownloadsToRunResult(created.workflowId, finalStatus, mergeDownloads, null, downloader);
+    }
+    /**
+     * Fire-and-forget: upload + create the merge workflow (wiring `webhook` into
+     * `callback_url` when given), return a client-bound {@link Handle}. Does NOT
+     * wait for terminal status. Mirrors {@link Recipe.submit}.
+     */
+    async submit(webhook) {
+        if (this.client === undefined) {
+            throw new GislConfigError('MergedRecipe.submit() requires a client; build the merge via gisl().files(...).merge(...) rather than constructing MergedRecipe directly.', { reason: 'no_client' });
+        }
+        const created = await this._uploadAllAndCreate(webhook, undefined);
+        return new Handle(created.workflowId, created.webhookSecret != null ? created.webhookSecret : undefined, this.client, null);
+    }
+    // ---------------------------------------------------------------------------
+    /**
+     * Upload every input (verbatim for a pre-uploaded id; uploading a path / blob
+     * otherwise, emitting `{phase:'upload'}` progress) then create ONE merge
+     * workflow. Rejects fewer than 2 inputs BEFORE any upload fires. Shared first
+     * half of {@link run} + {@link submit}.
+     */
+    async _uploadAllAndCreate(webhook, deadline, onProgress, signal) {
+        this.validatePreUpload();
+        const fileIds = [];
+        for (const input of this.inputs) {
+            _checkAborted(signal);
+            if (deadline !== undefined && Date.now() >= deadline) {
+                throw new GislTimeoutError('maxWait elapsed during merge uploads before all inputs were uploaded');
+            }
+            if (input.kind === 'uploadId') {
+                fileIds.push(input.fileId);
+            }
+            else {
+                const source = input.kind === 'path' ? input.path : input.blob;
+                const up = await this.client.uploadFile(source, {
+                    signal,
+                    ...(onProgress !== undefined
+                        ? {
+                            onProgress: (uploadedBytes, totalBytes) => {
+                                onProgress({ phase: 'upload', uploadedBytes, totalBytes });
+                            },
+                        }
+                        : {}),
+                });
+                fileIds.push(up.fileId);
+            }
+        }
+        _checkAborted(signal);
+        if (deadline !== undefined && Date.now() >= deadline) {
+            throw new GislTimeoutError('Uploads completed but maxWait elapsed before the merge workflow could be created');
+        }
+        const created = await this.client.createWorkflow(this.toWorkflowPayload(fileIds, webhook));
+        _checkAborted(signal);
+        return created;
+    }
+    /**
+     * Reject an invalid combine BEFORE any upload fires — mirrors the operation-
+     * first `MergeBuilder.planSequence()` bounds so a typo'd merge costs no
+     * bandwidth: 2–10 inputs (merge schema `min/max_inputs`), and an image merge
+     * must carry an explicit `output_type` (the server rejects image merges
+     * without one). Shared by {@link run} + {@link submit} via
+     * {@link _uploadAllAndCreate}.
+     */
+    validatePreUpload() {
+        if (this.inputs.length < 2) {
+            throw new GislConfigError(`merge requires at least 2 inputs to combine (got ${this.inputs.length}).`, { reason: 'too_few_inputs' });
+        }
+        if (this.inputs.length > 10) {
+            throw new GislConfigError(`merge accepts at most 10 inputs (got ${this.inputs.length}). Split the merge or reduce the input list.`, { reason: 'too_many_inputs' });
+        }
+        if (this.inferMediaKind() === 'image' &&
+            this.mergeOptions.output === undefined &&
+            this.mergeOptions.outputType === undefined) {
+            throw new GislConfigError('image merges require an explicit output_type — set MergeOptions output: "video" | "gif" (or outputType). ' +
+                'The server rejects image merge requests with no output_type.', { reason: 'image_merge_requires_output_type' });
+        }
+    }
+    /**
+     * Lower the post-combine chain by composing a single-file {@link Recipe} over a
+     * synthetic input whose extension matches the merged OUTPUT media — so
+     * `compress(optimize)` resolves the correct preset for the merged result (it
+     * needs a media hint, which a merge output carries no filename for). Reuses
+     * Recipe's `lowerStep` rather than duplicating it.
+     */
+    lowerPostSteps(mediaKind) {
+        if (this.postSteps.length === 0) {
+            return [];
+        }
+        const synthetic = fileInput.path(`merged.${this.outputExtensionFor(mediaKind)}`);
+        const recipe = new Recipe(synthetic, undefined, this.postSteps, this.presetDefaults, this.scopedPresetDefaults);
+        return recipe.toWorkflowPayload('merged').jobs[0].operations;
+    }
+    /**
+     * The merged-output media. Honours an explicit {@link MergeOptions.mediaKind};
+     * otherwise infers from the first PATH input's extension (mirrors
+     * {@link MergeBuilder}); defaults to video.
+     */
+    inferMediaKind() {
+        if (this.mergeOptions.mediaKind !== undefined) {
+            return this.mergeOptions.mediaKind;
+        }
+        // Sniff the first input carrying a media signal — a path extension or a
+        // Blob MIME type (mirrors the operation-first MergeBuilder.inferMediaKind,
+        // codex c2). Pre-uploaded ids carry no signal, so they are skipped.
+        for (const input of this.inputs) {
+            if (input.kind === 'path') {
+                const lower = input.path.toLowerCase();
+                if (/\.(jpe?g|png|webp|avif|gif|heic|tiff?)$/.test(lower))
+                    return 'image';
+                if (/\.(mp3|wav|flac|aac|ogg|m4a)$/.test(lower))
+                    return 'audio';
+                return 'video';
+            }
+            if (input.kind === 'blob') {
+                if (input.blob.type.startsWith('image/'))
+                    return 'image';
+                if (input.blob.type.startsWith('audio/'))
+                    return 'audio';
+                return 'video';
+            }
+        }
+        return 'video';
+    }
+    outputExtensionFor(mediaKind) {
+        // An image merge produces a video/gif output (output_type), so the
+        // post-combine media follows the output type when set.
+        const output = this.mergeOptions.output ?? this.mergeOptions.outputType;
+        if (mediaKind === 'image' && typeof output === 'string') {
+            return output === 'gif' ? 'gif' : 'mp4';
+        }
+        switch (mediaKind) {
+            case 'audio':
+                return 'mp3';
+            case 'image':
+                return 'png';
+            default:
+                return 'mp4';
+        }
+    }
+    withStep(step) {
+        return new MergedRecipe(this.inputs, this.mergeOptions, [...this.postSteps, step], this.presetDefaults, this.scopedPresetDefaults, this.client);
+    }
+}
+/**
+ * The single-output recipe you're in AFTER a fluent `files([...]).archive(...)`
+ * (FF3b). Archive bundles the N inputs into ONE downloadable archive (zip /
+ * tar.gz) — media-agnostic, inputs may mix types. Unlike {@link MergedRecipe},
+ * archive is TERMINAL: a zip is the final artefact, so there is no post-bundle
+ * chain — this exposes only `run()` / `submit()`.
+ *
+ * **Lowering (one workflow):** each input is uploaded once and wrapped in its
+ * own single-input `passthrough` source job (`src_N`); the `archive` job
+ * consumes those via `job_output` inputs (array order = entry order) and carries
+ * the single `archive` op. The archive job's id is `archive`, so {@link RunResult}
+ * projects ONLY its output. Mirrors the PHP `ArchivedRecipe`.
+ */
+export class ArchivedRecipe {
+    inputs;
+    options;
+    client;
+    constructor(inputs, options = {}, client) {
+        this.inputs = inputs;
+        this.options = options;
+        this.client = client;
+    }
+    /** The number of inputs being bundled (introspection / tests). */
+    get inputCount() {
+        return this.inputs.length;
+    }
+    /**
+     * Lower to the archive DAG: one `passthrough` source job per input + one
+     * `archive` job consuming them via `job_output`.
+     *
+     * @internal Consumed by {@link run} / {@link submit} (after uploading) and the
+     *   cross-language parity harness (with fixed ids).
+     */
+    toWorkflowPayload(fileIds, callbackUrl) {
+        const sourceJobs = [];
+        const inputs = [];
+        fileIds.forEach((fileId, i) => {
+            const srcId = `src_${i}`;
+            sourceJobs.push({ id: srcId, source: uploadSource(fileId), operations: [{ type: 'passthrough' }] });
+            inputs.push({ source: jobOutputSource(srcId) });
+        });
+        const archiveJob = {
+            id: 'archive',
+            inputs,
+            operations: [{ type: 'archive', options: this.wireArchiveOptions() }],
+        };
+        const jobs = [...sourceJobs, archiveJob];
+        return callbackUrl === undefined ? { jobs } : { jobs, callback_url: callbackUrl };
+    }
+    /**
+     * Execute end-to-end: upload every input, create the archive workflow, await a
+     * terminal state (SSE with poll fallback), then resolve ONLY the archive output
+     * into a {@link RunResult}. Throws {@link GislTimeoutError} on `maxWait`.
+     * Requires a client bound via `gisl().files(...).archive(...)`.
+     */
+    async run(options = {}) {
+        const signal = options.signal;
+        const onProgress = options.onProgress;
+        if (this.client === undefined) {
+            throw new GislConfigError('ArchivedRecipe.run() requires a client; build the bundle via gisl().files(...).archive(...) rather than constructing ArchivedRecipe directly.', { reason: 'no_client' });
+        }
+        const deadline = Date.now() + _parseMaxWait(options.maxWait ?? 300_000);
+        const created = await this._uploadAllAndCreate(undefined, deadline, onProgress, signal);
+        let finalStatus;
+        try {
+            finalStatus = await _consumeSseToTerminal(this.client, {
+                workflowId: created.workflowId,
+                deadline,
+                signal,
+                onProgress,
+            });
+        }
+        catch (err) {
+            if (!(err instanceof SseEndedWithoutTerminal || err instanceof GislNetworkError)) {
+                throw err;
+            }
+            finalStatus = await _pollToTerminal(this.client, {
+                workflowId: created.workflowId,
+                deadline,
+                signal,
+                pollIntervalMs: options.pollIntervalMs,
+            });
+        }
+        if (Date.now() >= deadline) {
+            throw new GislTimeoutError(`Workflow ${created.workflowId} reached terminal status but maxWait elapsed before downloads could be fetched`);
+        }
+        const downloads = await this.client.getWorkflowDownloads(created.workflowId);
+        // TDqmkWpX: re-check AFTER the downloads fetch so a slow getWorkflowDownloads
+        // cannot return a success past the advertised maxWait deadline.
+        if (Date.now() >= deadline) {
+            throw new GislTimeoutError(`Workflow ${created.workflowId} downloads fetch completed after maxWait elapsed`);
+        }
+        // Project ONLY the archive job's output — the `src_*` passthrough jobs
+        // re-expose the raw uploads, which are plumbing, not the deliverable.
+        const archiveDownloads = downloads.downloads.filter((d) => d.ref === 'archive');
+        const downloader = new LazyHttpDownloader();
+        return projectDownloadsToRunResult(created.workflowId, finalStatus, archiveDownloads, null, downloader);
+    }
+    /**
+     * Fire-and-forget: upload + create the archive workflow (wiring `webhook` into
+     * `callback_url` when given), return a client-bound {@link Handle}. Mirrors
+     * {@link MergedRecipe.submit}.
+     */
+    async submit(webhook) {
+        if (this.client === undefined) {
+            throw new GislConfigError('ArchivedRecipe.submit() requires a client; build the bundle via gisl().files(...).archive(...) rather than constructing ArchivedRecipe directly.', { reason: 'no_client' });
+        }
+        const created = await this._uploadAllAndCreate(webhook, undefined);
+        return new Handle(created.workflowId, created.webhookSecret != null ? created.webhookSecret : undefined, this.client, null);
+    }
+    // ---------------------------------------------------------------------------
+    async _uploadAllAndCreate(webhook, deadline, onProgress, signal) {
+        this.validatePreUpload();
+        const fileIds = [];
+        for (const input of this.inputs) {
+            _checkAborted(signal);
+            if (deadline !== undefined && Date.now() >= deadline) {
+                throw new GislTimeoutError('maxWait elapsed during archive uploads before all inputs were uploaded');
+            }
+            if (input.kind === 'uploadId') {
+                fileIds.push(input.fileId);
+            }
+            else {
+                const source = input.kind === 'path' ? input.path : input.blob;
+                const up = await this.client.uploadFile(source, {
+                    signal,
+                    ...(onProgress !== undefined
+                        ? {
+                            onProgress: (uploadedBytes, totalBytes) => {
+                                onProgress({ phase: 'upload', uploadedBytes, totalBytes });
+                            },
+                        }
+                        : {}),
+                });
+                fileIds.push(up.fileId);
+            }
+        }
+        _checkAborted(signal);
+        if (deadline !== undefined && Date.now() >= deadline) {
+            throw new GislTimeoutError('Uploads completed but maxWait elapsed before the archive workflow could be created');
+        }
+        const created = await this.client.createWorkflow(this.toWorkflowPayload(fileIds, webhook));
+        _checkAborted(signal);
+        return created;
+    }
+    /**
+     * Reject an invalid bundle BEFORE any upload fires — the archive schema allows
+     * 2–50 inputs (`min/max_inputs`), so a typo'd bundle costs no bandwidth.
+     */
+    validatePreUpload() {
+        if (this.inputs.length < 2) {
+            throw new GislConfigError(`archive requires at least 2 inputs to bundle (got ${this.inputs.length}).`, { reason: 'too_few_inputs' });
+        }
+        if (this.inputs.length > 50) {
+            throw new GislConfigError(`archive accepts at most 50 inputs (got ${this.inputs.length}). Split the bundle or reduce the input list.`, { reason: 'too_many_inputs' });
+        }
+    }
+    /**
+     * Project the archive options into the wire shape. Both fields are optional
+     * (the server defaults `format` to zip and `folder_structure` to flat), so an
+     * omitted option is dropped rather than sent.
+     */
+    wireArchiveOptions() {
+        const out = {};
+        if (this.options.format !== undefined)
+            out.format = this.options.format;
+        if (this.options.folderStructure !== undefined)
+            out.folder_structure = this.options.folderStructure;
+        return out;
+    }
+}