sw image "a lighthouse keeper with a teal coat, flat textured 2D" --aspect 16:9
sw video "he turns towards the sea" --start keeper.jpg --duration 4 --resolution 360p
sw music "moody indie folk, warm male vocal" --lyrics-file chorus.txt --length 1:00

Contents

• Showcase
• Why GnmyFlow
• Features
• Install
• Quick start
• Use it from AI agents
• Command reference
• Batch files
• Costs and safety
• How it works
• Troubleshooting
• Configuration
• Development
• Disclaimer

Showcase

The showcase below was inherited from StudioWright and made with StudioWright, from one small batch file and two chat messages.

{"id": "keeper", "kind": "image", "prompt": "a lighthouse keeper with a teal coat, standing, flat textured 2D illustration", "refs": ["style.png"]}
{"id": "keeper-close", "kind": "image", "prompt": "close-up portrait of the same lighthouse keeper, flat textured 2D", "refs": ["keeper"]}
{"id": "keeper-clip", "kind": "video", "prompt": "the lighthouse keeper turns his head towards the sea, gentle wind", "start": "keeper", "resolution": "360p", "duration": 4}

keeper · Nano Banana 2.1 · style from a reference image · 22 s	keeper-close · the same character, by referring to the first job · 21 s
keeper-clip · Omni Flash, 4 s, start frame from the first job · 32 s · 4 credits	Beacon in the Dark · written and sung by the Flow Music Producer · 37 s · 5 credits ▶ listen (MP4) · MP3

The song took two chat turns:

sw music chat "Write a four-line chorus about a lighthouse keeper who waits for the ships."
sw music chat "Moody indie-folk please. Make the song now, about one minute long." --session <id from the first reply>

Why GnmyFlow

Most ways to automate Google's media models need one of these: a paid API key, cookies copied out of your browser, a captcha-solving service, an unpacked browser extension, or a third-party relay that holds your session. GnmyFlow needs none of them.

What makes it different:

• It acts like you. GnmyFlow types, clicks and chooses files in Flow's real pages, in a Chrome you signed in to yourself. It only reads the replies those pages receive, to learn ids and download links. It never forges, replays or alters a request.
• One package, three front doors. The same engine is a CLI (sw), an MCP server (sw mcp) and a Claude Code plugin.
• Built for real production. Price checks before you spend (--dry-run, --max-credits), resumable batch files that chain results, sized downloads (2K images, 1080p/4K clips), pacing across processes, and a cooldown when Google asks for a break.
• Fast on a warm tab. About 20 s for an image, 30-60 s for a clip, 40 s for a song.

Features

Images (Google Flow)

• Nano Banana 2.1 (default), Nano Banana Pro and Nano Banana 2 Lite.
• Aspect ratios 16:9, 4:3, 1:1, 3:4, 9:16; one to four images per call.
• Reference images from files or earlier results; each file is uploaded once per project and then reused by content.
• Downloads at Flow's own sizes: 1K (original), 2K (upscaled by Flow), 4K where the plan allows.

Video (Google Flow)

• Omni Flash (default), Veo 3.1 Lite, Fast and Quality.
• Text only, start frame, start and end frames, or ingredients (several reference images).
• Omni: 360p / 720p and 4, 6, 8 or 10 s. Aspect 16:9 or 9:16.
• Downloads at 720p, 1080p or 4K (upscaled by Flow), or as a GIF.
• --dry-run shows Flow's price without submitting; --max-credits refuses anything more expensive.

Music (Google Flow Music)

• Songs with your lyrics (any language) or instrumentals; Lyria 3.5 or Lyria 3 Pro; BPM, length (1:00-3:00), seed, title, and cover art from a file or a prompt.
• Edit existing songs: extend, cover in another style, replace a section, variation, trim.
• Stems: drums, bass, vocals and other, each as its own file.
• Downloads: WAV (48 kHz), M4A or MP3; vertical MP4 videos from Flow Music's templates.
• The Producer agent: chat with it to write lyrics, plan songs, remix your own audio or make AI music videos; songs it makes are saved automatically.
• Your own audio: upload a file and extend, cover or split it like any other song.

Automation

• CLI with --json output, an MCP server with 16 tools, and a Claude Code plugin with a skill.
• JSON Lines batch files that resume where they stopped and pass results between jobs.
• A job history (sw jobs) and a screenshot of the page whenever a job fails.

Install

This fork is not published on PyPI yet. Install from the repository only after the rename changes have been pushed. The plugin pin v0.3.1 targets the first fork release and will only work after that tag is published. The existing v0.3.0 tag predates the rename and must not be moved or reused for the fork release.

Requirements: Python 3.12+, uv, Google Chrome, and a Google account with a Flow plan (images and video) and/or a Flow Music plan (music).

uv tool install git+https://github.com/anggi-susanto/gnmyflow
sw --version

The primary command is gnmyflow. Compatibility aliases sw and studiowright run the same CLI. Do not install both forks into the same environment: they share the studiowright Python namespace and legacy commands.

git clone https://github.com/anggi-susanto/gnmyflow.git
cd gnmyflow
uv sync
uv tool install --python 3.13 -e .

Updating

• Plugin (MCP server and skill): claude plugin update gnmyflow@gnmyflow, or in Claude Code /plugin → Installed → gnmyflow → Update now; then /reload-plugins. Each plugin version runs the GnmyFlow release of the same version. To update automatically, turn it on once: /plugin → Marketplaces → gnmyflow → Enable auto-update.
• Command line: uv tool install --force git+https://github.com/anggi-susanto/gnmyflow@vX.Y.Z with the release you want (see Releases).

Quick start

1. Sign in once. Chrome opens on a profile that belongs to GnmyFlow only. Sign in to Google, open Flow once, then close the window.

sw auth login
sw auth login --url https://www.flowmusic.app/    # only if you will make music

Run sw stop first if GnmyFlow's Chrome is already running.

2. Make something.

sw image "a red fox in the snow, flat 2D" --aspect 1:1
sw video "slow push-in on a quiet harbour at dawn" --model veo-lite --dry-run   # price only
sw music "calm documentary underscore, soft piano, warm strings" --bpm 80 --length 1:30

Files land in Downloads/studiowright (change it with --out), and each command prints their paths. Add --json for ids, sizes, timings and credits.

3. Keep Chrome in the background. sw window hide moves the window off-screen. Never minimize it: a minimized Chrome stops rendering and Flow stalls.

Use it from AI agents

Every client below runs gnmyflow mcp over stdio; sw mcp remains compatible. If a client cannot find gnmyflow, give it the full executable path.

Generating takes real time (an image ~20 s, a clip up to a few minutes), so give the client a generous tool timeout where it has one.

Claude Code plugin (recommended)

The plugin adds the MCP server and a skill that teaches Claude the options, prices and safe habits.

claude plugin install gnmyflow --marketplace anggi-susanto/gnmyflow

or inside Claude Code:

/plugin marketplace add anggi-susanto/gnmyflow
/plugin install gnmyflow@gnmyflow

The plugin starts its own matching GnmyFlow release with uvx, so it needs only uv and Chrome; installing sw separately is for the command line.

Then just ask:

Make a 16:9 image of a lighthouse keeper, then a 4-second clip where he turns to the sea. Check the price first.

Write a one-minute instrumental bed for a documentary intro and give me the stems.

Claude Code (MCP server only)

claude mcp add --scope user gnmyflow -- gnmyflow mcp

For long video jobs, raise the tool timeout, e.g. MCP_TOOL_TIMEOUT=900000 in your environment.

{
  "mcpServers": {
    "gnmyflow": { "command": "gnmyflow", "args": ["mcp"] }
  }
}

{
  "mcpServers": {
    "gnmyflow": { "command": "gnmyflow", "args": ["mcp"] }
  }
}

{
  "servers": {
    "gnmyflow": { "type": "stdio", "command": "gnmyflow", "args": ["mcp"] }
  }
}

codex mcp add gnmyflow -- gnmyflow mcp

[mcp_servers.gnmyflow]
command = "gnmyflow"
args = ["mcp"]
tool_timeout_sec = 900

gemini mcp add -s user gnmyflow gnmyflow mcp

{
  "mcpServers": {
    "gnmyflow": { "command": "gnmyflow", "args": ["mcp"], "timeout": 900000 }
  }
}

MCP tools

Command reference

sw --help and sw <command> --help show every option. The most used ones:

Images and video

sw image "PROMPT" [--model nano-banana-2.1|nano-banana-pro|nano-banana-2-lite]
                  [--aspect 16:9|4:3|1:1|3:4|9:16] [--count 1-4] [--ref FILE_OR_ID ...]
                  [--size 1K|2K|4K] [--project ID] [--out DIR] [--json]

sw video "PROMPT" [--model omni-flash|veo-lite|veo-fast|veo-quality] [--mode frames|ingredients]
                  [--start FILE_OR_ID] [--end FILE_OR_ID] [--ref FILE_OR_ID ...]
                  [--aspect 16:9|9:16] [--resolution 360p|720p] [--duration 4|6|8|10]
                  [--size 720p|1080p|4K|270p] [--max-credits N] [--dry-run] [--json]

sw upload FILE ...                 # upload references once, print their media ids
sw download MEDIA_ID ... [--size 2K|1080p]
sw project create "TITLE" | use ID | list
sw run BATCH.jsonl [--keep-going]
sw credits                         # Flow (as of the last clip) and Flow Music balances
sw jobs | sw pace | sw window show|hide | sw stop | sw mcp

Music

SONG is a song id, a unique id prefix, a flowmusic.app song URL, or last.

sw music "SOUND" [--lyrics TEXT | --lyrics-file FILE] [--bpm N] [--length 1:00-3:00] [--seed N]
                 [--model lyria-3.5|lyria-3-pro] [--title T] [--cover IMAGE | --cover-prompt TEXT]
                 [--format wav|m4a|mp3] [--out DIR] [--json]

sw music extend SONG [--from 0:45] [--until 2:30] [--prompt TEXT] [--lyrics-file FILE]
sw music cover SONG --prompt "slow jazz trio, brushed drums"
sw music replace SONG --start 0:20 --end 0:35 [--prompt TEXT] [--lyrics-file FILE]
sw music variation SONG
sw music trim SONG --start 0:05 --end 0:35
sw music stems SONG
sw music video SONG [--template simple|modern|player]
sw music download SONG ... [--format wav|m4a|mp3]
sw music chat "MESSAGE" [--session ID] [--audio FILE] [--image FILE] [--confirm-rights]
sw music upload FILE --confirm-rights
sw music list | sw music credits

sw music "SOUND" is short for sw music create "SOUND". Flow Music asks you to confirm that you have the rights to files you upload; --confirm-rights is that confirmation, and GnmyFlow never gives it on its own.

Batch files

A batch is JSON Lines (one job per line) or a JSON list. Jobs run in order, and a job can use an earlier job's result by its id.

{"id": "keeper", "kind": "image", "prompt": "a lighthouse keeper, flat 2D", "refs": ["style.png"]}
{"id": "clip", "kind": "video", "prompt": "he turns to the sea", "start": "keeper", "duration": 6, "max_credits": 12}
{"id": "bed", "kind": "music", "sound": "calm piano underscore", "length": "1:30"}
{"id": "bed-long", "kind": "music", "action": "extend", "song": "bed", "end": "2:30"}
{"id": "bed-stems", "kind": "music", "action": "stems", "song": "bed-long"}

sw run shots.jsonl

• refs, start and end take file paths (relative to the batch file), media ids or job ids. A music job's song takes a song id or a job id.
• Music actions: create (default), extend, cover, replace, variation, trim, stems, video. Extend, replace and trim take their times as start / end.
• Results are appended to shots.jsonl.results.jsonl. Running the batch again skips jobs that already finished; --keep-going continues past a failure.

Costs and safety

Prices as Flow and Flow Music showed them in October 2026. Your plan may differ.

Built-in guards:

• --dry-run / dry_run sets up a clip and reports Flow's price without submitting.
• --max-credits / max_credits refuses a clip that costs more.
• Submits are spaced (20 s plus jitter) per profile, across processes.
• When Google reports unusual activity, the profile cools down for 30 minutes and every call says so instead of retrying.
• One job at a time per profile (and one Flow Music job at a time), queued by a lock. Downloads run in a tab of their own beside generations, since they submit nothing.
• Upscaled downloads (2K/4K, 1080p) are spaced a minute apart. Flow refuses bulk upscaling with "unusual activity"; GnmyFlow then pauses upscales for an hour (generations go on) and saves new media at their original size meanwhile.

How it works

flowchart LR
    A["You, Claude or any MCP client"] -->|"sw … / MCP tools"| B["GnmyFlow"]
    B -->|"CDP on 127.0.0.1"| C["Your Chrome<br/>dedicated profile, signed in by you"]
    C -->|"types, clicks, chooses files"| D["Google Flow"]
    C -->|"types, clicks, chooses files"| E["Google Flow Music"]
    D -. "replies are read, never changed" .-> B
    E -. "replies are read, never changed" .-> B
    B --> F[("Downloads/studiowright")]

• GnmyFlow starts your installed Chrome on its own profile and attaches over the DevTools protocol on loopback. Chrome keeps running between commands, so every call finds a warm, signed-in tab.
• Generation uses page controls; selected project and download operations also use authenticated browser requests. Replies supply media ids, progress and download links.
• You sign in manually. Chrome retains authentication state in the dedicated profile. Loopback CDP has no application authentication: other local processes may control it.
• Job records contain prompts and local paths; failures can save page screenshots. Diagnostic scripts in spikes/ may capture session material. These files are sensitive; there is no automatic retention cleanup or application-level encryption. Do not share profiles, databases, screenshots or captures without reviewing them.
• No production telemetry destination was identified in the source audit. This is not a guarantee about Chrome, dependencies or third-party provider traffic.

Technical notes: Flow · Flow Music.

Troubleshooting

Configuration

The Python import namespace studiowright, STUDIOWRIGHT_* environment variables, and existing data/output directories are retained for compatibility. No profile or database migration is performed. The renamed plugin uses GNMYFLOW_SOURCE for its source override; update any previous STUDIOWRIGHT_SOURCE plugin override.

Several Google accounts

Each profile is one signed-in Google account with its own Chrome, queue and credits. Profiles run in parallel; jobs on the same profile take turns.

sw --profile account-1 auth login     # name it as you like; sign in, close the window
sw --profile account-1 image "…"      # or set STUDIOWRIGHT_PROFILE=account-1
sw auth list                          # each profile and the account it is signed in to
sw auth rename account-1 studio       # keeps the sign-in, projects and settings
sw auth logout studio                 # removes the profile from this computer

logout deletes only the profile's Chrome data and GnmyFlow's records of it. Downloaded files stay where they are, and the projects stay in your Google account. Both rename and logout wait until that profile's Chrome is closed (sw stop).

Development

uv sync
uv run pytest                        # unit tests, no browser needed
uv run ruff check src tests spikes

Manual release procedure (no CI)

GitHub Actions is currently disabled because the account has a billing lock, not because GitHub Free universally excludes Actions. The previous run stopped before any test step. Local validation passed on macOS with Python 3.12; the Linux/Windows and Python 3.13 matrix has not been verified. Live provider/browser behavior is a separate check.

Both registered workflows have been disabled on GitHub. Templates are retained under .github/disabled-workflows/*.yml.disabled, outside the active workflow directory. Do not restore them or enable publication without a separate review. Trusted Publishing OIDC is deferred; manual uploads use separate TestPyPI/PyPI API tokens.

The commands below assume macOS/Linux and a shell opened at the repository root. On Windows, use equivalent PowerShell commands and venv executables under Scripts/. Do not launch Chrome, run diagnostic spikes, sign in, or generate content for these checks. Installing dependencies accesses package indexes; the smoke test itself blocks socket connections and Chrome startup during the CLI checks. Package installation does not install Google Chrome or configure MCP/the Claude Code plugin. No Playwright browser binary installation is needed: this application attaches to separately installed Chrome.

1. Review and freeze the source

Obtain explicit approval for the release version and source commit. Commit/push only with permission; build from a clean reviewed checkout, not an uncommitted working tree. Check that package, runtime, plugin and plugin source pin all agree on 0.3.1:

git status --short
git rev-parse HEAD
uv sync --locked
uv run --locked pytest
uv run --locked ruff check src tests spikes scripts
git diff --check

Stop on any failure or unexpected working-tree change. Record the source commit, Python, OS and uv versions outside the package. Do not claim untested platforms are supported.

2. Build once, inspect and record hashes

Use a new temporary workspace so old artifacts are not mixed into the release. Keep this shell open: later commands reuse RELEASE_WORK and RELEASE_DIST.

RELEASE_WORK=$(mktemp -d "${TMPDIR:-/tmp}/gnmyflow-release.XXXXXX")
RELEASE_DIST="$RELEASE_WORK/dist"
uv build --out-dir "$RELEASE_DIST"
uv run --no-project python -c \
  'import runpy, sys; from pathlib import Path; runpy.run_path("scripts/audit_dist.py")["audit"](Path(sys.argv[1]))' \
  "$RELEASE_DIST"
uvx --from twine twine check --strict "$RELEASE_DIST"/*
(cd "$RELEASE_DIST" && shasum -a 256 gnmyflow-0.3.1* > "$RELEASE_WORK/SHA256SUMS")
cat "$RELEASE_WORK/SHA256SUMS"

The archive allowlist excludes profiles, session captures, incident screenshots, databases, .env, caches, diagnostic spikes and media showcase files. Inspect the source as well: an allowed Python file can still contain hard-coded secrets. Keep MIT copyright and upstream author attribution intact. Freeze these exact archives; do not rebuild between TestPyPI and PyPI. Any source/artifact change invalidates approval and requires revalidation.

3. Clean installation of both local archives

uv venv "$RELEASE_WORK/wheel" --python 3.12
uv pip install --python "$RELEASE_WORK/wheel/bin/python" "$RELEASE_DIST"/*.whl
"$RELEASE_WORK/wheel/bin/python" -I scripts/smoke_install.py
uv venv "$RELEASE_WORK/sdist" --python 3.12
uv pip install --python "$RELEASE_WORK/sdist/bin/python" "$RELEASE_DIST"/*.tar.gz
"$RELEASE_WORK/sdist/bin/python" -I scripts/smoke_install.py

Both must pass metadata/import checks and offline help/version/capabilities for gnmyflow, sw and studiowright. Dependency resolution here is fresh, rather than the development lockfile. Do not install alongside upstream in the same environment: the import namespace and legacy commands overlap. Additional tests on Linux/Windows and Python 3.13 require their own machines/environments; these commands do not prove them.

4. TestPyPI upload — explicit approval required

Create/manage the TestPyPI account and enable required account security (including 2FA) through the official website. PyPI is a separate account/index. Create a TestPyPI API token there; for the first upload a narrowly scoped project token may not yet be possible. If an account-wide token is needed, revoke it after initial project creation and use a project-scoped token thereafter. Never publish a placeholder to reserve the name.

Enter the token only into Twine's hidden terminal prompt (username __token__ if asked). Never put tokens in chat, commands/arguments, shell history, repository files or a plaintext .pypirc. Avoid terminal/session recording while entering secrets. The following command disables keyring lookup for that invocation without changing the user's global settings:

(cd "$RELEASE_DIST" && shasum -a 256 -c "$RELEASE_WORK/SHA256SUMS")
PYTHON_KEYRING_BACKEND=keyring.backends.null.Keyring \
  uvx --from twine twine upload \
  --repository-url https://test.pypi.org/legacy/ "$RELEASE_DIST"/*

Use an interactive terminal; do not use --non-interactive without configured credentials. If credential environment overrides are already set, clear them in that shell before upload so Twine prompts instead; never print their values. Do not use --skip-existing to hide errors. For a partial upload, inspect the index, compare existing file hashes, and upload only a missing file after review. Existing PyPI filenames cannot be overwritten.

5. Verify the TestPyPI package, not a mixed-index substitute

TestPyPI may lack dependencies. Install the exact package only from TestPyPI with --no-deps, then install dependencies separately from PyPI. Do not use a mixed --extra-index-url install as evidence that the target package came from TestPyPI.

uv venv "$RELEASE_WORK/testpypi" --python 3.12
uv pip install --python "$RELEASE_WORK/testpypi/bin/python" --no-deps \
  --index-url https://test.pypi.org/simple/ 'gnmyflow==0.3.1'
uv pip install --python "$RELEASE_WORK/testpypi/bin/python" \
  --index-url https://pypi.org/simple/ \
  'click>=8.5.0' 'filelock>=4.1.0' 'mcp[cli]>=2.3.0' 'pillow>=11.0' \
  'platformdirs>=4.12.4' 'playwright>=1.63.0'
"$RELEASE_WORK/testpypi/bin/python" -I scripts/smoke_install.py

Review https://test.pypi.org/pypi/gnmyflow/0.3.1/json: compare the uploaded filenames and digests.sha256 with the frozen manifest. Read metadata/README/license on the project page. TestPyPI success does not reserve the PyPI name or validate provider behavior.

6. PyPI upload — separate explicit approval required

Confirm the name is still usable, source commit is unchanged, and TestPyPI verification passed. Public HTTP errors are not evidence of name availability; a 404 is not a guarantee that PyPI permits registration. Use a real release, not a placeholder. Configure 2FA and an API token on PyPI, separate from TestPyPI; use project scope when available and revoke any bootstrap account-wide token afterward.

git status --short
git rev-parse HEAD
(cd "$RELEASE_DIST" && shasum -a 256 -c "$RELEASE_WORK/SHA256SUMS")
PYTHON_KEYRING_BACKEND=keyring.backends.null.Keyring \
  uvx --from twine twine upload \
  --repository-url https://upload.pypi.org/legacy/ "$RELEASE_DIST"/*

This uploads the same bytes already tested. Do not rebuild. Verify filenames and SHA-256 values at https://pypi.org/pypi/gnmyflow/0.3.1/json, then install from PyPI alone:

uv venv "$RELEASE_WORK/pypi" --python 3.12
uv pip install --python "$RELEASE_WORK/pypi/bin/python" \
  --index-url https://pypi.org/simple/ 'gnmyflow==0.3.1'
"$RELEASE_WORK/pypi/bin/python" -I scripts/smoke_install.py

7. Tag, GitHub release and updates

After verification and separate approval, create/push v0.3.1 at the exact reviewed source commit. Never move the existing v0.3.0 tag. Create the GitHub release with the source commit, artifact hashes, upstream attribution, compatibility details and platform/security limitations. The plugin's pinned source becomes usable only once the tag exists. Update the installation notice after publication; do not announce success prematurely.

Keep the archives and hash manifest as release evidence; remove only the disposable venvs when no longer needed. Users may install with uv tool install gnmyflow or pipx install gnmyflow, then update with uv tool upgrade gnmyflow or pipx upgrade gnmyflow. VCS installations use an explicitly selected release tag.

For a bad release, stop publication, yank the affected version when appropriate, and publish a new fixed version. PyPI releases cannot be overwritten; do not delete/reuse filenames or move tags. A yanked version can still be installed with an explicit pin, so publish a clear advisory. Reinstall a known-good version to roll back; preserve existing profiles/databases. Revoke compromised tokens immediately and rotate them outside chat.

References: Twine, TestPyPI, Trusted Publishing (future option).

• src/studiowright/flow/ drives Flow, src/studiowright/music/ drives Flow Music, and studio.py adds locking, pacing, job records and downloads on top.
• spikes/ holds the small scripts used to map each page. Most only open menus and read replies, at no cost; their docstrings say when one spends credits.
• Tested on Windows 11. The code also finds Chrome on macOS and Linux, but those platforms have not been tested yet.

Disclaimer

• GnmyFlow is not affiliated with, endorsed or sponsored by Google LLC. Google, Flow, Flow Music, Veo, Nano Banana and Lyria are trademarks of Google LLC.
• The MIT license covers this code only. It grants no rights to Google's services or to what they generate. Your use of Flow and Flow Music, and of their output, is governed by Google's terms and your plan.
• Automated use of a web app may conflict with its terms. Use your own account, at your own risk and at a human pace. GnmyFlow stops when Google asks it to.
• The web apps can change at any time and break GnmyFlow until it is updated.

License

MIT © 2026 Umut Koçak

Single-submit create contract (version 1)

gnmyflow capabilities --json is a local, read-only probe: it does not open Chrome, sign in, access an account, or generate anything. Its production_create_contract: 1 describes opt-in sw music create "SOUND" --single-submit behavior (not legacy defaults):

• At most one Generate trigger per invocation, including when no start is observed.
• No heuristic library recovery by title or creation time after an ambiguous failure. The original error is propagated; callers must not automatically retry paid creation.
• --cover-prompt is rejected because it can generate additional paid cover art. A local --cover image is allowed. Refilling a failed compose panel before Generate is still allowed.
• This is not an exactly-once server guarantee, cross-invocation idempotency, a credit reservation, or an assertion that an ambiguous submission spent no credits.

The JSON fields are production_create_contract (integer), single_submit (boolean), heuristic_recovery (boolean), create_flag (string), models and formats (arrays of accepted strings), and min_length / max_length (inclusive seconds). Models, formats, and length bounds come from request options. Existing create/edit/chat/batch/MCP callers retain legacy behavior unless a MusicRequest(single_submit=True) is explicitly used.