media-gen

name	media-gen
description	Generate or edit images, videos, or audio (text-to-speech) in the current task. Use whenever the user asks to create/generate/produce or edit/modify a picture / image / illustration / cover / poster / icon / artwork, a video / clip / animation, or speech / voiceover / narration / TTS — e.g. generate image, draw, design a cover, edit this image, change the background, text-to-video, generate speech; 画一张, 配图, 编辑图片, 改图, 换背景, 做个视频, 配音, 文字转语音. Also use when a document (slides, poster, README hero) needs an inline image.
disable-model-invocation	false
user-invocable	true
always-show	true

media-gen

Generate and edit images on demand by calling the local Clacky HTTP server, which dispatches to whichever image-generation model the user configured (type=image in their model settings). Editing (image-in → image-out) works with any image model that accepts image input — most current ones do.

Endpoint

POST http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/image
GET  http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/types

Step 1 — Verify a backend is configured

Before generating anything, confirm the user has a type=image model set up:

curl -s http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/types

If the response shows image.configured = false, stop and tell the user:

还没有配置生图模型。请打开设置页 → 添加模型 → 类型选 image（走 openclacky 官方网关时推荐 or-gemini-3-pro-image 或 or-gpt-image-2）。配好后再让我生图。

Do NOT try to fall back to terminal + a hand-written curl https://api.openai.com/... — that bypasses the user's configured backend and won't be billed correctly.

You do NOT configure models — the user does, in the settings page. Never edit the user's config.yml to add or change a model, and never invent a model name from memory (e.g. or-gpt-5.4-image-2 does not exist). The real, current model is whatever /api/media/types reports under image.model. If you think a different model is needed, tell the user which one to set in the settings page — don't touch the config file yourself.

Step 2 — Generate the image

The model does NOT honor exact pixel sizes

There is no size / width / height field — the only shape control is aspect_ratio (landscape / square / portrait), and even that is just a rough hint (ask for 576x96 and you may get 1408x768). When the user needs an exact pixel size, a grid, an icon at NxN, or a spritesheet, generate first at whatever size the model gives, then resize / crop / tile to the exact pixels with ImageMagick (magick). Verify with magick identify before reporting done.

Important: generation speed & concurrency

• Image generation can be slow — up to 2 minutes per image depending on the model. Before calling the API, warn the user that it may take a minute or two. The curl request blocks until the image is ready; do NOT run it in the background.
• One at a time only. Never generate multiple images concurrently (e.g. by running several curl commands simultaneously or in a script loop). Each call consumes significant server-side resources, and parallel requests will almost certainly cause timeouts. If the user wants several images, generate them sequentially, one after another.

curl -s -X POST http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/image \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A clean, modern hero illustration for a tech startup landing page. Soft gradient background, abstract geometric shapes in blue and purple, minimal style, 4K quality.",
    "aspect_ratio": "landscape",
    "output_dir": "'"$(pwd)"'",
    "session_id": "<%= session_id %>"
  }'

• The terminal blocks multi-line commands — write the request into a .sh file and run it, don't paste a multi-line curl.
• If a call fails with 400 / INVALID_ARGUMENT, drop the aspect_ratio field and retry once before reporting the error.
• If a call fails with unknown image model (400), the configured model name isn't recognized by its backend — tell the user to fix the model name in the settings page; do NOT guess another name and retry.

Request fields

Field	Required	Values	Notes
prompt	yes	string	Be detailed and concrete. See prompt tips below.
aspect_ratio	no	landscape / square / portrait	Defaults to landscape.
output_dir	yes	absolute path	Always pass $(pwd) so files land in the current session workspace. The image is saved under <output_dir>/assets/generated/.
session_id	yes	string	Current Clacky session ID. Always pass the rendered value shown in the request example.
image	no	file path / base64 / data URL	A single input image to edit. Triggers image-edit mode (see below).
images	no	array of the above	Multiple input images for a multi-image edit. Takes precedence over image.

Editing an existing image

To edit instead of generate from scratch, pass the existing image as image (a local file path is easiest — the skill reads and encodes it for you) plus a prompt describing the change. The configured image model receives the image alongside the prompt and returns an edited result.

curl -s -X POST http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/image \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "change the background to a starry night sky, keep the cat unchanged",
    "image": "/abs/path/to/input.png",
    "session_id": "<%= session_id %>"
  }'

• The result is a new edited image saved under assets/generated/ — the original file is never modified in place.
• For combining several inputs (e.g. "put the product from image 1 onto the background from image 2"), pass them as images: ["/path/a.png", "/path/b.png"] and describe the composition in the prompt.
• Same speed/concurrency rules apply: editing is as slow as generation, one at a time.

Response shape (success)

{
  "success": true,
  "image": "/abs/path/to/working_dir/assets/generated/img_20260525_011820_a1b2c3d4.png",
  "model": "<the configured image model>",
  "provider": "openclacky",
  "prompt": "A clean, modern hero illustration ...",
  "aspect_ratio": "landscape",
  "size": "1536x1024",
  "usage": {
    "prompt_tokens": 50,
    "completion_tokens": 4500,
    "cache_read_tokens": 0,
    "cache_write_tokens": 0,
    "total_tokens": 4550
  }
}

The image field is an absolute path on disk. To embed it in markdown, slides, or HTML, convert it to a path relative to the document you're writing.

usage may be absent when the configured backend doesn't return token counts. Treat it as optional.

Response shape (failure)

{
  "success": false,
  "image": null,
  "error": "Upstream 401: Invalid API key",
  "error_type": "api_error",
  "model": "...",
  "provider": "..."
}

Common error_type values: not_configured, auth_required, network_error, api_error, empty_response. Tell the user the error plainly; if it's auth_required or api_error 401/403, point them at settings to fix the api_key.

Step 3 — Show the image

Read does NOT show the image to the user — it only feeds it into your own context. To make the user actually see it, write a markdown tag in your reply:

![](file:///abs/path/from/response.png)

Take the image field from the response and prefix file:// (three slashes, since the path is absolute).

If you're also embedding it in a document (README, PPT, etc.), use a relative path: ![](./assets/generated/xxx.png).

Prompt writing tips

A good image prompt has 4 layers, in this order:

1. Subject — what is in the image, concretely. ("a golden retriever puppy", "a stylized icon of a rocket")
2. Style / medium — photo / illustration / 3D render / watercolor / flat vector / line art
3. Composition / lighting — close-up / wide shot / overhead / soft natural light / dramatic backlight
4. Mood / palette — minimal / playful / corporate / pastel / high-contrast monochrome

For PPT / slide decks specifically:

• Hero / cover slides: aspect_ratio: landscape, prompt should emphasise "clean", "minimal", "negative space" so text overlays well
• Section dividers: aspect_ratio: landscape, abstract or pattern-style works better than literal subjects
• Inline figures: aspect_ratio: square or portrait, more literal subject is fine

When the user gives a vague request like "给我配张图", ask one clarifying question (subject? style?) before calling the API — costs real money per image.

When NOT to use this skill

• The user wants a diagram / chart with specific data — use a charting library (matplotlib, mermaid, etc.) instead; image gen is for illustrations, not data viz
• The user asks for screenshots of real software — use the browser tool

Generating video (Veo)

The same /api/media/ namespace serves video generation. The user must configure a type=video model in settings (recommended: or-veo-3-1).

Endpoint

POST http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/video

Check GET /api/media/types first — if video.configured = false, tell the user to add a type=video model in settings before generating.

Video is slow and expensive

• A single clip can take 1–3 minutes (sometimes longer). Warn the user before calling, and run the curl in the foreground — it blocks until the MP4 is ready. Do NOT background it.
• One at a time. Never run multiple video generations concurrently.
• Each clip costs real money (billed per output-second). Confirm the prompt with the user before generating.

Request

curl -s -X POST http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/video \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A cinematic drone shot flying over a misty mountain range at sunrise, golden light, 4K.",
    "aspect_ratio": "landscape",
    "duration_seconds": 8,
    "output_dir": "'"$(pwd)"'",
    "session_id": "<%= session_id %>"
  }'

Field	Required	Values	Notes
prompt	yes	string	Same prompt-craft tips as images apply.
aspect_ratio	no	landscape / portrait	Defaults to landscape (16:9).
duration_seconds	no	4–8	Defaults to 8.
image	no	{ "b64_json": "...", "mime_type": "image/png" }	Optional first frame for image-to-video.
output_dir	yes	absolute path	Always pass $(pwd) so files land in the current session workspace. MP4 saved under <output_dir>/assets/generated/.
session_id	yes	string	Current Clacky session ID. Always pass the rendered value shown in the request example.

Response (success)

{
  "success": true,
  "video": "/abs/path/to/working_dir/assets/generated/vid_20260615_011820_a1b2c3d4.mp4",
  "model": "or-veo-3-1",
  "provider": "openclacky",
  "prompt": "A cinematic drone shot ...",
  "aspect_ratio": "landscape",
  "duration_seconds": 8,
  "cost_usd": 2.688
}

The video field is an absolute path on disk. Show it to the user with a markdown link or an HTML5 <video> tag pointing at the file:// path; embed it in documents with a relative path under ./assets/generated/.

Response (failure)

Same shape and error_type values as image generation, but with "video": null. not_configured means no type=video model is set up.

Continuous / long video (last-frame chaining)

A single Veo call maxes out at 8 seconds, and separate calls are visually unrelated (the character, lighting and framing jump between clips). To make several clips flow as one continuous shot, chain them: take the last frame of clip N and feed it as the image (first frame) of clip N+1. Veo's image-to-video then continues from exactly where the previous clip ended, so the seam is smooth.

Use the helper script (it only does the ffmpeg mechanics — you drive the generation with the same /api/media/video curl as above). The script's absolute path is given in the Supporting Files block; assign it once:

SEQ="SKILL_DIR/scripts/video_seq.sh"   # SKILL_DIR is provided in Supporting Files
# subcommands: lastframe | tob64 | payload | concat | probe

Workflow for an N-segment continuous video:

1. Plan the shots. Split the story into 4–8s beats. Write one prompt per beat; each prompt should describe the continuation, e.g. "The same girl keeps walking forward, the camera pushes in…". Keep subject, style and lighting wording consistent across prompts.
2. Segment 1 — normal text-to-video call. Save the returned mp4 path.
3. Extract its last frame (as JPEG — keep the .jpg extension):

   "$SEQ" lastframe seg1.mp4 /tmp/seg1_last.jpg
   

4. Segment 2 — build the request body with payload, then post it with curl --data @file. Do NOT inline the base64 into -d "{…}" — a frame's base64 is ~150KB+ and overflows the shell's argument limit ("Argument list too long"). The payload subcommand reads the frame, base64-encodes it, and writes a ready-to-send JSON file:

   "$SEQ" payload /tmp/seg2.json /tmp/seg1_last.jpg 8 landscape "$OUT_DIR" \
     "Continuing the same scene, the camera keeps pushing forward…" "<%= session_id %>"
   curl -s -X POST .../api/media/video -H "Content-Type: application/json" \
     --data @/tmp/seg2.json
   

   (payload <out.json> <frame> <duration_seconds> <aspect_ratio> <output_dir> <prompt> [session_id])
5. Repeat steps 3–4 for each subsequent segment, always chaining off the previous segment's last frame.
6. Stitch all clips in order into one file:

   "$SEQ" concat final.mp4 seg1.mp4 seg2.mp4 seg3.mp4
   

Rules & caveats:

• Strictly sequential. Generate one segment, wait for it, extract its frame, then start the next. Never run two video generations at once.
• Keep prompts consistent. The image carries visual continuity, but the prompt must not contradict it (don't switch the subject or scene mid-chain unless you intend a cut).
• Aspect ratio must match across all segments, or concat falls back to a slower re-encode (and may letterbox). Use the same aspect_ratio everywhere.
• Cost adds up linearly — N segments ≈ N × single-clip price. Confirm the number of segments and total length with the user before starting.
• For >30s or a true single-take >8s with no seam at all, this client-side chaining is the practical option today; Veo's native server-side extend (148s) is not wired into this endpoint yet.

Seedance (Volcengine Ark) — multimodal video

When the configured type=video model is a ByteDance Doubao Seedance model on Volcengine Ark (its Base URL is under *.volces.com, e.g. https://ark.cn-beijing.volces.com/api/v3), the same POST /api/media/video endpoint drives it. No separate endpoint — the server routes by Base URL automatically. Seedance adds richer inputs on top of the common fields; all are optional and only apply to Seedance:

Cost gate — ask before EVERY generation. Resolution is the main driver of Seedance's price (4k costs far more than 720p). So once you've confirmed via GET /api/media/types that the type=video Base URL is under *.volces.com, you MUST ask the user which resolution they want before EACH AND EVERY billable call — this covers not just a brand-new clip but also editing, multimodal reference, and extending/continuing an existing video (they all cost the same as a fresh render). Offer 480p / 720p / 1080p / 4k and state the default is 720p. Only after they answer (or explicitly say "use the default") do you proceed, passing their choice as resolution. Ask again every single time — a resolution the user picked for one clip is NEVER carried over to the next generation. Do not assume, do not reuse a prior answer, do not batch. One generation = one fresh resolution question. When editing or continuing/extending an existing video, default to that source video's resolution — never silently upgrade it (e.g. don't turn a 720p source into a 4k render). If the user gave no answer and you didn't ask, the server pins 720p. These Seedance-only fields (resolution, generate_audio, watermark, seed, first_frame, last_frame, reference_*) have NO effect on Veo or Qwen/DashScope backends — never send them unless the Base URL is *.volces.com.

Field	Values	Notes
aspect_ratio	landscape/portrait/square, or a raw Ark ratio like 16:9, 9:16, 4:3, 3:4, 21:9, adaptive	Raw ratios pass through unchanged.
duration_seconds	integer, or -1	-1 lets the model pick the length (Seedance 2.0 / 1.5 Pro).
resolution	480p / 720p / 1080p / 4k	Defaults to 720p when omitted (cost control). Ask the user before every generation — never reuse a prior answer. See the cost gate above. Model-dependent; unsupported values are rejected upstream.
generate_audio	true / false	Seedance 2.0 / 1.5 Pro can synthesize a synced audio track.
watermark	true / false
seed	integer	Reproducibility.
first_frame	media ref (see below)	First frame → image-to-video.
last_frame	media ref	Together with first_frame → first+last-frame video.
reference_images	array of media refs (0–9)	Reference images.
reference_videos	array of media refs (0–3)	Reference videos.
reference_audios	array of media refs (0–3)	Reference audio (background music / voice).

Which fields for which task — Seedance covers six capabilities; pick the fields by intent, and never mix the two families below:

Task	What you want	Fields to send
Text-to-video	a clip from a prompt only	prompt (no media)
Image-to-video (first frame)	animate a still image forward	first_frame
Image-to-video (first + last frame)	interpolate between two stills	first_frame + last_frame
Multimodal generation	new clip guided by reference images/videos/audio	reference_images / reference_videos / reference_audios
Edit an existing video	replace/add/remove/repaint something inside a given video	reference_videos: [<the video to edit>] (+ optional reference_images / reference_audios) + a prompt describing the edit
Extend / continue a video	prepend/append or stitch clips into one	reference_videos: [<clip1>, <clip2>, ...] (up to 3) + a prompt describing the join

🚫 Hard rule — the two families are mutually exclusive. first_frame/last_frame cannot be combined with any reference_* field; Ark rejects the request. If the user wants to edit or extend an existing video, that is a reference_videos task — do NOT fall back to extracting a frame and using first_frame (that produces a brand-new clip and silently loses the "edit the original" intent). The server also enforces this and returns a clear invalid_argument error if you mix them.

A media ref may be:

• a public http(s):// URL, or a data: URL, or
• a local file path (the server reads and base64-encodes it), or
• a { "b64_json": "...", "mime_type": "image/png" } hash.

Note: audio cannot be sent alone — pair it with at least one image or video. Prefer passing large videos/audios as public URLs; base64-encoding a big local file can exceed upstream size limits.

Example — first + last frame (image-to-video, no reference_*):

curl -s -X POST http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/video \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "First-person POV, a hand raises a cup of fruit tea toward the camera, bright and refreshing lighting",
    "aspect_ratio": "9:16",
    "duration_seconds": 8,
    "resolution": "720p",
    "first_frame": "'"$(pwd)"'/assets/frame_first.jpg",
    "last_frame": "'"$(pwd)"'/assets/frame_last.jpg",
    "output_dir": "'"$(pwd)"'",
    "session_id": "<%= session_id %>"
  }'

Example — edit an existing video (replace/add/remove something inside it; uses reference_videos, NOT first_frame):

curl -s -X POST http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/video \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Add a small wooden fishing boat with a warm lantern drifting slowly across the lake in the foreground, keep everything else unchanged",
    "resolution": "720p",
    "reference_videos": ["'"$(pwd)"'/assets/original.mp4"],
    "output_dir": "'"$(pwd)"'",
    "session_id": "<%= session_id %>"
  }'

Seedance is asynchronous — POST only submits, it does NOT return the video. Unlike Veo (which blocks and returns the mp4 in one call), the Seedance POST returns immediately with a task id:

{ "success": true, "status": "submitted", "task_id": "cgt-2024...-xxxx", "provider": "volcengine" }

status: "submitted" means the render is now running on Volcengine's servers and is already being billed — it does NOT mean it is done. You MUST now poll for completion: sleep ~15 seconds, then query the status endpoint, and repeat until it is succeeded (or failed):

curl -s "http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/video/status?task_id=cgt-2024...-xxxx&output_dir=$(pwd)&session_id=<%= session_id %>"

Status responses:

{ "success": true,  "status": "running" }                         // keep polling
{ "success": true,  "status": "succeeded", "video": "/abs/path.mp4" }  // done — this is the file
{ "success": false, "status": "failed", "error": "..." }          // give up, report to user

Only once you receive status: "succeeded" and the absolute video path may you present the result to the user. Do NOT end your turn while the task is still submitted/running — the user is waiting for the finished video.

A minimal poll loop:

TASK_ID=$(curl -s -X POST http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/video \
  -H "Content-Type: application/json" \
  -d '{"prompt":"...","resolution":"720p","output_dir":"'"$(pwd)"'","session_id":"<%= session_id %>"}' \
  | sed -n 's/.*"task_id"[[:space:]]*:[[:space:]]*"\([^"]*\)".*/\1/p')

while true; do
  sleep 15
  RESP=$(curl -s "http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/video/status?task_id=${TASK_ID}&output_dir=$(pwd)&session_id=<%= session_id %>")
  STATUS=$(echo "$RESP" | sed -n 's/.*"status"[[:space:]]*:[[:space:]]*"\([^"]*\)".*/\1/p')
  echo "poll: $STATUS"
  [ "$STATUS" = "succeeded" ] && echo "$RESP" && break
  [ "$STATUS" = "failed" ] && echo "$RESP" && break
done

Hard rules — a broken version of this once doubled a user's bill:

• ❌ Never POST the same generation twice. Once you have a task_id, the only valid next action is polling /api/media/video/status. A slow render is not a failed one.
• ⚠️ A timeout or error is NOT proof the task failed. The task keeps running and billing on Volcengine's side. Always query the status endpoint to find out the real state before doing anything else — never resubmit.
• ❌ Never kill the poll to "cancel" the job. Killing your curl/session does not stop the Volcengine task; it keeps running and billing. A running task also cannot be deleted upstream.
• ❌ Never bypass /api/media/* to call Volcengine's native API directly. All submission and status checks must go through this server (it meters cost). There is no reason to touch the raw Ark API.
• ⏱️ If polling exceeds ~15 minutes and status is still running, stop polling and tell the user the task is still rendering in the background, give them the task_id, and let them check again later — do NOT resubmit.

Generating speech (Gemini TTS)

The same /api/media/ namespace serves text-to-speech. The user must configure a type=audio model in settings (recommended: or-tts-gemini-2-5-flash, the cheap+fast default).

Endpoint

POST http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/audio/speech

Check GET /api/media/types first — if audio.configured = false, tell the user to add a type=audio model in settings before generating.

Request

curl -s -X POST http://${CLACKY_SERVER_HOST}:${CLACKY_SERVER_PORT}/api/media/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Hello and welcome to openclacky. Today we will explore...",
    "voice": "Kore",
    "output_dir": "'"$(pwd)"'",
    "session_id": "<%= session_id %>"
  }'

Field	Required	Values	Notes
input	yes	string	The text to speak. Plain prose works best; you can prefix with style cues like "Say cheerfully:" or "In a calm tone:".
voice	no	string voice name	Defaults to Kore. Common Gemini voices: Kore, Puck, Charon, Fenrir, Aoede.
output_dir	yes	absolute path	Always pass $(pwd) so files land in the current session workspace. WAV saved under <output_dir>/assets/generated/.
session_id	yes	string	Current Clacky session ID. Always pass the rendered value shown in the request example.

Generation typically takes 2–10 seconds depending on length. The request blocks until the WAV is ready.

Response (success)

{
  "success": true,
  "audio": "/abs/path/to/working_dir/assets/generated/tts_20260615_233522_4ff02705.wav",
  "model": "or-tts-gemini-2-5-flash",
  "provider": "openclacky",
  "input": "Hello and welcome to openclacky...",
  "voice": "Kore",
  "mime_type": "audio/wav",
  "usage": { "prompt_tokens": 13, "completion_tokens": 122, "total_tokens": 135 },
  "cost_usd": 0.000259
}

The audio field is an absolute path on disk. Output is mono 16-bit PCM at 24 kHz wrapped in a standard WAV container — playable by any browser, OS player, or <audio> tag without conversion.

To let the user hear it, write a markdown link in your reply:

[🔊 听一下](file:///abs/path/from/response.wav)

For embedding in HTML documents, use:

<audio controls src="./assets/generated/xxx.wav"></audio>

Response (failure)

Same shape and error_type values as image generation, but with "audio": null. not_configured means no type=audio model is set up.

Cost & length tips

• Gemini TTS bills by tokens (input text + generated audio). A typical one-paragraph narration costs well under $0.001.
• For long-form audio (>1 minute), split the script into paragraphs and generate each separately, then concatenate locally — avoids upstream truncation and gives you finer control over pacing.
• Voice consistency: Gemini TTS does not currently support voice cloning; use the same voice name across calls in one project to keep the narrator consistent.