Generate Media

Generate Media

One generation per invocation through a provider API. The file lands in $HOME/.cache/generate-media/.

Provider reference — read before the first API call:

• OpenRouter (default): providers/openrouter.md

Rules

1. Stop if the provider key is missing. For OpenRouter that is OPENROUTER_API_KEY — check .env in the project root, then the environment. Never write a key into a tracked file. If missing, point the user at https://openrouter.ai/settings/keys and stop.

2. Look up model IDs from the API. Never guess them. IDs change and a wrong one wastes a paid round trip.

3. Pick the cheapest model that meets the need, unless the user named one. Exceptions worth saying out loud: openai/gpt-image-2 for legible text inside an image, a -pro tier when the user asked for production quality.

4. Quote the cost before calling, and only quote a figure you can source. Video runs 10×–50× an image.

   • Video: use the per-second table in the provider reference. No row there means no published rate — say so and wait.
   • Image: unquotable before the call (pricing.image_output is per output token, not per image). Name the model, say the cost lands after the response, and go ahead — a still is normally a few cents.
   • Wait for the user when a sourced figure exceeds ~$0.50, or when a video price is unknown.

   Settle motion on a cheap draft before paying for a real render: bytedance/seedance-2.0-mini is ~$0.08 for 6 s at 720p, less at 480p. Price scales with width × height × duration, so halving the resolution is the cheapest way to test an idea.

5. Save to $OUT_DIR/<slug>.<ext>, where OUT_DIR="$HOME/.cache/generate-media". mkdir -p it first. Write $HOME, never a bare ~ — a quoted tilde does not expand and drops a literal ./~/ directory into the project. Never overwrite — add a numeric suffix. Never leave a result only in the API response. Copy into the project only when the user asks; raw output is 2–3 MB stills and 6 Mbps clips, which should not enter git.

6. Report the absolute file path and the real cost from usage.cost in the response, never an estimate.

Check before reporting done

• The file just written is non-empty: test -s "$OUT" — a failed base64 decode leaves 0 bytes. Check that one path, not the whole directory; it is shared across every project and holds old results.
• A video is an actual .mp4, not a JSON error body: file "$OUT".

Common mistakes

• Treating video as synchronous. POST /videos returns 202 with a polling_url. Poll ~every 30 s until completed, then download.
• Estimating cost from a pricing field. Read usage.cost from the response.
• Piping base64 straight into a file. Decode it: jq -r '.data[0].b64_json' resp.json | base64 -d > out.png.
• Sending a local path as a reference image. Reference inputs take an HTTP(S) URL or a base64 data URL.