OpenRouter Image Generation

Generate images from prompts using OpenRouter's image generation endpoint.

Requirements

• - OPENROUTER_API_KEY in environment
• INLINECODE1
• Python package: INLINECODE2

Install dependency:

CODEBLOCK0

FTUX preflight (run first)

CODEBLOCK1

If checks fail, tell the user exactly what is missing and provide copy/paste fix steps from the Fixups output.

Default model

• - INLINECODE4
• Optional alternatives (if enabled on your account): openai/gpt-5-image, INLINECODE6

Usage

CODEBLOCK2

Optional args:

CODEBLOCK3

Queue -> response pattern (avoid traffic jams)

When a user asks for multiple images/iterations, do not hold one long-running turn per image.
Do not block waiting for a "single message with all files" if the adapter does not support it.

Hard contract for queue mode:

• - NEVER run run_image_queue.py in the same foreground turn as enqueue for multi-image requests.
• ALWAYS enqueue + immediate queued ack first, then background handoff.

Use a queue + batched response flow:

1. 1. Enqueue each requested image quickly.
2. Immediately return with a short "queued" acknowledgement (do not wait for generation in the same turn).
3. Drain queue in background (preferred: sub-agent/session worker).
4. Send one consolidated completion status response when done.
5. Always attach generated image files (never send only paths).
6. If the messaging adapter allows only one media per send, post attachments as reply-chain messages under the consolidated completion status message (one file per message).

Enqueue command:

CODEBLOCK4

Background queue handoff (recommended):

CODEBLOCK5

Manual drain command (worker context only; not same foreground turn):

CODEBLOCK6

Batch-enqueue N variants with consistent file names:

CODEBLOCK7

Useful options:

CODEBLOCK8

Data transmission notice (external provider)

• - This skill sends prompts and generated/edit inputs to OpenRouter (openrouter.ai).
• If you pass --baseline-image, that image content is transmitted to the provider as part of the request.
• Do not submit sensitive/private images unless the user explicitly approves external transmission.

Notes

• - If generation fails due to model/provider mismatch, retry with --model openai/gpt-5-image-mini.
• Local baseline uploads are deny-by-default; pass --confirm-external-upload only when user explicitly approves sending that local file to provider.
• Local baseline files are restricted to png/jpg/jpeg/webp, non-symlink regular files, workspace-local paths, and max size threshold.
• For iterative work, prefer --image-size low; switch to medium or high for final renders.
• Use --clarify-hints to surface prompt-quality gaps early; use --strict-clarify for workflows that must fail on ambiguity.
• Keep prompts explicit for text rendering tasks.
• Save outputs into workspace paths, not /tmp, for durability.
• When user asks for generated images in chat, attach the generated file in the response (do not only send a path).
• For queue mode, read results from:

• - Edit/variant intent prompts fail fast if no baseline is supplied (--baseline-image) unless explicitly overridden.
• Resolve baseline deterministically in caller: current-message attachment > replied-message attachment > clarification request.
• Pass --baseline-source-kind current_attachment|reply_attachment|explicit_path_or_url for auditable provenance.
• When baseline is supplied, rails default to low-variation + locked palette/composition unless explicitly changed.
• Queue results persist provider metadata (generation id + provider response payload/path) and drift diagnostics (edit_intent_detected, baseline_applied, baseline_source, baseline_source_kind, baseline_resolution_policy, rails_applied) to help edits/debugging and smarter agent continuation.
• Queue worker writes handoff_mode + same_turn_drain_detected so you can enforce true async behavior in tests/ops.
• INLINECODE30 writes <prefix>-manifest.json with baseline, constraints, variant deltas, and output targets for reproducible reruns.