image-models

Image-model discovery recordings

record.sh compares real CLI binaries using openai images models --all against identical synthetic metadata. The before binary is pinned main, which does not have this command. The after binary shows exact IDs, visibility, retirement and dated versions. Separate offline and partial-failure scenes show unchecked names without a key and useful results after one metadata request fails. The fixture exposes no image generation route.

This is a real-PTY terminal replay rendered with asciinema 3.2.1 and agg 1.9.0, then assembled with ffmpeg. It is not native Apple Terminal, PowerShell or cmd.exe validation. The .tape file preserves an alternative VHS recipe; it is not the renderer used by record.sh.

Reproduce

Build the synthetic server from the repository root:

mkdir -p dist/demos/bin
go build -o dist/demos/bin/image-model-demo-api ./scripts/demos/image-models/main.go

Independently build openai at the comparison base and proposed feature commit. Verify binary provenance before recording: SHA arguments are recorded, but the runner cannot prove that a binary was built from the supplied commit. Put asciinema, agg, ffmpeg and ffprobe on PATH, then run:

bash scripts/demos/image-models/record.sh \
  /path/to/before/openai /path/to/after/openai \
  "$BEFORE_SHA" "$AFTER_SHA" /path/outside/repository/image-models-media

Both commits must be full 40-character IDs. Use main 37fe587ddf9e9c160f8d6c52e5f2b879f5623960 for this feature's before build; PR #238 is a behavior reference, not the comparison baseline. Use DEMO_API_BINARY to select a different build location for the fixture. The runner deliberately requires media outside the repository. It does not upload anything. Its temporary symlinks, shell state and local server are cleaned up on exit.

Each command runs under env -i with a temporary PATH exposing the chosen binary as openai, a fake API key, and a loopback URL. The offline scene removes the key. No user shell hooks, proxy settings or personal CLI environment are inherited. All four scenes have identical dimensions and styling: 110 columns by 34 rows, Menlo 22px, Dracula, line height 1.2, and a four-second final pause.

Fixtures and evidence

main.go binds only 127.0.0.1 on an ephemeral port. Its first argument receives the fixture origin; its second is a new request-log path. Supported base URLs:

• ORIGIN/normal/v1: known image IDs return metadata; chatgpt-image-latest returns 404; DALL-E metadata has the deliberately synthetic retirement date 2000-01-01. These statuses do not represent a live account or API.
• ORIGIN/partial/v1: the same data, with one 503 for gpt-image-1.
• ORIGIN/timeout/v1: the same data, with gpt-image-1 delayed ten seconds or until client cancellation. This optional manual fixture is not in the GIF.

Requests to /models and image-generation routes are rejected. The log records only recognized scenario/model names or a fixed rejection marker; headers, credentials, arbitrary URLs and bodies are never logged.

comparison.gif, before.png and after.png are the main comparison. offline.png / offline.gif and partial.png / partial.gif are supplemental. Each scene also has its original .cast, a text transcript and a final-frame GIF. The comparison cast and transcript retain both primary scenes. metadata.txt records supplied commits, binary/recipe hashes, actual statuses, request counts, OS, shell and tool versions. media.json records GIF dimensions and duration. requests.txt retains the synthetic metadata request log.

The runner checks exit statuses (before/partial: 1; after/offline: 0), request counts (before/offline: zero; after/partial: twelve), and expected output text. It waits for the fixture to shut down and close its request log successfully before copying the log and reporting success. Shutdown or close failures fail the recording. Inspect the resulting PNGs and GIF frames for readable text, clipping, timing and sensitive data before sharing. Regenerate media after changes to the demonstrated behavior. Keep binaries and media out of Git.

To use VHS separately, start the fixture, expose before/after directories with binaries named openai, and provide DEMO_API_URL, DEMO_BEFORE_DIR, DEMO_AFTER_DIR and a synthetic OPENAI_API_KEY in an isolated environment. Run vhs validate before rendering the tape from an external output directory.