# Claude refused-image recovery proof

Run each revision as the real web app with `T3CODE_DEV_SEED=0 vp run dev`, after `vp i`. Use separate checkouts and each checkout's own `.t3` database. Enable a `claude` provider instance with driver `claudeAgent`. The CLI must have working Claude access. No replay or mocked provider is used.

The driver takes: `node drive.mjs CHECKOUT RUN_DIRECTORY SERVER_ORIGIN ACTION [IMAGE_FILE]`. Give each run a new directory. `SERVER_ORIGIN` is the backend origin printed by that checkout, not the web port. Pair the browser with its web port, and open the thread returned in RUN_DIRECTORY/state.json. The scenario runner is unsuitable because its replay provider speaks generic ACP, not Claude's background-work protocol.

1. Run action `setup`. It creates a synthetic git project, starts Claude Haiku, and asks Claude to start a real 600-second Bash background command. It captures that command's PID in the sample project. Verify BACKGROUND READY and pending background work.
2. Start a browser recording. Run `refuse IMAGE_FILE`. It persists the PNG through the app's assets RPC and sends the image request with Haiku 4.5 / thinking disabled, which requires replacing the live process. Observe the real background-work refusal. The driver sends a follow-up and observes another refusal.
3. Run `native` to record the synthetic session's native identity. Stop only the dev runner you started, using its captured PID. The server shuts down its Claude processes and background commands.
4. With that server stopped, run `unavailable`. It backs up and empties only this synthetic native transcript. An existing transcript without conversation lines makes the real Claude adapter reject resume before prompting and triggers the fresh-session fallback. Do not delete the file instead: missing files defer detection to the CLI and exercise a different recovery route.
5. Restart the same app with `T3CODE_DEV_SEED=0 vp run dev`. Run `recover`. It asks for the code from the missed image without attaching a new image or including the code in the text. Wait for Claude's answer.
6. Capture the completed chat. Blur the composer and take two screenshots at least 500 ms apart; confirm they are identical. Stop the recording.
7. Optionally inspect the actual native prompt: `python3 inspect-native.py CHECKOUT RUN_DIRECTORY IMAGE_FILE`. It emits only counts, booleans, and SHA-256 hashes, not private transcript metadata.

Both runs: Linux web client, Chromium, 1280×800 CSS viewport, 2× device scale, light theme. Published captures crop the chat at CSS x=256..986, y=0..800 to omit machine identity and unrelated panels; videos use the same crop and scale. No chat text is modified.

Base: missed text is present in the accepted replacement-session prompt; image count 0; Claude answers IMAGE NOT AVAILABLE.
Head: missed text is present; image count 1; its bytes hash identically to the original PNG; Claude reads ORCHID-7392. Claude expresses uncertainty about the image's provenance, which the native prompt hash resolves.

Exploratory limitation: deleting the transcript entirely, instead of leaving an empty transcript, caused a later CLI error asking for a resend. Both commits answered IMAGE NOT AVAILABLE after that resend. The successful paired proof covers the earlier adapter-detected resume failure.

Prepared by GPT-6.1 Sol through the Codex harness. No repository changes or commits.
