# Antigravity saved thinking level: manual real-app proof

Both sides ran their own dev server in an isolated worktree, with T3CODE_DEV_SEED=0. No live database was read or copied. Browser: Chromium, 1280x800 CSS pixels, device scale 2, light theme. Runtime: agy 1.5.13. API-key mode used the supplied Gemini-compatible test hub.

## Recorder

The attached Python recorder is a transparent stdin/stdout relay to the real agy executable. It never changes ACP requests or responses. It logs only method names, configuration values, agent version, text output and stop reasons; authentication secrets and filesystem paths are omitted. The public script replaces the two machine-local paths with AGY_REAL_EXECUTABLE and ACP_EVIDENCE_LOG environment variables. Its executable filename must be `agy` for T3 runtime discovery.

Configure its path as Antigravity's custom binary on each isolated app. The JSONL artifacts contain the complete selected title-helper process, including every client method in order. Provider discovery traffic, which does not prompt a model, is excluded. The PNGs combine a crop from real Settings screenshots with a condensed rendering of these JSONL records; they are explicitly labelled composites, not an additional app view.

## Browser actions performed on each side

1. Open the startup pairing URL, then /settings/providers.
2. Add provider -> Antigravity -> Next -> Next. Set Sign-in method to Gemini API key, set a nonempty test key, and set the recorder as Binary path. Add instance, then Connect.
3. Open /settings/threads. In the Text generation model row, open [data-chat-provider-model-picker], search Gemini 4 Neon, and select its option. Open the row's button[aria-haspopup=menu] and select menuitemradio High.
4. Confirm settings.json contains textGenerationModelSelection={instanceId:"antigravity",model:"gemini-4-neon",options:[{id:"thought_level",value:"high"}]}.
5. Open an unused thread in the app's automatically registered server project. Hide Thread details to keep the host label out of captures. In the composer model picker search Showcase, select the Replay (dev) option, and close the picker.
6. Fill [role=textbox][contenteditable=true] with Add a dark mode toggle to the settings page and click button[aria-label="Submit message"]. Wait for title generation to complete.
7. Read the recorded helper ACP exchange. Base selects the model then prompts while thought_level remains low. Head selects the model, sends thought_level=high, receives high in configOptions, then prompts. Both return end_turn and a generated title.
8. Reopen /settings/threads and scroll the Text generation model row into view. For each still, capture twice at least 500ms apart and compare pixels; both pairs were identical.

The flow videos show composer model selection, identical user message and completed title; the head video also shows saved settings. The PNGs show saved High on both sides. Privacy masks cover host labels and the conversation during streaming tool output, which includes workspace paths. The masks leave the title and composer visible, and disappear from the conversation after the Replay output collapses. The first head run also used the dev replay shortcut during setup; the retained head video is a second run on an unused thread using the same composer steps as base. Both head title helpers set High before prompting.

## Observation

The base title was Add Dark Mode Toggle to Settings. The head title was Settings Dark Mode Toggle. Title wording is nondeterministic and is not the proof: the ACP configuration immediately before session/prompt is.

The runtime accepted High. Invalid-parameter rejection fallback was not induced, so that branch is not claimed as verified.
