# Local ACP flavor parity runtime proof

Environment: Linux web development app, Chromium, 1280×800 CSS pixels (screenshots at device scale 2), light theme. Each commit ran from its own checkout with a fresh `.t3` home and T3CODE_DEV_SEED=0. Only synthetic data was used. No adapter test harness or unit tests were run.

Base: ed203c46cb3d6d1d53dd85ac58852425b8434b6a
Head: 637e017e11fb989e3e6fe8defe04332c6b9b3a4b

The scenario runner documentation excludes provider-specific behavior; this proof uses real local ACP subprocesses instead. `acp-fixture.py` implements the protocol and chooses behavior from its executable filename. Copy it to three executable files named `glm`, `droid`, and `vibe-acp`. All three provider instances use driver `acpRegistry`, config.source `local`, commandPath pointing to that executable, no arguments, customModels `["fake"]`, enabled true. All real providers are disabled.

For each commit:

1. Start `T3CODE_DEV_SEED=0 vp run dev --home-dir .t3` with the same local provider settings. Add an empty synthetic Git project using `node apps/server/src/bin.ts project add <project> --title 'ACP local parity' --base-dir .t3`.
2. Start GLM: `node apps/server/src/bin.ts thread start <project> hi --provider glm --model fake --title 'GLM token usage' --wait --json --base-dir .t3`. The subprocess returns protocol v1 prompt usage: 2500 input, 500 output. Inspect the actual persisted provider-turn projection through read-only SQLite with `collect.py`. Base marks usage unavailable; head records complete usage with both counts.
3. Start Droid the same way with `--provider droid --title 'Droid paired form'`, without `--wait`. Open its thread in the real browser, collapse Thread details, and record. Select Plain text, click Next, and type Markdown for the second answer. Base incorrectly displays Question 2 (the first field's companion), then requires an additional Alternative at 3/4. Head correctly displays Alternative at 2/2, accepts Submit, and the ACP subprocess prints the exact response it received: question_1=option_1, question_2=other, question_2_other=Markdown. The captures retain this difference rather than bypassing the extra base controls.
4. Start Vibe with `--provider vibe-acp --title 'Vibe rate limit'`. After the initial turn, open that thread, collapse Thread details, start recording, type `hi`, and press Enter. The same subprocess sends three `_session/retrying` notifications with category rate_limited and HTTP 429 detail, then rejects prompt with -31001. Base shows no retry notices and ends with Provider error. Head shows Retrying provider, then Usage limit reached / Limited. Both initial and recorded turns are included in the runtime capture.
5. `python3 collect.py <checkout> <before|after> <output.json> <commit>` reads only that checkout's synthetic database. The JSON files contain selected projection fields, not fabricated assertions. The videos show actual browser interaction. Stills were accepted only after matching screenshots at least 500 ms apart.

GLM's token counts are shown in the runtime JSON, not in the fixture's assistant message (the message is merely a description of the input). Droid's received response message is produced by the ACP subprocess from the JSON-RPC answer it actually received.
