[Bug]: antigravity-acp server orphans 1.16 GB _MEI temp dir per launch, fills OS drive #11204

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

antigravity-acp tool bundle (~/.t3/tools/antigravity-acp)

Steps to reproduce

• Open T3 Code (Alpha) 0.0.40 on Windows 11 with the Antigravity provider connected
• Start an Antigravity-agent session (spawns agy_acp_server.exe from ~/.t3/tools/antigravity-acp/win32-x64/versions/47cb50…/)
• End the session / cancel a prompt / close or kill the server process
• Check %TEMP% — a ~1.16 GB _MEI* dir (e.g. _MEI00002a002, containing google3/.../antigravity_extensions/acp_server, agy_acp_licenses.txt) is left behind
• Repeat — every launch orphans another one (10+ in a single day observed)

Expected behavior

• The PyInstaller bundle extracts once to a persistent dir (or cleans up its _MEI* dir on exit), so repeated sessions don't accumulate gigabytes in %TEMP%

Actual behavior

• Each launch extracts ~1.16 GB to a new %TEMP%\_MEI* dir that is never removed
• 39 orphans (~37 GB) accumulated and filled C: to 0 bytes free
• .t3/userdata writes and _MEI creation share timestamps to the minute (13:03, 13:06, 13:11, 13:12), confirming the 1:1 link with Antigravity-agent sessions
• Code-level mechanism (verified in repo main): the managed binary is Google's official PyInstaller-onefile agy_acp_server.exe (AntigravityInstallation.ts downloads/verifies it, no temp handling); sessions spawn it via buildAntigravityAcpSpawnInput with no TEMP/runtime-tmpdir override (antigravityAuthSupport.ts); on teardown/cancel-timeout/interrupt, retireRuntime runs child.kill({ forceKillAfter: "1 second" }) (acp/AcpSessionRuntime.ts) — a force-kill the PyInstaller bootloader cannot clean up after, so every killed session guarantees one orphaned _MEI*
• Reproduced locally (controlled, orphans removed after): spawned the bundled exe directly, waited for extraction (1,192.3 MB _MEI*), force-killed → orphan remained. Second run: closing stdin does NOT exit the server (ignores EOF), so it also ends in kill → orphan. Every launch orphans, graceful or not.
• Worse than disk waste: killing the bootloader leaves the server CHILD process running headless (observed 2 ghost agy_acp_server processes holding the _MEI DLL locks, blocking even manual deletion until killed). So orphaned sessions may also linger as ghost processes, not just dead disk.

Impact

Critical on small OS drives — 0 bytes free bricks the system (apps fail to write, updaters fail)

Version or commit

0.0.40 (tool bundle release 47cb50eef14f0a4655d78cfcfda869bcea7aaee5f9787e936bc2935ea612c3b8, agy_acp_server.exe 410 MB)

Environment

Windows 11, win32-x64

Logs or stack traces

Screenshots, recordings, or supporting files

No response

Workaround

Manually delete %TEMP%\_MEI* orphans (reclaimed ~36 GB; kill any ghost agy_acp_server processes first or files stay locked) and/or move user %TEMP%/%TMP% to a data drive