# Proof audit, pass 1 — Antigravity validation cleanup

PR [#32462](https://git.cafe/pingdotgg/t3code/pull/32462), fixing [#32457](https://github.com/pingdotgg/t3code/issues/32457).

## Findings

**No missing capture row. Existing linked evidence is sufficient for the affected server behavior, with native Windows explicitly unverified.** The diff changes one Windows-only finalizer from `Effect.orDie` to `Effect.ignore` and adds one regression test. There are no changed rendering components, props, layout rules, client contracts, or payload shapes. The observable change is that a validated installation reaches the existing success state instead of the existing failure state.

Reviewed the supplied local, unpushed merge [bec70fef2e](https://git.cafe/pingdotgg/t3code/commit/bec70fef2e4a27edd963ca49601fb6475fa0f61d) against local `origin/main`, [20c3dd3689](https://git.cafe/pingdotgg/t3code/commit/20c3dd3689782bfcb1a04332086d08017028c718), using `git diff origin/main...HEAD` in the supplied checkout. Both changed files are byte-for-byte identical to the public evidence revision [b8bb93b76d](https://github.com/pingdotgg/t3code/commit/b8bb93b76d4ae5952d5466910f7a2bc4ebe79480); this was checked against the raw files at that revision. Thus the merged main has not altered either changed module or its regression fixture. This audit did not rerun tests or reproduce the pre-fix failure; test execution and pre-fix failure are existing evidence, not new audit results.

At the final read-only check, the shared checkout had advanced to merge [ebb02a9066](https://git.cafe/pingdotgg/t3code/commit/ebb02a9066542f8e8b4b66704d7c10d73cd40f60). `git diff bec70fef2e4a27edd963ca49601fb6475fa0f61d HEAD` was empty: the whole tree remains identical to the requested audit head. The checkout remained clean.

Evidence read in full:

- [Earlier fixture matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9fcbe2e7a9d8557e/fixture-matrix.md): its no-render-change conclusion matches this independently derived matrix.
- [Focused test log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/27d633120ce598d9/antigravity-tests.log): one file, 30 tests passed. The current code contains the described default-validation regression with injected `win32`, Windows ZIP fixture, real Node ACP initialize responder, and a kill wrapper that awaits actual child exit before calling the real spawner kill.
- [Runtime transcript](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7880ee37a0772ae2/kill-already-exited-child.log): exit code 0 followed by the same typed `ChildProcess.kill` failure. It proves Linux spawner behavior after exit, not native Windows process-tree behavior.
- PR description from `cafe pr view 32462 --repo pingdotgg/t3code --json`, `data.resource.description`: reports the regression failing with the old finalizer on main, the fixed run, and `cannot-verify windows`. No separate linked pre-fix execution log was supplied; that failure is a reported result, also described by the earlier audit.

## Paths derived from code

[AntigravityInstallation.ts](https://github.com/pingdotgg/t3code/blob/b8bb93b76d4ae5952d5466910f7a2bc4ebe79480/apps/server/src/provider/AntigravityInstallation.ts#L387) supplies default `validate` only when `options.validate` is absent. The production server layer constructs the service; tests also construct it. The extra finalizer is registered only for `HostProcess.Platform === "win32"`; Linux and macOS use the ordinary spawner.

All validation callers are inside `install`: existing completed destination; newly extracted runtime pair; and a completed winner discovered after a failed publish rename. Fresh install, retry, update and reinstall converge on those three sites. All scope validation before activation. The kill remains attempted with the same one-second force-kill setting. Only typed kill failures are discarded; the installed Effect implementation preserves defects. ACP initialize validation, interruption/timeout, profile removal and publish/activation errors remain independent outcomes. The checks for runtime name/version, protocol 1 or 2, load/resume/logout support and oauth-personal remain unchanged. Windows profile removal retains its existing retries for locked/busy files.

Production entry is `provider.install.start` in `ws.ts` -> `ProviderInstallation.start` -> `AntigravityInstallation.start` -> `install`. Subscribe, cancel and remove use separate existing RPCs. `AntigravityDriver` uses resolve/acquire, not this installation validator. The server refreshes matching providers when `installedVersion` changes; the start service also attempts model/account refresh for enabled instances after success. These can change existing provider summaries and sign-in availability; success does not establish that authentication or real Antigravity sessions work.

[ProviderSetupSection](https://github.com/pingdotgg/t3code/blob/b8bb93b76d4ae5952d5466910f7a2bc4ebe79480/apps/web/src/components/settings/ProviderSetupSection.tsx#L48) is unchanged. Its only production caller is the Antigravity editor in `ProviderSettingsPanel`, reached through `/settings/providers?environmentId=<environment>&instanceId=<instance>`. Desktop wraps the same web UI. Native mobile has no consumer of the install subscription/start command; it displays refreshed provider/account availability. No additional Antigravity managed-install entry was found in onboarding, keybindings, the command palette, or MCP. Local, remote, relay and tunnel use the destination server's platform; the browser's OS does not select this branch.

## Rendering axes and interactions

The set of changed components and changed render paths is empty. To account for downstream visibility, the following existing branch inputs were inspected:

- Route/card: connected versus absent environment, list versus editor, enabled/disabled, read-only, missing provider/setup metadata, environment label wrapping. Environment/instance ids choose the query and keyed action component.
- Runtime: null query; idle/downloading/extracting/verifying/succeeded/failed/cancelled; downloadedBytes; totalBytes null/zero/positive; operationId absent/present; version and installedVersion null/equal/different; provider.installed; canInstall; canRemove; message absent/equal to status/different. Previous installedVersion remains visible on failed or cancelled updates.
- Guards: trimmed custom binary empty/present, enabled/disabled custom runtime, unsupported install, active auth, pending command, query/command error. Custom-path starts are rejected before this validator. Query errors expose Retry setup status; command failures expose an alert. Read-only and missing setup metadata render their existing unavailable/update rows.
- Interactions: Enable; Install/Retry/Update/Reinstall; Cancel with current operationId; Remove with confirmation and return to idle; retry setup queries. The unchanged auth sibling can enable sign-in after provider refresh. Auth methods, account text, method counts, browser/terminal/credential interactions and auth errors are not altered by this finalizer.
- Layout: narrow/wide setup container, settings-row 32rem breakpoint, small-screen alignment, light/dark theme and forced-colors progress styles are unchanged. Byte/version/message lengths and counts introduce no new render case. These axes do not multiply the capture requirement for the server fix.

## Reduced fixture matrix

Common affected-row seed: a selected environment with an Antigravity instance, custom binary empty, canInstall=true, operationId present, no pending/query/command error, auth inactive, positive totalBytes, version=`fixture-new`. The service fixture uses the Windows ZIP and initialize response with name `antigravity-acp`, matching version, protocol 1, loadSession/resume/logout, and oauth-personal. Set provider.installed=false and installedVersion=null unless overridden. The table separates distinct terminal render outcomes; unchanged guards above are not expanded into an unrelated full settings regression suite.

`recording` marks a fixture whose demonstration needs an interaction or time. It does not require a capture for this server-only change. No app or provider was started in this audit.

| Row / fixture | Required data and visible result | Diff effect | Real-app seed / substitute | Medium | Evidence status |
| --- | --- | --- | --- | --- | --- |
| A1 — Fresh false failure | Valid initialize; child exits; kill fails with typed PlatformError; no previous version. Before fix: failed, generic cleanup failure message, Retry installation, Remove, missing runtime. | Direct: this false failure is eliminated. | Actual Windows install can encounter the race but cannot deterministically seed it from Linux. Pre-fix default-validator fixture forces exit before real kill; a UI-fed state would only illustrate it. | recording | Covered by reported pre-fix failure and real Linux kill transcript. No missing capture. |
| A2 — Update false failure | As A1; previous installedVersion differs from version, provider may remain installed. Installed status plus failure message, Update Antigravity, Remove. | Direct: same finalizer on update/reuse/collision paths. | Previous managed release plus newer pinned asset on Windows. Linux substitute: previous=true / seeded completed release and filesystem collision harness. | recording | Covered by common validator reasoning and existing previous-release/activation tests; exited-child update combination is inferred, not separately demonstrated. No missing capture. |
| A3 — Reinstall false failure | As A1; previous installedVersion equals version. Installed plus failure message, Reinstall Antigravity, Remove. | Direct: same finalizer on existing-destination path. | Existing completed pinned release; start reinstall on Windows. Linux existing-release harness exercises reuse and denied activation. | recording | Covered by shared validator and reuse/retention tests; this exact exited-child prior-release combination is inferred. No missing capture. |
| A4 — Validated success | Same typed kill failure; succeeded, installedVersion=version, message=null, canRemove=true. Installed, Reinstall Antigravity, Remove. Successful kill already reaches this state. | Direct: existing success now reached for A1–A3. | Real Windows install; Linux substitute is the added default-validation test with real Node child and injected win32. It asserts succeeded and fixture-new. | recording | Covered by linked 30/30 test result and identical current fixture. Native acceptance remains W1. |
| C1 — Idle | idle, installedVersion=null, message=null, canRemove=false; size hint and Install Antigravity. | Unaffected; before validation. | Real app with isolated empty managed directory and configured instance; merely viewing idle requires no Antigravity executable. | still; recording for Install | Existing UI; capture not required. |
| C2 — Query loading | install query null; Checking the Antigravity runtime. | Unaffected. | Open setup while subscription loads; deterministic hold needs delayed query harness. | recording | Existing UI; capture not required. |
| C3 — Download with known size | downloading, positive totalBytes, downloadedBytes within it; MB denominator/progress, Cancel. | Unaffected. | Native pinned download; deterministic hold uses existing streamed-body Deferred harness, with explicitly labelled UI data if desired. | recording | Existing focused download coverage; capture not required. |
| C4a — Download without size | downloading, totalBytes=null; counter without denominator/progress. | Unaffected; production start uses known asset size. | Synthetic subscription state; contract/UI input unchanged by the finalizer. | recording | Branch inspected; capture not required. |
| C4b — Download with zero size | downloading, totalBytes=0; formatted zero denominator, no progress. | Unaffected; production start uses positive known asset size. | Synthetic subscription state; contract/UI input unchanged by the finalizer. | recording | Branch inspected; capture not required. |
| C5 — Extract | extracting, operationId present; Extracting Antigravity, Cancel, no progress. | Unaffected. | Native install or existing filesystem-sink Deferred harness. | recording | Existing focused coverage; capture not required. |
| C6 — Verify / publish | verifying, operationId present; Checking the downloaded runtime, Cancel. Existing/new/collision validators share this view. | Cleanup outcome changes after this state; rendering is unchanged. | Native install; Linux default-validator fixture reaches it. Holding with a Deferred validator bypasses the changed finalizer and illustrates only the UI. | recording | Service transition covered; capture not required. |
| C7 — Genuine failure | Initialize rejects/fails, timeout, profile deletion or publish/activation fails; failed with message. Fresh: Retry. Prior unequal/equal version: Update/Reinstall and retained installed status. | Preserved independent failure gates. | Native bad runtime or locked files; synthetic ACP/filesystem fixtures on Linux. | recording | Existing archive/ENOSPC/denied-pointer tests and source inspection; no claim of native NTFS coverage. No new capture needed. |
| C8 — Cancel | Cancel active operation; cancelled, prior version retained, cancellation message; fresh Retry or prior Update/Reinstall. | Interruption still cancels. | Cancel via current operationId; existing Deferred-gated download/extract/verify tests await actual interruption and resource release. | recording | Existing focused coverage; capture not required. |
| C9 — Remove / retry | Confirm Remove -> idle, installedVersion=null, canRemove=false; Retry -> active flow. | Unaffected. | Isolated managed directory, remove RPC and confirmation; existing removal/reinstall fixture. | recording | Existing focused coverage; capture not required. |
| C10 — Non-Windows host | Linux/macOS; default validator uses unwrapped spawner. | Unaffected; changed finalizer absent. | Linux tests; native Antigravity runtime required to claim real installation. | recording for install | Platform branch inspected; capture not required. |
| W1 — Native Windows acceptance | Actual Windows host, pinned Google runtime, native taskkill, PyInstaller descendants, profile release and NTFS publishing. | Target platform; native lifecycle remains unverified. | Cannot faithfully seed on these Linux VMs. Injected win32 and real Node child substitute for the typed kill failure, not taskkill/tree/DLL/NTFS behavior. | recording plus native runtime logs if available | **cannot-verify windows**, already explicit in PR. Accepted stated boundary; no fabricated substitute capture. |

**Minimum capture set (viewport × theme × row): empty — 0 captures.** Every affected nonnative row is covered by existing behavioral proof or the same-validator/retained-state reasoning above. No capturer must add a row. A labelled optional UI recording could illustrate failure-to-success, but no viewport/theme combination would establish native Windows cleanup. W1 remains disclosed rather than claimed verified.

Coverage completed: full two-file diff, all three validator call sites, production service/RPC callers, settings caller and render conditions, web/desktop/mobile boundaries, provider refresh consequences, linked matrix/log/transcript, and evidence/current-file identity. No worktree edits, pushes, comments, tests, dev servers or new worktrees. Scratch downloads and this report were the only files written; scratch is removed after successful upload.

FRICTION:
- tool | Web open could not access all three evidence attachments; Python urllib received HTTP 403; curl fetched them successfully | Use curl for these file-drop attachments.
- forge | GitCafe reported head fc3290f1bd while the origin branch advertised b8bb93b76d and the requested local merge was bec70fef2e | Pin the supplied local head and compare the evidence-revision files directly, rather than attribute remote CI to the local merge.

VERDICT: SUFFICIENT
