Base: pingdotgg/t3code main at e9d61252f6fd. Candidate: a2483937. Environment: Windows 10 build 19045, x64, Node 26.12.1, Vite+ 1.3.0.
Before applying the implementation, the new test was run against the base resolver:
vp test run packages/shared/src/shell.test.ts -t "yields while discovering a Windows executable"
FAIL: expected true to be false
1 failed; 44 tests outside the selected scope
It checks whether a native Promise microtask can run before executable discovery finishes. It uses no duration assertion, sleep, polling, retry, or custom timeout. The patched test passes. Other coverage checks installation immediately after a miss, PATH/PATHEXT precedence, inaccessible directory enumeration, Windows short-name fallback, existing cache behavior, shim escaping, and async provider lookup.
Three trials alternate baseline/candidate order. The baseline is the unmodified shared shell module from the base commit, imported beside its original HostProcess module. The candidate is the checked-in shell module. Both receive the same environment and fixture. No running T3 instance, database, provider credential, or conversation is used.
The benchmark calls the actual resolver and server ProcessRunner. Missing commands are intentionally unique and must produce ProcessSpawnError. Successful process runs execute a hard link to the test runtime's real node.exe and must return exit 0 and stdout ok. Each ProcessRunner invocation gets a fresh successful-lookup cache so it exercises a cold lookup. A 5 ms Node interval records the longest callback gap; this is a responsiveness diagnostic, not an application latency or disk-throughput measurement.
Medians of three trials, milliseconds:
| Work | Baseline elapsed | Candidate elapsed | Baseline max callback gap | Candidate max callback gap |
|---|---|---|---|---|
| 32 uncached misses, actual inherited PATH | 4256.7 | 1044.3 | 4256.9 | 13.4 |
| 32 successful lookups, executable in last of 30 fixture directories | 494.4 | 219.4 | 494.5 | 19.7 |
| 8 real ProcessRunner missing-command attempts | 1100.4 | 382.4 | 1101.5 | 21.6 |
| 8 real ProcessRunner node starts | 1102.2 | 1137.5 | 92.6 | 56.9 |
Every missing-command sample reported the expected misses/errors; every successful lookup/child run succeeded. Successful process startup elapsed time was comparable rather than consistently faster. Raw lookup batches deliberately bypass the existing 30-second positive cache. Other tests were running during measurement, so absolute times and callback gaps depend on host load. These results do not establish that this fixes every reconnect or measure high-concurrency agent throughput.
To reproduce, save the attached benchmark as apps/server/.t3-pr-benchmark.ts in the candidate checkout and materialize the original shared shell beside its source:
git show e9d61252f6fd:packages/shared/src/shell.ts > packages/shared/src/.spawn-baseline.ts
node apps/server/.t3-pr-benchmark.ts
Use a full Node 26.12.1 runtime on Windows. Create ../../outputs relative to the checkout before running; that is where the script writes its JSON. The script removes its temporary fixture in finally. Remove the two untracked source files after the experiment. Do not commit this PR-only evidence.
The repository's unmodified scripts/check-changed.ts ran with base e9d61252f6fd. On Windows it launches a bare vp, which cannot execute the installed .cmd shim with Node spawn. The attached external preload only translates that exact command into process.execPath plus the installed Vite+ entry point; it does not change any check or arguments.
node --require ../t3-pr-vp-launcher.cjs scripts/check-changed.ts e9d61252f6fd
pass format
pass lint
pass knip 1/2
pass knip 2/2
pass GraphQL generation check
pass typecheck (29 affected packages)
test: 79 passed, 1 platform skip, 15 failed
All 15 failures are in AcpRegistrySupport.test.ts. A separate run using its unmodified source and fixtures from the base commit, with imports redirected to the unmodified baseline shell module, reproduced exactly the same 15 failing test names on this Windows host. They involve Unix fixture paths/scripts and runner availability. They were not skipped or changed to pass. The other three changed test modules passed, including the new Windows regression and short-name case.
Linux/WSL follow-up: Node 26.12.1 with the existing real npm/npx CLI available on PATH, running from packages/provider-acp-registry:
vp test run --config ../../vite.config.ts src/server/AcpRegistrySupport.test.ts
31 passed (31)
An earlier minimal Linux runtime without npm/npx had two runner-availability failures; providing those real tools resolved them without source/test changes.
The existing Claude replay-helper unused-function lint warning and existing Effect diagnostic suggestions remain non-failing. No full-repository checks or production build were run.
Model: GPT-6. Harness: Codex desktop.