# Claude startup failure reproduction

Run the real web/server dev stack on each specified commit with a fresh isolated T3 home and T3CODE_DEV_SEED=0. Install dependencies with vp i. No existing database is copied.

Create this executable shell stub and configure a Claude provider instance's binaryPath to it:

```sh
#!/bin/sh
if [ "$1" = "--version" ]; then
  echo "2.1.0 (Claude Code)"
  exit 0
fi
echo 'error: An internal error occurred (EPERM)' >&2
exit 1
```

The version branch lets executable detection succeed; every query exits immediately without SDK messages.

1. Add an empty synthetic git project using the app CLI's project add command.
2. Pair the web browser with the isolated server.
3. Choose the configured Claude instance. Open a fresh thread in the synthetic project through its breadcrumb.
4. Close the thread details panel and dismiss the provider account-status warning (the stub intentionally cannot answer account-status requests).
5. Type `Reply with exactly: OK` and press Enter.
6. Observe the failed turn. Base shows the generic provider-start failure; head identifies Claude Code exit code 1 and directs the user to server logs. Neither client message contains EPERM.
7. Inspect the real server's `orchestration V2 provider turn start failed` log entry. Base loses the startup cause; head retains the EPERM stderr line in its nested cause.
8. Dismiss transient error/provider warnings for still captures. Compare two screenshots taken at least 500 ms apart, then retain the identical conversation region.

Environment for both captures: real web client in Chromium, Linux, 1280×800 CSS viewport at device scale 2, light theme. Screenshots show the 1024×600 conversation region; recordings show that same region with everything below y=405 masked to exclude runtime model metadata. No behavior or failure text was edited. Recordings show typing, sending, and the resulting failure in a fresh thread. Server log excerpts redact checkout paths and run IDs.

AI-assisted; verified with the tests and tools listed above.
