# WSL keep-awake runtime proof

Environment: Microsoft Windows Server 2025 Datacenter; WSL2 Ubuntu 24.04, systemd enabled; kernel 6.18.58.1-microsoft-standard-WSL2; Node 26.13.1.

Base: 134580067f3c811f32630667bd57f048d0bff9ee (merge-base of origin/main and fix).
Head: c5227a77601bb13e1ee4b7172300e73d10e96c4c.

The application source is unmodified at both revisions. Each run has a fresh T3 home and an equivalent synthetic git project. No desktop host or native Windows T3 server runs alongside the WSL server. The keep-awake setting retains its default, true.

Run the same driver inside WSL for each checkout:

```sh
python3 drive.py SIDE RUN_NAME CHECKOUT OUTPUT_ON_WINDOWS_MOUNT FRESH_T3_HOME FRESH_PROJECT
```

The driver starts the actual server, adds the project through its CLI, and runs:

```sh
node apps/server/src/bin.ts thread start --provider devReplay --model command-output --wait PROJECT go --base-dir FRESH_T3_HOME
```

It samples the persisted orchestration run status through a read-only SQLite connection, the server's keep-awake log messages, and the real Windows `powercfg.exe /requests` output. It stops its own server after the turn completes and the inhibitor release is observed. The CLI's `--wait` means it awaits the replay turn; it does not simulate one.

For capture, run the monitor on the logged-in Windows desktop:

```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File monitor.ps1 SIDE RUN_NAME OUTPUT_DIRECTORY
```

The monitor displays actual runtime observations. Elapsed time identifies the original sample. It preserves a running sample if a brief replay turn falls between monitor refreshes. It does not change the application or generate sleep requests. Each saved screen must match a second capture taken 500 ms later. Videos encode the saved screens with their capture-time intervals (filesystem timestamp granularity is two seconds). Incomplete monitor reads are omitted from the sampled recording; all raw runtime samples remain available. A narrow border crop removes unrelated desktop windows; the application observations are unedited.

`base-samples.jsonl` and `head-samples.jsonl` contain the runtime observations, including powercfg output. Server logs containing the temporary pairing credential are excluded. All project data is synthetic.

The proof establishes the same OS-level condition as triage: no Windows SYSTEM sleep request on base, a Windows PowerShell SYSTEM sleep request during a running turn on head, and no request after head releases its inhibitor approximately 60 seconds after the turn. It does not put the host to sleep or exercise mixed-mode Electron orchestration.
