Three more pull requests merge into pi ahead of the next release

Three pull requests merged into the pi coding agent on September 2, 2026, a smaller follow-up to the nine-PR batch from September 1. Like that batch, all three sit on main and no published release includes them yet.

The feature is pull request #9004 from AppleDannyClegg: an opt-in vllmPriority compat flag for OpenAI-compatible models, injecting vLLM’s top-level priority request field (vLLM 0.28 and newer; a lower value means earlier handling, default 0). The flag is off by default and absent from the generated model catalog — custom providers set compat.vllmPriority in models.json. The use case is a local vLLM box serving interactive sessions and background runs under --scheduling-policy priority: setting a numerically higher priority on the background model defers its long prefills behind interactive sessions instead of stalling an answer mid-stream while a large background prefill runs.

Pull request #8997 from hobostay fixes an agent hang (issue #8996): streamProxy could leave consumers waiting forever when the proxy server closed the SSE connection without a terminal done or error event. Two defects compounded — the residual line buffer was never processed after the read loop, so a final data: line without a trailing newline was silently dropped even when it carried the done event, and stream.end() without a result never resolved EventStream.result(), leaving the agent loop awaiting response.result() indefinitely. The fix flushes the decoder and residual buffer at EOF and synthesizes an error event, “Connection closed by proxy server before the response completed”, when the stream ends without a terminal event. Regression tests cover both failure modes.

The third is documentation: pull request #9009 from rwachtler updates the extension API docs to spell out that pi.setModel() and pi.setThinkingLevel() are session-scoped — since pull request #8356 neither persists its change to settings.json, the documentation gap issue #8976 reported.

All three live on main only; the next pi release after v0.84.4 picks them up.