v0.3.17

Answers stream again in Claude Code and Codex against a --api claude / --api codex-cli seat.

Fixes

• cli-seat: stream text again when a request carries tools (cd172d5)
  The seat dropped every text delta whenever a request merely offered tools, and released the
  whole answer at the end. Claude Code and Codex always send their tool roster, so every turn
  arrived in one burst — the client showed nothing at all while the model wrote. On a long turn that
  is minutes of a frozen screen: measured on a live grid, the gap between the first chunk and the
  second was 344s on an opus turn, equal to the whole request.

  Text before the first { cannot be part of a tool call, so it now streams as it arrives; only the
  rest waits for the parse that decides which bytes were the call. A tool call still never reaches
  the client as prose. Measured on the same request with and without one tool attached: first text
  delta 16.7s → 2.9s. All three wires — /chat/completions, /messages and /responses — share
  one rule.

  This shortens time-to-first-token, not the turn itself.

Full Changelog: v0.3.16...v0.3.17