v2.7

## Release Notes

- **Response tool routing fix** — Native Responses output text is now routed through the normalized response-tool executor instead of ending the monologue directly, preserving active goal overrides, proper state cleanup, and rejecting empty response payloads.

- **Streaming resilience** — Partial tool arguments no longer raise errors during streaming; both text and legacy message response fields are rendered, and live response logs stay synchronized with incomplete snapshots.

- **Prompt JSON fence handling** — Fenced JSON examples in prompt Markdown are now stripped of fence markers at render time so the assembled system prompt matches the required wire format. Non-JSON fences are preserved, and all bundled profiles are covered.

- **Block final responses from parallel calls** — Response tool calls are now kept at the top level to correctly end the parent message loop, preventing unnecessary follow-up turns when issued inside parallel call groups.

- **Add Cerebras model provider** — Cerebras is available as a native LiteLLM provider with Chat Completions support, onboarding API-key guidance, a production default model, and a bundled logo.

- **Browser proxy support** — The internal Docker browser now supports project-scoped HTTP and SOCKS proxy settings, including bypass rules and optional authentication. Proxy changes are passed through Playwright's native proxy option and trigger a browser runtime restart. Settings are grouped in a collapsible accordion in the UI.

- **Harden remote Linux computer use** — Agents must verify window focus before target-bound typing, report computer-use receipts from observed evidence, use scoped snapshots for focus verification, and avoid repeating identical failed actions.

- **Fix Codex OAuth reasoning effort forwarding** — The legacy `reasoning_effort` field is now normalized at the Codex Responses proxy boundary so it cannot reach upstream endpoints in an unsupported shape, while explicit native reasoning objects are preserved.