Repository navigation
v1.6.1
Immutable
release. Only release title and notes can be modified.
v1.6.1
Fix MCP tool requests wrapped in <escape>...</escape> becoming literal final answers on follow-ups. Valid wrapped requests now execute through the existing tool pipeline, while malformed requests reach bounded protocol repair. This closes a prefix-recognition gap also present in v1.5.1; it is not identified as a new v1.6.0 regression.
Fixes
- Buffer partial escape tags and tool prefixes across SSE snapshots. Keep wrapped requests out of final-answer text, streaming deltas and the visible transcript while the tool exchange completes.
- Require the current nonce, matching tool markers and a closing escape tag before validating the complete JSON batch. Ordinary escaped text and commands quoted in prose or explicitly marked final answers retain their original text.
- Apply the existing hidden correction budget to malformed JSON, incomplete wrappers and stale wrapped markers. Strict mode still rejects invalid requests without executing tools; repeated failures stop at the configured limit.
- Preserve application approval, progress events and durable outcome records. Confirmed calls execute once; uncertain calls remain pending and block continuation until explicitly resolved.
Validation and downloads
- Add 83 regression cases covering fresh chats, follow-ups, mixed and resubmitted history, stream boundaries, three-call batches, approval and uncertain outcomes. Typed Chat Completions and Responses clients are checked with both streaming and collected results.
- Require at least 1,444 core/drop-in SDK cases plus 18 native SDK, 167 automation and 12 extension cases. Four platform test jobs and all bundled-browser checks remain required before publication.
- Update SDK packages, browser extensions, installation instructions and download links to 1.6.1. The OpenAI compatibility assembly retains version 2.14.0.0.
See the verification record for reproduction and test details. All commits are linked in the collapsed appendix below.