v0.15.1
·
4 commits
to main
since this release
This release focuses on making tool-call permission decisions reach the Analyzer reliably, particularly for Codex sessions using Code Mode, and adds a safety limit to the grading API.
Live
- Added: The Watcher client now sends a
watcher-live/<version>identifier on every request to the server; self-hosted deployments can still override it with their own User-Agent. - Changed: Watcher client logs are now structured instead of freeform text, which may affect anyone parsing the client's log files directly.
- Fixed: Permission decisions are now retried until they successfully reach the server, so a decision made after a session's last upload, or lost to a transient failure, no longer gets stuck on the machine and missing from the Analyzer.
- Fixed: Codex sessions using Code Mode now correctly report tool-call permission decisions to the Analyzer; previously, most decisions made inside Code Mode never reached the Analyzer at all.
- Fixed: Codex Code Mode permission decisions are now recorded accurately: a tool call denied inside a Code Mode program now shows as denied rather than unknown, and a program that aborts no longer causes unrelated later tool calls to inherit a denial they never received.
API
- Changed: On Apollo's hosted API, a single grading request that would produce more than 10 grades is now rejected with an error instead of processing an unbounded number of grades; split large batches into smaller requests.