Repository navigation
Immutable
release. Only release title and notes can be modified.
This release changes how enforce mode handles an unavailable Watcher client and gives reviewers access to flagged-call evidence without opening the full transcript.
Breaking changes
- Changed: In enforce mode, calls that need grading now wait when the Watcher client is unavailable, then are denied if it does not answer. The wait is about 10 seconds, or up to two minutes while the client starts or updates. The grading-failure policy does not override this denial. Keep the client running in watched containers and on developer machines. To pause reviews, use snooze, an allowed
.watcher-ignorefile, or an administrator's enforcement-mode change. - Changed: The built-in org-viewer role and admins under the organization content restriction can now read flagged calls' inputs and outputs. Routine calls and the surrounding transcript keep their existing restrictions. Review org-viewer assignments before rollout. Self-hosted deployments can retain findings-only access by removing
tool-calls:content:read:flaggedfrom the role's grants. Custom roles and API keys need that grant to use the new access. - Changed: User access now comes from role permissions in a shared YAML policy. WorkOS user-token grants no longer grant Watcher access, and missing or unknown proxy roles grant nothing. Self-hosters must send a known proxy role and include every intended grant in each role override. Invalid role YAML prevents startup. Clients reading managed settings need
org:managed-settings:readororg:settings:read. API keys retain explicit scopes, but cannot grant reserved internal capabilities. Deploy the prepared configuration with the new image and refresh credentials after changing role assignments or WorkOS widget grants.
Live
- Fixed: Transcript uploads refresh expired credentials and retry once after an authentication rejection. Pending recordings remain bound to their original server, account and organization.
- Fixed: Claude Code calls left at an approval prompt when the process is killed retain their escalation decision for delivery to the server.
Analyzer
- Added: The Flagged tool calls view shows calls from the selected trajectory and its ancestors. It preserves the selected call when switching to the full transcript, refreshing, or using Back and Forward. Links outside the selected trajectory open a separate linked-call card.
API
- Fixed: Long transcript uploads, decision recording and batch grading use fewer database queries.
Other
- Added: A Custom instructions guide explains how to write monitor instructions, what blocking monitors can see, and when to use command rules.