Minor release (0.57.0): /clear now resets the agent conversation (not just the screen), clearer Edit-tool guidance, Cloud Workstations OAuth support, and updated privacy opt-in copy.
Changelog
[changed]/clear starts a new session. Running /clear clears the active agent conversation context and begins a fresh session, in addition to clearing the visible terminal screen and scrollback. Ctrl+L remains a visual-only clear: it redraws the UI but keeps the current conversation. See the commands reference.
[changed]Edit tool guidance. When an edit is missing a usable old_string, the model is told that parameter is required, that Edit does localized search-and-replace, and to call ReadFile on the target range if it has not viewed that section recently so the match is exact.
[new]Cloud Workstations-aware OAuth. Support for the GOOGLE_CLOUD_WORKSTATIONS environment so OAuth redirect handling works correctly when running on Google Cloud Workstations (dynamic proxy redirect URI resolution).
[changed]Product-improvement data permission copy. The privacy prompt now states clearly how to opt out (“No, deny permission…”) or opt in (“Yes, grant permission…”) for using data to improve Google’s products and services.
[changed]Vertex AI auth docs.GOOGLE_CLOUD_LOCATION documentation now links to the official Vertex AI locations guide for choosing a region.
Models & prompts
Edit-tool instructions emphasize exact old_string matches and reading the file again before patching unfamiliar sections, reducing failed or imprecise edits.
No new public model IDs or default model changes in this release.
Prefer /clear when you want a true new conversation; Ctrl+L alone will not reset agent context.
On Cloud Workstations, ensure the environment exposes GOOGLE_CLOUD_WORKSTATIONS as expected so browser OAuth redirects resolve through the workstation proxy.