Gemini CLI 0.55.1

• [fix] Capacity exhaustion is treated as a terminal condition so the client stops retrying and no longer hangs.
• [fix] Empty or invalid model streams surface clearer InvalidStreamError detail in the UI, with more specific guidance when the response is empty.
• [fix] A new user message no longer fuses into an unanswered tool response, which could previously corrupt turn structure.
• [fix] /compress session reload and quota-fallback paths no longer drop tool responses.
• [fix] When thought parts are stripped from history, functionCall thoughtSignature values are preserved so follow-up tool use stays valid.
• [fix] Nested gaxios streaming errors are unwrapped and parsed from the cause message for more accurate failure reporting.
• [fix] On macOS, the CLI falls back to embedded Seatbelt profiles if the expected profile files are missing.
• [fix] Session IDs rotate on model fallback to avoid stateful API errors after a switch.
• [fix] Merged function-response turns are skipped when detecting the active loop, improving loop tracking.
• [fix] Thought parts are filtered out of history retrieval when context management is disabled.
• [fix] Google credentials auth enforces HTTPS to prevent cleartext credential leakage; file keychain tags use explicit length checks and validation.
• [fix] A2A proposed content normalizes CRLF line endings to LF.
• [new] Contributor documentation for behavioral evaluations and the Eval Development Kit (EDK), including inventory, validate, and report workflows.
• [new] npm run eval:report aggregates local eval report.json artifacts and summarizes pass rates per model.
• [changed] Release tooling tolerates npm registries that forbid dist-tag deletion instead of failing the release flow.

Models & prompts

No new model IDs or default model changes in this release. History handling is stricter around thought parts and tool-call signatures so multi-turn tool use remains consistent when context management or thought stripping is in play.

Under the hood

• Recognition of Google Cloud Workstations via GOOGLE_CLOUD_WORKSTATIONS for environment-aware behavior.
• Continued work on the Antigravity / PR-generator stack (agent runner and prompts, Firestore dual-locking, orchestrator state machine, and Cloud Run / Workflows packaging) used for automated PR workflows.
• Caretaker triage improvements: comments before auto-closing issues, and issue titles sanitized under untrusted context.
• Sandbox image tags aligned to 0.55.1.

Upgrade notes

• macOS users relying on custom or missing Seatbelt profile files should see more reliable sandbox startup via the embedded profile fallback; no action required in the common case.
• Contributors writing agent evals can follow the new Behavioral evaluations guide and use npm run eval:report alongside the existing inventory and validate scripts.