auto-compression summarizer failure aborts the entire user turn instead of degrading

Finding: In google-gemini/gemini-cli, per an external issue report, automatic chat compression can abort the user's whole turn when the auxiliary summarize or verify model call throws after retries. The reported environment is main at commit 5411f113c; a later commenter reports reproducing it on main at fb972b2f. No released version is identified. Source: https://github.com/google-gemini/gemini-cli/issues/29025 Reported mechanism (source audit, a hypothesis not confirmed by a maintainer): compression runs two calls through BaseLlmClient.generateContent, which uses retryWithBackoff and throws when attempts are exhausted. The issue says compress(), tryCompressChat() and the processTurn() call site have no try/catch. A transient failure of the utility model, such as a 5xx or timeout, would therefore fail the prompt even if the main model is healthy. The issue notes that existing statuses such as COMPRESSION_FAILED_EMPTY_SUMMARY and COMPRESSION_FAILED_INFLATED_TOKEN_COUNT already express graceful degradation for non-exception failures. (https://github.com/google-gemini/gemini-cli/issues/29025) Reproduction reports: the issue gives steps. Use a session large enough to trigger auto-compression, make generateContent reject with a 5xx-shaped error so retries run out, then send a message. The reporter expects compression to be skipped and the conversation to continue. A commenter on 2026-10-04 reports that a 503 from either the summary or verification call escapes compress() and prevents the main turn from proceeding. (https://github.com/google-gemini/gemini-cli/issues/29025#issuecomment-5982576379) Status: the issue is open. The supplied content has no maintainer response and no confirmed cause. The commenter describes a local candidate that is automatic-only. It returns an API-error status and preserves history, remembers the failure so later attempts use the existing truncation fallback, and leaves cancellation and manual /compress behavior unchanged. The commenter reports focused tests passing. They also report two existing Windows locale failures, no full preflight, no live-model run and no PR. This is a proposed change only, not a confirmed fix or release. Limitations: this record relies on external reports. KFA did not reproduce the issue and has not verified the candidate patch. No supported user-side workaround is given, so no actionable repair is offered. Anyone affected should track the issue for maintainer confirmation and a merged, released fix.