openclaw - 💡(How to fix) Fix Vertex beta INVALID_ARGUMENT can wedge long Enterprise sessions without actionable recovery

Summary

OpenClaw Enterprise can leave a long-running session stuck or repeatedly failing when the active LiteLLM Vertex beta route returns a generic 400 INVALID_ARGUMENT response after a large conversation/context buildup.

Observed provider error:

litellm.BadRequestError: Vertex_ai_betaException BadRequestError - b'{
  "error": {
    "code": 400,
    "message": "Request contains an invalid argument.",
    "status": "INVALID_ARGUMENT"
  }
}'

User-facing error:

Something went wrong while processing your request. Please try again, or use /new to start a fresh session.

After this happens, the session can appear dead: messages may receive no response, briefly recover, then fail again with the same generic error.

Why this is not a duplicate

This is related to existing session/context recovery work, but it is not fully covered by the current open issues/PRs:

• Related: #84536, #84602 - preemptive context overflow can silently kill embedded sessions.
• Related: #82754 - context overflow should fall back to larger-context models.
• Related: #76806 - irreducible overflow should not loop forever.
• Related: #74822 - recurring generic "Something went wrong" gateway failures.
• Related historical PR: #14314 - Gemini INVALID_ARGUMENT recovery attempt, closed without landing.
• Related merged PR: #58856 - provider error normalization, but it does not appear to cover this Vertex beta failure mode.

The observed failure is specifically a provider-side Vertex beta/LiteLLM INVALID_ARGUMENT response that is surfaced as a generic failure and does not leave the Enterprise session in a clearly recoverable state.

Expected behavior

OpenClaw should classify this as a recoverable provider/session request-shape error instead of leaving the session wedged:

• Detect Vertex/LiteLLM INVALID_ARGUMENT responses from the beta route.
• Preserve the useful provider details in structured logs with sensitive fields redacted.
• Surface a user-actionable message instead of only the generic /new fallback.
• Mark the failed turn/session lane as cleanly failed so later messages are not stuck behind the bad turn.
• Where possible, attempt a safer recovery path: compaction, smaller request shape, non-beta route, or configured fallback model.

Actual behavior

• The provider returns INVALID_ARGUMENT.
• The UI shows only a generic "Something went wrong" message.
• The user cannot tell whether this is a context limit, provider request-shape bug, model routing issue, or local gateway problem.
• The same session can remain unreliable until a fresh session is started.

Notes

This was observed during normal OpenClaw Enterprise use in a long session suspected to be near a large context window. It should be handled separately from Feishu upload permissions and duplicate-message behavior; those are channel/config symptoms, while this issue is at the model provider and session recovery boundary.