[BUG] Safety classifier intermittently unavailable, blocking tool execution in long sessions

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Environment: Claude Code desktop, Windows 11, v2.1.209. Models: claude-opus-4-8 and claude-sonnet-5.

Symptom: Before executing a tool (Bash/PowerShell/Write), the pre-execution safety classifier repeatedly returns "<model> is temporarily unavailable, so auto mode cannot determine the safety of <Tool> right now" and the tool call is blocked.

Observations:

• Occurs on both opus-4-8 and sonnet-5, so it appears to be the classifier service itself, not a specific model.
• Read-only tools are unaffected; only execution/write tools are blocked.
• Intermittent — retrying usually succeeds but recurs frequently.
• Markedly more frequent in long sessions with hundreds of tool calls over several hours.
• No data loss or wrong execution, but causes delay and repeated retries that re-process a large context and waste tokens.

Questions:

1. What causes the classifier to become intermittently unavailable?
2. Is there a fix or ETA?
3. Any recommended mitigation for long, tool-heavy sessions (e.g., correlation with session length / context size)?

What Should Happen?

The pre-execution safety classifier should return a decision so that execution/write tools (Bash, PowerShell, Write) run without being blocked.

Error Messages/Logs

claude-opus-4-8 is temporarily unavailable, so auto mode cannot determine the safety of Bash right now. Wait briefly and then try this action again.
(same for claude-sonnet-5)

Steps to Reproduce

1. Run a long Claude Code session with many tool calls (hundreds of Bash/PowerShell/Write invocations over several hours).
2. Keep issuing tool-executing commands.
3. Intermittently, before a tool runs, the classifier returns "<model> is temporarily unavailable, so auto mode cannot determine the safety of <Tool>".
4. The tool is blocked; retrying usually succeeds but the failure recurs frequently.

Note: read-only tools (file read/search) are never affected — only execution/write tools.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.209

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

Session had hundreds of tool calls over ~6 hours; both opus-4-8 and sonnet-5 were affected. Notably, the gh CLI call to file this very issue was itself repeatedly blocked by the same classifier.