Auto mode blocked intermittently for ~20h: safety classifier (claude-opus-4-8[1m]) unavailable, rejecting Bash/PowerShell/Workflow calls

Summary

In auto permission mode, state-changing tool calls (Bash, PowerShell, Workflow, ScheduleWakeup, Agent) are intermittently blocked for extended stretches because the safety-classifier model is unavailable. The error persists across retries for minutes-to-hours at a time and recurred across a ~20-hour window, repeatedly stalling an autonomous long-running engineering session.

Environment

• Claude Code desktop app on Windows (win32), Windows 11 Pro 10.0.26200 (no claude CLI on PATH — desktop-app harness / Agent SDK session)
• Main model: claude-fable-5
• Permission mode: auto
• Session ID (for telemetry correlation): c3aa611a-f843-4e92-986a-10f46b25a7bd

Exact error (verbatim)

claude-opus-4-8[1m] is temporarily unavailable, so auto mode cannot determine the safety of Bash right now. Wait briefly and then try this action again. If it keeps failing, continue with other tasks that don't require this action and come back to it later. Note: reading files, searching code, and other read-only operations do not require the classifier and can still be used.

The same message appears with PowerShell, Workflow, and ScheduleWakeup in place of Bash.

Frequency / evidence (from the session transcript JSONL)

• 30 distinct transcript records containing the error, spanning 2026-07-16T20:15:29Z → 2026-07-17T16:39:09Z (~20 hours, intermittent clusters rather than one outage).
• Pattern-match breakdown by named tool (some records contain multiple matches): Bash 32, PowerShell 16, Workflow 9, ScheduleWakeup 2.
• During outage windows, back-to-back retries fail identically (e.g., two consecutive Workflow launches seconds apart both rejected), then the same call succeeds later unchanged — consistent with upstream classifier-endpoint availability/capacity rather than anything about the specific commands.

Impact

• The agent cannot execute any state-changing work (shell, subagent workflows, scheduled wakeups) during the windows; only read-only tools function.
• In long autonomous sessions this stalls multi-phase pipelines mid-flight; combined with app restarts it caused background agents to need manual resume.

What would help

1. Failover/secondary classifier model (or automatic retry with backoff inside the harness) so a single classifier model's availability doesn't hard-block auto mode.
2. Surfacing outage state distinctly from per-command denials, so agents/users can distinguish "classifier down" from "action denied."
3. Optionally: queue the blocked tool call and auto-execute when the classifier recovers, instead of forcing manual retry.

🤖 Generated with Claude Code