Auto mode: intermittent 'classifier temporarily unavailable' errors across tool types (weeks-long, not firewall-related)

Bug report: Auto mode — "classifier temporarily unavailable" recurring for weeks

Summary

Auto mode intermittently fails with:

claude-sonnet-5 is temporarily unavailable, so auto mode cannot determine the safety of <Tool> right now.

This has been happening persistently for several weeks, is not tied to any single tool (seen on both Bash and WebFetch in the same session, minutes apart), and is not resolved by retrying immediately — though it sometimes clears after a short wait.

Environment

• macOS 26.5.2 (Build 25F84)
• Claude Code CLI 2.1.227
• Terminal: iTerm2
• Little Snitch v6 (outbound firewall) — ruled out as the cause, see below

What I ruled out

I initially suspected Little Snitch (an outbound connection firewall) was blocking the classifier's network calls. Investigation showed:

• No Little Snitch alert ever fires when the failure occurs (checked in Alert Mode).
• A pre-existing, broad "Allow any outgoing connection" rule scoped to the iTerm process (created May 2026, 21,000+ uses) covers all traffic launched from the terminal, including Claude Code's CLI subprocess — so Little Snitch has not been blocking this traffic.
• Disabling Little Snitch entirely, and separately adding a dedicated Code-ID-based allow rule for the com.anthropic.claude-code binary (Team ID Q6L2SF6YDW), made no measurable difference — failures still occurred with the firewall fully open.
• Failures reproduced identically with the firewall disabled, in silent-allow mode, and in alert mode.

This strongly suggests the cause is not local network/firewall configuration.

Reproduction / observed pattern

In a single test session (2026-08-11, ~16:15–16:40 BST):

| Time (BST) | Tool | Result |
|---|---|---|
| 16:15:59 | Bash | ✅ pass |
| 16:19:xx | Bash | ✅ pass |
| 16:20:35 | Bash | ✅ pass |
| 16:24:44 | Bash | ✅ pass |
| 16:30:49 | Bash | ✅ pass |
| 16:37:15–16 | Bash x3 | ✅ pass |
| 16:37:2x | WebFetch (api.anthropic.com) | ❌ fail |
| immediately after | WebFetch (retry, same URL) | ❌ fail |
| 16:37:47 | Bash | ✅ pass |
| same batch | WebFetch (docs.anthropic.com) | ❌ fail |
| ~16:40 | Bash x3 (consecutive) | ❌ fail all three |

Pattern: failures cluster in bursts across arbitrary tool types (not one specific tool), interspersed with runs of clean successes, with no apparent local trigger (same machine, network, and Little Snitch state throughout).

Impact

Auto mode becomes unusable during these windows — actions requiring the classifier (Bash, WebFetch, Edit, etc.) are blocked, forcing a fallback to manual approval mode. This has been ongoing for weeks with no clear pattern I can find on my end (time of day, network, etc. all checked).

Ask

Given local causes (firewall, network config) have been thoroughly ruled out through direct testing, requesting investigation into whether the classifier service itself is experiencing intermittent availability issues, and whether there's a way to get more diagnostic detail (e.g. a request ID) surfaced in the error message to help correlate with server-side logs.