[Bug] Anthropic API 529 Overloaded causes subagent termination without retry; permission classifier outage blocks all tool execution

Status Open
Reported on v2.1.221
Maintainer reply None cached
Activity 0 comments · opened Aug 5, 2026

Bug Description
[529 Overloaded + permission-classifier outage — cascading degradation, 2026-08-05 ~17:30-19:30 CEST]

Session: claude.ai/code/session_01Ace5qeyKVVFWFbPTnRCLaS (main model: claude-fable-5)

Two distinct failure modes hit simultaneously during an Anthropic API overload window:

1) Background subagents killed mid-task by 529.
A read-only verification subagent (sonnet profile) was terminated early TWICE with
"API Error: 529 Overloaded", losing its in-progress verification both times. There
appears to be no automatic retry/backoff at the harness level for subagents on 529 —
the agent just dies and reports partial results. Expected: transparent retry with
backoff on 5xx, or at minimum a resumable checkpoint.

2) Permission classifier outage hard-blocked ALL tool execution in auto mode.
Every Bash, Monitor, and MCP tool call failed for ~2 hours (flapping — occasional
calls got through) with: "claude-sonnet-5[1m] is temporarily unavailable, so auto
mode cannot determine the safety of <tool> right now."
This blocked even trivially safe commands (ls, cp of a local backup, a read-only
curl to our own CMS API). File tools (Read/Write/Edit/Grep) kept working, which we
used as a partial workaround.

Impact: an opus builder subagent running a video-render pipeline on a remote box was
left stranded with ~7 files existing only in the ephemeral session scratchpad
(recovered manually via Read+Write); a production remediation plan (blog quality
gate deploy) is still queued.

Suggestion: when the safety classifier is unavailable, degrade to interactive
user-approval prompts (or a cached allowlist for previously-approved command
patterns) instead of hard-failing every call. A prolonged classifier outage
currently converts "auto mode" into "no execution at all" even for a user sitting
right there who could approve manually.

Environment Info

  • Platform: darwin
  • Terminal: iTerm.app
  • Version: 2.1.221
  • Feedback ID: 8e63b87d-02a5-4d8a-9b82-314ca41af812

Errors

[]

View original on GitHub ↗