Recurring HTTP 529 Overloaded + Rate Limit errors disrupting active sessions (45 errors in 50 sessions)

Summary

Persistent API errors breaking active Claude Code sessions. Analyzed last 50 sessions from ~/.claude/projects/ — 45 real API errors across 24 affected sessions.

Error breakdown

| Error type | Count |
|---|---|
| HTTP 529 Overloaded (API Error: 529 Overloaded) | 27 |
| Rate Limit (429) | 18 |
| Total | 45 |

Worst affected sessions

| Session ID | Errors | Date/Time |
|---|---|---|
| d608053f | 16 | 2026-06-22 01:06–01:25 UTC |
| 580b36d3 | 5 | 2026-06-23 15:23–15:26 UTC |
| 7355f6a3 | 4 | 2026-06-22 |
| 04fb2365 | 4 | 2026-06-22 |
| 6c1743bf | 4 | 2026-06-22 |

Verbatim error messages (from JSONL transcripts)

API Error: 529 Overloaded. This is a server-side issue, usually temporary — try again in a moment. If it persists, check https://status.claude.com.

These appear in tool_result blocks — the harness receives the error from the API and surfaces it mid-task, breaking workflow, losing context, forcing session restarts.

Worst burst example

Session d608053f, 2026-06-22:

01:06:01 — API Error: 529 Overloaded
01:06:08 — API Error: 529 Overloaded
01:06:41 — API Error: 529 Overloaded
01:08:39 — API Error: 529 Overloaded
01:24:52 — API Error: 529 Overloaded
01:25:20 — API Error: 529 Overloaded

6 errors in 19 minutes. Session destroyed mid-task.

Environment

• Model: claude-opus-4-8 (1M context variant, model: opus[1m])
• Effort: high (configured in settings.json)
• Platform: macOS Darwin 25.5.0 (Apple Silicon)
• Claude Code: latest CLI

Impact

• Loss of work mid-session
• Forced context rebuild on every error (expensive with 1M context Opus)
• 24 out of the last 50 sessions affected — this is not occasional

Ask

1. Investigate why 529 Overloaded is this frequent for Opus 4.8 [1m] + effort:high users
2. Retry at harness layer — exponential-backoff on transient 529s so they don't surface as hard failures
3. Status bar visibility when API is degraded — users shouldn't think it's their local setup
4. Auto-fallback to Sonnet on repeated 529s rather than killing the session

Notes

• 100% server-side errors (Anthropic infrastructure), not user configuration
• Rate limit errors (18) may be related to effort:high + opus[1m] consuming tokens faster than quota refresh cycle
• status.claude.com did not always reflect the degradation during these incidents