Recurring 529 Overloaded bursts (322 events, request IDs included) + session-terminating malformed tool-call output corruption (23+ forced restarts)

Summary

Two recurring reliability problems on Claude Code (Windows / VSCode extension environment), verified from local session transcripts (~/.claude/projects/**/*.jsonl):

1. Bursts of API Error: 529 Overloaded — 322 logged error events across 13+ days (2026-06-08 → 2026-07-09), repeatedly blocking work mid-task. The latest burst today (2026-07-09, 03:22–03:48 UTC) produced 31 events and interrupted an active work session.
2. Session-terminating output corruption ("malformed" tool calls) — the assistant's output stream corrupts (e.g., raw tool-call XML fragments like antml:parameter name="command" leaking into plain text, or Edit tool calls arriving malformed), after which the session becomes unusable and must be restarted. We counted 23+ distinct sessions whose first user prompt documents a forced restart after such a crash (we keep hand-over notes per crash in our internal knowledge base, so each occurrence is individually documented).

Environment

• Claude Code v2.1.205 (crashes span multiple versions from early June 2026 to now)
• Windows 11 Pro (10.0.26200), VSCode native extension environment
• Model: claude-fable-5 (also observed on opus-class models in earlier sessions)

Issue 1 — 529 Overloaded bursts

Message shown: 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.

Daily distribution of logged 529/overloaded events (from session transcripts):

| Date | Events |
|---|---|
| 2026-06-08 | 14 |
| 2026-06-11 | 12 |
| 2026-06-15 | 19 |
| 2026-06-17 | 61 |
| 2026-06-22 | 15 |
| 2026-06-26 | 20 |
| 2026-06-29 | 21 |
| 2026-07-01 | 20 |
| 2026-07-02 | 56 |
| 2026-07-09 (today) | 31 |

Sample request IDs from today's burst (2026-07-09 03:22–03:48 UTC; 26 distinct, full list of 259 distinct request IDs across all days available on request):

req_011Ccqkp99rFq6F5yn7UMQ5z
req_011CcqkpJoH17zwJCNUZ4TRg
req_011CcqkpUPjV9kbcmTGGxSTw
req_011CcqkpgYUxp2nfGhjbPwmW
req_011CcqkpzwX562mx6RYqUXz1
req_011CcqkqVR2VDLEA6q39pqw9
req_011CcqkrN346C7JYBrAtNyQG
req_011CcqkstdLfujsYffTjaU7N
req_011CcqkviB82pyn8zAaiyWai
req_011CcqkyPfj58vscr7sBxhas
req_011Ccqm2R8cVDo5BLyEQbaxc
req_011CcqmLT3xPvAXYEjFbewhS
req_011CcqmLbFa9cCFTHcaWvWNS
req_011CcqmLnjdXaAKWjmjDghbf
req_011CcqmLzkSkMKMtK83EawsB
req_011CcqmMHdwb7nZTPRoZFV4F
req_011CcqmMp8kZwGh5m8fNM8zv
req_011CcqmNiYvTiCCZWnfK3etf
req_011CcqmQK1sZfiDCdSBsuNte
req_011CcqmSyGK6zvfsqTR7QgSS
req_011CcqmVuKmD98gkuanZZ6y9
req_011CcqmYf9ozB1PXHyRBjJgJ
req_011Ccqmp5UCzebqRD8iy1sj5
req_011CcqoF4Ccz1Kz9sZFh1Dzq
req_011CcqoKJJ4KRQqtETMYJFox
req_011CcqoMN1rcRxAKnFH6e4tM

Issue 2 — session-terminating output corruption ("Claude bug" restarts)

Symptoms observed repeatedly since early June:

• Edit/Write tool calls arrive malformed and fail; retrying degrades until the session stops making progress.
• Raw tool-call markup leaks into the visible text stream (e.g., a user-captured fragment: 클로드 버그: call antml:parameter name="command" ... — i.e., antml-namespaced XML printed as plain text).
• After the corruption starts, the session is effectively dead; the user must restart Claude Code and re-establish context.

Impact evidence: 23+ distinct session transcripts begin with the user reporting a restart after a crash (Korean prompts such as "클로드버그로 재시작함" = "restarted due to Claude bug"), each requiring a written hand-over note and context re-establishment. Examples of first-prompt timestamps/sessions available (session UUIDs + our internal hand-over record IDs are retained and can be provided).

Each crash costs roughly 15–40 minutes of re-work (hand-over note authoring, restart, context recovery), in addition to the token cost of re-establishing context in the new session.

Expected behavior

1. 529 bursts should be rare and brief; when they occur mid-agentic-loop, Claude Code should retry/resume gracefully without losing the in-flight turn.
2. Malformed tool-call output should never corrupt an entire session; a single bad generation should be retried transparently rather than requiring a manual restart.

Ask

• Investigation of the recurring 529 bursts (request IDs above) and the output-corruption crash pattern.
• Guidance on claiming usage credit for the interrupted usage windows — we are also filing a support request with the request-ID list; please advise if additional data (full 259 request IDs, session UUIDs) is needed.