Conversation becomes unrecoverable after API 500/529 errors during parallel agent execution

Description

A Claude Code conversation became completely unrecoverable (cannot resume, restart, or continue) after hitting multiple Anthropic API 500/529 errors while parallel background agents were returning results.

Steps to reproduce

1. Launch multiple background agents in parallel (3 reviewer agents + 2 builder agents)
2. While agents are completing and returning results, Anthropic API starts returning 500 (Internal server error) and 529 (Overloaded) errors
3. User messages sent during this period all fail with API errors
4. After the API stabilizes, the conversation cannot be resumed - it's effectively dead

Expected behavior

The conversation should be recoverable after transient API errors resolve. At minimum, it should be possible to continue the thread even if some agent results were lost.

Actual behavior

The conversation is permanently corrupted and cannot be restarted. This has never happened before despite heavy usage across many sessions with large context windows.

Environment

• Model: claude-opus-4-6 (1M context)
• Platform: macOS (Darwin 25.3.0)
• Date: 2026-03-18 evening (CET)
• API errors observed: req_011CZ9KLnYtD3o79FuwmdJwu (500), req_011CZ9KPKcyQFmTR2aftzf8V (500), req_011CZ9KRvJqxmjcL7s8HDha4 (500), req_011CZA69WfqebGArbws7Kxht (529)

Hypothesis

When multiple background agents complete simultaneously during API instability, the context may end up with malformed tool call results (incomplete JSON from interrupted responses). This makes the compressed context unparseable, permanently breaking the conversation.

Impact

• No work was lost (commits were pushed before corruption)
• But the conversation history and context is gone
• User has never experienced this despite extensive usage

Additional context

The session had heavy parallel agent usage: 3 reviewer subagents (data, UI, security) followed by 2 builder subagents, all running concurrently. The corruption happened when agents were completing while user messages were hitting API errors.