HTTP 500 cascade during long-running autonomous session (output generation crash, not context overflow)

Summary

Claude Opus 4.6 via Claude Code hit 4 consecutive HTTP 500 Internal Server Errors during a coding session. The crash occurred during output generation, not input processing — the preceding tool call returned successfully.

Request IDs

• req_011CYFmpPc7ChW182KHj5kwo
• req_011CYFmvAWybKgz7R6z7rQCT
• req_011CYFmzPUhiKwSkwDTiSz8G
• req_011CYFn6Cb2JrE9kcHFsWiuP

Reproduction pattern

• Task: Iterative code review and security hardening of a Python file (~3,800 lines) — the model was reading code, identifying issues, and writing fixes
• Last successful tool call: shell running sed -n '2615,2640p' on the file — output returned normally
• Crash point: Generating the model's next response after receiving the tool output

Why this is not routine context overflow

1. Context length exceeded errors return as 4xx with a clear message. These were raw 500 Internal Server Error with no explanatory payload.
2. The tool call succeeded — input processing (reading the context, executing the tool) completed normally. The failure was specifically in output token generation.
3. All 4 retries failed against the same context, suggesting a persistent broken state rather than a transient infrastructure blip.
4. The shared request ID prefix (req_011CYFm*) suggests rapid sequential retries hitting the same failure mode.

Environment

• Claude Code CLI (latest as of 2026-02-18)
• Model: Claude Opus 4.6 (claude-opus-4-6)
• macOS Darwin 25.2.0, Apple Silicon
• MCP tools active (tool calls via MCP server)

What we're NOT reporting

This is not a complaint about rate limiting, context limits, or expected error handling. Those work fine. This is specifically about unhandled 500s during output generation triggered by a legitimate usage pattern. Filing because your engineering team can trace the request IDs to see exactly what failed internally.