API Error 400 "tool use concurrency issues" caused by orphan tool_result blocks after context compaction

Description

After a long, tool-heavy session, Claude Code starts returning API Error: 400 due to tool use concurrency issues on every new message. The session becomes completely unrecoverable — even /rewind doesn't fix it.

Root Cause (from JSONL analysis)

The conversation's JSONL file contains orphan tool_result blocks — user messages with tool_result entries that reference tool_use IDs which no longer exist in any preceding assistant message. The Anthropic API correctly rejects this as invalid input (every tool_result must have a matching tool_use in a prior assistant message).

11 orphan tool_result references were found scattered across the conversation history, while all 32 resolved tool_use/tool_result pairs were properly matched. This means the assistant messages containing those 11 tool_use blocks were removed at some point, but the corresponding user messages with tool_result blocks were not.

How It Happened

1. Long session with heavy tool usage (~32+ tool calls including Bash, Read, Grep, Glob, Agent, ToolSearch, and MCP tools)
2. At some point, context management (likely auto-compaction) removed assistant messages containing tool_use blocks
3. The paired user messages containing tool_result blocks for those same IDs were not removed
4. This created an invalid message history where tool_result blocks reference non-existent tool_use IDs
5. On the next API call that included these orphaned messages, the API returned 400
6. Every subsequent message also fails because the corrupted history is still sent

Reproduction Context

• Claude Code version: 2.1.86
• Model: claude-opus-4-6 (also used claude-sonnet-4-6 during the session)
• Platform: macOS (Darwin 25.4.0)
• Session characteristics:
• ~130 JSONL entries
• 32+ tool calls (Bash, Read, Grep, Glob, Agent subagents, ToolSearch, MCP tools)
• Background agents were used (task-notifications present)
• Session was exported and resumed at one point (but export is read-only, shouldn't affect state)
• /compact was not manually triggered — compaction happened automatically

Evidence from JSONL

The /context command output showed:

claude-opus-4-6 · 16k/200k tokens (8%)
Compact buffer: 3k tokens (1.5%)

The low 8% usage with a compact buffer present confirms auto-compaction had already occurred. The queue-operation entries (enqueue/dequeue) appear immediately before the first error, suggesting possible timing issues.

The error message in the JSONL:

{
  "type": "assistant",
  "message": {
    "role": "assistant",
    "content": [{"type": "text", "text": "API Error: 400 due to tool use concurrency issues."}]
  },
  "error": "invalid_request",
  "isApiErrorMessage": true
}

Expected Behavior

When context compaction removes an assistant message containing tool_use blocks, it should also remove the corresponding user message(s) containing the paired tool_result blocks — maintaining valid tool_use/tool_result pairing at all times.

Workaround

Starting a new session (/clear) is the only workaround. /rewind does not help because the orphan tool_result blocks are deep in the conversation history, not in the last turn.