[Bug] Parallel Agent calls can lose tool_results

Summary

When Claude Code launches 3 Agent tool calls in a single assistant turn, some tool_result blocks can go missing even though the session itself remains usable.

In the UI this shows up as:

• [Tool result missing due to internal error]

In the debug log, the key signal is:

• ensureToolResultPairing: repaired missing tool_result blocks

In my case, one assistant message emitted 3 Agent call IDs, but the following user-side tool_results block only contained 1 of them. The other 2 results were effectively dropped and surfaced as internal errors in the UI.

This looks like a parallel Agent result aggregation / pairing bug in Claude Code, not a hook failure and not a prompt issue.

Reproduction

1. Use Claude Code CLI on macOS
2. In a single assistant turn, launch 3 Agent tool calls concurrently
3. Wait for all 3 to finish
4. Observe that one or more Agent results may show up as:

• [Tool result missing due to internal error]

1. Inspect the debug log and see:

• ensureToolResultPairing: repaired missing tool_result blocks

My concrete case was a /simplify run that launched 3 review agents in parallel.

Observed behavior

• Assistant turn contains 3 Agent tool uses
• The next user tool_result turn only contains 1 matching tool_result
• Missing results are shown as internal errors in the UI
• Session continues, but the affected Agent outputs are lost

Relevant debug evidence:

ensureToolResultPairing: repaired missing tool_result blocks

And the repaired message structure showed the mismatch explicitly:

assistant(... tool_uses=[call_uHD..., call_Tu..., call_VKK...]);
user(tool_results=[call_VKK...])

So 2 of the 3 tool results were missing from the paired result turn.

Expected behavior

One of these should happen reliably:

1. All Agent tool_result blocks are preserved and paired correctly, or
2. The whole multi-Agent turn fails cleanly with a structured error, without dropping individual results

The system should not silently lose a subset of parallel Agent results and then repair the turn with missing blocks.

Why I think this is not a hook/config issue

In this session:

• PreToolUse:Agent hooks succeeded
• PostToolUse:Agent hooks succeeded
• The affected agents did run
• OAuth/API requests continued successfully afterward

There was also a separate log entry:

Auth error: No API key available

But this appears unrelated, because OAuth checks completed and subsequent API calls still succeeded. The failure pattern here is specifically about parallel Agent tool_result pairing, not auth.

Environment

• Claude Code CLI: 2.1.85.29d
• Platform: macOS
• OS: Darwin 25.3.0
• Shell: zsh

Workaround

Current workaround on my side:

• Avoid 3 parallel Agent calls
• Use 2+1 instead
• If a result is missing, rerun only the missing Agent(s) serially
• If serial rerun still fails, fall back to the main thread

Related issue

This looks related to, but distinct from, anthropics/claude-code#31328.

That issue is about session corruption / resume problems after dropped entries.
This issue is about a live parallel Agent turn where the session stays usable, but some Agent results are lost and surfaced as internal errors.