Bug

Recurring corruption where a tool_use block gets persisted to session JSONL but its matching tool_result does not (likely due to crash/timeout mid-tool-call). This makes every subsequent API call to that session fail Anthropic's message validation:

messages.251: `tool_use` ids were found without `tool_result` blocks immediately after: call95378438

Once corrupted, the session is permanently broken — gateway restarts don't help because the history is persisted to disk. The only fix is manually deleting/moving the session JSONL file, which loses all conversation history.

Occurrences

• 2026-02-10: Session session:agent:kalki:main:thread:986992, orphaned toolu_01XXejQsDDVRysXnYjVyJmue at messages.322
• 2026-02-17: Session kalki-bd/9397c2ef...topic-1771220855, orphaned call95378438 at messages.251

Proposed Fix

Add a validation/repair layer that runs before sending messages to the API:

1. On session load: Scan for tool_use blocks without a matching tool_result in the next message
2. Auto-repair: Inject a synthetic tool_result with an error like "Tool call interrupted — session recovered"
3. Log a warning when auto-repair kicks in so it's visible in diagnostics

This should be a lightweight pass over the message array before API submission — no need to modify the JSONL on disk.

Workaround

Move/delete the corrupted session JSONL file and restart the gateway. The agent starts a fresh session but loses all prior context.