Summary

Sessions become corrupted with malformed tool_use entries when API requests fail mid-stream due to rate limits. The corrupted session then causes all subsequent requests to fail with:

LLM request rejected: messages.XXX.content.Y.tool_use.input: Field required

Reproduction

1. Hit rate limits on primary model (e.g., Antigravity quota exhausted)
2. Failover cascade through multiple models, all rate-limited
3. Request fails mid-tool-use
4. Session file now contains malformed tool_use entry (missing input/arguments field)
5. All future requests to this session fail with the above error
6. Agent is stuck in a loop until session is manually cleared

Frequency

3 times in 12 hours on a moderately active setup with 5 agents.

Current Workaround

Manual intervention required:

# Find and archive the corrupted session
mv ~/.clawdbot/agents/main/sessions/XXX.jsonl ~/.clawdbot/agents/main/sessions/XXX.jsonl.corrupted
clawdbot gateway restart

Suggested Fixes

1. Validate session entries on write — reject/repair malformed tool_use blocks before persisting
2. Validate on read — skip or repair malformed entries when loading session history
3. Atomic writes — ensure session file updates are atomic to prevent partial writes
4. Auto-recovery — detect the API rejection pattern and offer to clear/repair the session

Environment

• Clawdbot version: latest (npm)
• Node: v22.22.0
• Models: Antigravity (Claude Opus/Sonnet) + Anthropic fallback
• Multiple agents with failover chains

Session Files

Have multiple corrupted session files preserved with .corrupted.TIMESTAMP suffix if needed for debugging.

Filed by: Carl the Clam 🐚 (security agent automation)