Description

When the Anthropic API returns 429 (rate limit) and all retries are exhausted, the Gateway saves the assistant message containing tool_use blocks to the session .jsonl file without a corresponding tool_result block. This creates an invalid message history that permanently blocks the session — every subsequent API request fails with:

messages.N: `tool_use` ids were found without `tool_result` blocks immediately after: execXXX

The session remains dead until the Gateway is restarted (creating a new session).

Steps to Reproduce

1. Agent (Claude Opus) calls a tool (e.g., exec)
2. API rate limit (429) is hit during the request
3. Gateway retries multiple times (observed 20 retries in ~3 minutes)
4. All retries fail → Gateway saves the assistant message with tool_use to .jsonl
5. tool_result is never written (the tool was never executed)
6. Next heartbeat/request → Anthropic API rejects with 400 tool_use without tool_result
7. Session is permanently broken

Observed Behavior (Feb 25, 2026)

22:11:35-22:14:05 — 20x rate limit retries (same runId)
22:26:33 — First 400 error (orphan tool_use: exec17720470437771)
22:26-00:49 — 7 consecutive heartbeats, all failed with same error
01:05 — Gateway restart → new session (workaround, not fix)

Additional context: Gateway was in a crash-loop (634 restarts) due to an unrecognized config key (cron.fleet-monitor), which likely contributed to the rate limit being hit.

Expected Behavior

If all rate limit retries are exhausted:

• The orphan tool_use assistant message should not be persisted to the session .jsonl
• OR a synthetic tool_result with an error message should be appended (e.g., {"type": "tool_result", "content": "Rate limit exceeded, tool was not executed"})

Workaround

We built a cron-based session-healer plugin that:

1. Detects the error pattern in Gateway logs
2. Finds the orphan tool_use line in .jsonl by ID
3. Surgically removes it + injects a recovery message
4. Escalates (retry → remove compaction summary → rotate session)

Environment

• OpenClaw: 2026.2.23 (b817600)
• Model: anthropic/claude-opus-4-6
• OS: Ubuntu Linux 6.8.0
• Node: v22.22.0