Summary

When using kimi-coding/k2p5 (Anthropic Messages API) in OpenClaw, tool calls intermittently arrive with empty arguments ({}), causing validation failures such as:

• Validation failed for tool \"exec\": command: must have required property 'command'
• Similar behavior also observed for tools requiring required fields (e.g. web_search.query, gateway.action).

Switching the same agent/session behavior to openai-codex/gpt-5.4 significantly reduces the issue.

Environment

• OpenClaw: 2026.3.23-2
• OS: Linux 5.15.0-100-generic
• Node: 22.22.1
• Model with issue: kimi-coding/k2p5 (api: anthropic-messages)
• Comparison model: openai-codex/gpt-5.4 (issue much less frequent)

Representative logs

From session transcript:

Validation failed for tool "exec":
  - command: must have required property 'command'

Received arguments:
{}

There are multiple consecutive occurrences in a single session where the assistant emits toolCall for exec with arguments: {}.

Repro pattern (not 100% deterministic)

1. Use kimi-coding/k2p5 as agent model.
2. Run in a long-context conversation (memory-heavy/system-heavy prompt).
3. Ask for short imperative actions (e.g., "execute", "check status", etc.).
4. Observe occasional tool calls emitted without required arguments.

Expected behavior

If model tool-call output is missing required fields, OpenClaw should ideally:

• optionally auto-repair/retry once with strict schema reminder, or
• fail gracefully with a model-side fallback strategy.

Actual behavior

OpenClaw correctly rejects invalid calls, but repeated {} tool calls can loop/flood errors in a session.

Request

Could maintainers confirm whether this is a known compatibility issue for specific model providers, and whether OpenClaw should add a resilience layer (auto-retry / required-field guardrail) for malformed tool arguments?