Summary

After updating from OpenClaw 2026.2.9 to 2026.2.12, the embedded agent using Kimi K2.5 via Moonshot Direct (OpenAI-compatible API at api.moonshot.ai) consistently generates malformed tool calls. This worked correctly on 2026.2.9.

Symptoms

1. read tool called without path — The model sends an object but omits the path (and file_path) parameter entirely. Happens systematically (18+ occurrences in a single day):

   [agent/embedded] read tool called without path: toolCallId=read:40 argsType=object
   

2. edit tool missing newText/new_string — The model calls edit without providing either accepted parameter name:

   [tools] edit failed: Missing required parameter: newText (newText or new_string)
   

3. edit exact match failures — Downstream effect when the model eventually sends valid params but wrong content:

   [tools] edit failed: Could not find the exact text in /Users/.../.openclaw/openclaw.json
   

Environment

• OpenClaw version: 2026.2.12 (worked on 2026.2.9)
• Model: moonshot/kimi-k2.5 via Moonshot Direct (api.moonshot.ai, OpenAI-compatible API)
• OS: macOS (darwin 25.1.0)
• Node: 22.22.0

Analysis

The tool schemas come from @mariozechner/pi-coding-agent and are wrapped by patchToolSchemaForClaudeCompatibility() which adds aliases (file_path → path, old_string → oldText, new_string → newText). Despite these aliases, Kimi K2.5 appears to be sending parameters under names that none of the aliases cover.

The patchToolSchemaForClaudeCompatibility function name suggests these schemas are optimized for Claude. Non-Claude models using the OpenAI-compatible API may interpret the patched schemas differently, especially after the 2026.2.12 schema changes.

Reproduction

1. Configure OpenClaw 2026.2.12 with moonshot/kimi-k2.5 as primary model
2. Send any message that triggers file reads or edits
3. Observe read tool called without path and edit failed: Missing required parameter in logs

Expected Behavior

Tool calls should work with OpenAI-compatible providers (Kimi K2.5 via Moonshot) as they did in 2026.2.9.

Suggested Fix

Consider one of:

• Adding broader parameter name aliasing for OpenAI-compatible models (e.g., also check filename, file, filepath for read; replacement, text, content for edit)
• Providing a config option to use simplified/legacy tool schemas for non-Claude providers
• Testing tool schema changes against OpenAI-compatible models before release