Bug type

Regression (worked before, now fails)

Summary

After upgrading to OpenClaw 2026.3.2 (85377a2), sessions running on nvidia/moonshotai/kimi-k2.5 show severe tool instability: tool calls intermittently fail with “Tool not found”, time out, or produce malformed tool-call behavior. Switching the same session to OpenAI GPT‑5.2 immediately restores reliable tool execution (read, web_fetch, exec, etc.).

This appears to be a regression specific to Kimi 2.5 tool-calling compatibility (or OpenClaw’s adapter for it) rather than the tools themselves.

Steps to reproduce

Reproduction Steps (high-level)
Run OpenClaw 2026.3.2 (85377a2).
Use a Discord direct chat session with model set to nvidia/moonshotai/kimi-k2.5.
Attempt basic tool calls:
read a local file (e.g. workspace markdown)
web_fetch a URL
exec simple shell command (echo, whoami)
Observe intermittent failures (“Tool not found”, timeouts, inconsistent behavior).
Switch the same agent/session model to GPT‑5.2.
Re-run the same tool calls; observe stable success.

Expected behavior

Tool calls should behave consistently regardless of model choice, assuming the same OpenClaw runtime + permissions.

Actual behavior

Symptoms / Observed Behavior
When model is nvidia/moonshotai/kimi-k2.5:
Tool calls intermittently fail with errors like:
“Tool not found”
timeouts on tool calls (e.g., cross-session send)
inconsistent tool invocation behavior (some tools work once, then fail again)
Failures observed with:
read
exec
web_fetch
browser (attempted)
messaging/session tools sometimes timing out depending on agent session state

When switching model to GPT‑5.2:
Tool calls become stable and consistent.
Verified successful:
read (local workspace files)
web_fetch (external docs fetch)
exec (shell commands)
extended workflows (Firecrawl install/auth, CLI tests, etc.)

OpenClaw version

2026.3.2 (85377a2)

Operating system

macOS Monterey 12.5.x

Install method

npm

Logs, screenshots, and evidence

Tool not found
timeout
exec denied
allowlist miss

Impact and severity

Agents become unreliable / unusable for automation because tool calls fail unpredictably.
Significant time lost debugging gateway/tool configuration, but issue resolves by changing model.

Additional information

Notes / Hypothesis
This looks like a tool schema / tool-call formatting regression affecting Kimi 2.5 in OpenClaw 2026.3.2 (or the specific provider adapter for nvidia/moonshotai/kimi-k2.5). The exact same tools + permissions worked immediately after model swap.

Environment
OpenClaw version: 2026.3.2 (85377a2)
Host OS: Darwin 21.6.0 (x64) (macOS)
Node: v22.22.0
Channel/surface where observed: Discord direct chat (provider: discord)
Models compared:
Problematic: nvidia/moonshotai/kimi-k2.5
Stable: OpenAI GPT‑5.2 (openai/gpt-5.2)

Workaround
Switch tool-using agents from nvidia/moonshotai/kimi-k2.5 to a model with stable tool calling (e.g. openai/gpt-5.2).

Logs / Additional Info
If maintainers want, I can provide:
session transcript paths from ~/.openclaw/agents/.../sessions/*.jsonl
exact error strings as they appeared in Discord
exact tool call payloads during failure vs success