Model outputs '[Historical context]' / tool-call JSON instead of a normal reply

Symptoms

• The bot replies with internal-looking text like:
  • [Historical context: a different model called tool "..." with arguments: {...}. Do not mimic this format - use proper function calling.]
• Tools don’t run (or run inconsistently), and you see “tool call” syntax as plain text in the chat.
• This is most noticeable with some “reasoning-heavy” or “non-tool-native” models (often on Telegram), but it can happen on any channel.

Cause

The selected model/provider is not reliably using OpenClaw’s native tool calling interface. Instead of emitting a structured tool call, it outputs internal prompt artifacts (like the “Historical context” guard text) as normal assistant text.

Common triggers:

• A model that doesn’t support function/tool calling on your provider.
• A provider integration that disables or degrades tool calling for that model.
• A “thinking/reasoning” model producing extra metadata that leaks into the final output stream.

Fix

1) Switch to a known tool-capable model (fastest)

In chat, switch models (exact command may vary by channel UI), for example:

• /model list
• /model <provider>/<model>

Then retry the same prompt that previously leaked “Historical context”.

2) If you are using Gemini / Antigravity, try a different Gemini tier

If you’re on a Gemini model that frequently leaks tool metadata:

• Try a “flash” / less reasoning-heavy variant for tool-heavy workflows, or
• Temporarily switch to a different provider/model for tasks that require tools.

3) If you are using a local backend (LM Studio / Ollama), confirm tool calling is enabled

Local backends often require an explicit “function calling / tool use” toggle, and not every model checkpoint supports it.

Checklist:

• Enable function/tool calling in the backend UI (if available).
• Switch to a model that explicitly supports tool calling.
• Restart the gateway after changing model/provider settings.

Verify

• Ask the agent to do a trivial tool action (for example, read a local file or run a simple diagnostic) and confirm:
  • You do not see [Historical context: ...] in the chat
  • The tool actually executes (not just described)
• openclaw models status --probe succeeds for your selected model.

Related

• GitHub issues (examples): see related.githubIssues in frontmatter.