Bug: Calling non-existent tools returns synthetic context_info content instead of error

[peremajoral]

Description

When calling a non-existent tool (e.g., mcp_context_info), instead of receiving a "tool not found" error, the system returns the synthetic context_info content (the prunable-tools list).

Steps to Reproduce

1. Have DCP plugin installed and enabled
2. Try to call any non-existent tool, e.g., mcp_context_info
3. Instead of an error, you receive the <prunable-tools> list

Expected Behavior

Calling a non-existent tool should return an error like "tool not found" or similar.

Actual Behavior

The synthetic context_info tool output (containing the prunable-tools list) is returned, making it appear as if the tool "succeeded" but with unexpected content.

Root Cause

Found in lib/messages/utils.ts line 69:

// createSyntheticAssistantMessageWithToolPart function
return {
    parts: [{
        type: "tool",
        callID: SYNTHETIC_CALL_ID,
        tool: "context_info",  // <-- This synthetic tool name
        state: {
            status: "completed",
            output: content,  // <-- prunable-tools list
            title: "Context Info",
        },
    }],
}

The DCP plugin creates synthetic tool results with tool: "context_info" to inject the prunable-tools list into the conversation. When a user tries to call a non-existent tool, this synthetic content gets returned instead of a proper error.

Suggested Fix

Option 1: Rename the synthetic tool to something clearly internal, like:

• _dcp_internal_context_info
• __synthetic_context_info
• dcp:context_info

Option 2: Prefix with underscore to indicate internal use: _context_info

This would prevent confusion when users accidentally try to call non-existent tools.

Environment

• DCP version: latest (@tarquinen/opencode-dcp@latest)
• OpenCode version: 1.0.216+

[Tarquinen]

I don't really understand the problem, I tried to reproduce as you said, this looks normal?

When a tool use fails in opencode, the output describes that fail back to the model. Here's exactly what was sent to the model in that session:

I think what you're saying is when a tool use fails, the model may believe the context_info injections are the output of the failed tool, but I don't see how that's possible considering opencode sends back a clear fail message in the output