[BUG] Agents unable to call MCP tools when using ChatGPT

[HitByBricks]

Problem (one or two sentences)

Roo Code is unable to detect or call enabled MCP tools when prompted, and instead either ask how to proceed or automatically think its way around using said tools.

Context (who is affected and when)

Happens when GPT-5.2 is selected as model (OpenAI is selected as the provider).

Reproduction steps

1. macOS Tahoe 26.2, using the following MCP servers:
   -git
   -context7
   -Context Portal (conport), with custom instructions set for all mode to view and update the MCP automatically.
2. Prompt Roo Code to make a tool call (or, for conport, have it initialize the database automatically).
3. Roo Code determines that it, for whatever reason, cannot make such a call and either asks the user for more detailed information how to use the MCP server, or attempts to figure out how to use the tool call on its own (e.g. use the "create-mcp-server" skill).

Expected result

The task calls "use-mcp-tool"

Actual result

Either determines it cannot use the tool, or skips usage to complete the initially described task.

Variations tried (optional)

-I've tried both with default and custom instructions, but it still fails.
-With Gemini 3 Pro or Opus 4.6 enabled, the task calls the tools as expected, making it highly likely it's an issue with OpenAI models.

App Version

3.47.0

API Provider (optional)

OpenAI

Model Used (optional)

ChatGPT-5.2 High

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

I've analyzed this issue and identified a likely root cause. Here's my analysis and proposed implementation plan:

Root Cause Analysis

In PR #10895 (commit 84f7409f3, Jan 24 2026), the MCP SERVERS section was removed from the system prompt. That section previously contained:

1. A description of what MCP servers are and how they work
2. A list of all connected MCP servers with their tools, descriptions, and input schemas
3. Server-specific instructions (from server.instructions)
4. Explicit guidance to use use_mcp_tool / access_mcp_resource

After removal, the system prompt only has a brief mention in the capabilities section: "You have access to MCP servers that may provide additional tools and resources."

MCP tools are now presented only as native function definitions with auto-generated names like mcp--context7--resolve-library-id. While Claude and Gemini handle these well (they are good at calling tools based on native tool definitions alone), OpenAI models may need more explicit system prompt context to understand:

• That mcp--serverName--toolName functions represent MCP server tools
• When and why to use these tools
• Any server-specific instructions the MCP server provides

Additionally, MCP server instructions (server.instructions) are loaded from the MCP protocol but are no longer surfaced to the model anywhere in the prompt.

Proposed Plan

Re-add a lightweight MCP servers section to the system prompt (in src/core/prompts/sections/) that:

1. Lists connected MCP servers and their tool names (no full schemas -- those are already provided via native tool definitions)
2. Includes MCP server-specific instructions for servers that provide them
3. Explains the mcp--serverName--toolName naming convention so models understand these are callable tools
4. Provides brief guidance telling the model to prefer MCP tools when they match the task

This is intentionally more lightweight than the old removed section -- no duplicated tool schemas or descriptions (already in native tool definitions), just the mapping context and server instructions that are currently missing.

Files to modify:

• src/core/prompts/sections/mcp-servers.ts (re-create, lightweight version)
• src/core/prompts/sections/index.ts (re-export the new section)
• src/core/prompts/system.ts (include the new section in the prompt)
• Tests and snapshots as needed

This approach preserves the benefits of native tool calling (no schema duplication in the prompt) while restoring the contextual information that helps models like GPT-5.2 understand and use MCP tools.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[roomote]

I've seen the comment and will analyze this further. The previous analysis identifying the removed MCP servers system prompt section as the root cause looks solid. I'll proceed with implementing the lightweight MCP context re-addition and open a PR to share progress.

[roomote]

Thanks for the clarification, @HitByBricks. Good question -- the reason other providers like Gemini and Claude are not affected is likely because those models are better at inferring tool usage from native function/tool definitions alone. When the MCP servers section was removed from the system prompt (in PR #10895), the detailed context about MCP servers, their tool naming convention (mcp--serverName--toolName), and server-specific instructions was lost. Claude and Gemini models can still figure out when to call these tools just from the native tool definitions, but OpenAI models appear to need more explicit system prompt context to understand the mapping and when to use them.

I'll proceed with implementing the fix -- re-adding a lightweight MCP servers section to the system prompt that includes the server/tool name mapping and any server-specific instructions, without duplicating the full schemas (which are already in the native tool definitions). Will open a PR shortly.

[roomote]

Opened PR #11322 with my attempt at addressing this. The PR re-adds a lightweight MCP servers section to the system prompt that lists connected servers, their tool name mappings (the mcp--serverName--toolName convention), and server-specific instructions -- without duplicating tool schemas. This should give OpenAI models the context they need to understand and call MCP tools. Feedback welcome!

[roomote]

Opened PR #11323 with my attempt at addressing this. The fix re-adds a lightweight MCP servers section to the system prompt that gives models explicit context about the mcp--serverName--toolName naming convention and connected servers, without duplicating tool schemas. All tests pass (207 tests, 0 regressions). Feedback welcome!