Write tools (edit, write_file, create_file) not exposed to OpenAI models in function calling namespace

[bdlst]

Description

When using the GitHub Copilot SDK 0.2.1 with OpenAI models (gpt-5.3-codex, gpt-5.4), the write tools (edit, write_file, create_file, edit_file) are not visible to the model in the function calling namespace, even though they are correctly registered in the session.

Claude models (claude-sonnet-4.6) using the exact same session configuration can see and use all write tools without any issue.

Context

Using the Copilot SDK. I can switch between Claude and GPT models. The SDK session is created with identical parameters for both models.

Steps

1. Send a message asking to modify a file (e.g., "add ScaleControl to MapProvider.tsx")
2. Claude (claude-sonnet-4.6): calls edit and create tools successfully
3. GPT (gpt-5.3-codex / gpt-5.4): states write tools are not available

Evidence from GPT-5.4 session logs

GPT internal reasoning:

"I'm wondering if hidden tool names outside of the functions namespace are causing this 'cannot call' mistake. It seems I can only send tool calls with the available namespace."

GPT explicit statement about tools:

"Claude a pu utiliser edit, puis ensuite le runtime a annonce: create, edit, grep non disponibles"

Sub-agent running on GPT-5.3-codex:

"I'm unable to apply the requested code changes because this environment only exposes view and delete_file tools, and does not provide any file-editing tool (edit_file / write_file / create_file)."

Key observation

The system prompt (text-based tool documentation) is identical for both models. The difference appears to be in the function calling schema — the structured JSON tool definitions that the SDK sends alongside the system prompt. GPT models report that write tools are not in their callable namespace, while Claude models see and use them.

Expected behavior

All 75 registered tools (including write tools: edit, write_file, create_file, edit_file, delete_file) should be available to ALL models in the function calling namespace, regardless of whether the model is Claude or GPT.

Environment

• SDK: @github/copilot-sdk@0.2.1
• Node.js: 22.x
• OS: Windows 11
• Models tested: gpt-5.3-codex (fail), gpt-5.4 (FAIL), claude-sonnet-4.6 (OK)

[github-actions]

Investigation Findings

Thanks for the detailed report. I've investigated the relevant SDK code and want to clarify the architecture here.

What I investigated

• nodejs/src/client.ts — createSession and resumeSession RPC calls, specifically how tools are sent to the CLI
• nodejs/src/types.ts — SessionConfig, availableTools, excludedTools, and Tool type definitions
• nodejs/README.md — documented behavior for tools and model configuration

What I found

The SDK is not responsible for tool visibility in the model's function-calling namespace. Here's why:

The SDK acts as a thin client. When you call createSession, it sends your configuration (tools, availableTools, excludedTools, model name) to the Copilot CLI over JSON-RPC. The CLI is the component that then communicates with the model provider (Anthropic or OpenAI) and decides how to format and send tools in the function-calling schema.

The tools you're seeing as "not available" — edit, write_file, create_file, edit_file, delete_file — are built-in CLI tools, not custom tools defined via the SDK's tools: [...] option. The SDK has no visibility into which built-in tools the CLI exposes to a given model provider, and the SDK applies no model-specific tool filtering of any kind.

The relevant code in client.ts confirms this:

tools: config.tools?.map((tool) => ({  // Only user-defined custom tools are mapped here
    name: tool.name,
    description: tool.description,
    parameters: toJsonSchema(tool.parameters),
    ...
})),
availableTools: config.availableTools,   // Passed to CLI as-is
excludedTools: config.excludedTools,     // Passed to CLI as-is

Why this is a question, not an SDK bug

The observed difference in tool availability between Claude and GPT models is most likely caused by one or more of the following, all of which are CLI-level or model-provider-level concerns:

1. Model-provider-specific tool routing in the CLI — The CLI may send tools differently when targeting the Anthropic API vs. the OpenAI API.
2. OpenAI function-calling tool limits — OpenAI's API has documented limits on the number of tools per call. With 75 tools, some may be silently dropped depending on the model.
3. Built-in tool availability per model — The CLI may intentionally restrict certain built-in tool categories for specific model families.

Recommendation

This issue is best directed at the Copilot CLI team rather than the SDK. If you'd like to work around it at the SDK level, you can use the availableTools session config parameter to explicitly list only the tools you need — this will restrict the set of built-in tools the CLI exposes to those you specify, potentially avoiding any implicit limits.

Reclassifying as a question since the SDK is behaving as designed.

Generated by Bug Handler for issue #1008 · ◷