Configure GitHub Copilot in VS Code

Connect GitHub Copilot to a BullSequana AI model and verify file tool calls.

Agentic Friendly

Use GitHub Copilot's Bring Your Own Key (BYOK) support to connect VS Code to a model through the BSQAI API.

Prepare the connection

Before you add the model, prepare the following:

  • VS Code 1.142 or later, with GitHub Copilot Chat enabled. The screenshots use VS Code 1.143.1.
  • Your deployment's HTTPS API URL, such as https://api.<platform-domain>/v1/responses.
  • A BSQAI API key with chat.message.create. Include tenant.read if you use the key to list models.
  • A model name from /v1/models that your key's owner can invoke. Choose a model with tool calling enabled.
  • The model's input and output token limits, supplied by your platform operator.

For a new key, follow Create and use an API key. If the model is absent from /v1/models, ask your Tenant Administrator to grant model access.

VS Code supports BYOK without a GitHub account or Copilot plan. For an organization-managed setup, confirm that your administrator permits custom endpoints before you continue.

Add the Custom Endpoint provider

Add the provider through the VS Code prompts:

  1. Open the Command Palette and run Chat: Manage Language Models.

    VS Code Command Palette with Chat: Manage Language Models selected

  2. Select Add Models, then Custom Endpoint.

    Add Models menu with Custom Endpoint selected

  3. In Group Name, enter BullSequana AI and press Enter.

    Group Name prompt with BullSequana AI entered

  4. Enter your BSQAI API key in the masked API Key prompt and press Enter.

    BullSequana AI API Key prompt with the key masked

  5. Select Responses as the API type.

    API Type prompt with Responses selected

  6. In the chatLanguageModels.json file that VS Code opens, keep the generated provider entry and its apiKey line.

    chatLanguageModels.json with a generated secret reference and stateless Responses settings

The configuration screenshot uses example token limits and enables image input. Use your deployed model's limits and capabilities instead of copying those values. The generated apiKey reference belongs to your VS Code profile. Do not copy the screenshot's reference.

Configure stateless Responses requests

In your provider entry, keep apiType set to "responses". Inside models, set the following fields in each BullSequana AI model entry. This fragment belongs inside a model entry, not in place of the complete configuration file.

{
	"id": "<model-name-from-/v1/models>",
	"url": "https://api.<platform-domain>/v1/responses",
	"toolCalling": true,
	"streaming": true,
	"statefulResponses": false,
	"zeroDataRetentionEnabled": true,
	"requestHeaders": {
		"X-Api-Key": "${apiKey}"
	}
}

Before you save, complete the model entry:

  1. Replace id with the model name from /v1/models.
  2. Replace url with your deployment's full /v1/responses URL. Keep url at model level, not at provider level.
  3. Set name to the label you want in the model picker.
  4. Set maxInputTokens and maxOutputTokens to your deployed model's limits.
  5. Set vision to true only if the model accepts image input. Otherwise, set it to false.
  6. If the model entry has an apiType override, remove it or set it to "responses".
  7. Save chatLanguageModels.json.

Keep ${apiKey} unchanged in requestHeaders. VS Code substitutes the key from its secret storage. Do not put the key itself in the JSON file or a shared workspace.

Keep both statefulResponses: false and zeroDataRetentionEnabled: true. Copilot then resends the conversation with store: false, without previous_response_id. The BSQAI API does not restore earlier turns from previous_response_id, so Copilot must resend the conversation. These settings do not disable platform logs, usage records, or Copilot's local chat history.

Verify a file tool call

Use a folder with no private files for the check:

  1. Create a folder containing one file named fixture.txt with this content:

    BSQAI_COPILOT_FILE_OK
  2. Open that folder in VS Code.

  3. Start a new Copilot chat in Agent mode.

  4. Select your BullSequana AI model from the model picker.

  5. Send this prompt:

    Read fixture.txt with a file tool and quote its contents. Do not modify any files.

If VS Code asks for permission, approve only the file read. The chat must show a file-read tool action and an answer containing BSQAI_COPILOT_FILE_OK. An answer without a visible tool action does not verify the tool call.

In the same chat, send a second prompt:

Read fixture.txt again with a file tool and tell me whether its contents changed. Do not modify any files.

Require a second file-read action and a completed answer without user_message_missing. After any configuration change, start a new chat instead of reusing a conversation that began in stateful mode.

Resolve setup failures

Restore an unavailable model

If the model does not appear, run Developer: Reload Window and select the model again.

If Copilot reports that it cannot obtain a Copilot token, use a Copilot-enabled account. For an organization-managed account, ask your administrator to check the BYOK policy. For a personal setup not subject to organization policy, sign out of GitHub through VS Code's Accounts menu. Then reload the window and start a new chat.

For a personal account, signing out restores provider registration in VS Code 1.143.1 when the account cannot obtain a Copilot token. The API connection still needs the file-read check. Do not sign out to work around an organization's BYOK policy.

Fix a tool-result continuation error

If a tool-result continuation fails with user_message_missing, check both stateless settings in the selected model entry. Set statefulResponses to false and zeroDataRetentionEnabled to true. Save the file. Start a new chat before you repeat the file-read check.

Report an invalid tool definition

If invalid_tool_definition names tools[N].name, ask your platform operator to check the deployed API's tool-name limit. Some VS Code extensions generate long tool names. Report the VS Code version, the tool name, the error code, and param. Do not send your API key or a workspace transcript.

Check an empty reply

If the reply is empty, check that the selected model uses apiType: "responses" and streaming: true. Then start a new chat and repeat the file-read check.

On this page