-
Notifications
You must be signed in to change notification settings - Fork 3
FAQ
简体中文 · Back to documentation home
First check whether the selected model and channel support tool calling, then make sure the tools are not disabled in GrayCode settings. Some channels require tool support to be enabled explicitly; custom OpenAI-compatible channels must also return tool-call fields in the expected format.
If the problem affects only one conversation, try a new one to rule out stale context or channel state.
GrayCode decides when to request approval based on tool risk, workspace boundaries, and your confirmation policy. You can adjust that policy in settings, but keeping appropriate confirmation for side-effecting actions such as file writes and command execution is recommended.
File tools are designed around the current workspace. Paths outside it are subject to the relevant safety restrictions and confirmation flow. If a task spans several projects, consider adding them to a single VS Code multi-root workspace.
Ask it to pass multiple targets in one read_file call, or use search first and then read the relevant files as a batch. This is usually faster and uses less context than repeated single-file round trips.
- Enable or manually trigger context compaction.
- Start a new conversation after asking the model to summarize the important task state.
- Keep only files, terminal output, and references relevant to the current problem.
- Store durable conventions in prompt files or Memory instead of pasting them repeatedly.
See Context and Prompts for more details.
After the model writes files, GrayCode presents a diff. You can accept it, reject it, or ask for another revision. For larger changes, review the whole diff first and pay particular attention to configuration, dependencies, deletions, and edge cases.
Check GrayCode’s notification settings, VS Code notification permissions, and the operating system’s focus or do-not-disturb mode. Sounds and desktop notifications may be controlled separately by extension settings and system permissions.
Verify, in order:
- The API endpoint, key, and model name.
- Whether the channel supports the selected model and tool-call format.
- Network, proxy, service quota, and provider status.
- Whether the error occurs only with long contexts, images, or a particular tool request.
If the issue persists, open a GitHub issue with reproduction steps, the error message, and relevant environment details. Never include API keys or other secrets.