Approvals & Diffs

KnightCode prioritizes developer security and transparency. Every operation that modifies your files or executes commands on your operating system is gated by an interactive approval flow.

The Approval Card

When the agent attempts to call a security-sensitive tool (like writing a file or executing a shell script), KnightCode pauses execution and renders an approval card:

• Approve (Y): Accepts the action. The agent proceeds with execution.
• Reject (N): Rejects the action. The agent is notified and continues without executing the tool.
• Provide Feedback: You can type a response (e.g., "No, use the existing database schema instead") to guide the agent to a different approach.

Visual Diff Viewer

When the agent uses editing tools (like Edit or MultiEdit), KnightCode doesn't silently rewrite files.

• It displays a side-by-side or unified diff directly inside the TUI showing the exact line insertions, deletions, and modifications.
• This lets you review the code changes hunk-by-hunk before hitting Approve.

Auto-Allowed Commands

In BUILD mode, constantly approving the same safe commands (like git status or ls) can be repetitive. (In AUTO mode every tool call runs without prompting — the one exception is AskUserQuestion, which always surfaces to you since it is the model explicitly asking for your input, not a permission gate — so the allowlist mainly benefits BUILD mode.)

KnightCode provides an Allowlist for commands:

• Open the /allow dialog.
• You can review or add command prefixes for trusted commands (e.g., git status, npm test, bun run lint).
• A Bash command runs automatically when it exactly matches an allowed prefix, or begins with one followed by extra arguments. As a safety guard, a command whose extra arguments contain shell metacharacters (;, &, |, `, $, (, ), redirects, newlines) is not auto-allowed and still prompts.