Shell tool executes destructive commands (e.g. `rm -rf`) with no approval/confirmation gate

Start with the shell execution and banned-builtins checks in internal/core/shell.go, then read the approval patterns in examples/extensions/confirm-destructive.go and examples/extensions/permission-gate.go. Work out how a default gate should classify dangerous commands and behave in interactive and non-interactive contexts. Done means risky commands require approval or are blocked safely, including in --quiet, CI, and GitHub Action contexts.

Summary

The shell core tool (internal/core/shell.go) runs any command the model requests with no human-in-the-loop confirmation step, including destructive commands such as rm -rf. The only existing restriction is a regex blocklist of a handful of shell builtins (alias, bg, bind, builtin, command, set, trap, ulimit, etc. — see internal/core/shell.go ~L227-229). Any external binary — rm, curl, nc, ssh, scp, aws, git push --force, etc. — executes immediately once requested, with no approval prompt and no per-command risk classification (~L343-345).

Reproduction

In an interactive TUI session:

1. Start kit in a project directory.
2. Prompt: "Please execute the command rm -rf ~/test."
3. The model issues Shell rm -rf ~/test.
4. The command runs immediately — no confirmation dialog, no "allow once / always / deny" prompt — and the transcript shows only (no output) before the agent reports back and marks the turn "Done".

Why this matters

Kit is designed to also process content that may not be fully trusted — model output steered by file contents, fetched web pages, MCP tool results, or (via the GitHub Action integration) third-party issue/PR text. Because the shell tool has no confirmation gate and no dangerous-command classification, any prompt-injected instruction that reaches the model can translate directly into destructive local command execution (data deletion, exfiltration via curl/nc, credential theft, etc.) with no opportunity for the user to intervene.

Note that examples/extensions/confirm-destructive.go and examples/extensions/permission-gate.go already demonstrate an approval-gate pattern as an opt-in extension — this suggests the gap is that no form of this is enabled by default in core, rather than the capability being entirely absent from the codebase.

Suggested remediation

• Add a confirmation/approval gate to the core shell tool, at minimum for commands matching a dangerous-pattern list (recursive delete, network tools, credential-path access, force-push, sudo, etc.), prompting the user before execution.
• Consider promoting the existing confirm-destructive / permission-gate extension pattern to a default-on core behavior rather than leaving it opt-in.
• Ensure whatever gate is added fails closed (blocks/prompts) rather than open in non-interactive contexts (--quiet, CI, GitHub Action), since those are also the contexts most likely to encounter untrusted input.

Relevant code

• internal/core/shell.go (banned-builtins check ~L227-229, execution path ~L343-345)
• examples/extensions/confirm-destructive.go, examples/extensions/permission-gate.go (existing opt-in pattern)