AI agent
Agent approvals and safety boundaries
Use the shell-command approval card and kit drafts without assuming Kition has a universal per-tool permission editor.
Updated
By Kition Docs
Verified against Kition 0.2.0 (7d49036f) on 2026-11-06
Shell commands have a resumable approval flow
The Agent runs shell commands with exec_command inside the workspace. A command that the execution policy does not allow pauses the turn, and Kition shows a Shell execution requires approval card with the exact command and the reason.
- Allow once — run this exact command once
- Always allow — shown only when the runtime suggests a command prefix; remembers that prefix
- Deny — tell the Agent not to run the command and to continue without it when possible
Heads up
Read the full command, its arguments, redirects, and shell operators before approving. A safe-looking prefix can still permit risky later arguments.
Execution policy
A small set of read-only commands, such as ls, cat, grep, git status, and git diff, run without asking, and commands such as sudo are always refused. Command lines that use substitution, redirection, or background jobs always ask.
Always allow saves an allow rule for the prefix in execpolicy.toml in the .kition folder of your home directory. A command not covered by a saved rule asks again. The app has no screen for viewing or removing saved rules.
There is no universal tool permission editor
Kition does not ship Settings → Agent → Permissions, a workspace .kition/permissions.json, or allow/ask/deny rules for every document, kit, web, MCP, and filesystem tool.
Outside shell commands, Agent tools run without a confirmation step, including file removal, document writes, and MCP tools. Do not assume writes always ask or destructive tools always deny.
Review controls that exist
- Kit edits in Draft mode stay out of the document until you Merge them
- Live kit edits can be undone in the kit editor
- Files modified lists the files a chat changed
- Version history records a version after an agent edits a document
- The Stop button ends the current turn but does not undo completed actions
Practical safety rules
- Back up the workspace before broad edits
- Use least-privilege provider, website, database, and external-service accounts
- Treat fetched web pages and external tool output as untrusted prompt input
- Verify filesystem and remote-system changes at the source
- Deny commands you cannot explain and ask for a safer alternative
Implementation sources
- kition@7d49036f: contracts/runtime/agent-shell-approval.schema.json
- kition@7d49036f: crates/kition-agent/src/execpolicy.rs
- kition@7d49036f: crates/kition-agent/src/tools/exec.rs
- kition@7d49036f: crates/kition-agent/src/tools/mod.rs
- kition@7d49036f: src/features/agent/components/AgentShellApprovalCard.tsx
- kition@7d49036f: src/features/kits/kition/KitDraftReview.tsx
Related pages
Agent tool calls
Read the runtime tool names, how availability is decided, and the inputs, outputs, errors, and changed files shown in the Agent timeline.
Plans, drafts, and review
Distinguish the informational Agent plan card from the review steps that actually hold changes back: kit drafts and shell approvals.
Web research for the Agent
Let the Agent search the web through the model provider, fetch public pages, and save articles into the workspace.