AI coding tools

Kilo Code Operating Card

Use Kilo Code modes, agent rules, approvals, and snapshots as a scoped workflow with independent Git and test verification.

Core concept
CurrentNext review due 2026-08-25Content version 0.36.0
Format
Reference
Level
Beginner
Audience
Practitioner, Developer
Owner
Project42 Editorial
Review cadence
Every 30 days
Prerequisites
A supported Kilo Code client; An approved model-provider path; A version-controlled workspace
01

Choose a mode whose tools match the phase

Kilo Code provides planning, coding, debugging, orchestration, and custom agent modes across editor and CLI surfaces. Choose the least-capable mode that fits the current phase and review any requested mode expansion.

Custom modes can restrict commands and files. Snapshots or checkpoints support experimentation but do not replace repository Git history or release controls.

02

Keep approval patterns narrow

Start interactively, inspect command and file approvals, and save only patterns that are safe for the current workspace. Avoid a blanket auto mode for unfamiliar repositories, tools, or network targets.

Version project agents and rules, review the complete Git diff, and run deterministic checks outside the model's narrative.

Kilo Code task card
text
Task: [one bounded outcome]
Scope: [workspace, paths, and mode]
Permissions: [files, shell patterns, network, tools, subagents]
Rules: [project agents and constraints]
Snapshot: [enabled recovery aid]
Verification: [Git diff and exact checks]
Recovery: [snapshot plus authoritative Git checkpoint]
03

Expected evidence and verification

Expected evidence includes the selected mode, permission decisions, rules, snapshot state, final Git diff, and test output. Confirm an automatic mode switch did not broaden file or tool access unexpectedly.

If the task crosses scope, stop approvals, preserve the trace, restore from a snapshot or Git checkpoint as appropriate, and retry in a narrower mode.