How Code works - Cortex Docs

A Code turn is a tool loop like Chat’s, with two differences that define the product: the tools operate on a real repository inside an isolated guest, and the mode you choose decides which tools the model is offered at all.

The workspace catalog

Tool	What it does
`read_file`	Read a file from the repository
`write_file`	Create or replace a file
`edit_file`	Make a targeted edit
`glob`	Find files by pattern
`grep`	Search file contents
`bash`	Run a shell command in the guest — builds, tests, git
`todo_read`, `todo_write`	Keep a task list for the turn
plan tool	Write a plan — with a mermaid diagram — in Plan mode

All of these run inside the attached guest. python is a Chat tool; a Code turn runs the guest’s own Python through bash.

Attach before the first command

If the session names a repository and no runtime is attached, the harness starts one, attaches it, and prepares the environment before the first tool runs. An unattached runtime is never the first thing a turn does. See Cloud runtimes.

Modes decide the tools

Mode	Tools offered
Ask	Read-only: `read_file`, `glob`, `grep`
Plan	Read-only plus the plan tool
Agent	The full catalog, including `write_file`, `edit_file`, `bash`

Ask and Plan withhold the mutating tools in the harness. The model cannot edit or run anything in those modes because it is never offered the tools. See Ask, Plan, Agent.

Stopping

A turn ends when the model stops calling tools or you press Stop (Esc in the CLI). Code does not copy Chat’s eight-round budget; a coding task legitimately takes many tool rounds. Cancel aborts the stream and tells the runtime to stop the turn.

Compaction

Long sessions are compacted with Code’s own keep-set — the current task, open files, recent diffs — so the model keeps fitting in context. Compaction fails closed: an unavailable summariser is an error, not a silent loss of history.

The CLI on top

The CLI drives the same session API with streaming tokens and first-class tool rows. It adds harness-side locks — Plan and Spec keep mutating tools blocked until you switch back or accept the plan — plus approvals, a sandbox policy, and a persisted long-horizon /goal. See Modes and permissions and Goals.

Cortex CLI composer with the Goal chip, rendered in a 40 by 12 terminal — A CLI session with an active goal: the composer border reads Agent — Goal · 2/8, and the agent keeps working toward it across turns.

Compared with Chat and Bot

	Chat	Code	Bot
Where tools run	Snippet sandbox	The session’s guest	The bot’s computer
Mutations gated by	—	Mode (Ask / Plan / Agent)	Approvals (Allow / Deny / Always)
Stops on	8 tool rounds	Model stops or Stop	15-minute deadline or six identical calls
Transcript	Streamed reply	Turn timeline	User-visible bubbles

See How Chat works and How Bot works.

Sessions — reading the timeline this loop produces.
Ask, Plan, Agent — the modes that gate mutating tools.
Cloud runtimes — the guest the tools execute in.
Code troubleshooting — when the agent will not edit.

⌘I

​The workspace catalog

​Attach before the first command

​Modes decide the tools

​Stopping

​Compaction

​The CLI on top

​Compared with Chat and Bot

​Related