Kimi Code 1.1.1: reload a session after changing project rules

Kimi Code is an AI coding desktop client. Developers can open a local project in a graphical interface, ask the assistant to read or change code, and inspect the result.

Desktop version 1.1.1 was released on October 21, 2026. The official changelog adds /reload and a session-menu reload action to apply external changes to configuration, AGENTS.md, and skills immediately in the current session. AGENTS.md is a project file containing working instructions for a coding assistant, such as the permitted scope, checks to run, and output format.

The update also adds a workspace-trust restriction: declining trust for a project directory blocks sending messages. Model switching gains a confirmation step offering a new session, a switch after compacting context, or a direct switch.

The point here is the explicit reload action. Earlier releases already described automatic monitoring of some external configuration changes, so 1.1.1 should not be treated as the first version able to notice changed project rules.

Suppose you add “Explain the affected scope before changing styles” in your editor. The next task given to AI should start under that new instruction. An explicit reload connects saving the file with rereading the session’s rules, without restarting the whole application.

A workspace is the local project folder used by the session. Checking it before reloading helps avoid editing project A’s rules while asking questions in project B. Trust determines whether work can begin in the directory; it does not approve every later command. Individual tasks may still produce their own approval cards.

Reloading does not retroactively revise the results of an earlier task. It provides a way to update session rules before continuing. Actual changes still need to be checked through file diffs and project checks; an assistant saying “understood” does not complete that verification.

Use the official getting-started guide to install or update to a desktop version that supports /reload. Kimi’s official models require a Kimi membership, with model access varying by tier. After signing in, check the account’s actual access and usage. The desktop client also supports configured third-party model providers, which require the corresponding credentials and model settings. The example below uses an official model available to your account. Try it in your own test project: it is a suggested procedure, not a record of a completed client test.

1. Confirm the project. After creating a session, click the folder name above the input. Choose a recent folder or use “Select folder…” to open the target project. Read the trust prompt. If you do not trust the directory, do not begin a task there.
2. Save the rule externally. In an editor, update the project’s AGENTS.md—for example: “Before changing README, list the headings you plan to edit. Do not change source code in this task.” Save it and confirm that it belongs to the current workspace.
3. Wait until idle, then reload. Enter /reload, or open the menu next to the session title and select the session-reload action. The project and session documentation says reload is unavailable while content is being generated or in an archived session. Wait for the current turn to finish; an archived session is not an eligible target for this step.
4. Assign a small task. Try: “Check the installation instructions in README. First list the headings you plan to change and wait for my confirmation. Do not change source code.” Compare the plan with the saved rule before continuing. This prompt repeats the task boundary to make the result easy to inspect; on its own, it does not prove that the rule file was read.
5. Inspect the output. If you proceed with editing, review each file in the changes panel and check that only agreed files changed. Before running checks, confirm the terminal directory. If the scope is wrong, stop further modifications, inspect the project, rule file, and session, then restate the next task with a clear scope.

The output should be a clearly scoped plan and, after your confirmation, inspectable file differences. Reload makes updating instructions more direct; those actual outputs still determine whether the work followed them.