Copilot CLI adds settings commands and keeps your latest prompt

An assistant can write the right code for the wrong request if it loses your last correction. GitHub’s Copilot CLI 1.0.92, dated October 5, 2026, reports a fix for retaining the latest prompt during compaction when requests exceed context limits. Its release also adds practical setup controls. We checked the versioned notes on October 6. Release 1.0.92.

This is the terminal application, separate from the Copilot code review API update we covered earlier.

Make the setup visible before work starts

The new copilot config subcommands let you list, read, set and remove settings. Before a conversation, Ctrl+E opens an environment picker for local or cloud runs. Versioned change list.

For someone moving between a local prototype and a longer cloud task, the useful question is where the work will run before the agent begins. A saved preference is easy to forget; inspecting it is cheaper than discovering the mismatch after files have changed. That is our workflow interpretation, not a measured reduction in mistakes.

GitHub’s documentation lists CLI access with all Copilot plans, subject to the organization’s CLI policy when access comes from an organization. Cloud and local sandboxing remain a public preview. This update does not promise additional included usage or remove those controls. Access and sandbox documentation.

A longer conversation still needs a clear handoff

Compaction condenses an old conversation so a task can continue within its context window. GitHub also reports recovery-context fixes for the latest requests and says oversized Anthropic requests can retry with smaller images or fewer attachments. These are vendor-described fixes; we have not reproduced their reliability. Release notes.

Preserving a prompt does not establish that every detail in an attachment survives. Keep the current acceptance criteria in a small project note: what should change, what must remain intact and the check that will establish completion.

Next action: after updating through GitHub’s supported installation route, inspect the settings and choose the intended environment. On a disposable task, restate one correction and check that the resulting diff follows it. This is a proposed check, not one we executed. Our guide to evaluating AI output explains how to judge the result rather than the assistant’s confidence.