Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Kimi Code 2.6.0 canary: read the label

Written November 12, 2026

A 2.6.0 canary is not a 2.6.0 release. The newest entry on Kimi Code's releases page is a pre-release with a real change list, and the stable line is still 2.5.3. Stay on stable unless you are testing for Moonshot.

What a canary means here

Moonshot publishes Kimi Code canaries as pre-releases on the Kimi Code releases page: the 2.6.0 canary line has been rolling since early November, with dated tags such as 2.6.0-canary.20261109 and 2.6.0-canary.20261111 in the page's own changelog links, and the newest entry sits above them, marked Pre-release. A canary is a channel build for people willing to hit unfinished changes early. It is not the version a team should roll out, and nothing in it is promised to survive into a stable 2.6.x.

Keep the lines separate. The 2.5.x line is the CLI's stable line, and 2.5.3, which added steering, is still its current release. The 1.3.x line is Kimi Code Desktop, a different product surface. A 2.6.0 canary headline says nothing about either.

What the newest canary lists

Two minor changes lead the entry. File suggestions (fs:suggest) now align with local fd completion: bounded scans, at most 20 results, optional streamed previews, hidden and gitignored paths included by default, and regular files placed first at equal match quality. In fullscreen mode, consecutive tool calls and thinking now fold into summary lines, and the Activated skill card no longer shows for skills the model loads. Both folds can be set under Tool call density and Thinking folding in /settings, or with tool_call_density = "ungrouped" and thinking = "ungrouped" in tui.toml to keep the previous layout.

The patch list is long. The useful parts, in the order a daily user would notice them:

Server API changes in the same entry

Builders driving Kimi Code through its server API get breaking changes, which is exactly what a canary is for. Sessions now receive their agent profile, model and thinking level when created, and prompts no longer accept profile: pass agent_config.profile when creating the session instead. Context compaction is reported as a compacting flag on the WebSocket agent frame, and a manual compaction requested while another is running now returns an error instead of being silently ignored. File-history queries take turn_number rather than turn_id. The server also stops sending stack traces in error responses, and reports a missing provider, credential or model as a setup error instead of an internal error. Duplicate background task notifications after a resume or undo are prevented.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Kimi Code 2.5.3: steer a running session