* Sat Aug 08 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 0.34.0:
* Show a cache-expiry reminder when resuming a long-idle session
or submitting after a long idle stretch
* Restore how the last turn ended (completed, cancelled or failed)
when a session is resumed after a server restart, and keep it in
session listings
* Fix "kimi -p" exiting right after the main turn instead of
waiting for background tasks and subagents to finish
* Keep live sessions stable when an MCP server is removed from the
workspace config or uninstalled with its plugin: its tools stay
registered but calls fail with a removal notice; servers added
mid-session take effect only in new sessions or after /new or
/reload
* Read UTF-16 LE/BE text files by transcoding them to UTF-8
instead of refusing them as binary
* Show retry progress (attempt N of M) during automatic retries
after a failed model request
* /feedback now works for any signed-in user regardless of the
active model
- Changes from version 0.33.0:
* Run the CLI surfaces (interactive TUI, kimi -p, kimi acp,
kimi export, kimi provider) on the agent-core-v2 engine by
default; set KIMI_CODE_LEGACY_FLAG=1 to fall back to the legacy
engine
* Add Kimi Computer Use and Kimi WebBridge as built-in official
marketplace entries in the v2 CLI
* Ask whether to trust the current folder on startup
* /fork no longer switches to the forked session; the current
session stays active with its background tasks running
* Fix MCP OAuth re-authorization always failing with "Invalid
redirect URI" because a stale dynamic client registration
recorded the first flow's callback port
* Surface the MCP structuredContent field and _meta server
metadata to the model instead of dropping them silently
* Wait for MCP startup to finish before the first request while
still opening the interface immediately
* Start the interactive TUI without creating a session
* Add /bug as an alias for /feedback
- Changes from version 0.32.0:
* Add the TurnStarted, UserPromptQueued, TaskStarted and
SessionHeartbeat hook events, enrich hook payloads with the
session title and client type, include the model and profile in
SessionStart, and report SessionEnd as archive for archived
sessions
* Add a [token_counting] config section to choose how context
token counts are derived: measured+estimated (default),
measured or estimated
* Rename [loop_control] max_retries_per_step to
max_attempts_per_step and max_steps_per_run to
max_steps_per_turn; the old keys no longer take effect on the v2
engine and warn at startup
* Fall back to the built-in models.dev catalog snapshot when the
public catalog is unreachable
* Fix submitting answers to interactive question prompts when the
provider returns tool call IDs containing colons
* Drop the KIMI_SNAPSHOT_READER, KIMI_SNAPSHOT_TIMEOUT_MS and
KIMI_SNAPSHOT_CACHE_LIMIT environment knobs
* Mon Aug 03 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 0.31.1:
* Replace the "kimi server" command tree with "kimi web": servers run
in the foreground and stop with Ctrl+C, the background daemon and
OS-service lifecycle commands are gone, and any "kimi server"
invocation now prints a deprecation notice and exits 1
* Add Markdown-defined custom agents with frontmatter, usable as the
main agent or a sub-agent and discoverable from plugins, together
with global tool gating, per-agent delegation restrictions and a
user-level override of the main system prompt
* Add the /secondary_model command and secondary-model bindings so
subagents can run on a different model than the main session
* Let plugins contribute agent system-prompt instructions and custom
agents through kimi.plugin.json
* Add a customizable footer status line, configured via the
[status_line] section of tui.toml
* Select a thinking effort level from ACP clients instead of only an
on/off toggle, and take the offered levels from each model's
declared capabilities so unsupported ones are no longer shown
* Import many more providers from the models.dev catalog, honour
per-model gateway protocol and endpoint overrides, filter out
deprecated and alpha models, and detect the reasoning field per
endpoint so thinking content is no longer lost with newer vLLM
* Deliver videos attached to a prompt together with the prompt
instead of through an extra tool round trip, falling back to
inline delivery when the provider has no upload channel
* Add environment overrides for the agent loop and background task
limits, global default MCP server timeouts, and web search and
fetch configuration without an OAuth login
* Reconnect a dropped MCP server automatically when one of its tools
is called, and retry that call once
* Fix a range of session and turn failures: interrupted requests
being silently retried, sessions wedged after a content-filtered
response, goal turns broken by the per-turn step limit, messages
rejected while a goal was running, /undo not restoring todo lists
and plan mode, and invalid tool calls being retried indefinitely
* Fail fast when the account quota or balance is exhausted instead
of retrying for about three minutes
* Improve TUI performance and resume speed on long sessions, and cut
down full-screen redraws
* Rework the bundled web dashboard: Monaco based code highlighting,
corrected code block sizing and line numbers, a checkerboard
backdrop for transparent images, and permission modes ordered from
safest to most permissive
* Remove the never-functional [platforms] configuration section and
the provider.platformId field