MoonshotAI/kimi-code @moonshot-ai/kimi-code@2.6.0

on GitHub

Minor Changes

• #4752 5aa81c5 Thanks @sailist! - Align fs:suggest with local fd completion, using bounded scans and at most 20 results with optional streamed previews, while including hidden and gitignored paths by default and placing regular files first at equal match quality.

• #4618 e6ae8ed Thanks @Grapedge! - Fold consecutive tool calls and thinking into summary lines in fullscreen mode, and no longer show the Activated skill card for skills the model loads. Choose how much folds under Tool call density and Thinking folding in /settings, with a preview of each choice, or set tool_call_density = "ungrouped" and thinking = "ungrouped" under [transcript] in tui.toml to keep the previous layout.

Patch Changes

• #4751 b64a016 Thanks @Grapedge! - Show how long the current turn has been running in the activity bar while the model writes, and how long the last turn took and when it finished once it completes.

• #4663 1dd5578 Thanks @wbxl2000! - Report context compaction as a compacting flag on the WebSocket agent frame.

• #4839 afadca2 Thanks @wbxl2000! - Server API sessions now get their agent profile, model, and thinking level when created; prompts no longer accept profile, so pass agent_config.profile when creating the session instead.

• #4836 5ad9aa1 Thanks @sailist! - Prevent duplicate background task notifications after resume or undo, and preserve prompt lifecycle completion during turn finalization.

• #4755 322d036 Thanks @Grapedge! - Outside SSH and mosh, copying now goes only to the system clipboard instead of also through the terminal (OSC 52); set KIMI_CODE_CLIPBOARD_OSC52=1 to keep the old behavior.

• #4773 881e61b Thanks @wbxl2000! - Report an error instead of silently ignoring a manual compaction requested while another compaction is running in the web UI.

• #4663 1dd5578 Thanks @wbxl2000! - Stream each subscribed agent's conversation over the server WebSocket as turns, turns.removed, turns.failed, turn.*, item.* and notice frames, and page older turns with GET /sessions/{id}/agents/{agent_id}/turns; subscribe_v2 takes agent_ids instead of transcript grades, prompts with bundled skills and skill activations accept a client prompt_id, and the transcript frames and the GET /sessions/{id}/transcript endpoints (with /ops, /plan and /user-messages) are no longer available.

• #4618 e6ae8ed Thanks @Grapedge! - Fix Ctrl+O expanding too little of the latest turn after the model loads several skills.

• #4708 d0cd33c Thanks @RealKai42! - Stop forbidding multiple Edit calls to the same file and stop requiring a re-read of the file before each Edit.

• #4777 b9c1897 Thanks @wbxl2000! - Stop sending server stack traces in error responses, and report a missing provider, credential, or model as a setup error instead of an internal error.

• #4665 8d1af37 Thanks @Grapedge! - Keep the terminal UI responsive when a lot of tool output is expanded with ctrl+o.

• #4766 adc46cc Thanks @pvzheroes125! - Reload skills when files in explicitly selected skill directories change.

• #4663 1dd5578 Thanks @wbxl2000! - Report a prompt that fails before its turn starts with status failed instead of queued.

• #4826 13590cd Thanks @wbxl2000! - The server API's file-history queries no longer accept turn_id; pass turn_number instead.

• #4822 4ac656c Thanks @liruifengv! - Fix context compaction being cancelled when system reminders arrive during compaction, and keep those reminders in the compacted history.

• #4800 4b4cf3a Thanks @kimi-agent-bot! - Return correct HTTP errors for failed file downloads.

• #4719 bae3211 Thanks @pvzheroes125! - Fix duplicate tool definitions and empty tool announcements after tool updates.

• #4720 df90d90 Thanks @pvzheroes125! - Fix outdated tool definitions for models without dynamic tool loading.

• #4563 1eb3c53 Thanks @kimi-agent-bot! - Fix file previews for non-ASCII filenames.

• #4825 41b7acc Thanks @pvzheroes125! - Fix crashes when starting commands in long Windows paths and file searches in long Windows directories.

• #4832 6dd56ba Thanks @sailist! - Preserve prompt cache when forking an agent after its available tools change.

• #4867 8aacf88 Thanks @sailist! - Preserve tool result notes when forking conversations so side chats keep the same model-visible history and prompt-cache prefix.

• #4755 322d036 Thanks @Grapedge! - Fix copying in fullscreen mode and /copy not working in tmux and on Linux.

• #4755 322d036 Thanks @Grapedge! - Add a Copy on select setting for fullscreen mode; turn it off in /settings to copy only when you press Ctrl-C.

• #4755 322d036 Thanks @Grapedge! - In fullscreen mode, you can now copy selected text with Ctrl-C or Cmd-C.

• #4771 c01a04a Thanks @Grapedge! - Default the TUI to fullscreen mode instead of regular mode when KIMI_CODE_EXPERIMENTAL_FLAG=1 is set.

• #4814 f2ce60f Thanks @Grapedge! - Show NotifyUser updates in fullscreen mode with a primary-colored bullet instead of the Update badge.

• #4827 5518d54 Thanks @7Sageer! - Preserve Gemini thought signatures and native tool context across turns.

• #4735 bc6fa89 Thanks @7Sageer! - Import google-genai providers from a custom registry, including endpoints that already end in /v1beta.

• #4618 e6ae8ed Thanks @Grapedge! - Stop showing the ctrl+o expand and collapse hint in the fullscreen footer while tool or thinking grouping is on.

• #4663 1dd5578 Thanks @wbxl2000! - Answer a prompt resubmitted with the same prompt_id and request with the result of the first submission instead of a conflict error.

• #4618 e6ae8ed Thanks @Grapedge! - Show live thinking as a highlighted Thinking title with an elapsed timer instead of the thinking… spinner.

• #4757 c4f81f8 Thanks @7Sageer! - Apply [[permission.rules]] from config.toml when a session starts, including deny, ask, and allow rules in print mode.

• #4848 9ae2757 Thanks @sailist! - Prevent multiple app instances from restoring the same session and running its pending task turns concurrently.

• #4831 392505c Thanks @sailist! - Fix desktop retry status, partial response history, and background task scheduling.

• #4690 27bd9ba Thanks @liruifengv! - Keep same-name workspace skills selectable in a dedicated TUI Skills group, with directory names or global and built-in labels distinguishing their sources.

• #4550 53d1caf Thanks @sailist! - Persist user and project permission rules and apply them when sessions start.

• #4644 8985b71 Thanks @7Sageer! - Preserve UTF-8 byte order marks when editing files.

• #4800 4b4cf3a Thanks @kimi-agent-bot! - Preserve Unicode filenames when downloading uploaded files.

• #4842 421388b Thanks @wbxl2000! - The server API's session profile endpoint now ignores the reserved metadata keys cwd, goal, parent_session_id, and child_session_kind, as session creation already does.

• #4662 8fc5e22 Thanks @sailist! - Replace protocol details in session upgrade prompts with a concise upgrade message; reader versions remain in error details and diagnostic logs.

• #4699 2a6d1cf Thanks @7Sageer! - Record session title generation outcomes, timing, and title length in anonymous telemetry.

• #4698 2be062e Thanks @7Sageer! - Use the bound working directory in /status and Markdown export metadata, and read session-list git details from each session's environment.

• #4427 1c259c8 Thanks @tpoisonooo! - Allow primary as an ordinary [secondary_model] model alias and remove the reserved PRIMARY_SUBAGENT_MODEL_CHOICE SDK export; Tower reviewers now use the pool default instead of the primary model and no longer accept the special model: "primary" escape hatch, so primary-model reviews require explicitly selecting a pool alias for the primary model or setting it as default_model for all spawns.

• #4855 fb8ea02 Thanks @wbxl2000! - Fix the partial reply of a model request that failed mid-stream staying on screen next to the reply from its automatic retry.

• #4767 97637e5 Thanks @pvzheroes125! - Avoid repeatedly uploading the same MCP tool image across model requests.

• #4617 6974ac8 Thanks @Grapedge! - Keep the file list in expanded Grep and Glob results indented when it wraps.

• #4663 1dd5578 Thanks @wbxl2000! - Remove the source field from POST /api/v1/search responses.

• #4778 c87878a Thanks @wbxl2000! - Apply the agent settings and custom metadata passed when creating a session through the server API; a creation request that turns on Tower mode is now rejected, so turn it on through the session profile once the session exists.

• #4799 11d9fda Thanks @liruifengv! - Fix the skill catalog being omitted from the system prompt when the Skill tool is disabled.

• #4713 cd48f5e Thanks @RealKai42! - Direct models to invoke skills through the Skill tool instead of reading skill files directly.

• #4854 265783c Thanks @wbxl2000! - Fix stopped tool calls showing a generic interruption notice instead of the output they produced before being stopped.

• #4730 ce38ebf Thanks @wbxl2000! - Streamed reasoning and message items no longer get an early item.ended when the next block of their step starts; they now end once, after the step is persisted, so reasoning durations no longer change after the fact.

• #4778 c87878a Thanks @wbxl2000! - Return a validation error instead of an internal server error when a server API request asks for a thinking level the model does not support.

• #4862 7980593 Thanks @sailist! - Fix duplicate tool declaration errors after reloading dynamic tools.

• #4689 51b47db Thanks @liruifengv! - Follow the current primary directory for default local workdirs after a primary swap; use an explicit environmentCwd to retain the previous pinned-directory behavior, and keep tools, project configuration, and session indexes in sync.

• #4772 aa12538 Thanks @wbxl2000! - Fix running tool calls showing as failed after you stop a turn in the web UI.

• #4427 1c259c8 Thanks @tpoisonooo! - Add artifact missions with key-deliverable review and whole-tip human acceptance, limit the LLM-facing TowerStatus detail lists to the newest 20 entries, and show mission kinds in status and mission views.

• #4834 bd62933 Thanks @sailist! - Truncate oversized Skill tool content and preserve the full content in a file.

• #4870 bd75a36 Thanks @Grapedge! - Add the KIMI_CODE_TUI_ANIMATION_FPS environment variable to change the frame rate of the shimmer on in-progress titles. Set it to a whole number from 1 to 60.

• #4663 1dd5578 Thanks @wbxl2000! - POST /sessions/{id}:undo now takes the turn_id and optional agent_id of the turn to undo and returns removed_turn_ids, instead of a count and a page of messages.

• #4846 ac09c02 Thanks @wbxl2000! - Fix two simultaneous undo requests for the same turn also removing the turns before it.

• #4770 823c159 Thanks @pvzheroes125! - Reload skills when files change in extra skill directories.

• #4871 54087b8 Thanks @kimi-agent-bot! - web: Improved interactions and fixed known bugs.

• #4747 79ff88b Thanks @hunshcn! - Fix slow workspace operations when many projects are registered.

• #4823 7a0205c Thanks @wbxl2000! - Fix server websocket connections being closed with a heartbeat timeout after a server stall or a system clock jump.