Kimi Code changelog: what's new each month
Every stable Kimi Code release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 7 months covered; the current month updates daily.
November 2026 · so far
14 releases: 2.5.0 → 1.4.2
November brought major improvements to session management, file handling, and agent control, with new support for external agent integration and Amazon Bedrock providers. The month focused on stability, performance, and expanding integration capabilities.
- Text-only /feedback submissions no longer available; feedback dialog now requires attachments.
Highlights
- Fork running sessions and resume sessions open in other clients for better collaboration.
- Amazon Bedrock provider support and experimental next-prompt predictions in the terminal.
- Steering now interrupts active model requests and continues in the same turn via new API endpoint.
- Improved file suggestions and remote file operations for faster performance.
- Connect external agents like Claude Code and Cursor to hand off sessions and collaborate across turns.
New
- Amazon Bedrock provider support with AWS_REGION configuration.
- Fork running sessions and fork from selected historical turns.
- POST /api/v1/sessions/{session_id}/steer endpoint for one-request message steering.
- Creation-time SSH, Docker, and custom-command environments with --environment flags.
- Experimental next-prompt predictions after completed turns in the terminal.
- Follow-up behavior setting to steer messages into running turns instead of queueing.
- /tasks command to view running foreground agents and their activity.
- kimi server command to run local server without opening browser, with --open option to open UI.
Improved
- Merge skills from every configured directory in multi-directory projects with precedence handling.
- Agent status, settings, context usage, goal and todo list sent over WebSocket frames.
- Significantly faster remote file reads, edits, writes, search, and completion.
- fs:suggest now uses bounded scans with up to 20 results and optional streamed previews.
- Fold consecutive tool calls and thinking into summary lines in fullscreen mode.
- Show turn duration in activity bar while model writes and completion time after turns finish.
- Plugin marketplace interactions improved.
- Allow /model and /effort commands while agent is working.
Fixed
- Repeated HTTP 400 errors from empty system messages in dynamic tool schemas.
- Session titles not generating automatically after first turn of new sessions.
- Local images not displaying when paths contain spaces.
- Missing image and video previews in live and reloaded tool results.
- Long sessions stalling with empty replies instead of compacting when tool results include notes.
- Multi-second loading delays when reopening long sessions.
October 2026
7 releases: 2.2.0 → 2.4.1
October brought major improvements to session management, swarm orchestration, and user control over agent behaviour. The month focused on giving users more explicit control over when agents enter special modes and better visibility into multi-client sessions.
- Agent swarms now start only when you explicitly enable swarm mode; the agent can no longer launch AgentSwarm by itself.
- Plan mode now starts only when you explicitly enable it; the agent can no longer enter Plan mode by itself.
- single file reads capped at 100,000 characters instead of up to 500,000 characters.
Highlights
- Sessions now unload from memory after 30 minutes idle and resume transparently, reducing resource use.
- Plan mode and swarm mode now start only when explicitly enabled, preventing agents from entering them automatically.
- Model switching in active sessions now prompts you with options to start fresh, compact and continue, or continue directly.
- Import sessions from Claude Code and Codex from the last 15 or 30 days with /import-sessions.
- Single file reads are now capped at 100,000 characters to prevent loading overly large files.
New
- NotifyUser progress update feature shows the agent's mid-turn progress with optional title.
- kimi app [path] command opens a new desktop chat in the current or specified directory.
- Tower mode: sending a message while the agent runs aborts the turn and runs as its own turn.
- KIMI_CODE_TRUST_WORKSPACE environment variable marks a workspace as trusted.
- Sticky user message summaries in fullscreen mode that jump to the original message when clicked.
- Ctrl-O expands thinking during streaming.
- /provider refresh [providerId] re-fetches provider model lists at runtime.
- OSC 7501 terminal status reporting for main-session and subagent status including pending interactions.
Improved
- Tower mode now has deterministic orchestration with TowerRebase and LLM-free status, with goal/tower mutual exclusion and workspace protection.
- Plugin skills now display with qualified <pluginId>:<skillName> names so same-named skills stay visible.
- Remote control streaming now supports up to 100 MiB transfers with concurrent transfers bounded, up from 10 MiB.
- Multi-client sessions show an occupancy prompt with retry/cancel options when a second client sends a message.
- Session titles are automatically generated after the first turn and shown in the terminal window title.
- Tower's default agent concurrency limit raised from 16 to 256.
- Terminal footer no longer blocks while reading git status.
- Session events now continue after a session is archived and restored.
Fixed
- Workspace trust failures when Git Bash is unavailable on Windows.
- Sessions occasionally remaining occupied after application crashes on macOS and Linux.
- Failed swarms no longer display when AgentSwarm calls are rejected before any subagent starts.
- Cancelled questions showing as answered in session history.
- Sessions left locked by an exited process can now be taken over and used again.
September 2026
11 releases: 0.40.0 → 2.1.1
September brought major updates to Kimi Code including a jump to version 2.0, new collaboration features, and numerous stability improvements. The month focused on refining the user experience, improving performance, and introducing experimental multi-agent modes.
- Subagent model pool ([secondarymodel]) now always enabled; KIMICODEEXPERIMENTALSECONDARYMODEL environment variable removed.
- Turn-level file history experimental flag removed; feature always enabled.
- kimi acp subcommand no longer honors KIMICODELEGACYFLAG; always runs on default agent engine.
Highlights
- Tower multi-agent collaboration mode added (experimental) for coordinating multiple agents on tasks.
- AI session titles now generated automatically after the first turn, with ability to regenerate.
- Subagent model pool enabled by default for improved agent capabilities.
- Media rail in the web composer lets you preview and manage images and videos while composing.
- Significant performance improvements for startup time, session loading, and long conversation handling.
New
- Plugins panel in Settings for browsing, installing, enabling, disabling, and removing plugins.
- Tower mode with /tower command for experimental multi-agent collaboration with base branch support.
- Session rating prompt appears occasionally above the input box.
- Experimental Updates panel showing paginated progress messages from main and subagents.
- Code wrap toggle in the diff panel.
- Desktop app installer via /desktop command and kimi install-app subcommand.
- Mermaid diagram rendering in terminal code blocks.
- TUI fullscreen layout mode (experimental) with click-to-fold interaction for folded blocks.
Improved
- Turn-level file history now always enabled; experimental flag removed.
- Bash tool's cwd parameter no longer restricted to workspace.
- Config.toml updates now preserve comments, key order, and formatting.
- Background question answers delivered directly to agent instead of via saved output file.
- Auto permission mode no longer blocks dangerous or statically unanalyzable commands.
- Conversation scrollbar made thicker and easier to click and drag.
- Finished tool calls collapsed in transcript to header plus outcome row with optional output expansion.
- Session startup and loading performance significantly improved with incremental search indexing.
Fixed
- New content not appearing after auto-compaction.
- Models and providers transiently disappearing when config.toml saved non-atomically.
- Background questions being cancelled when agent finished its turn.
- Session titles in forked sessions being overwritten by auto-generated titles.
- Occasional crash when cancelled turn was running.
- Steered messages appearing twice after turn interruption.
August 2026
12 releases: 0.32.0 → 0.39.1
August brought major workflow improvements including multi-skill activation, session management enhancements, and experimental features for remote access and multi-agent orchestration. The month focused on refining existing features and expanding mobile and web UI capabilities.
- Loopcontrol config keys renamed: maxretriesperstep to maxattemptsperstep, maxstepsperrun to maxstepsperturn.
- CLI surfaces (interactive TUI, kimi -p, kimi acp, kimi export, kimi provider) now run on agent-core-v2 engine by default; set KIMICODELEGACYFLAG=1 to revert.
- Edit and Write tools now require reading an existing file before modifying it and reject writes when the file changed on disk since last read.
Highlights
- Activate multiple skills in a single prompt and use the new WaitFor tool to have agents wait for background tasks to complete within the current turn.
- Web sessions can now be marked as done to keep the open list focused, with a new cross-workspace session management page for triage.
- Mobile support improved with bottom-sheet panels for slash commands and file mentions, plus a flat/by-workspace session list view.
- Windows native CLI now supports automatic updates, and Windows users can install the built-in Kimi Computer Use capability.
- Experimental features added for remote access and multi-agent tower mode orchestration.
New
- Multiple skill activation: type / after whitespace to insert additional skill tokens in a single prompt.
- WaitFor tool allows agents to wait for background task completion within the current turn instead of ending and being re-invoked.
- Remote Control experimental feature for accessing local web sessions remotely (enabled via environment variable).
- Tower mode experimental feature for multi-agent orchestration (enabled via environment variable).
- Session Done/Reopened status to keep the open sessions list focused, with cross-workspace session management page.
- Windows native CLI automatic update support.
- Windows support for built-in Kimi Computer Use capability on x64 systems.
- OAuth support for multiple login methods (kimi.ai and kimi.com).
Improved
- Web UI now shows flat and by-workspace tabs in the mobile session list.
- Right sidebar redesigned as a multi-tab panel with improved organization.
- Slash command and @-mention panels converted to bottom-sheet grab handles on mobile for better accessibility.
- Running foreground Bash commands or subagents can now be moved to background with a button.
- LaTeX math formulas ($ / $$) now render as Unicode formulas in messages.
- Image and video tool results open in fullscreen preview with zoom support on click.
- Background agent panel renamed from Subagent and shows live work progress of subagents.
- Web UI stutter reduced while AI responses stream in long sessions.
Fixed
- Token counts after compaction now match real context size instead of reporting far below.
- Multi-second freezes at startup resolved when large search index loads, replays, or rebuilds.
- Sessions no longer hang on approval prompts when using self-hosted OpenAI-compatible endpoints.
- Pasted images and videos now reach the model on first send.
- Config file entries no longer lost when file has syntax errors or external edits.
- Bare URLs in chat output no longer absorb following CJK characters that broke links.
July 2026
27 releases: 0.21.0 → 0.31.1
July brought major engine standardization, improved image handling, and extensive stability fixes. The local server now runs on the modern agent-core-v2 engine by default, and custom agents can be defined via Markdown files.
- kimi server replaced with kimi web; background daemon and OS-service lifecycle commands removed.
- selecttools model capability renamed to dynamicallyloadedtools.
- Micro-compaction now defaults off (was on by default).
Highlights
- Plugins can now provide slash commands and custom agents; agents can be defined as Markdown files with optional restrictions on sub-agent delegation.
- Images are automatically compressed before reaching the model, with original and compressed details sent together; videos in prompts now reach the model directly without extra tool round trips.
- Thinking effort persists across turns for Kimi models by default; support added for thinking level selection in IDE clients via ACP.
- Web UI redesign with new design system, colors, typography, and improved responsive layout; session export for diagnostics.
- Foreground-only web server lifecycle: kimi web now always starts fresh, and background daemon commands are removed.
New
- Server-side key-value store API for persisting web UI preferences.
- Click-to-enlarge for images in web chat; any file type can be attached and uploaded to server.
- Archived sessions page in Settings to browse and restore archived sessions.
- Swarm footer replaced with inline tool card showing live subagent progress.
- /copy slash command to copy last assistant message to clipboard.
- /export-debug-zip command and session diagnostic export to download session and logs as ZIP.
- AskUserQuestion now feeds answers back as structured data with question text and option labels.
- Global tool gating to constrain which tools agents may use, with per-session override.
Improved
- Conversation compaction now keeps only recent user prompts plus summary; micro-compaction defaults off.
- Compaction summaries show forward plans instead of only immediate next steps; visible in TUI with Ctrl-O.
- Model alias overrides survive provider catalog refreshes; set under [models."<alias>"].overrides.
- Thinking effort routing improved: non-Kimi providers preserve configured values, Kimi models validate runtime selections.
- Default per-step LLM retry budget increased from 3 to 10 attempts with exponential backoff.
- Subagent timeout increased from 30 minutes to 2 hours, configurable via [subagent] timeoutms.
- Foreground Bash commands that hit timeout move to background instead of being killed.
- Goal completion and updates produce final outcome summaries; goal wall-clock budgets enforced during work.
Fixed
- TUI rendering issues causing blank screens and missing input boxes; screen now clears fully on session switch.
- Sessions with interrupted tool calls between call and result no longer silently drop later messages.
- Strict provider errors from non-adjacent tool calls/results and alternating-role violations.
- Duplicate transcript content in scrollback during streaming.
- Image compression note leaking into visible messages and session titles.
- Thinking content rendering as blank bullet lines; empty thinking steps in preserved-thinking history no longer stuck sessions.
June 2026
26 releases: 0.7.0 → 0.20.3
June brought major feature expansions across the CLI and a new web interface, with improvements to goal management, provider support, and session handling throughout the month.
- Project-level local config moved to .kimi-code/local.toml (recommend adding to .gitignore).
- Anthropic-compatible providers no longer read ambient Anthropic shell credentials and custom headers.
- YOLO mode no longer prompts before writing or editing files outside the working directory.
Highlights
- Launched a server-hosted web UI accessible via the browser with chat, session management, and real-time collaboration features.
- Added shell mode for running shell commands directly and expanded agent swarms with concurrency control.
- Introduced multi-directory workspace support with /add-dir command and local project configuration in .kimi-code/local.toml.
- Implemented Agent Client Protocol (ACP) support so IDEs like Zed and JetBrains can drive Kimi Code.
- Made goals, background questions, and sub-skill discovery available by default without experimental flags.
New
- Web UI server with REST and WebSocket APIs, session list, auto-scroll, and browser-based chat interface.
- Shell mode for executing shell commands as a first-class interaction.
- All-sessions picker view with search, pagination, and clipboard-ready resume commands across directories.
- Session visualizer launched via kimi vis command in your browser.
- Custom color themes defined as JSON files in ~/.kimi-code/themes/ or generated via /custom-theme skill.
- Multi-directory workspace support with /add-dir command and project-level .kimi-code/local.toml configuration.
- Agent swarms via /swarm command with live progress and rate-limit-aware retries.
- Import from Claude Code and Codex via /import-from-cc-codex command.
Improved
- Provider model metadata refreshes in the background so new models appear without restarting.
- Micro compaction enabled by default while keeping opt-out flag available.
- Goal queue system allows preparing multiple goals for sequential execution with /goal next.
- Built-in skills now appear as direct slash commands and group ahead of external skill commands.
- File tool guidance routes incremental edits through Edit for better incremental workflows.
- Web login flow redesigned with single "Authorize in browser" button and embedded device code.
- Session snapshot loading accelerated with direct disk reader and request timeout safeguard.
- Glob now uses ripgrep, respects .gitignore by default, and supports brace patterns.
Fixed
- External editor (Ctrl+G) on Windows no longer depends on /bin/sh and uses platform-aware shell quoting.
- Device login keeps URL and code visible when browser cannot be opened.
- Goal marker text no longer overflows terminal width.
- Provider requests no longer fail when restored conversation history contains empty text content blocks.
- Image format detection now reads from file contents, so mismatched filename extensions don't produce rejected data URLs.
- ACP editors in Zed and JetBrains can now successfully start new threads.
May 2026
3 releases: 0.4.0 → 0.6.0
May brought major improvements to plugin management, session handling, and export capabilities for Kimi Code. The month also introduced auto permission mode, scheduled tasks, and better support for custom model configurations.
- Default per-turn step limit of 1000 removed; users can set maxstepsperturn in config for custom limits.
Highlights
- Install plugins directly from GitHub URLs with trust level indicators and manage them interactively.
- New auto permission mode streamlines tool access decisions with /auto command and --auto CLI flag.
- Export sessions as Markdown or debug ZIP archives directly from the terminal interface.
- Run Kimi Code against any specific model via KIMIMODEL environment variable.
- Improved background agent status visibility shows real terminal states and recovery information.
New
- User-global plugin installation with interactive management and plugin-provided skills.
- Plugins can now own MCP servers.
- Scheduled tasks feature for automation.
- KIMIMODEL environment variable lets you specify model provider, base URL, API key, context size, and capabilities.
- /export-debug-zip slash command exports current session as debug archive.
- /export-md slash command exports current session as Markdown.
- /auto slash command and --auto CLI flag for auto permission mode.
- Query sessions by sessionId or workDir in listSessions with helpful cd command when resuming from different directory.
Improved
- Paste markers now expand back to original content when pasting again on a marker.
- Tool permissions reworked: reads outside current directory no longer prompt, session approvals match exact calls, and path-based rules are case-insensitive.
- Full Bash command visible in expanded tool card view with ctrl+o.
- Session title in terminal window/tab shortened from 80 to 32 characters for better readability.
- Inline todo panel capped at five rows with +N more indicator.
- Session export manifest includes more diagnostic information.
- Message compaction handles edge cases and improves retry logic.
- Official datasource tools now preserve complete responses and write returned result files.
Fixed
- TUI no longer crashes when pull request lookup fails during startup.
- Thinking spinner no longer leaks past turn end from empty thinking deltas.
- Official datasource tools preserve complete responses.
- Failed compaction retries now back off by fixed slice of model context window.
- Context overflow errors handled consistently across provider responses.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Kimi Code: pricing, features and alternatives · All changelogs