Template v1.0 2026-09-12 — As of Kimi Code v0.42.0

Content / Kimi Code Releases

November 2026 v2.5.0 – v2.6.3 8 releases

Kimi Code v2.6.3 2026-11-17

  • FIX! — fixed duplicate output and a lingering loading spinner after a shell command finished in the foreground
  • FIXStop — fixed external hooks not firing after every completed turn
  • FIX[background] — fixed background notifications incorrectly stopping Tower tasks
  • FIX[background] — restored background task notifications no longer start a new model request when a session is resumed or forked
  • FIXFixed Windows WebBridge updates failing while its daemon is running

Kimi Code v2.6.2 2026-11-14

  • ADD[providers."<id>"] — a custom-registry (api.json) entry can declare optional search_url / feedback_url fields for WebSearch and /feedback
  • CHG/feedback — the text-only “No attachment” option is gone; it now shows the session’s files before confirmation and includes the session’s request_id
  • FIXMulti-second loading delays reopening long sessions, extending the 2.6.1 fix to sessions with many tool calls
  • FIXCtrl-V — a failed or unfinished Kimi media upload is now rejected instead of being cached as an unusable file reference
  • FIX/swarm — fixed the panel missing live subagent updates during execution

Kimi Code v2.6.1 2026-11-13

  • FIXMulti-second loading delays reopening long sessions
  • FIX/compact — long sessions no longer stall with an empty reply instead of auto-compacting when tool results carry notes, errors, or timing headers

Kimi Code v2.6.0 2026-11-12

  • ADD[transcript].tool_call_density / [transcript].thinking — fullscreen mode folds consecutive tool calls and thinking into summary lines by default and hides Activated skill cards for model-loaded skills; set either to "ungrouped", or adjust folding in /settings, to keep the previous layout
  • ADD@ file-path completion now includes hidden and .gitignore-ignored paths by default, ranking non-ignored matches first at equal match quality
  • ADDSame-name workspace Skills now stay individually selectable in a dedicated picker group, labelled by directory, global, or built-in source
  • ADDFullscreen mode — Ctrl-C / Cmd-C copies selected text; toggle with Copy on select in /settings
  • CHG[[permission.rules]] — user and trusted project rules are now persisted and reapplied automatically whenever a session starts, including in print mode, with deny still taking precedence over ask and allow
  • ADD[background] — Tower now tracks artifact missions with key-deliverable review and human acceptance, shown in status and mission views
  • CHG[secondary_model] — primary can be listed again as an ordinary pool alias; Tower reviewers now use the pool default instead of the main model
  • ADDKIMI_CODE_CLIPBOARD_OSC52 — outside SSH and mosh, only the system clipboard is used by default; set to 1 to also send OSC 52
  • ADDKIMI_CODE_TUI_ANIMATION_FPS — adjust the frame rate of in-progress title shimmer animations
  • FIX--environment <id> — the local environment’s default workdir now follows the current primary directory after a primary swap; pass an explicit --environment-cwd to keep the previous pinned-directory behavior

Kimi Code v2.5.3 2026-11-10

  • CHGCtrl-S — steering now interrupts an active model request immediately and continues in the same turn, instead of waiting for the next interruptible point
  • FIXmcp.json — fixed stdio servers configured with script commands such as npx failing to start on Windows

Kimi Code v2.5.2 2026-11-06

  • FIX[mcp] default_deferred — fixed duplicate tool declaration errors in select_tools when MCP tool definitions change

Kimi Code v2.5.1 2026-11-05

  • FIX/sessions — sessions belonging to deleted workspaces no longer appear in the list

Kimi Code v2.5.0 2026-11-04

  • ADDSwitching the model while the agent is working — the change applies from the next turn; a turn that already started calling the model keeps its own settings
  • ADD[environments] / --environment <id> / --environment-cwd <path> — create sessions whose tools run on a configured SSH, Docker, or custom-launcher remote environment
  • ADDProject-level Skills now merge across every configured source directory instead of only the primary one; the primary directory still wins name conflicts
  • ADD[workspace].require_trust — set to false to treat every workspace as trusted and skip the startup trust prompt; takes effect on /reload
  • CHG[secondary_model] force — now ignored with a warning; a lone default_model already pins every subagent to it, and the pool no longer offers an implicit primary choice — list the main model explicitly to use it
  • CHG[database] global_search — now defaults to false; the full-text search index does not build and POST /api/v1/search reports it unavailable until enabled (KIMI_CODE_GLOBAL_SEARCH)
  • ADD/fork — can now fork a still-running session, and fork from a selected historical turn instead of only the latest
October 2026 v2.2.0 – v2.4.1 7 releases

Kimi Code v2.4.1 2026-10-30

  • CHG[read] max_chars — a single file read now returns at most 100,000 characters (~25,000 tokens) by default, down from 500,000; raise both limits here
  • CHGShrank built-in tool and subagent prompts, including repeated goal and plan-mode reminders

Kimi Code v2.4.0 2026-10-28

  • ADD/import-sessions / kimi import-sessions — import Claude Code and Codex sessions from the last 15 or 30 days
  • ADD/provider refresh [providerId] — re-fetches one provider’s model list, or every provider’s with no argument
  • CHGCtrl-O — also expands or collapses streaming thinking output
  • ADDmain-session and subagent status reporting to compatible terminals via OSC 7501, including pending interactions and retained results
  • CHGTower (experimental) — supports per-task model selection and reuses max_running_tasks as a combined worker/reviewer cap; raises the default agent concurrency from 16 to 256
  • ADD[mcp] reconnect_initial_delay_ms / reconnect_max_delay_ms (and the matching KIMI_MCP_RECONNECT_INITIAL_DELAY_MS / KIMI_MCP_RECONNECT_MAX_DELAY_MS env vars) — workspace MCP servers now reconnect automatically with backoff after a dropped or failed connection
  • CHGPlan mode can now only be enabled by the user, not entered automatically by the agent; every plan requires review
  • CHGNever Ask — no longer auto-approves plans
  • CHG-p, --prompt — now skips default_plan_mode and cannot approve a plan when the session is already in Plan mode
  • CHGFullscreen UI — improved approval-option display, mouse selection, and focus switching
  • DELCtrl-T and the old input-area todo panel — removed; todo lists now live in expandable transcript cards
  • CHG/import-from-cc-codex skill improved

Kimi Code v2.3.3 2026-10-26

  • CHG/swarm — turning swarm mode on now shows a high-usage warning
  • FIXAgentSwarm — rejecting a call before any subagent starts (such as while swarm mode is off) no longer shows as a failed swarm; the agent now points to /swarm instead
  • CHG/model — reworded the model switch confirmation prompts
  • CHGswarm/goal start confirmation — merged the two repeated permission-mode notice lines into one sentence each

Kimi Code v2.3.2 2026-10-25

  • CHGAgent — handles tasks directly by default instead of proactively launching subagents; delegation now needs an explicit request in your message, AGENTS.md, a skill, or an enabled multi-agent mode
  • CHGAgentSwarm — the agent can no longer launch it on its own; swarms now start only when swarm mode is turned on, so they’re unavailable in kimi -p or ACP clients such as Zed

Kimi Code v2.3.1 2026-10-23

  • ADDKIMI_CODE_SESSION_LOCK — set to 0 to turn off the cross-process lock that stops two processes writing the same session
  • CHG/sessions / /resume — after a retry, offers to take over a session still held by an exited process; session-in-use errors name the holding client
  • FIXsessions occasionally remaining occupied after an application crash on macOS and Linux
  • FIXworkspace trust prompt — no longer fails when Git Bash is unavailable on Windows

Kimi Code v2.3.0 2026-10-21

  • ADDauto_session_title — auto-generates a session title after the first turn and shows it as the terminal window title; a global config.toml gate and an independent tui.toml opt-out
  • ADD[mcp] default_deferred — defaults every MCP server’s tools to on-demand loading when the experimental tool-select flag is on
  • CHG/model — switching models mid-session with history can show a confirmation dialog, gated by cloud configuration
  • CHG/compact — now queues while the agent is busy instead of failing with an error
  • CHG/skill:<name> — plugin skills are listed and invoked only under their qualified plugin-id:skill-name form; a bare plugin-skill name no longer resolves
  • CHGdisable_feedback_survey — moved from tui.toml to config.toml
  • CHG[permission].dangerous_command_guard — set to false now disables every built-in guard, not only the dangerous-command check
  • FIX[[hooks]] — hook commands now run in the session’s project directory instead of the server’s startup directory

Kimi Code v2.2.0 2026-10-14

  • ADDkimi app [path] — opens a new desktop chat in the current or a given directory; needs a desktop app version that supports it
  • ADDKIMI_CODE_TRUST_WORKSPACE — marks the current workspace as trusted for this process, like the interactive trust prompt
  • CHGCtrl-P / Ctrl-N Updates panel — now shows opt-in mid-turn progress updates, off by default and toggled from /experiments
  • CHGCtrl-S — also sends a highlighted skill prompt, replacing the old skill-suggestion cards
  • CHGtui_mode fullscreen — a scrolled-out user message gets a sticky one-line summary; clicking it jumps back to the original message
  • CHG[background] — the agent’s in-turn wait for a task is capped at 90 seconds and discourages repeated waits; a new message ends the wait early
  • CHGCtrl-C — forces an immediate exit during shutdown hooks instead of waiting for the hook timeout; a “Shutting down” hint shows meanwhile
  • FIXUserPromptSubmit hook — fixed turn ordering and rendering when the hook injects text; hook JSON without a message field is no longer added to the conversation
  • FIXforked sessions — cron tasks from the source session no longer fire inside a fork
  • DELexecutor = "kaos" — removed from MCP stdio configuration (never implemented); configs using it now fail validation at load time
September 2026 v0.42.0 – v2.1.1 7 releases

Kimi Code v2.1.0–v2.1.1 2026-09-24

  • ADDtui_mode (/settings → TUI mode) — experimental fullscreen TUI layout; click a folded block to toggle it, with a clickable “Jump to bottom” indicator
  • DELKIMI_CODE_TUI_FULL_SCREEN — removed now that fullscreen is a tui_mode setting instead of an env flag
  • ADDKIMI_CODE_REPEAT_BREAKER — set to 0 to stop reminders and forced stops for repeated tool calls
  • CHG[workspace].additional_dir — rejects directories that resolve to the home directory or filesystem root
  • FIXhardened workspace security: file tools can no longer escape the working directory via symlinks, project-local config only applies once the workspace is trusted, and background git operations can no longer run commands from repository git config
  • CHGreduced the CLI’s startup time and memory usage
  • CHG[watch] / KIMI_CODE_WATCH — briefly defaulted to off in 2.1.0; reverted to on by default in 2.1.1

Kimi Code v2.0.2 2026-09-19

  • FIX/resume — new messages no longer land at an old position in the conversation after resuming a session
  • FIX/compact — no longer fails after switching to a model with a smaller context window
  • CHGthe agent no longer assumes the current working directory is the project root
  • FIXa message sent while the agent was running no longer sometimes appears twice in the chat

Kimi Code v2.0.1 2026-09-18

  • ADD[watch] / KIMI_CODE_WATCH — disable filesystem watchers for local.toml, AGENTS.md, skills and MCP config entirely
  • CHG[providers."<id>"] — api_key_env names an environment variable to read the API key from, as an alternative to api_key
  • ADD[database] — searchSyncSessionCap / searchSyncDebounceMs tune the incremental background search-sync pass (marked unverified — not yet in the config reference)
  • CHG/yolo / -y, --yolo — Ask When Needed no longer prompts for a Bash command that cannot be statically analyzed
  • CHGkimi install-app — renamed to kimi install-desktop; the old name keeps working as a hidden alias

Kimi Code v2.0.0 2026-09-17

  • ADD/desktop / /install-desktop — opens the Kimi Code desktop app page in your browser
  • ADDkimi install-app — prints and opens the desktop app page (renamed to kimi install-desktop in 2.0.1)
  • ADD[markdown].mermaid — renders fenced mermaid code blocks in the terminal as diagrams (marked unverified — not yet in the config reference)

Kimi Code v0.43.1 2026-09-15

  • CHG/copy — native clipboard write path for Linux X11 sessions
  • FIXCtrl-C while subagents are running now interrupts only the subagents, instead of exiting the whole CLI

Kimi Code v0.43.0 2026-09-14

  • ADD/sessions — Ctrl-X then y deletes the selected session from the picker
  • ADDkimi upgrade / kimi update — -y, --yes skips the confirmation prompt and installs directly
  • ADDloop_control.compaction_max_attempts — caps total attempts for a failing compaction request; default 5
  • ADDKIMI_CODE_PERMISSION_MODE_REMINDER env var — set to 0 to stop injecting auto permission-mode reminders into the model context
  • CHGCtrl-S — a steering message can now interrupt a wait for a background task
  • CHG/goal — the time budget no longer counts time spent with the session closed, and the 24-hour limit is removed

Kimi Code v0.42.0 2026-09-09

  • ADDcard-keys — initial reference: the seven keyboard-shortcut contexts, from general input to the approval panel (19 rows)
  • ADDcard-slash-core — initial reference: account, session, mode and status commands with their aliases (33 rows)
  • ADDcard-slash-tools — initial reference: swarm and goal control, extensibility, and the built-in skill commands (16 rows)
  • ADDcard-cli — initial reference: all 14 top-level flags and all 12 subcommands (25 rows)
  • ADDcard-settings — initial reference: config.toml, tui.toml and .kimi-code/local.toml (37 rows)
  • ADDcard-env — initial reference: core paths, model overrides, runtime switches and diagnostics (22 rows)
  • ADDcard-skills — initial reference: skill and agent directories, the four AGENTS.md locations, plugins (18 rows)
  • ADDcard-hooks — initial reference: 18 hook events, hook exit codes, mcp.json and MCP permissions (20 rows)
  • DELKIMI_CODE_EXPERIMENTAL_REMOTE_CONTROL and KIMI_CODE_EXPERIMENTAL_SECONDARY_MODEL — Remote Control and the subagent model pool are always on in 0.42.0, so both flags were removed
  • CHG[database] replaces the experimental minidb flags; the read model and search worker are now controlled by KIMI_CODE_PERSISTENCE_MINIDB_READMODEL and KIMI_CODE_SEARCH_WORKER
  • ADD[read] — configurable character limits and resumable long-line file reads, replacing repeated output truncation
  • CHGCtrl-O — finished tool calls collapse to a header plus one outcome row; hidden output is counted and revealed here

Template / Structure

  • ADDv1.0 — template derived from claude-code-placemat v1.2
  • CHGPer-agent localStorage keys (pm-kimi-code-seen-version, pm-kimi-code-collapsed); theme and density stay shared across every placemat on this origin