Skip to main content

Kimi Code

Kimi Code support is measured on one live 0.42.0 session (2026-09-09): the install path, the three hook registrations, and the print-mode gap below all come from that run, not from a fleet. Ruling pre-action checks are not wired for this host yet.

Install​

Adds daimon's capture -> inject loop to Kimi Code from the released package:

daimon hooks install kimi

This copies three hook scripts, plus the modules they share, into the Kimi config directory, then appends [[hooks]] entries to ~/.kimi-code/config.toml (or $KIMI_CODE_HOME/config.toml when that variable is set). The config file is backed up before daimon writes to it. The writer only ever appends or removes daimon's own blocks. It never rewrites the provider or model tables.

Re-run daimon hooks install kimi after every uv tool upgrade daimon-briefing, same as every other host, so the installed scripts match the CLI version. Hooks load once, at session start: after installing, start a new Kimi session. The one already running will not pick up the change.

Requires the daimon CLI on PATH:

uv tool install 'daimon-briefing[pretty]'

What each script does​

  • daimon-kimi-user-prompt-submit.py: UserPromptSubmit hook. This is the only channel Kimi gives a hook to put text into the session, so it carries the briefing. Kimi has no session-start injection point (see Limitations below). The model sees the briefing as a user message wrapped in a hook_result tag, attached to your first prompt.
  • daimon-kimi-session-end.py: SessionEnd hook. Serializes the finished session when an interactive Kimi session exits, including the turns after the last Stop capture. Bytes Stop already captured are not re-serialized: the CLI checks the transcript against the last checkpoint before any model call.
  • daimon-kimi-stop.py: Stop hook. Runs a throttled capture after each turn. This is crash insurance for an interactive session, and it is the only capture path in print mode, where SessionEnd never fires (see Limitations).

Limitations​

  • No briefing at session start. Kimi's SessionStart hook is observation only, its stdout is dropped. The briefing arrives with your first prompt instead, through the UserPromptSubmit hook above.
  • Print mode never closes. kimi -p sessions stay resumable and never fire SessionEnd, measured twice on a live session. The Stop hook is what captures them. On an interactive exit SessionEnd runs its own capture regardless of a recent Stop, so the turns after the last Stop are not lost; the CLI compares the transcript against the last checkpoint first, so bytes Stop already captured cost a file read, not a second model call.
  • A resumed session is not briefed again. The briefing marker is keyed on the session id, so kimi -r on a session that already received its briefing gets the per-prompt recall injection but no second briefing. Resume itself has not been measured on a live session.
  • No pre-action checks yet. Kimi's deny channel has not been measured on a live session, so this release ships no check profile for it.
  • Hooks load at session start only. Installing does not reach a session already running. Start a new one to pick up the change.
  • macOS records the resolved path. Kimi stores the working directory it resolves to, so a session started under /tmp/x is recorded as /private/tmp/x. Keep that in mind when you look sessions up by directory.

Transcripts​

Kimi writes each session's transcript to ~/.kimi-code/sessions/<workspace>/<session id>/agents/main/wire.jsonl. The hook payload carries the session id, not a path, so daimon resolves the transcript by globbing for that session id under the sessions directory. Subagent transcripts are not merged into the main transcript in this release.

Kimi is a registered source host: daimon why <item-id> --source and daimon audit quotes resolve a Kimi checkpoint's transcript the same way they do for Claude Code, Codex, and Windsurf. Set $KIMI_CODE_HOME and both commands look under it, matching the install path above. Checkpoints captured before this host was registered were stamped with an unresolvable source and stay that way; only checkpoints captured afterward resolve.

why --source goes one step further for Kimi than it does for Codex and Windsurf: every folded message (each context.append_message/lifecycle row, and each tool.result) carries a stable id, so a verbatim item captured after this landed discloses the exact quoted message, the same way it does for Claude Code, instead of falling back to the stored quote. daimon does not read a per-message id from Codex or Windsurf transcripts yet, so their disclosure still falls back. A Kimi checkpoint captured before this landed carries no message ids either and keeps falling back too.

MCP​

daimon hooks install kimi registers the read-only MCP server automatically: it merges an mcpServers.daimon entry into ~/.kimi-code/mcp.json (or $KIMI_CODE_HOME/mcp.json when set), the file Kimi's own docs describe for this scope — separate from config.toml, which carries the hook registration. That entry points at a copied wrapper script under the Kimi hooks directory, audited by daimon hooks status the same way as the three hook scripts. daimon hooks remove kimi takes the mcp.json entry and that wrapper back, together with the config.toml hook entries. Once the mcp.json entry exists, the recall hint (the per-prompt "you worked on this before" pointer) names the daimon_recall tool instead of the daimon recall "..." shell command.

(An earlier version of this page said Kimi reads a project-root .mcp.json in the same format Claude Code uses. Checked against a live install: the shape is right — {"mcpServers": {...}} — but the path is not; Kimi reads ~/.kimi-code/mcp.json, never a project-root file.)

Teach the agent the protocol​

daimon skill install kimi              # ~/.kimi-code/skills/daimon/SKILL.md
daimon skill install kimi --project    # <repo>/.kimi-code/skills/daimon/SKILL.md

Install delivers two skills: daimon (the protocol) and daimon-end, the in-session /daimon-end checkpoint flow, written next to it as ~/.kimi-code/skills/daimon-end/SKILL.md.

Kimi scans both locations for skills. The probe measured where it looks, not how it ranks the two when both hold a daimon skill, so install to one scope. Re-run install after upgrading daimon to refresh the content.

Remove​

daimon hooks remove kimi

This takes daimon's [[hooks]] entries back out of config.toml and leaves everything else in the file as it was, line endings included. The one exception: a file whose last line had no newline gains one, and remove cannot tell it from one you wrote. The three hook scripts stay where they are, inert once unregistered. The MCP wrapper (see MCP above) is the exception: it is deleted along with the mcp.json entry that pointed at it, since nothing else legitimately points at that file. Checkpoints under ~/.daimon/ are untouched.

Verify​

daimon status

daimon status reports capture health for Kimi the same as any other host. A capture made through the Stop throttle counts the same as one made through SessionEnd; nothing in the report singles out print-mode sessions.