v3.25.0

Changed

• Kimi backend rewritten onto the Kimi Code ACP protocol — Moonshot replaced the Python kimi-cli with Kimi Code, a different product with a different CLI surface, so every flag the old backend spawned with (--print, --input-format, --output-format, --work-dir, -r) is gone and Kimi sessions died at argv parsing. The backend now speaks ACP (Agent Client Protocol) JSON-RPC over stdio via kimi acp, on a new id-correlated transport layer. Note the rebrand moved the version backwards (1.41.0 → 0.26.0), so checkRequirements() probes for the acp subcommand rather than comparing semver — a version comparison would conclude the exact opposite. Old kimi-cli is no longer supported; the failure message points at kimi upgrade.
• Kimi skills install to .kimi-code/skills/ — Kimi Code no longer discovers the legacy .kimi/skills/ directory, so mode skills silently failed to load even once the transport was fixed. The manifest's skillsDir and the matching watcher / shadow-git / evolution path rules all move to .kimi-code/. AGENTS.md remains Kimi's instructions file. Leftover .kimi/ directories in pre-existing sessions become visible to the file watcher once (harmless one-time noise).

Added

• Kimi gains real permission and tool-progress support — ACP exposes a blocking session/request_permission round trip and tool_call / tool_call_update status transitions carrying structured rawInput / rawOutput, so the backend's permissions and toolProgress capabilities are now genuinely true instead of declared-false. Modes that ask for approval (permissionMode: "default") get the same approval flow as Claude Code; the builtin modes all declare bypassPermissions, which maps to ACP's yolo and preserves today's behaviour exactly.
• Kimi model switching and interruption go through the protocol — the model list arrives with session/new's configOptions (no static fallback), switching uses session/set_model, and interrupt is now the session/cancel notification instead of an out-of-band SIGINT.