Skip to content
Proposals/Switch /model and /effort mid-run in Kimi Code 2
proposaldeveloperP5Worth a lookKCKimi Code

Switch /model and /effort mid-run in Kimi Code 2.5.0

Kimi Code 2.5.0 lets you run /model and /effort while the agent is working (in-flight turn keeps its settings; Compact and continue waits until it ends), fork a live session or a historical turn, merge skills across multi-directory projects, and create sessions with --environment <id> for SSH, Docker, or custom-command environments.

Why this loop

Kimi Code 2.5.0 removes the wait-for-idle tax on model and effort changes. A turn that has already started calling the model keeps its settings; /model and /effort apply from the next turn, and Compact and continue runs only after the current turn ends, so you can steer cost and depth without aborting work in flight. Forking a running session or a selected historical turn lets you branch an alternate approach while the original continues. Creation-time SSH, Docker, and custom-command environments are selected with --environment <id> and optional --environment-cwd <path> when the session is created, so remote work starts in the right workspace. Skills now merge from every configured directory in multi-directory projects, with the primary directory taking precedence on name conflicts. If you still used [secondary_model] force, drop it: the key is ignored with a warning, and a lone default_model still pins every subagent. Global full-text search is off by default; set [database] global_search = true or KIMI_CODE_GLOBAL_SEARCH=1 only if you need POST /api/v1/search.

Proposed actions

  1. In the live Kimi Code session, send /model while the agent is still working; do not cancel the turn—the new model applies from the next turn only.
  2. In the live session, send /effort while the agent is working. If you also want Compact and continue, select it now; it will run after the current turn ends.
  3. Fork this running session, or fork from a selected historical turn, and continue the alternate approach in the fork without stopping the original.
  4. Create the next session with --environment <id> and --environment-cwd <path> so SSH, Docker, or custom-command environments are attached at creation time.
  5. Put skills in every configured project directory (primary directory wins on name conflicts). Remove [secondary_model] force from config. If you need global session search, set KIMI_CODE_GLOBAL_SEARCH=1 or [database] global_search = true.

Agent prompt

Paste into your agent or query via MCP (get_agent_prompt) — free, no extra AI cost

AGENTS.md / CLAUDE.md / GEMINI.md rule update

DevAgentRadar → Repo Harness Rule Patch

Goal: update our repository's permanent harness instructions based on this release.

Update the file that tool actually reads. Do not dump everything into one file.

  • Claude Code → CLAUDE.md plus .claude/ (skills, hooks, settings, agents, commands). It does not read AGENTS.md natively; start CLAUDE.md with @AGENTS.md.
  • Codex, Copilot, Cursor, Factory Droid, Grok Build, Roo Code, Goose, OpenCode, Amp, Zed, Aider → AGENTS.md.
  • Gemini CLI / Antigravity → GEMINI.md. AGENTS.md only if context.fileName is set.
  • Cursor glob-scoped rules → .cursor/rules/*.mdc (plain .md is ignored), not a second constitution.
  • Codex MCP → .codex/config.toml, not .mcp.json.
  • Copilot extra instructions → .github/copilot-instructions.md; custom agents → .github/agents/.
  • Windsurf → .windsurf/rules (do not assume AGENTS.md).
  • Cline → .clinerules.
  • Procedures → a skill (SKILL.md). Enforcement the model must not skip → a hook. Isolated roles → subagents.
  • Do not fork the same rule into three tool files.

Context

Assistant: Kimi Code Proposal: Switch /model and /effort mid-run in Kimi Code 2.5.0 Summary: Kimi Code 2.5.0 lets you run /model and /effort while the agent is working (in-flight turn keeps its settings; Compact and continue waits until it ends), fork a live session or a historical turn, merge skills across multi-directory projects, and create sessions with --environment <id> for SSH, Docker, or custom-command environments. Primary source: https://github.com/MoonshotAI/kimi-code/releases/tag/%40moonshot-ai/kimi-code%402.5.0

Why it matters

Kimi Code 2.5.0 removes the wait-for-idle tax on model and effort changes. A turn that has already started calling the model keeps its settings; /model and /effort apply from the next turn, and Compact and continue runs only after the current turn ends, so you can steer cost and depth without aborting work in flight. Forking a running session or a selected historical turn lets you branch an alternate approach while the original continues. Creation-time SSH, Docker, and custom-command environments are selected with --environment <id> and optional --environment-cwd <path> when the session is created, so remote work starts in the right workspace. Skills now merge from every configured directory in multi-directory projects, with the primary directory taking precedence on name conflicts. If you still used [secondary_model] force, drop it: the key is ignored with a warning, and a lone default_model still pins every subagent. Global full-text search is off by default; set [database] global_search = true or KIMI_CODE_GLOBAL_SEARCH=1 only if you need POST /api/v1/search.

Suggested actions

  1. In the live Kimi Code session, send /model while the agent is still working; do not cancel the turn—the new model applies from the next turn only.
  2. In the live session, send /effort while the agent is working. If you also want Compact and continue, select it now; it will run after the current turn ends.
  3. Fork this running session, or fork from a selected historical turn, and continue the alternate approach in the fork without stopping the original.
  4. Create the next session with --environment <id> and --environment-cwd <path> so SSH, Docker, or custom-command environments are attached at creation time.
  5. Put skills in every configured project directory (primary directory wins on name conflicts). Remove [secondary_model] force from config. If you need global session search, set KIMI_CODE_GLOBAL_SEARCH=1 or [database] global_search = true.

Config surfaces this release may change

  • skills — check your repo before applying

After you finish

Do not report this as applied to DevAgentRadar. You cannot write the visitor's loop.

Tell the human: open https://devagentradar.com/proposals/kimi-code-moonshot-ai-kimi-code-2-5-0-switch-model-and-effort-mid-run-in and mark Applied, Skipped, or Failed. Proposal id: 4bd687f5-c6d8-485e-9f6f-409cd7c291e3

Instructions:

  1. Read the existing context files and settings for the tools in this repo.
  2. Draft an explicit Git diff. Update ONLY agent harness rules or tool configuration.
  3. Call out deprecated flags, obsolete habits, or changed permission boundaries.
  4. Do not touch application logic.

Start with the proposed diff and say which layer and which tool's file it belongs in.

agentmcpmodelbreakingRelease source ↗

Your loop

This browser · no sign-in · not shared as “you”

After you run the prompt

Only you can mark this. Agents cannot write your loop.

Your decision stays on this device. A public tally appears after a few votes.

Originating release signal

KCKimi Code@moonshot-ai/kimi-code@2.5.0Nov 4, 2026

@moonshot-ai/kimi-code@2.5.0

#4621 1b13934 Thanks @Grapedge! - Allow /model and /effort while the agent is working; a turn that has already started calling the model keeps its settings, the switch applies from the next turn, and choosing Compact and continue runs after the current turn ends. · +34 more changes
Verified excerpt — the source's own words

Minor Changes

  • #4621 1b13934 Thanks @Grapedge! - Allow /model and /effort while the agent is working; a turn that has already started calling the model keeps its settings, the switch applies from the next turn, and choosing Compact and continue runs after the current turn ends.

  • #4544 e24e1e1 Thanks @sailist! - Add support for forking running sessions and forking from a selected historical turn.

  • #4520 3ed0e68 Thanks @liruifengv! - Merge skills from every configured directory in multi-directory projects, with the primary directory taking precedence on name conflicts.

Excerpt ends here — this release continues at the source ↗.

Primary source ↗