plan-split

Runs the main loop on one model in plan mode and on another in every other mode, switched by itself, set with /plan-split and off until /plan-split on.

Plan with a model that thinks harder and write the code with another: by default plan mode runs on Fable and every other mode on Opus. You do not pick the model by hand when you enter or leave plan mode; the mod changes the model of each main-loop request to the side of the mode it runs in.

What it does

• Plan mode (plan): each main-loop request goes to the plan model.
• Every other mode (default, acceptEdits, auto, dontAsk, bypassPermissions): each main-loop request goes to the code model.
• No known mode: the request keeps the session's model. A session learns its mode with its first prompt.
• A subagent keeps its own model. The mod changes main-loop requests only.

The mode reaches the mod from three places:

1. The prompt that starts a turn carries the mode the turn starts in, so Shift+Tab between two prompts applies at the next prompt.
2. Each tool call carries the mode after it, so Shift+Tab during a turn applies after the next tool call.
3. An accepted EnterPlanMode or ExitPlanMode moves the side at once. After you approve a plan, the next request of the same turn already runs on the code model. A rejected ExitPlanMode keeps plan mode.

Each change of model shows in the sidebar section plan-split (plan · claude-fable-5-1[1m], and the two settings under it). With the sidebar closed or not installed, a change writes one transcript line:

plan-split: plan mode: the main loop runs on claude-fable-5-1[1m]

Models

A setting is a Claude Code alias (fable, opus, sonnet, haiku) or a model id (claude-fable-5-2). An alias reads the same way Claude Code reads it: with ANTHROPIC_DEFAULT_FABLE_MODEL (and the OPUS, SONNET, HAIKU variables) set, the alias stands for that id, [1m] suffix included; without the variable, the alias goes to the engine as written and answers on the newest model of that family (measured on 2.1.322: sonnet answered as claude-sonnet-5-5). So an alias follows new model releases, and a model id stays fixed.

Cost of a switch

A model reads only its own prompt cache. Measured on 2.1.322: the first request after a switch wrote the whole conversation to the new model's cache (about 94,000 tokens, none read), and the request after it read the cache normally. Switching back to a model that ran earlier in the session read that model's old cache, when it was still warm. So each plan-to-code switch costs one full cache write on the code model, and a long plan conversation makes that write larger.

Command

/plan-split               the status: on or off, both models, the mode the main loop is in now
/plan-split on            route by mode from the next request on
/plan-split off           every request keeps the session's model again
/plan-split plan <model>  the plan mode model, default fable
/plan-split code <model>  the model of every other mode, default opus

The settings live in the store every window shares and are read at each prompt, so a change in one window applies in the others at their next prompt.

Install

claude plugin marketplace add KilimcininKorOglu/claude-code-mods
claude plugin install plan-split@kilimcininkoroglu-mods

Function hooks are early access. Claude Code 2.1.288 and later load them by default, so there is nothing to switch on.

After installing

1. Restart Claude Code.
2. Run /plan-split on. The mod is off after an install and changes nothing until then.
3. To use other models, run /plan-split plan <model> and /plan-split code <model>.

What it can reach

Validated with claude plugin validate on Claude Code 2.1.322:

❯ ./register.ts hooks: session.start, command.run{command=plan-split}, prompt.submit, classic.UserPromptSubmit, classic.PostToolUse, tool.call{tool=/"^(EnterPlanMode|ExitPlanMode)$"/}, turn.step
❯ ./register.ts calls: $.command.register, $.env.get (via aliasEnv), $.sidebar.clear (via dropSection), $.sidebar.set (via toPerson), $.store.get (via readSettings), $.store.set (via setModel, switchTo), $.ui.log (via toPerson)

Reach L2: it changes the model of Claude's own requests.

1. Reads:    the permission mode of each prompt and tool call, the result of EnterPlanMode and ExitPlanMode, the model of each main-loop request, and the four ANTHROPIC_DEFAULT_*_MODEL variables
2. Runs:     nothing
3. Sends:    nothing of its own; it names the model each main-loop request goes to, through Claude Code's own API connection
4. Persists: in $.store, the on/off setting and the two model settings
5. Hostile input: a model setting must be an alias or a model id (letters, digits, `.`, `-`, `[`, `]`); any other text is refused

Limits

• Each switch writes the whole conversation to the new model's cache once (see "Cost of a switch").
• The model named for a side must be one your account can use. A model the API refuses fails that request as any wrong model would.
• A plan model with a smaller context window than the conversation fails the request. Keep the [1m] suffix on both sides when the session runs with a 1M window.
• Shift+Tab during a turn applies only after the next tool call, because no hook reports the mode before that.
• effort-auto changes the effort of the same requests; both work together, and the effort applies to whichever model plan-split names.

Development

make install     # eslint, typescript-eslint, typescript
make lint        # complexity limit 10, the build fails above it
make typecheck   # needs .claude/types/ from /plugin-types
make validate
make test        # claude plugin test