Chief routing

With Chief v2 on, Chief does not guess where a task goes or which model does it. It follows fixed rules, and it tells you which rule it used.

Where a task goes

The first rule that applies wins:

1. The matching session is waiting on a question: Chief asks you to answer it first.
2. The task changes files another open session owns: it goes to that session, now if it is idle, queued if it is working.
3. The matching session is idle, its cache is warm and its context is under 100K: the task is added there.
4. The matching session is cold and over 100K: a new session starts from its handoff.
5. Nothing matches: a new session, lean and read-only when the task only reads.
6. A read-only task with a matching session: a subagent there answers it.

A correction you made in the last 30 days for the same kind of task wins over rules 3 to 6.

Which model

• Fable plans architecture and new projects, and reviews. Its sessions get only read tools: it never edits and never runs commands.
• Opus writes the plan for every other task, and builds complex work.
• Sonnet builds only when the plan passes the gate: a list of files to change, an exact command to check the work, 1 to 8 numbered steps, no hotfile, and a low complexity score.
• Haiku runs the check commands after a build, without MCP servers, and writes a short receipt into the task's record.

Fable reviews a build only for architecture work, a hotfile or shared code, or a second failed round. After a first failed round the same session gets a critique and tries again.

Change models, efforts and limits under Routing on the host screen, or with melta-chief routing.

Common questions

Can Fable edit files?

Never. Its sessions start with read tools only.

Why did Opus build a small task?

The plan missed one of the gate's checks. Chief says which one.

Can I pick the lane myself?

Yes. Tell Chief, or pass --lane to melta-chief new.

What is a hotfile?

A file many tasks touch, from the project map. Changes there get Opus and a review.