Skip to Content
AutoMode

Delegated reads

AutoMode's optional read gate: a full read of a large file stops once with a free outline and a cheap Haiku extract tool. How it works and stays safe.

Delegated reads ship off. Set "delegation": "on", or shadow to record what the gate would do. The change takes effect in the next session.

A large file read into a long Opus chat costs that file's tokens on every later request. The read gate offers a cheaper way to get what the model needs from it.

How it works

The first full read stops

A full read of a text file at or above delegationGateTokens stops once per file and task. The model gets a free outline of the file instead. Size comes from bytes, not lines, so a minified file counts.

The model picks a path

It can read the file again, and the second read always runs, so the exact text is always available. Or it can call mcp__automode__extract with a question.

Haiku extracts

extract asks Claude Code whether each path may be read, so your Read deny rules hold. It makes one Haiku call with no tools.

Code checks the evidence

A finding is kept only if its quote is in the file at or near the cited lines. A kept claim is still Haiku's reading, so the model must read those lines before it relies on them.

When the gate stands aside

  • A file edited in the task.
  • The debug, reconsider and review stages.
  • Any task with a risk rating.

The gate only saves money, so it fails open. Shadow mode registers no tool, because a tool definition changes the request.

What it costs and saves

At the shipped 5k-token gate, only 3 of 76 benchmark tasks ever hit a full read that large. Full reads in the audit's bloat list were about 1.6% of 30 days of spend. The canary runs test safety, not a saving. See bench/README.md.

Last updated on

On this page