Axon Shaft docs

Model, effort and turns

How auto picks the model and effort from the diff, the turn limit by size, the time limit, and how to see what a run used.

The model and effort settings are auto by default. A model or effort set on the dashboard turns auto off for that one only. An automation can set its own model and effort, which win over the project's.

The model and effort on auto

For a review, the action picks both from the diff, after the skip globs. The first row that matches wins:

The diffModelEffort
Touches a focus path or a sensitive one: auth, crypto, payments, billing, workflows, migrations, or the dependencies in a package.jsonopushigh
Any diff on a deep reviewopushigh
Over 1,000 changed lines or over 30 filesopusmedium
Only docs, comments, lockfiles or confighaikulow
Anything elsesonnetmedium

On auto, /ask runs on sonnet at medium effort, and /fix and /axonshaft on opus at high effort.

  • If the Claude plan can't use the model auto picked, the run falls back to the next smaller one, and the footer's table has a Fallback row.
  • A deep review (depth Deep) adds subagents for the correctness, breakage and security checklists. They run on sonnet when the model is auto, and on the set model when it isn't.
  • Every model checklist that runs shares one model session.

The turn limit

A turn is one step of the model: a tool call and the thinking around it. With no turn limit set, the action picks one from the review's changed lines, after skip:

The diffTurn limit
Under 100 lines12
Under 500 lines20
500 lines or more, or a deep review30

A limit set on the dashboard always wins. /ask and /fix use 30 when none is set. When the limit is reached, the run fails with review_failed.

The time limit

A run may take 15 minutes unless the account's admin changes the time limit, from 1 to 120 minutes. Past it, the run fails with review_timed_out. The turn limit and the time limit are set on the account only.

Seeing what a run used

The summary comment's footer names the model, the effort and the tokens, and its table says why auto picked them, such as 60 lines, up to 12 turns. The run's page on the dashboard shows the same, with the tokens split into input, output and cache.

Larger models, higher effort and more turns use more tokens. A run on a Claude plan counts against the plan's usage. On an API key, the footer, the Runs list and the run's page show what each run cost, and the monthly spending cap can limit the total.

Last updated on