Can a cheaper model do the work?

Since 2026.11.10, part of it. A capable model writes an issue's plan as self-contained steps, each with the files it touches and how it is checked. A cheaper model then carries them out one at a time: it takes the next step, does it, closes it, or reports it blocked and stops. The plan's author or a lead verifies each done step, or reopens it with a note.

Quasar holds the queue and records who ran each step on which model. Since 2026.11.11:

  • a step cites what the executor must read, a section of the issue or a document or one of its headings, checked when the plan is written, so a small model reads only the passages it needs;
  • the executor can be kept to the plan: an assistant connected as a worker is offered nothing but the plan's queue, and can change nothing but the step it holds.

Since 2026.11.12 Quasar also says how likely a small model could do it. Every pre-analysis estimates it from 0 to 100, with its reasons, before any plan exists, and every step of a plan states its own; the issue then shows the steps' mean, the share of the work a small model could take. Every open issue shows it as a robot beside its work kind: green from 90, then one colour every 10 points, red under 60 (since 2026.11.13). It is the planner's judgement, gates nothing, and is not yet compared with how the steps went.

Since 2026.11.13 Quasar also helps the small model stay on its step:

  • a worker is told how to work by Quasar itself: the loop, the rules (stay on the checked-out branch, never commit, change only the lines the step names, run only its check) and how to edit, so a person gives it only the issue's name;
  • a step with a check is closed with its proof: the command run, its exit code, which must be 0, and the tail of its output. The reviewer sees a command that differs from the planned one marked «rewritten». Quasar does not run the check itself: the proof is what the executor reports;
  • the planner is warned when a step suits a small model badly, for instance a check that is not one exact command, an invisible character in the text, a step too long, or an estimate that its own hazards contradict. The warnings never refuse the plan.

Since 2026.11.15 a step the executor took but will not work, one more than it was asked to do, goes back to the queue with release, instead of being held for 30 minutes or reported blocked; the plan's author or a lead can release it too, saying why.

Which model runs is the client's setting. The guide Plans: a planner, an executor, a reviewer shows how to set it in Claude Code, and what makes a piece of work suit a small model.