Skip to content

Pin the sweep to a Sonnet worker and the review to Opus - #30

Merged
thorwhalen merged 1 commit into
masterfrom
docs/sweep-model-tiering
Sep 15, 2026
Merged

thorwhalen merged 1 commit into
masterfrom
docs/sweep-model-tiering

Conversation

@thorwhalen

Copy link
Copy Markdown
Member

Refs #16 (WP6).

Why

WP6's fleet sweep was dispatched as four parallel sessions on the largest model at high effort. The brief template had been copied from the earlier design work packages (core v2, validate), which legitimately needed that; nobody made a per-task model decision. The four sessions hit the five-hour limit and burned about a quarter of a week's allowance in an afternoon.

The sweep is the opposite of design work. epythet-repair-migrate and docs-migrator already contain every decision it makes — baseline, repair, coverage, correctness, completeness, theme, docsrc/, exit gate, PR. A larger model spends its cost rediscovering a method that is written down.

What changed

  • docs-migrator: model: sonnet in the frontmatter, plus one sentence in the prompt saying why — the skill is the procedure, so what the sweep needs is care and throughput; one repository at a time; the only place a larger model earns its cost is the bounded review of its own diff.
  • docs-reviewer: model: opus. This one stays on the big model deliberately. It is the judgment half of the sweep — reading a rendered page and deciding whether a docstring's meaning survived — and it is bounded: one diff, one narrow question, run once. Its own prompt already argues that a two-thirds-precision judge is disqualified as a gate and exactly right for discovering patterns; that is the work worth paying for. Sonnet would make it cheaper and worse at precisely the step whose whole value is precision.
  • epythet-repair-migrate, Delegation section: one sentence stating the tiering, so anyone writing a dispatch brief from the skill reads it before choosing a model.

The matching fix upstream — so this is not re-derived per fleet — landed in crowsnest-dispatch as a new "Choose the model before the brief" section: thorwhalen/crowsnest#61.

559 tests pass locally. Merging publishes a release; expected.

https://claude.ai/code/session_01Ej6PBTNwxnY4oHcyqcs6TH

The WP6 fleet sweep was dispatched as four parallel sessions on the largest
model at high effort, because the brief template carried the flags of the
earlier design work packages. It burned about a quarter of a week's allowance
in an afternoon. The sweep is procedural: this skill and the docs-migrator
subagent already contain every decision it makes.

- docs-migrator: `model: sonnet` in the frontmatter, and one sentence saying
  why -- the skill is the procedure, so the sweep needs care and throughput,
  not a larger model rediscovering a written-down method; one repository at a
  time.
- docs-reviewer: `model: opus`, and one sentence saying why -- it is the
  judgment half, bounded to one diff and one narrow question, run once. Its
  own prompt already argues that a two-thirds-precision judge is right for
  discovering patterns and wrong as a gate; that is the work worth a big model.
- epythet-repair-migrate: the Delegation section now states the tiering up
  front, so anyone writing a dispatch brief from the skill sees it.

Refs #16.
@thorwhalen
thorwhalen merged commit 2b4826e into master Sep 15, 2026
12 checks passed
@thorwhalen
thorwhalen deleted the docs/sweep-model-tiering branch September 15, 2026 11:30
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant