Skip to content
Komputa Practice
Esc
↑↓navigate↵open⌘Jpreview
On this page

Model selection

How each agent's model is chosen.

One file holds every model choice: ~/.claude/pstack-models.md, the sheet. The operator edits it. Nothing in this repo names a model.

Two layers

Agent Who sets its model How
A pstack subagent (swarm, arena, how, why, interrogate, …) pstack, per call The skill reads its own role line in the sheet, or its default when the sheet has none.
An hc worker (owner, implementer, fixer, reviewer, verifier) The one who starts it Read the role’s line in the sheet, and start claude --model <m> [--effort <e>].

A claude with no --model runs on the "model" in settings.json, which is not a role choice.

Roles and lines

Role Line in the sheet
owner, implementer feature
fixer bug-fix
reviewer judgment and prose
verifier verifier, else judgment and prose

A tier 2 verifier runs as verifier. reviewer stays for adversarial review. Both use the reviewer template.

A value is a model with an optional effort, as in <model> @<effort>. Without an effort, default effort: applies.

The move-up rule

A worker or a subagent runs on its role’s model. It moves up only in one of two cases:

  • (a) The feature needs more than the role line: a subtle algorithm, a cross-cutting design or a hard review.
  • (b) Two fix rounds on the role’s model failed the same check.

Say in your reply which one you moved up and why, as feature: <reason> or fix-rounds: <n> failed on <check>. The rule itself is in rules/working-rules.md.

What enforces each part

Part What holds it
The herdr skill and templates name no model scripts/check.sh, step no model names
A worker’s model comes from the sheet Prose only
A move-up has a reason Prose only

Until 10 November 2026 a launcher held the last two: it read the role’s line, refused a move-up with no reason and wrote each launch to a log. The launcher is in archive/herdr/scripts/, and archive/herdr/model-selection.html draws that flow.

Was this page helpful?