Models
Which model does which job, from subagents to Codex arms.
Codex tiers (luna, terra, sol, astra) are not models a Claude subagent can run. A Codex role is a Codex arm the main agent fires itself. When the harness can run a Codex model as a child task, as T3 Code does with delegate_task, the arm goes through it, shows up as its own thread and reports back when it finishes. Everywhere else it is one background Bash call to the Codex CLI, read back from a file.
Codex Arms
codex-arms.md holds each tier's model and reasoning effort, how to pick between harness delegation and the CLI, the T3 Code fields, the exec and review invocations, and the rules every prompt restates. Pick the tier deliberately for the task. Never quietly drop to the cheapest tier for work that needs judgment.
| tier | use |
|---|---|
| luna | mechanical work, renames, boilerplate, lookups |
| terra | everyday implementation from a clear spec, the default |
| sol | hard implementation, complex reasoning, long-context investigation, the review arm |
| astra | critical work only, in a seat that names it |
The review arm reads staged, unstaged, and untracked changes together. Through harness delegation it is a read only delegate briefed with the reviewer template, so it gets the intent and rubric, except in a seat that asks for a cold read, which cuts the intent. Under the Codex CLI it runs review --uncommitted, which takes no instructions. It runs sol, or astra only on critical work. Critical means critical: true in the plan frontmatter or work you call critical.
Arm Roster
Every seat that names a tier goes to Codex, then OpenCode at the roster's model for that tier or the nearest tier above it, then a Claude stand in, among the families chosen at install. The implementation seat is the exception: the lead writes the code, and an arm writes it only in the cases the Implementation section of codex-arms.md names. Those are a family set by ARMS_IMPLEMENTATION, a seat you route to Codex by name, and independent slices of M or L work. The arm roster records those choices, and the playbook's Subagents and Codex arms section states the rules every seat follows.
Claude Roles
agents/ holds the Claude agents only. Judgment, taste, prose, and vague intent go to fable-judgment. A second opinion is opus-review plus the review arm. The comment sweep is comment-sicko. See Claude Code for the full agent list.
Ad Hoc Routing
The playbook skill sets disable-model-invocation: true. This routing reaches a session only when /plans do or /playbook sends it there. For subagents and Workflows you spawn ad hoc, point AGENTS.md or CLAUDE.md at the playbook's Subagents and Codex arms section.