[JacobStephens2]

Summary

Setup now offers Codex. Its Model and Effort questions show Codex's catalog and are checked against it, the same way a Run checks them.

 ask_harness
-  harness.unavailable()           # codex: "not supported yet"
+  harness.installed()             # only "not installed" is refused
   default = first installed       # claude when both, codex when only it
   Claude → ask_claude             # test call, as before
+  Codex  → Catalog::read()        # codex debug models
+           ├─ fails → say why; harness settings stay as they are
+           └─ ask_codex
+                say "Codex's Models: gpt-6.1-sol (GPT-6.1-Sol), …"
+                ask Model  until catalog.settle(model) ok    # unknown → asked again, choices shown
+                say "Efforts gpt-5.5 supports: low, medium, high, xhigh"
+                ask Effort until catalog.settle(model, effort) ok
+                write the slugs and Efforts as Codex names them

In harness.rs, Catalog is now public, with read, listed, efforts and settle. Choice::check_codex reads the catalog through it too, so a Run and Setup can't disagree about a Model or Effort.

Evidence

• Before: Harness … claude or codex (not supported yet) [claude]: codex → codex is not supported yet: Codex sessions come in a later thirdshift.
  After: these new tests in tests/setup.rs fail on origin/issue-136 (they time out waiting for Model for codex) and pass here:
  • codex_is_offered_with_its_models_and_the_chosen_models_efforts: codex, gpt-5.5, high are written as default = "codex", model = "gpt-5.5", effort = "high".
  • with_only_codex_installed_codex_is_the_default_answer: the prompt is claude (not installed) or codex [codex].
  • a_codex_display_name_and_capitalised_effort_are_written_as_codex_names_them: GPT-6.1-Sol, Max are written as gpt-6.1-sol, max.
  • an_unknown_codex_model_or_unsupported_effort_is_asked_again_with_the_choices: gpt-7 gets choose one of gpt-6.1-sol, gpt-6-luna, gpt-5.5, and max on gpt-5.5 gets choose one of low, medium, high, xhigh.
  • the_codex_model_and_effort_default_to_the_current_ones_as_codex_names_them
  • when_codexs_catalog_cant_be_read_the_harness_is_left_as_it_was
• cargo test: every binary is green. cargo clippy --all-targets -D warnings and cargo fmt --check are clean.

Merge Danger

Door: two-way

Only the interactive Setup questions changed. The User config's format is the same.

Blast Radius: Setup

Someone with codex installed now gets Codex's Model and Effort questions if they choose it. A Run's Codex errors also read slightly differently: codex debug models failed, so Codex's Models can't be read: ….

Unaddressed findings

Standards

• Catalog is now public with four methods (interface widening). Skipped: Setup has to list the Models and Efforts, not just check them. Keeping CatalogModel private and returning choices keeps the surface small.
• Catalog::read runs codex itself (dependency created rather than accepted). Skipped: parse is still the seam the unit tests use, and the integration tests use the fake codex.
• listed() formats slug (Display name) in the domain type (Feature Envy). Skipped: formatting there keeps CatalogModel private. Otherwise its fields would have to be exposed.
• Per-Harness match in ask_harness repeats the one in Choice::check (Repeated Switches). Skipped: one asks and the other checks, and with two Harnesses a trait would be speculative.
• ask_claude/ask_codex loops share an ask → validate → retry shape (Duplicated Code). Skipped: Claude re-asks both settings together after one test call, while Codex re-asks each on its own. A shared helper would hide that difference.
• CODEX_MODEL/CODEX_EFFORT sit beside MODEL/EFFORT in the tests. Skipped: renaming MODEL/EFFORT would touch dozens of unrelated tests.

Spec

• Hidden Models are accepted but not listed. No change: the spec asks for answers to be checked "as a Run checks them", and a Run accepts them.
• With no Model, the Effort question lists every Model's levels. No change: this matches how a Run settles an Effort without a Model.
• gpt-5.5 (GPT-5.5) repeats itself when the names differ only in case. No change: the spec asks for "slug and display name".

Closes #396

🤖 Generated with Claude Code