umputun / ralphex / 26301508386

83%

Build Type

push

github

Committed by web-flow

Commit Message

feat(codex): per-phase model selection for codex executor (#357)

Under the codex executor (`--codex` or `executor = codex`), the `--task-model` / `--review-model` flags were silently ignored. They only fed the claude executor, so codex model and effort could be set via config file (`codex_model` / `codex_reasoning_effort`) but not per run or per phase.

This wires `--task-model` / `--review-model` (and their config equivalents `task_model` / `review_model`) into the codex executor with full per-phase support:

- `ResolveCodexModelEffort` resolves a `model[:effort]` spec against `codex_model` / `codex_reasoning_effort` as the defaults - an empty spec keeps the defaults, each populated half overrides its default (`--task-model=:high` changes effort only)
- `buildCodexExecutors` builds a separate review `CodexExecutor` when the resolved review model/effort differs from task, mirroring `buildClaudeExecutors`; otherwise the task executor handles task, review, and finalize
- `--task-model` sets the task phase; `--review-model` sets the review phase and falls back to `--task-model` when unset, so tasks and reviews can run on different codex models
- the `max` effort level is claude-only - a spec requesting it under `--codex` keeps the default effort and the startup banner emits a warning
- the banner shows the resolved task model/effort, plus a separate `review model` / `review reasoning effort` line when the review phase resolves differently

Verified end-to-end with the toy project: `--codex --task-model=gpt-5.5:high --review-model=gpt-5.5:low` produces a banner showing `reasoning effort: high` and `review reasoning effort: low`, and the task phase runs through a separate codex executor from the review phases.

Docs updated: `README.md`, `CLAUDE.md`, `llms.txt`.

Related to #356

Coverage Stats

68 of 93 new or added lines in 2 files covered. (73.12%)

4 existing lines in 1 file now uncovered.

7500 of 8998 relevant lines covered (83.35%)

230.72 hits per line