Codex Evolve v0.3.0

Codex Evolve v0.3.0

This release makes Codex Evolve aware of the GPT-5.6 Sol, Terra, and Luna model family, improves prompt-cache reuse, and lets long-thinking workers continue past their first timeout without restarting.

Highlights

• Route strong work to Sol with high reasoning, balanced work to Terra with high reasoning, and cheap work to Luna with extra-high reasoning.
• Start with the strong profile by default, or use --init cheap for Luna/xhigh and --init mid for Terra/high.
• Override each profile's model and reasoning level through max. Ultra is excluded because its automatic delegation conflicts with Evolve's direct worker-call boundary.
• Put shared task content before per-worker variation so repeated candidates and recombinations can benefit from prompt-cache reads.
• Continue the same live worker across --timeout-extensions grace windows instead of restarting its task. One extension is enabled by default.
• Report timeout extensions in the execution trace while preserving the N + M*T direct-call upper bound.

Validation

• 18 tests pass locally and in GitHub CI on Node.js 18.3 and Node.js 22.
• All 22 changed execution paths are covered; the complete runner reports 92.42% line and 82.32% branch coverage.
• Timeout tests cover successful continuation, multiple exhausted windows, zero-extension failure, cancellation during grace, recombination fallback, and subprocess cleanup.
• A live self-review exercised Luna/xhigh initialization and Sol/high recombination, reproducing the former 600-second cutoff that motivated timeout continuation.
• Release commit and v0.3.0 tag are signed and verified.

Install

Plugin:

codex plugin marketplace add ukint-vs/codex-evolve --ref v0.3.0
codex plugin add codex-evolve@codex-evolve

Skill only:

npx skills add ukint-vs/codex-evolve/plugins/codex-evolve/skills/codex-evolve --agent codex -g -y