codex-flow

Claude + Codex dual-model development, layered on superpowers without forking it. Claude orchestrates and reviews; Codex (gpt-5.6: sol/terra) is the independent adversarial reviewer and the fast implementer. The skills call the superpowers skills as sub-skills and weave Codex in at four seams.

Requirements

• The superpowers plugin installed — these skills call superpowers:brainstorming, superpowers:writing-plans, superpowers:subagent-driven-development.
• The OpenAI Codex CLI installed and authenticated — command -v codex, codex login. Models gpt-5.6-sol / gpt-5.6-terra. Each skill degrades to plain superpowers if Codex is absent.

Install

/plugin marketplace add hugo-clemente/codex-flow
/plugin install codex-flow@codex-flow

(Private repo — fetched with your gh / git credentials.)

Skills

skill	role
brainstorming-codex	superpowers:brainstorming + Codex adversarial spec critique → chains to writing-plans-codex
writing-plans-codex	superpowers:writing-plans + Codex adversarial plan critique (findings JSON, Claude adjudicates) → chains to sdd-with-codex-implementer
sdd-with-codex-implementer	the superpowers SDD loop with Codex as the per-task implementer (terra/low) + Claude reviewer kept + a final Codex challenge on the branch diff
pr-triage-codex	red PR → green: triage review comments + CI blockers with the user (batched AskUserQuestion, ELI5 per item), Codex implements the fixes + challenges the diff, push, watch CI until green
codex-lanes	shared reference: the exact codex exec invocation — guards, terra/sol lanes, result schema, prompts. Loaded by the others.

Kick off

"brainstorm <feature> with the codex flow"

brainstorming-codex runs; the chain carries through plan → implementation, with human approval gates between phases.

Model lanes

lane	model	effort	sandbox
implementer	gpt-5.6-terra	low	workspace-write
review	gpt-5.6-sol	xhigh (escalate max/ultra)	read-only — findings JSON, Claude adjudicates + applies

Fast mode: pass --fast to any codex-flow skill → every codex exec call in the whole flow (reviews included) runs with service_tier="fast" + features.fast_mode=true (~1.5× speed, ~2.5× credits). Sticky across hand-offs, opt-in.

Claude lanes (Agent dispatches, see codex-lanes §3b): reviewers (per-task + final) opus · user-facing implementer sonnet · escalated implementer opus · mechanical chores sonnet/low. Defaults, not limits — bad output gets rerun on a smarter model, no asking.

Versioning

Any change under skills/ bumps version in .claude-plugin/plugin.json, in the same PR. CI enforces it.

Change	Bump
wording, typo, clarification	patch (0.6.0 → 0.6.1)
behaviour, new guard, moved or renamed file	minor (0.6.0 → 0.7.0)
skill removed or renamed	major (0.6.0 → 1.0.0)

Changes to plans/, scripts/, .github/ or README.md alone do not need a bump.

Checks

./scripts/check.sh — asserts the skill files agree with each other (frontmatter names, cross-references, lane/manifest consistency). No dependencies. Runs on every push via .github/workflows/check.yml.

Built and verified with superpowers:writing-skills (RED→GREEN→REFACTOR).