About codex-build
codex-build is a compatibility skill that directs implementation to Codex while relying on the shared workflow in claudex-loop. It covers both delegation from Claude Code and implementation in a Codex session, with separate inspection and proof checks.
Key features
- 01Selects Codex as the implementation provider
- 02Uses claudex-loop shared build workflow
- 03Preserves spec, log, proof, and model arguments
- 04Supports delegated and in-session builds
- 05Requires independent inspection and proof checks
Use cases
- 01Delegate a planned change from Claude Code to Codex
- 02Implement a work order in a Codex session
- 03Verify changes against a specified proof command
- 04Prepare a completed diff for sign-off
Install
Copy the command and run it in a terminal.
npx skills add https:/ / github.com/ chaseai-yt/ claudex-loop --skill codex-buildAfter installing, type/codex-buildin Claude Code to confirm it loaded.
Details
From the project READMECodex Build
Compatibility entry point: the builder is explicitly Codex. The shared workflow and executable live in the sibling claudex-loop skill; install that skill alongside this one.
Load the shared build reference and runtime reference. Preserve SPEC_FILE (mapped to the runner's --plan), LOG_FILE, PROOF_CMD, MAX_FIX_ROUNDS, and explicit model/effort arguments. The spec may have any filename; never substitute PLAN.md silently.
- In Claude Code, Claude coordinates and delegates implementation to Codex through the runner's
build --host claude --builder codexpath. Claude then inspects all changes and independently runs proof checks. Log evidence and limitations. A fresh Claude CLI inspector is also available through the shared runner if useful. - In Codex, implement in the current host session with its normal tools, then use a fresh Claude inspector through
inspect --host codex --builder codex. Do not ask Codex to certify its own changes as independent review.
Use the current valid plan approval when this follows claudex-loop. An explicitly requested standalone work order can use --unreviewed-spec, with that status recorded. If the spec still needs consequential decisions, settle those before delegating; do not build by inventing missing requirements.
Preserve unrelated changes, record the baseline, independently run the proof, inspect staged and new files as well as the unstaged diff, and keep fix rounds bounded. After a Claude takeover, require fresh Codex inspection of Claude's edits; if both contributed, log authorship and have each inspect the other's changes. Never present an earlier review as covering later fixes. Present the completed diff for any remaining sign-off required by the user's authorization.