Repository navigation
Releases
v1.5.3
Fix double test re-runs on small/medium tasks
Root cause: git diff --stat only shows tracked files. When Codex creates new test files (untracked), Opus doesn't see them in the diff stat, concludes they're missing, investigates, and re-runs tests — wasting 3-4 extra turns.
Fix: Added git ls-files --others --exclude-standard to Step 7 so Opus sees both modified and newly created files immediately.
This was the fixed overhead making 80/400 LOC tasks uneconomical. The 1060 LOC task was unaffected because grep found clear "Verification passed" text in the Codex log.
v1.5.2
Fixes based on v1.5.1 benchmark data
v1.5.1 cache reads went UP (+5-22%) because of two issues:
- CODEX_LOG mktemp suffix — same macOS bug as PLAN_FILE,
mktemp /tmp/codex-log-XXXXXX.txtfails. Removed.txtsuffix. - Opus was reading the full log file — Step 8a said "Scan the Codex stdout/stderr" which made Opus
catthe entire 600-line log, putting it right back in context. Now explicitly usesgrep | tail -5and says NEVER read the full log.
v1.5.1 — Two bug fixes
Bug fixes
1. Codex stdout no longer bloats conversation context
Previously codex exec ... 2>&1 dumped 600+ lines directly into the Claude conversation. Every subsequent API turn re-sent those lines as cache read (~200-300k extra tokens per session). Now pipes to a temp file.
2. macOS mktemp compatibility
mktemp /tmp/opus-plan-XXXXXX.md fails on macOS (doesn't support characters after the X template). Fixed with mktemp + mv rename.
v1.5.0 — Token optimization
Performance optimizations based on benchmark data
Benchmarked Pure Opus vs Opus+Codex across 3 task sizes. The post-Codex review phase was the biggest cost driver. This release cuts review overhead significantly:
Changes:
- Diff-only review — read
git diffoutput, never individual files. Fixes the 1M+ cache read explosion on large tasks. - No double testing — parse Codex stdout for test results. NEVER re-run pytest/npm test.
- Consolidated git commands — one
git diff --stat+ onegit diff. Nogit status, nogit log, no redundant calls. - Leaner plans — intent-based descriptions instead of full code snippets. 50-70% smaller plans.
- Auto-cleanup — removes Codex sandbox artifacts (
sitecustomize.py,__pycache__,.codex). - Plan cleanup step — every plan now ends with instructions to delete unplanned files.
Estimated impact: ~$0.35-0.50 savings per run on large tasks (from $1.40 → ~$0.90-1.05).
v1.4.4
Changes
- Stronger duplicate test prevention: Added CRITICAL/NEVER language to prevent Opus from re-running tests that Codex already executed, saving unnecessary token spend
- Auth caching: Preflight auth check now caches verification for 7 days at
~/.opus-codex/auth-verified, skipping the slowcodex execprobe on subsequent runs
v1.4.3 — Benchmark-driven optimizations
What's new
- Skip duplicate test runs — if Codex already ran tests successfully during execution, Opus no longer re-runs them in Step 8. Saves tokens and time.
- Cached auth check (v1.4.2) — auth verification cached for 7 days instead of running a full
codex exectest on every invocation. - Preflight check (v1.4.0) — verifies Codex CLI + auth before spending Opus tokens on planning.
- Standalone (v1.4.1) — no gstack dependency.
Benchmark findings
These optimizations came from head-to-head benchmarking Pure Opus vs Opus+Codex. Key insight: for small tasks (<20 LOC), the planning serialization overhead makes Opus+Codex more expensive. These fixes reduce that overhead.
Upgrade
cd ~/.claude/skills/opus-codex && git pull origin mainv1.4.1 — Standalone
What's new
- Removed gstack dependency — the skill no longer invokes
/reviewfrom gstack. Opus readsgit diffand reviews the changes itself. opus-codex now works completely standalone. - Preflight auth check (v1.4.0) — verifies Codex CLI is installed and authenticated before spending any Opus tokens on planning.
Upgrade
cd ~/.claude/skills/opus-codex && git pull origin mainv1.3.0 — Interactive upgrade prompts
What's new
-
Interactive upgrade prompts — when an update is available, users choose:
- A) Upgrade now — pull and continue
- B) Always auto-upgrade — persisted in
~/.opus-codex/config, silent updates from now on - C) Not now — snooze for 24h
- D) Never ask again — disables update checks (re-enable with
update-config set update_check true)
-
Config persistence — user preferences stored in
~/.opus-codex/config -
Snooze mechanism — "not now" respects a 24h cooldown before asking again
Upgrade
If you have auto-upgrade enabled, this version will apply automatically on next /opus-codex invocation. Otherwise:
cd ~/.claude/skills/opus-codex && git pull origin mainv1.2.0 — Auto-upgrade
What's new
- Auto-upgrade: skill now checks for updates on every invocation (cached 60 min) and auto-pulls when a new version is available
- Auto-review:
/reviewruns automatically after Codex completes — no more "want me to review?" prompt
How upgrades work
You don't need to do anything. When you run /opus-codex, the preamble checks GitHub for a newer VERSION file. If one exists, it runs git pull automatically and tells you it updated.
v1.1.0 — Initial Release
opus-codex v1.1.0
Initial public release.
Features
- Opus produces detailed implementation plans with file paths, code snippets, edge cases, and verification steps
- Hands off to
codex execfor autonomous execution (full-auto or manual approval modes) - Auto-runs
/reviewon completion — no manual prompt - One-command install via paste into Claude Code
Requirements
- Claude Code
- OpenAI Codex CLI (
npm i -g @openai/codex) - Claude Opus model (
/model opus)