Claude Code Codex Skill

Claude Code skills. Four skills:

• codex-subagent — the primitives: one-shot Codex runs and steerable Codex app-server sessions, driven directly. Defaults to gpt-5.6-sol; gpt-6-astra is available for the hardest problems.
• codex-scheduler — a job system built on top of those primitives: submit jobs with a slug, dependencies, and per-job model/effort/fast-mode; a global daemon runs them with a configurable parallelism limit and notifies the submitting Claude session natively (via Claude Code's background-task mechanism) when a job finishes or fails, instead of requiring Claude to keep polling. Runs jobs on Codex or Cursor (Composer, or Grok 4.6 with a chosen reasoning level) — a batch can mix both engines freely. Includes a live local dashboard.
• issue-tracker — a per-project issue tracker with a live kanban board: Open, Planned, In Progress, Deployed, Done, Cancelled. Claude files, moves, and flags issues from the CLI; the board is a read-only view at localhost:2345, live over SSE.
• logbook — a template for a Twitter-style status feed, published as an Artifact, where Claude reports outcomes to its user on a loop (hourly, plus ad-hoc alerts and questions). Ships as an empty starter — fill in the project name and artifact URL, and adjust the writing rules to your user, before using it for real.

What Is Included

• .claude/skills/codex-subagent/SKILL.md documents the base skill.
• .claude/skills/codex-subagent/scripts/codex_appserver.py runs a single Codex app-server turn.
• .claude/skills/codex-subagent/scripts/codex_session.py runs an interactive file-controlled Codex session.
• .claude/skills/codex-scheduler/SKILL.md documents the job-scheduler skill.
• .claude/skills/codex-scheduler/scripts/db.py — SQLite schema + query helpers (WAL mode).
• .claude/skills/codex-scheduler/scripts/appserver_client.py — the app-server client each job runs on, embeddable in the daemon.
• .claude/skills/codex-scheduler/scripts/scheduler_daemon.py — the global orchestrator (dependency graph, parallelism limit, hang/crash recovery).
• .claude/skills/codex-scheduler/scripts/scheduler_cli.py — the CLI Claude actually calls (submit/list/show/wait/steer/stop/rm/edit/reorder/config/ui).
• .claude/skills/codex-scheduler/scripts/cursor_client.py — the Cursor engine driver (cursor-agent), interface-compatible with appserver_client.py so the daemon drives either engine through identical code.
• .claude/skills/codex-scheduler/scripts/dashboard.py + scripts/static/index.html — the live local dashboard.
• .claude/skills/issue-tracker/SKILL.md documents the issue-tracker skill.
• .claude/skills/issue-tracker/issues.js — the CLI Claude calls (add/start/deploy/done/cancel/flag/list/show/...).
• .claude/skills/issue-tracker/store.js — the file-backed store (~/.claude/issues/<project>/<n>.json, one file per issue).
• .claude/skills/issue-tracker/server.js + board.html — the live kanban board, autostarted by the CLI.
• .claude/skills/logbook/SKILL.md documents the logbook skill — a template, fill in PROJECT_NAME and ARTIFACT_URL before using it.
• .claude/skills/logbook/logbook.html — an empty starter feed with one entry; publish it once with the Artifact tool to get your ARTIFACT_URL.

Requirements

• Claude Code with local skill support.
• OpenAI Codex CLI installed and authenticated (for codex-subagent / codex-scheduler only).
• Cursor CLI (cursor-agent) installed and authenticated, only if you want codex-scheduler's Cursor engine (--engine cursor) — Codex-only use doesn't need it.
• Python 3.9 or newer (stdlib only) for codex-subagent / codex-scheduler.
• Node.js (no dependencies) for issue-tracker.

The bundled recipes assume the Codex CLI is available as codex and that authentication is already configured for the local user.

Installation

Symlink the skill directories you want into your Claude skills directory:

mkdir -p ~/.claude/skills
ln -s "$(pwd)/.claude/skills/codex-subagent" ~/.claude/skills/codex-subagent
ln -s "$(pwd)/.claude/skills/codex-scheduler" ~/.claude/skills/codex-scheduler
ln -s "$(pwd)/.claude/skills/issue-tracker" ~/.claude/skills/issue-tracker
ln -s "$(pwd)/.claude/skills/logbook" ~/.claude/skills/logbook

Restart Claude Code after installing so it can discover the new skill metadata. Use a real symlink (not a copy) — each skill is meant to be edited in place here and picked up live.

Usage

Open each skill file for its full operating rules and verified command patterns:

cat .claude/skills/codex-subagent/SKILL.md
cat .claude/skills/codex-scheduler/SKILL.md
cat .claude/skills/issue-tracker/SKILL.md
cat .claude/skills/logbook/SKILL.md

codex-subagent covers:

• read-only Codex second opinions
• bounded workspace-write coding tasks
• resuming Codex sessions
• steerable app-server sessions with a file-based control plane

codex-scheduler covers:

• submitting single jobs or a dependency DAG in one batch
• a global daemon (one shared instance across every workspace/session) with a configurable parallelism limit, at ~/.claude/codex-scheduler/
• getting notified natively (no polling) when a job finishes or fails, via wait launched as a backgrounded Bash command
• steering or stopping a running job, and removing/editing/reordering queued ones
• a live dashboard, auto-started at http://localhost:1234 the first time the daemon starts in a session (no need to run scheduler_cli.py ui yourself unless you want a different port or it isn't already up) — showing every job's status, a "Working On" summary, and a live log tail, with a Cancel button per running job
• safe concurrent use from multiple Claude sessions and sub-agents at once
• submit --worktree: a per-job git worktree at <workspace>/.claude/worktrees/<slug> on branch lane/<slug>, so two lanes on one repo never overwrite each other. It is prepared to be runnable: vendor is cloned from the main checkout (APFS clone on macOS, reflink on Linux, plain copy otherwise) and composer dump-autoload is re-run inside the worktree, node_modules is symlinked, and .env is copied with any sqlite DB_DATABASE= repointed at the worktree's own database. --no-symlinks skips the whole preparation step. vendor must not be a symlink: Composer bakes $baseDir = dirname($vendorDir) from the real path into vendor/composer/autoload_psr4.php, so a symlinked vendor autoloads the main checkout's classes and every test inside the worktree quietly tests the wrong code. Node has no such problem (it resolves relative to the real file), so node_modules stays a cheap symlink.
• Codex can message the submitting Claude session mid-run (notify), delivered through the same wait channel as job completion
• a daily npm install -g @openai/codex before the first job launches each day, so Codex stays current
• two engines, one interface: --engine codex (default) or --engine cursor, selecting Cursor's Composer or Grok 4.6 — --effort picks Grok's reasoning level (low/medium/high/xhigh), resolved and validated against Cursor's own live model list before the job ever launches. Cursor jobs can't be steered or asked (no mid-turn channel — cursor-agent -p is one-shot), but their reasoning is visible in the log, unlike Codex's single marker.

issue-tracker covers:

• filing bugs and features, and moving them through Open → Planned → In Progress → Deployed → Done → Cancelled — none of the steps are mandatory
• flagging any active issue for the user's attention, with a required reason shown right on the card; clears only when the issue is actually resolved, not just moved
• per-project boards, numbered independently (#01, #02, ...), auto-detected from the git repo containing the working directory
• a live kanban board at http://localhost:2345, autostarted by the CLI on first use, updating over SSE with no reload — six columns side by side, each independently collapsible
• Done and Cancelled show only the last 12 hours; every other column is never filtered
• every column sorts newest-created first
• one JSON file per issue under ~/.claude/issues/<project>/, atomic writes, safe for concurrent sessions

logbook covers:

• a Twitter-style feed, published as a Claude Artifact, that Claude writes to on a /loop tick or ad-hoc — reports, alerts, and questions, each its own entry type
• writing rules aimed at outcomes over process: short entries, screenshots for visual changes, never editing a published entry after the fact
• a pinned "open questions" box, separate from the main timeline, so anything waiting on the user doesn't get lost in the feed
• ships as a template — an empty starter feed with one entry, PROJECT_NAME and ARTIFACT_URL placeholders to fill in, and writing rules meant to be adjusted to the user rather than used as-is

License

MIT License. See LICENSE.