Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/dwgx/claude-codex-subagent/codex-subagentnpx skills add dwgx/claude-codex-subagent --skill codex-subagentgit clone --depth 1 https://github.com/dwgx/claude-codex-subagentWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/dwgx/claude-codex-subagent/codex-subagent)<a href="https://agentmods.dev/skills/dwgx/claude-codex-subagent/codex-subagent"><img src="https://agentmods.dev/badge/skills/dwgx/claude-codex-subagent/codex-subagent.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00179 | $0.05348 |
| Opus 5 | $0.00089 | $0.02674 |
| Sonnet 5 | $0.00036 | $0.01070 |
| Haiku 4.5 | $0.00018 | $0.00535 |
Grade D, and why
codex-subagent scanned grade D with 3 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Reads agent configuration directoriesmediumAgent snooping
.claude/, .codex/, .gemini/ hold keys, settings and other credentials a mod has no legitimate need for.
- `-p <PROFILE>` — if the user has profiles defined in `~/.codex/config.toml` Recursive force deletehighDestructive command
rm -rf with a variable or a broad path is one typo away from removing the wrong tree.
**Truly destructive ops** (`rm -rf` outside workspace, dropping a Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
| **Anything needing shell network** (`curl`, `pip install`, `npm view`, `gh api`, package installs, remote API calls) | bypass/full access | `--dangerously-bypass-approvals-and-sandbox` | How it starts
The opening of the file, as written. The whole thing — 498 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Codex as Worker Subagent
Claude is the orchestrator. Codex is a worker subagent with its own fresh context window, its own network access, and its own sandboxed shell. You hand Codex a well-scoped task and take back a short answer. Codex's thinking, tool calls, and file reads happen in Codex's context, not yours — that is where the token savings come from.
Thin-forwarder contract
Treat each Codex call like a function call:
- You write the prompt (the "arguments")
- Codex runs its own full agent loop inside its own context
- You take Codex's stdout as the authoritative return value
- No Claude-side freelancing — don't "edit" what Codex said in your head and then act on a phantom version. If Codex's answer is wrong or under- specified, resume the session with a follow-up or dispatch a fresh call with a better prompt. Do not pretend you know what it meant.
When to delegate
Delegate when any of these apply:
- Network needed. Web search, fetching docs, scraping an API,
checking a package version,
gh api,npm view,pip install. You (Claude) pay steep token cost for WebFetch/WebSearch; Codex's network calls happen in its own context. - Large reads. Scanning a repo, digesting a long file/log, cross- referencing many sources.
- Long output that just needs a conclusion. E.g. "find all TODOs and tell me which are stale" — Codex emits the list internally and returns only the verdict.
- Writing long files. Codex writes in its sandbox; you only see "wrote foo.py (240 lines)".
- Running tools you'd rather not stream. Tests, builds, linters, data-processing scripts.
- You're stuck. If your own attempt has stalled or you want a fresh
second pass with no prior context, dispatch Codex with a clean problem
statement. That's explicitly the best-case use from the
codex-plugin-ccdesign — "rescue" mode.
When NOT to delegate
- Quick decisions you can make from existing context.
- Tasks that depend tightly on conversation state that would be painful to re-explain.
- One-line edits you could
Editdirectly with less overhead than composing a Codex prompt. - The user is clearly expecting you to do it interactively ("let's walk through this together").
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 498 lines · 179 tokens per session scan D 925f7b0949aa
codex-subagent is a skill published in the GitHub repository dwgx/claude-codex-subagent (3 stars, last pushed 1mo ago), licensed MIT. It adds 179 tokens to every session and 5,348 once invoked, about $0.0009 per session on Opus 5. A static security scan graded it D with 3 findings (reads agent configuration directories, recursive force delete, makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
image-to-code
Elite website image-to-code skill for Codex. For visually important web tasks, it must first generate the design image(s) itself, deeply analyze them, then implement the website to match them as closely as possible. In Codex, it must prefer large, readable, section-specific images instead of tiny compressed boards…
edgeone skill scanner
Scan any agent skill for security risks before you install or use it. Powered by Tencent Zhuque Lab A.I.G (AI-Infra-Guard). 100% local static analysis — no file contents or credentials leave your device. Compatible with CodeBuddy, Cursor, Windsurf, Claude Code, OpenClaw and more. Triggers on: 这个 skill 安全吗, skill 安全扫描…
convex-billing
Add Stripe billing/payments to the Convex app via @convex-dev/stripe (checkout + webhook + gating).
dev-workflow
The complete development workflow for SkillHub contributors including local dev, staging validation, testing, and PR creation. Ensures agents follow the correct sequence of steps.
convex-performance-audit
Audits Convex performance for reads, subscriptions, write contention, and function limits. Use for slow features, insights findings, OCC conflicts, or read amplification.
video-frames
Extract a single frame from a local video at the first frame, a timestamp, or a zero-based frame index using FFmpeg.