🤖 OpenAI Codex Now Runs Inside Claude Code

The holiday weekend barely ended and GitHub is already on fire.

A repo called system_prompts_leaks hit 49,785 stars and the top of GitHub Trending. It contains extracted system prompts from Claude Fable 5, Opus 4.8, Claude Code, Claude Design, ChatGPT 5.5 Thinking, GPT-5.5 Instant, Codex, Gemini 3.5 Flash, and more. The full list of what these models are actually told to be, written out, sitting in a public repo.

I've been saying for months that the black box is cracking. This is the crack.

What's useful for operators: these aren't theoretical. You can read exactly what guardrail language looks like at production scale, what persona-locking instructions look like, and how competing labs structure their default behaviors. That is real intelligence for anyone building on top of these models. It tells you what you're working with and where the walls are.

The second thing worth naming: openai/codex-plugin-cc is trending with 25,343 stars. OpenAI shipped an official Claude Code plugin that lets you delegate tasks to Codex from inside your Claude Code session. A competitor's agent, accessible as a plugin in your workflow. That is a real shift in how the tooling layer is composing.

Both repos are in Drops today. So are eight others. Let's get into it.

[Repo] asgeirtj/system_prompts_leaks, 49,785 stars and climbing. Extracted system prompts from Claude Fable 5, Opus 4.8, Claude Code, ChatGPT 5.5 Thinking, GPT-5.5 Instant, Codex, Gemini 3.5 Flash, and more. Read what these models are actually told to be; the guardrail language alone is worth an hour of your time.

[Repo] openai/codex-plugin-cc, 25,343 stars. Official OpenAI plugin that wires Codex into Claude Code for code review and task delegation. A competitor's model, accessible as a tool in your workflow; that composability is the story.

[Repo] Zackriya-Solutions/meetily, 16,633 stars. Privacy-first meeting assistant running Parakeet/Whisper transcription at 4x speed, with speaker diarization and Ollama summarization, 100% local, no cloud. If you run sensitive client calls, this is the one.

[Repo] langfuse/langfuse, 30,466 stars. Open-source LLM observability: evals, metrics, prompt management, playground, datasets, OpenTelemetry integration. The standard answer when someone asks "how do I know if my agent is degrading in production?"

[Repo] microsoft/autogen, 59,508 stars. Microsoft's agentic framework for building multi-agent systems. If you are serious about orchestrating more than two agents, this is where most of the serious architectural thinking is happening in open source.

[Repo] invoke-ai/InvokeAI, 27,557 stars. Leading creative engine for Stable Diffusion: node-based workflow editor, full-stack generation, professional-grade. The move if you want image generation you own and control.

[Repo] FunAudioLLM/CosyVoice, 21,977 stars. Multi-lingual voice generation with full inference, training, and deployment stack. Local TTS that actually sounds like something you'd ship.

[Skill] realkimbarrett/advertising-skills, 680 stars. Direct response, paid ads, funnels, and copy systems packaged as Claude Code skills for Open Claw and AI agents. The skills are organized around the actual job: write the ad, build the funnel, close the conversion.

[Skill] nizos/tdd-guard, 2,241 stars. Automated TDD enforcement for Claude Code. It stops Claude from shipping untested code. If you have ever watched your agent confidently delete a test suite to make the tests pass, you know exactly why this exists.

[Skill] lishix520/academic-paper-skills, 994 stars. Systematic framework for planning and writing academic papers using Claude Code, with a strategist skill for planning and a composer skill for writing, plus quality checkpoints between them. Useful beyond academia: the structure works for any long-form structured document.

[Repo] punkpeye/awesome-mcp-servers, 90,330 stars. The canonical MCP server list. If you are wiring a new tool into Claude today, this is the first place you check. Not a tool itself, but the directory that saves you two hours of searching.

[Repo] VoltAgent/awesome-ai-agent-papers, 1,520 stars. Curated 2026 agent research papers covering engineering, memory, evaluation, workflows, and autonomous systems. When you want to know what the actual researchers are concluding, not what Twitter summarized, this is the stack.

[Repo] luckeyfaraday/athena-loops, MIT-licensed, backend-agnostic agent orchestration loop in Python. The pattern is orchestrator → worker → reviewer as a deterministic harness, driveable through an MCP server and CLI. The non-obvious move: the reviewer step is what makes outputs actually trustworthy instead of just fast.

- Mechanical Turk is closing to new customers. Amazon stopped accepting new signups for MTurk, which was the default answer to "how do I get labeled training data" for a decade. Operators running fine-tuning pipelines should note the direction: human labeling marketplaces are contracting while synthetic data generation is expanding. TechCrunch

- Simon Willison shipped sqlite-utils 4.0rc2, mostly via Claude Fable, for $149.25 in model costs. The real signal is not the dollar figure, it is the methodology: a named maintainer of a real production library handed a release candidate to an AI and tracked exactly what it cost and what broke. That is a reproducible benchmark. Simon Willison

- Vercel Sandbox now supports FUSE-based filesystems, letting you mount S3 buckets, network filesystems, or any FUSE-compatible driver as a regular directory inside a running sandbox. If you are building agents that need persistent or remote storage inside an isolated execution environment, this is the config that unblocks it. Vercel

- PamStealer is a new macOS infostealer using tradecraft designed to survive detection. Named for how it abuses PAM (Pluggable Authentication Modules). If you run Claude Code on a Mac and your session handles API keys or OAuth tokens in plaintext, the threat model just got more specific. Ars Technica

This week's technique: context and cost control. When to /clear, how compaction works, and how to keep your session sharp across a long build.

Most operators treat the context window like it's free. It is not free, and more importantly it degrades. The model's attention spreads thinner as the window fills, old decisions get relitigated, and cost compounds fast on a multi-hour session.

1. Know the /clear moment. When you finish a discrete task (a feature, a refactor, a file migration) and the next task is unrelated, /clear is almost always right. You lose nothing useful and you buy back a full, sharp window. The mistake is holding onto context "just in case." 2. Understand compaction. When the context window fills, Claude Code compacts it automatically: it summarizes the history and replaces it with a compressed version. You can also trigger this manually. The summary is good but lossy. Anything you need preserved across compaction should live in CLAUDE.md, not just in the conversation. 3. Run cost checks. After a long session, run /cost to see what the session spent. If the number surprises you, audit how many tool calls happened and whether /clear points earlier in the session would have gotten the same outcome for less.

You will know it worked when a session that used to run $4-6 in model costs on a complex feature lands under $2, and the output quality is the same or better because the context is tighter.

Claude Code docs: memory and context

The move: wire a local reviewer into your Claude Code loop before anything ships.

The problem is that Claude Code ships fast. Too fast. It will confidently push a change that breaks a downstream dependency, because it optimized for the task in front of it and not the system around it. The fix is not a prompt, it is a structural step.

1. After Claude Code writes or modifies a file, invoke a second pass: either /review with a specific rubric in CLAUDE.md, or spin a subagent with instructions scoped only to critique (no write access). The reviewer gets the diff and a question: "What does this break?" 2. Gate the commit on the reviewer's output. If it flags something, Claude Code addresses it before the write is finalized. This is the orchestrator → worker → reviewer pattern from athena-loops above: deterministic, not hopeful. 3. Set the scope of the reviewer in CLAUDE.md so it runs consistently, not just when you remember to ask.

You will know it worked when Claude Code catches its own dependency conflicts before you do, and your commit history stops carrying the "oops, fix the fix" entries.

We build The AIgent's engine in the open. An honest look at what we are making, what broke, and where it is headed. FlowStack, the machine that dreams in pictures. Part five: the machine runs.

Here is what is easy to lose sight of in four parts of bugs: it works.

A machine takes an idea, writes a script, generates a voice that pronounces Ezekiel correctly, paints scenes where the same character stays the same character, moves the camera with intention, scores it, mixes in sound you can actually hear, lays in the slow drift across an image, writes a title for the algorithm, and publishes the finished thing to the world. Across multiple platforms. Across a dozen distinct creative identities. Day after day.

The kind of operation that would take a studio full of people runs as a queue of jobs, each one stamped with its own identity so a dozen of them never step on each other.

It stumbles. Sometimes Veile's image comes back and it is not quite her, and you shrug, because the upload still went through, and the upload going through is what matters. That is not failure. That is a system mature enough to fail gracefully and keep moving.

For a long stretch we were not sure a machine could really do all of it. That part is settled now. It is, genuinely, a machine that dreams in pictures.

Next: where it goes from here, and the one lesson the machine could not teach us.

Got something running end to end, even ugly? Hit reply and tell us what your machine does while you sleep. We read every one.

The system prompt leak is the most honest thing that happened in AI this week. Fifty thousand developers starred a repo because they wanted to see what the models are actually told. That instinct, to look past the interface and understand the machine, is exactly the right one. Keep it.

See you Tuesday.