tstack is a GitHub repository by Theo (t3dotgg) containing nine agent skills for Claude Code and Codex. The README describes it as "the agent skills I use every day, cleaned up so anyone can use them." It showed up on GitHub's trending list for week 44 of 2026 with about 1,591 new stars since October 24, and the repository page showed roughly 1.6k stars, 88 forks and just three commits when we read it on October 27.
The headline skill is full-send: say the phrase and the agent takes a change all the way to a merged pull request while you are not watching. That is the part worth understanding before you copy anything into your skills folder. This post covers what each skill does, how to install the pack, how it compares with other skill packs we have covered, and the risks of letting an agent merge code.
TL;DR: the questions people ask first
| Question | Answer |
|---|---|
| What is it? | Nine markdown skills (SKILL.md files) for PR-centered agent work |
| Who made it? | Theo, GitHub user t3dotgg |
| License | MIT |
| Which harnesses? | Claude Code and Codex; some skills use T3 Code features when present and fall back to native tools |
| What does it need? | An authenticated GitHub CLI (gh), and python3 for the babysit-pr watcher |
| Does it need other providers? | council and sol-loop need subagents on other providers' models |
| Biggest risk | full-send is explicitly allowed to merge and to force-push the PR branch |
| Safer entry point | full-stop, which stops when the PR is ready for you to merge |
What is in tstack?
A shelf of skill cards with one pulled out, showing the tstack agent skills pack for Claude Code
The README lists nine skills, each triggered by a short phrase you say to the agent. We read the README and the council and full-send skill files directly. We did not run the pack, so for the other skills we describe only what the repo says they do.
| Skill | You say | What the README says it does |
|---|---|---|
| full-send | "full send" | Takes a change to a merged PR with nobody watching: proves it, reviews it, files it, babysits it, merges it |
| full-stop | "full stop" | Same as full-send, but stops when the PR is ready for you to merge |
| council | "council" | Gets review from three models from different providers, and loops until they agree |
| sol-loop | "sol loop" | Reviews a change with GPT Sol in a loop until both agents agree |
| file-pr | "file a PR" | Files a short PR that leads with the problem, not an inventory of changes |
| takeover | "take over" | Finishes someone else's PR, branch, or agent thread, and keeps their credit |
| babysit-pr | "babysit" | Watches a PR through CI and review bots, and fixes the real findings |
| modernize | "modernize" | Rebases an old PR and decides if it should still exist |
| leaving-pr-comments | automatic | Formats comments that an agent posts on your behalf |
Several skills call each other, so installing one is rarely enough. Per the README, full-send uses sol-loop, file-pr and babysit-pr; full-stop uses full-send; takeover uses modernize, file-pr, babysit-pr and leaving-pr-comments; and file-pr, babysit-pr and modernize use leaving-pr-comments. The simplest path is to copy all nine.
How do I install it?
The README gives two copy commands, one per harness:
git clone https://github.com/t3dotgg/tstack
cp -R tstack/skills/* ~/.claude/skills/ # Claude Code
cp -R tstack/skills/* ~/.codex/skills/ # Codex
Then confirm gh auth status shows an authenticated session. If you want the babysit watcher, make sure python3 is on your path. Skills load by name and description, so the trigger phrases in the table above are what activate them. If you are new to the format, our guide to agent skills explains how a SKILL.md file is discovered and loaded, and Anthropic documents the Claude Code side in its skills documentation.
One design note from the README: personal settings, like which models sit on your council, belong in your CLAUDE.md or AGENTS.md, not inside the skills, "so you can update the skills without losing them." That is good hygiene for any shared skill pack.
What does full-send actually let the agent do?
The full-send skill file is the most interesting document in the repo because it is a permission contract written in plain English.
It opens by saying the user triggered it "because they are not watching this thread" and that updates "go unread until long after you stop." It then adds a guard: if the user did not explicitly request full send, the agent should not assume full autonomy. When it does apply, the user grants three permissions: merge the PR once the merge gate passes, force-push the PR's own branch with --force-with-lease after a rebase, and use whichever test methods the proof step needs without asking first.
The workflow has seven steps:
- Finish the change with the smallest edit that works.
- Prove it with evidence the agent observed itself: a reproduction, command output, or a screenshot.
- Run the
sol-loopreview. - File the PR with
file-pr, rebased on the latest default branch. - Babysit the PR through CI and review bots.
- Merge when the gate passes.
- Report the PR link, merge commit, proof and every decision made on your behalf.
The merge gate
The gate is five conditions on the latest commit: the proof covers every change in the diff, required checks are green, review bots have finished with each finding fixed or dismissed in writing, no human review still requests changes, and the branch is up to date and mergeable. The skill says to merge with the repo's usual method and never to use --admin or other bypasses.
Two details stand out. First, "Proof is evidence you observed yourself in this session," and "green mocked tests do not prove that the user experience works." Second, the skill tells the agent to stop and ask rather than guess when it cannot prove a change. Asking is done through the harness's question tool, because "a question in a normal message goes unseen, and the thread looks done."
These are sensible rules. They are also instructions to a language model, not enforced controls, which matters for the risk section below.
What does council do, and does it need three subscriptions?
Two reviewers marking one sheet with a green check, like the council skill's cross-provider review
The council skill brings in three delegates, "one strong model from each of three different providers, at high reasoning effort." Unless you name members, it picks the latest flagship from each provider, with Claude Opus, GPT Sol and Gemini Pro as examples.
For a code review, each delegate returns a score out of 10, each full pass counts as a round, and a fresh review after fixes starts a new round. If the work ends in a PR, the skill writes a "Council review" table into the PR description with one column per round. It also says: if a delegate gives no score, write "no score" and do not invent one.
The practical catch is in the requirements. In T3 Code the skill uses delegate_task to start one delegate per model. Elsewhere it needs "the harness's subagent tools," and the README says council and sol-loop need subagents on other providers' models. If it cannot reach three providers, the skill says to report that instead of substituting. So in plain Claude Code with only an Anthropic login, council will stop and tell you it cannot run. You need access to the other providers, through a harness or router that can call them.
The skill also includes a brake: "Do not let repeated reviews result in scope creep. It is your job to fight to keep scope limited." Anyone who has watched a review loop grow a one-line fix into a refactor will see why that line is there.
How does it compare with other skill packs?
tstack is narrow where other packs are broad. It does not try to teach an agent testing discipline, design taste or research methods. It standardizes one thing: how a change travels from a finished diff to a merged PR, and who reviews it on the way.
| Pack | Focus | Our coverage |
|---|---|---|
| tstack | PR lifecycle: file, review, babysit, merge | This post |
| Superpowers | Agentic skills framework with TDD, subagents, worktrees | Superpowers install guide |
| pstack-claude | Router plus workflow playbooks across seven harnesses | pstack-claude guide |
| ponytail | Push agents toward less code | Ponytail skill |
| ECC | Large harness with many agents and skills | ECC guide and install risks |
If you already use a planning or TDD pack, tstack slots in at the end of the pipeline. If you use none, it is a small, readable set of nine files you can audit in an afternoon, which is a point in its favor compared with packs holding hundreds of skills.
What should you check before installing?
Treat any skill pack as code that runs with your permissions, even though it is markdown.
- Read the skills. Each is a short document.
full-sendgrants merge and force-push rights, so read exactly what it grants before you say the trigger phrase. - Scope your GitHub token. The agent acts through your
ghlogin. If that login can merge to protected branches or touch repos you did not intend, the skill inherits that reach. Use a token limited to the repos where you want autonomy. - Keep branch protection on. The skill says never to bypass it, but your repository settings are what actually enforce it. Required checks and required reviews should be configured on the server, not just requested in a prompt. Our write-up on how Anthropic red-teamed auto mode and Codex monitors shows why prompt-level promises are a weak last line of defense.
- Start with full-stop. It runs the same pipeline and stops at a ready PR, so you keep the merge decision.
- Watch the cost. Council runs three flagship models at high effort, in loops. That multiplies token use. Our look at what a harness buys in tokens is a useful frame for estimating it.
- Check provider terms. Sending your code to three vendors for review is a data-handling decision. Make sure your employer or client allows it.
Running unattended agents against live infrastructure is also an agent-safety question. AgentBeam, the agent security platform from the explainx.ai team, is built to stop AI agents before they take dangerous actions, which is the layer you want when a skill says "merge without asking."
What are people asking about it?
Is this an official Anthropic or OpenAI project? No. It is a personal repository by Theo. The skills are plain markdown that works in both harnesses.
Why does it mention GPT Sol and T3 Code? The skills were written around the author's own setup. sol-loop names GPT Sol directly, and several skills use T3 Code features, such as delegate_task and watch_pull_request, when they exist. The README says they fall back to native tools otherwise. Expect some friction if your harness lacks those tools, and edit the skills: the README itself says "These are starting points. Edit them."
Is a three-commit repo mature enough to depend on? It is young. The star count reflects interest in the author's workflow, not production hardening. We found no tagged releases on the page. Pin to a commit if you copy it into a team setup.
Does the review loop actually improve code? The repo gives no measured results, and we have not tested it. The idea of cross-provider review is plausible, since different models fail in different ways. But the skill's own scoring is the models grading each other, so it should supplement human review rather than replace it. For evidence on AI review in practice, see our coverage of a team that put an AI reviewer in its code review process.
What this means for you
If you ship code with Claude Code or Codex, the useful takeaway is the pattern, not only the repo: write the merge gate down, define what counts as proof, and tell the agent when to stop and ask. Those three ideas port to any skill pack. Install it, read it, run full-stop on a low-risk repo first, and only then decide whether you trust full-send with your default branch.
Related reading
- Superpowers: agentic skills framework
- pstack-claude workflow playbooks
- What are agent skills? Complete guide
- Codex Auto-Review: a reviewer agent and fewer approvals
- Ponytail: the lazy senior dev skill
- Anthropic red-teaming of auto mode and Codex monitors
- Skill sources: tstack on GitHub
Repository details, star counts and skill contents are accurate as of October 27, 2026 and may change as the project evolves.