tstack: nine agent skills that take a change to a merged PR in Claude Code and Codex
tstack (t3dotgg/tstack) is a free, MIT-licensed set of nine agent skills for Claude Code and Codex that take a code change through proof, cross-model review, filing, CI babysitting and merge. Install is a git clone and a copy.
tstack (GitHub: t3dotgg/tstack) is a set of nine agent skills for Claude Code and Codex that handle the pull-request side of coding with an agent: proving a change, getting it reviewed by other models, filing the PR, watching CI and review bots, and merging. The README calls them the skills its author uses every day, cleaned up for anyone; the MIT licence names Theo Browne as copyright holder. The repository was created on October 24, 2026 and had about 3,800 stars and 293 forks on November 15. Most skills start on a phrase: say "full send" and the agent owns the change until it is merged; say "council" and three models from different providers review it. leaving-pr-comments applies automatically whenever the agent comments on GitHub.
Quick facts
- What it is: nine skills: full-send, full-stop, council, sol-loop, file-pr, takeover, babysit-pr, modernize and leaving-pr-comments.
- Works with: Claude Code and Codex; some skills use T3 Code tools when present.
- Requirements: an authenticated GitHub CLI (gh), python3 for babysit-pr's pr-wait script, and for council and sol-loop a harness that can run models from other providers.
- Form and licence: Markdown SKILL.md files (babysit-pr ships a Python script, pr-wait; five skills add agents/openai.yaml display metadata for Codex). MIT licence, free; no package, build step or release.
- Activity (November 15, 2026): 6 commits, last push November 11; 6 open issues, 11 open PRs.
How to install
The README's only install method is clone and copy:
git clone https://github.com/t3dotgg/tstack
cp -R tstack/skills/* ~/.claude/skills/ # Claude Code
cp -R tstack/skills/* ~/.codex/skills/ # CodexThere are no releases or tags, so you get whatever is on main. Some skills call others, so copy the dependencies too: full-send needs sol-loop, file-pr and babysit-pr; full-stop needs full-send; takeover needs modernize, file-pr, babysit-pr and leaving-pr-comments; file-pr, babysit-pr and modernize also use leaving-pr-comments. Copying the whole folder, as the commands above do, covers that. Read full-send before copying it: it merges and force-pushes unattended. Open PR #6 proposes a Claude Code plugin marketplace and open PR #4 adds README instructions for npx skills; neither was merged on November 15.
First setup
Authenticate the GitHub CLI first, since the PR steps depend on gh. Without python3, babysit-pr has no watcher outside T3 Code. council needs one model from each of three providers; if it cannot reach three, or a named member is unavailable, it reports that instead of substituting. sol-loop needs a route that runs GPT Sol (T3 Code's delegate_task, a native Sol subagent, or in Claude Code a custom subagent set up to run Sol) and otherwise reports a blocker.
The README also advises you to keep your personal choices, such as which models sit on your council, in your own CLAUDE.md or AGENTS.md rather than in the skill files, so you can pull updates without losing them. The skills are meant as starting points to edit.
What stands out
- A spoken phrase is a permission grant. full-send tells the agent the user is not watching, then grants it three things: merging once the merge gate passes, force-pushing its own branch with --force-with-lease after a rebase, and using the test methods that proof needs without asking, even if a project file says to ask first (rules on how to test, and harness approval prompts, still apply). The agent must still ask before deleting data, touching production, changing things outside the repo, closing a PR or force-pushing any other branch, and the merge gate forbids --admin or other bypasses. With no question tool or no answer, it leaves the PR open and unmerged.
- Proof means evidence seen in the session. The skill wants the original failure reproduced and then passing, the checks the repo's CI runs, and for UI work the app actually opened, because green mocked tests do not prove the user experience works. The merge gate also requires required checks green, every bot finding fixed or answered with a reason, and the branch mergeable and up to date.
- full-stop is the cautious twin. It does the same work but never merges. It also turns off existing auto-merge and pulls the PR out of any merge queue, checking again before each push; before handoff it re-checks the live PR's head, base, checks, reviews and mergeability, and hands back a PR marked ready, waiting for approval, or blocked.
- Review by other models, with a budget. council runs one flagship model from each of three providers (the example names Claude Opus, GPT Sol and Gemini Pro) at high reasoning effort, checks each result as it arrives, and, if the work ends in a PR, writes each delegate's score per round into the PR description. sol-loop uses a single GPT Sol reviewer for three rounds by default; it extends to five only for security, permissions, auth, data loss or state/concurrency while a blocker or should finding stays open with evidence, or when the user asks, and stops early when only nits remain. Both skills push back against review-driven scope creep.
- A watcher instead of sleep loops. babysit-pr forbids home-made polling. Outside T3 Code it runs the bundled pr-wait script, which prints one line per real event (a failed check, a new review, the base branch moving under the PR's files) so Claude Code's Monitor tool wakes the agent only when something happens. It exits 0 when the condition is met, 1 when the agent has to act, and 2 on bad input or when another pr-wait already watches the PR. Its first run records what already exists and reports only newer events, so early silence is normal.
file-pr wants a description that opens with the problem in plain words rather than a list of changed functions, and ends with the model and harness that wrote the change. modernize picks MODERNIZE, REWORK or CLOSE before pushing anything, and never closes a PR without permission. takeover by default imports the original commits onto a new branch for a replacement PR, preserving authorship and co-author trailers. leaving-pr-comments puts a note callout naming the model and the user on agent comments, but not on PR descriptions, commit messages or text you asked to post verbatim.
How it was built with Claude
Of the six commits on main, all authored by Theo Browne, three carry a Claude Co-Authored-By line: two name Claude Opus 5.5 (the initial publish on October 24 and a full-stop trigger fix on October 27) and one names Claude Fable 5.2 (a November 11 council change that has the agent act on each review as it arrives). The repository has no CLAUDE.md, AGENTS.md or docs folder recording how the work was split.
Try it with your AI. You can borrow the full-stop idea without installing anything. This prompt asks your agent to take one change to a reviewed, ready-to-merge PR, and to stop before merging.
Take one change to a PR that is ready for me to merge. Do not merge it.
Repo default branch: [main]
The change: [describe the bug or feature]
First check: run git status --short and confirm gh is authenticated; if it is not, stop
and tell me. If the tree has unrelated changes, stop and ask me. Check whether a PR
already exists for this branch.
1. If we are on the default branch, create a branch. Make the smallest change that fixes
the case I described.
2. Prove it: reproduce the failure, apply the fix, show the same reproduction passing,
and run the checks the repo's CI runs (tell me which). For UI changes, run the app
and look. If you cannot prove it, stop and tell me.
3. Run git fetch origin, then review your diff against git merge-base HEAD
origin/[main]. Fix real bugs in the intended case; list nits as follow-ups.
4. Draft a PR title and description: the problem in plain words, the fix, one line on
proof. Check it for secrets and private details. Show me the draft and wait for my
confirmation before running gh pr create.
5. After filing, if auto-merge is on or the PR is in a merge queue, turn it off and tell
me. Then check gh pr checks and new review comments. Verify each finding against the
code first. Push only to this PR's branch; if a rebase needs a force push, show me
and wait for my confirmation, then use --force-with-lease. If I do not answer, stop.
6. Never merge, enable auto-merge, push to the default branch, close any PR, or delete
anything.
Finish with: the PR link, whether it is ready, waiting for approval or blocked, your
proof, reviewer findings and what you did, open risks.Known limits
- Broad permissions by design. full-send overrides ask-first testing rules in your project files and merges without a human. Use it only on repos where an unattended merge is acceptable, and read the grant before copying it.
- The cross-model skills need more than one provider. In T3 Code, council and sol-loop use delegate_task; elsewhere you need a harness whose subagents can run other providers' models.
- babysit-pr assumes specific review bots. pr-wait's default ready mode waits for Greptile or Macroscope, and an open issue reports that it never finishes in repos with neither; the list is hardcoded in the script, so CLAUDE.md cannot change it. The skill says to use --until green there; fix PR #16 is open.
- Open reports on sol-loop: one issue says a review round has no deadline and an agent waited 9 to 10 hours several times in one night; another says Sol runs at its default reasoning effort, which the reporter found to be low. Fix PRs #9, #17 (deadline) and #15 (effort) are open; the effort issue suggests asking for high effort in your global agent instructions.
- No Windows support for pr-wait: it imports fcntl, which Windows lacks; fix PR #11 is open.
- Privacy: file-pr has the agent check PR titles, descriptions, quotes and attachments for secrets and private details, and leaving-pr-comments has it check each whole comment, including quotes and attachments, before posting. These are agent instructions, not a filter.
- Licence: if you copy and redistribute the skills, MIT requires keeping the copyright and permission notice.
Related on claudefox
- Product Engineering: 45 free agent skills for shipping a product with Claude Code and Codex
- Floorplan: a Claude Code and Codex plugin that maps your codebase in plain English
- PhoneBox turns an old Android phone into a desk panel and control deck for Windows, Linux and macOS
- BAMA: Claude Code and Codex skills that plan, design and build native macOS apps
Problems and feedback
Use the repository's public GitHub issues and pull requests; there is no CONTRIBUTING or support file. Fixes for the problems above (#9, #11, #14, #15, #16, #17) were still open on November 15, so check them before filing a new issue.
FAQ
Does tstack work without T3 Code? Yes, for most skills: babysit-pr uses pr-wait instead of watch_pull_request, council uses the harness's own subagents, and full-send asks questions with AskUserQuestion in Claude Code or request_user_input in Codex.
What is the difference between full-send and full-stop? Both prove, review, file and babysit the PR. full-send merges once the merge gate passes; full-stop never merges and hands the PR back as ready, waiting for required approval, or blocked.
Which models does council use? The ones you name in your CLAUDE.md or AGENTS.md; otherwise the latest flagship from each of three providers. If it cannot reach three providers, it reports that instead of substituting.
Source: GitHub: t3dotgg/tstack, tstack README, tstack full-send skill, tstack council skill, tstack babysit-pr skill, tstack commit history, tstack full-stop skill, tstack sol-loop skill, tstack pr-wait script, tstack file-pr, modernize, takeover and leaving-pr-comments skills, tstack issues and pull requests
This story first appeared in the claudefox daily briefing for November 15, 2026.