t3dotgg/tstack

This is a small collection of prompts-as-skills for people already driving coding agents through pull requests, not a new agent runtime. Its value is the opinionated workflow around reviewing, filing and shepherding changes with little attention.

The problem

You have a change ready and want it merged, not just written. The usual routine is to run the agent, read the diff yourself, write a PR description, wait for CI and review bots, then come back to fix whatever they flag. Each of those steps is a separate context switch, and an agent left alone can open a PR without proving the change works.

The skills package that routine into named workflows. Saying "full send" asks the agent to prove the change, review it, file the PR, watch it through CI and merge it; "full stop" does the same but hands the merge back to you. The result is fewer manual handoffs between writing code and landing it.

What's different

Instead of one monolithic agent command, each skill is a short, editable instruction set with a single job: file-pr writes a problem-led PR description, babysit-pr watches CI and fixes findings, modernize rebases old PRs and asks whether they still belong, takeover continues someone else's work while keeping their credit. Some skills compose: full-send calls sol-loop, file-pr and babysit-pr, so you install the dependency chain rather than one file.

The multi-model pieces are the most distinctive. council asks three models from different providers for review and loops until they agree; sol-loop runs a review loop between agents until both agree. That is useful when you want disagreement surfaced before merging, but it depends on a harness that can reach other providers' models. The README notes T3 Code does this via delegate_task; otherwise you supply your own setup. The skills also use T3 Code features when present and fall back to native tools, so the same instructions do not require that environment.

The trade-off is that these are starting points, not a finished product. The author explicitly tells you to edit them and put personal settings in CLAUDE.md or AGENTS.md, which implies the review criteria and model choices are yours to own. Everything also hinges on the agent faithfully following prose instructions.

Who should try it

A good fit is a developer who already uses Claude Code or Codex on GitHub repositories and wants a repeatable PR lifecycle: file, review, babysit, merge. It suits teams willing to read and tune the skill text to match their review standards, especially those who want a second opinion from models on other providers.

A poor fit is someone looking for a standalone tool with a UI or a guaranteed merge pipeline. Without an authenticated gh and the skill dependencies installed, full-send and takeover cannot function as described. If you only work locally and do not use GitHub pull requests, the core skills have little to act on.

Requirements and limitations

You need an authenticated GitHub CLI, python3 for babysit-pr's wait script, and a harness that can run models from other providers for council and sol-loop project instructions. The install is a clone plus copying the skills directory into ~/.claude/skills/ or ~/.codex/skills/; there is no package manager or update mechanism described README.

Because skills call each other, partial installs produce broken workflows. The README lists the dependencies, but you must track them yourself when editing or copying individual skills. The project is MIT licensed license.

HotRepo's take

HotRepo first recorded the repository on 2026-10-26, two days after creation, at 829 total stars and #18 on the NEW-REPOSITORY board (新项目榜, not the overall Trending board); this is a snapshot, not daily growth, and there is not enough comparable board history to judge persistence project page. Treat the interest as a single dated observation. Before adopting, check whether your agent harness can actually run the cross-provider models that council and sol-loop depend on; without that, the most distinctive skills are the ones you cannot use.