Theo's agent skills, boxed up and given away

What it does

tstack is a set of agent skills from Theo (of T3 fame) for Claude Code and Codex. Each skill is a named routine you invoke with a phrase like “full send” or “babysit” — and it handles a chunk of the PR lifecycle: writing, reviewing, filing, watching CI, and merging. It’s glue, but well-organized glue: the skills compose, with full-send calling sol-loop, file-pr, and babysit-pr under the hood.

The interesting bit

The review skills are the standout. council gets three models from different providers to review a change and loops until they agree; sol-loop does the same with two agents. It’s a cheap way to get adversarial review without scheduling a human. Also notable: takeover finishes someone else’s abandoned PR and keeps their credit, which is a nicer etiquette than most agent workflows bother with.

Key highlights

• full-send takes a change from idea to merged PR autonomously; full-stop does the same but stops for your approval before the merge
• file-pr writes PRs that lead with the problem, not a changelog dump
• babysit-pr watches CI and review bots, fixing real findings
• Skills degrade gracefully — they use T3 Code features like delegate_task when available, native tools otherwise
• Designed to be edited: personal settings live in CLAUDE.md/AGENTS.md, so updating the skills doesn’t clobber your tweaks

Caveats

• Requires an authenticated GitHub CLI, and council/sol-loop need a harness that can run models from multiple providers (T3 Code’s delegate_task or your own setup)
• babysit-pr needs python3 for its pr-wait script
• Skills have interdependencies — you’ll need to install the ones each skill calls

Verdict

Worth a look if you live in Claude Code or Codex and want agents to handle the tedious middle of the PR pipeline. If you don’t already have multi-provider model access, the review-loop skills — the best part — won’t do much for you.