Agent can wave off human review

A review bot asks for "needs human review". The merge gate in tstack's `full-send` skill says the agent can ignore that if it has high confidence in its changes. The skill runs with nobody watching and merges the PR itself.

tstack is a set of agent skills for Claude Code and Codex that cover the boring middle of shipping a change: filing the PR, watching CI, answering review bots, and merging. The author, Theo (t3dotgg), says they use these skills every day. Each skill is a folder with a SKILL.md, triggered by a phrase like "full send" or "council". The problem it targets is an agent that stops early and leaves a PR sitting while checks and bots are still pending.

How it works

The skills are plain markdown instructions, and several call each other. full-send is the big one. Per its SKILL.md, it takes a change from the current state to a merged PR with nobody watching. The steps are: finish the change, prove it, get review, file the PR, babysit it, merge, then report.

"Prove it" has a specific meaning in the file. Proof is evidence the agent observed in that session, such as command output, a reproduction, or a screenshot. The file says green mocked tests do not prove that the user experience works.

Review comes from sol-loop, which sends the diff to GPT Sol and loops. The default budget is 3 rounds. It extends to 5 only for security, auth, data loss, or concurrency issues. The file tells the agent to make the smallest change that satisfies the reviewer and to push back on scope creep.

council asks three models from different providers for review, collects a score out of 10 from each, and puts a table in the PR description. It repeats until the delegates are satisfied.

babysit-pr waits with a watcher instead of a sleep loop. In T3 Code that is watch_pull_request. Elsewhere it is a pr-wait script. The merge gate in full-send requires green required checks, finished review bots, no open change requests, and a mergeable branch. The file says never to use --admin or other bypasses.

Try it

git clone https://github.com/t3dotgg/tstack
cp -R tstack/skills/* ~/.claude/skills/   # Claude Code
cp -R tstack/skills/* ~/.codex/skills/    # Codex

Copy all the skills, because they depend on each other. For example, full-send uses sol-loop, file-pr, and babysit-pr.

Worth knowing

The README lists requirements. You need an authenticated gh, and python3 for the pr-wait script. council and sol-loop need subagents running on other providers' models. T3 Code does this with delegate_task. Otherwise you need a harness that can run those models. The sol-loop file says that if no route runs Sol, it reports a blocker and does not substitute a model.

full-send gives the agent permission to merge and to force-push its own branch with --force-with-lease. It also overrides ask-first testing rules in a project for that task. It does not cover production changes or data deletion, which the file says to ask about.

The README calls the skills starting points and suggests putting personal settings, like council members, in CLAUDE.md or AGENTS.md. The license is MIT.

This is for developers who already run agents on PRs and want a written process for the review and merge loop. The repo data shows no benchmarks or test results.