Autoresearch Self Loop

What you get

• An agent loop on one metric. The agent edits the code, the plugin runs your benchmark and checks, and every run is appended to a session log in .auto/. The log lets a loop survive restarts and context resets.
• Keep rules the plugin enforces. A kept run is committed with a Result: trailer, and its commit is logged. A keep is refused when the last benchmark or its checks failed, or when nothing outside the session files changed.
• Reverts for discarded runs. A discarded, crashed, or checks-failed run reverts every change outside the session files and deletes new untracked files that Git does not ignore. The session files stay.
• A thread panel. It shows the metric, the baseline, the best kept run, and the confidence, with the run table and the Start/Resume, Stop, Finalize, and Export buttons.
• A port of pi-autoresearch. This is a close BB port of pi-autoresearch v1.9.0 (MIT License) by Tobi Lutke and David Cortés. After setup, the loop runs without a sequence of manual approval steps. BB permissions and repository protections still apply.

How it works

1. During setup, the agent writes the benchmark script and the session brief into .auto/. The first run of a segment is the baseline.
2. Each run executes the benchmark and reads its METRIC name=value lines. If the benchmark passes and .auto/checks.sh exists, the checks run next. The agent then logs the run as keep, discard, crash, or checks_failed.
3. The loop stops when you press Stop, or when the maxIterations limit in .auto/config.json is reached.

Requirements

• BB 0.46 or later.
• A Git checkout with at least one commit. Starting a session needs a clean working tree. The session files do not count as changes.
• A committer identity (user.name and user.email) in the checkout, because each keep makes a commit.
• A benchmark that prints METRIC name=value lines. The benchmark and checks are your repository's own commands, and the plugin runs them as written.
• On a remote environment host, the plugin works through the BB server, so the server must be running. The host needs POSIX with sh, bash, and git.