autoresearch-suite

Define a measurable goal, a scope, and a verify command, then invoke the loop. Each iteration makes one focused change, commits it, runs verification, keeps improvements, and reverts regressions, logging every result until the cap is reached.

Any mechanically verifiable target works: test coverage percentages, benchmark scores, bundle size in kilobytes, p95 latency, lint or type error counts. The verify command must output a comparable number or a pass/fail result.

Yes. The loop commits before every verification so failed experiments can be reverted with git revert, and it reads git log and git diff before each iteration. Repositories without git history are explicitly unsupported.

Only if the user explicitly opts in and can interrupt. By default you set an iteration cap with Iterations: N, which the loop honors strictly to prevent unbounded autonomous changes.

Avoid it when no mechanical verification exists or can be defined, and for subjective work like copy tone or visual design without metrics. For subjective refinement, use the reason subcommand with its blind-judge panel instead.

The security audit is read-only by default, performing STRIDE, OWASP, and red-team analysis without modifying code. The --fix mode that applies changes must never run without explicit user confirmation.