improvement-loops

Designs what an improvement or optimizer loop should climb for a given project and goal, and what must not get worse while it does: goal metrics measured on the real path, guardrails, and counter-checks aimed at the cheapest way to fake each gain, so no single number can be reward-hacked. Proves each proxy tracks the outcome it stands for, measures its noise, writes the scenarios as data in the loop state, and proposes the fill for AGENTS.md's Measurement section. Rerun in review mode when the goal changes. User-invoked only.

Runs one improvement pass over a project in the direction its north star points, without breaking what it does today: faster in real use, less code, a truer account of what it does, rules that now run as checks. Works with the repo's AGENTS.md sections (North star, Measurement, Reporting, Loop state); state carries between runs in a ledger, bookmarks and a scoreboard; each change lands as its own commit with evidence on a worktree branch; the report leads with what needs the owner. User-invoked only, because a run spends hours and many subagents.

Speeds up a codebase as far as it will go in one session, against a hard target (every primary benchmark at least 1.2x faster than the base commit by default) without regressing correctness, output quality or safety. Works with the repo's AGENTS.md sections (North star, Measurement, Loop state); base and branch worktrees, exact counts proven against wall-clock time, read-only reviewers one lens each, a re-check of every headline result on the final commit. User-invoked only, because a run spends hours and many subagents.

Tightens a repository's always-loaded agent instructions (AGENTS.md, CLAUDE.md, unconditional rules): every line must prevent a mistake, and each one that doesn't is moved to the narrowest home that loads when it matters (a path-scoped rule, a skill, a hook or test, the status doc, the user-level file) or cut. Fills the gaps a newcomer falls into: commands, done criteria, the stops wanted, hazards, settled decisions. Re-points everything that cited the old text, and verifies with a fresh-context diff review and spot-tested tasks. User-invoked only.