improve-codex

One agent should not grade its own work

One agent can audit a codebase, propose a fix, implement it, and declare itself correct. That is convenient. It also lets one set of blind spots survive every stage of the work.

I split those jobs. One agent audits and plans. A bounded executor implements one plan. Other agents review the plan and the diff. A human decides whether to merge.

The plan comes first

The expensive reasoning happens before implementation. improve-codex begins with a read-only audit, turns each finding into a self-contained plan, then reviews that plan before an executor receives it.

That order matters. A cheap executor with a precise plan can do good work. A powerful executor with a vague plan can produce a large, polished mistake. Spending intelligence on scope, done criteria, and repository-specific verification is usually the better trade.

Worktrees are a boundary

Every portable execution runs in its own git worktree. Browser tools, MCP servers, plugins, watch processes, and interactive surfaces are removed from the worker. CPU priority and runtime are bounded. Report paths are validated before the run starts, and an executor that exits without a usable status is treated as failed.

Isolation is not theatre here. It keeps one plan from contaminating another, protects the main checkout, and leaves a branch that can be inspected or thrown away without archaeological work.

Incomplete review fails the run

The workflow reviews both the plan and the diff. Reports cannot silently overwrite an earlier round, critic verdicts are tied to the run that produced them, and incomplete completion reports are rejected instead of guessed around.

Review is capped at two rounds. If a major finding survives the second, the plan is too broad or the approach is wrong. It gets split or retired rather than pushed through a third ceremonial pass.

Those rules came from real failure paths: vacuous assertions, incomplete reports, stale local facts, and reviewers unable to verify a claim against a live local service. Each one made the system stricter in the place where the previous version had been too trusting.

It stops before merge

The skill can return approve, revise, or block. It never merges or pushes an implementation branch. It is less autonomous than the usual agent demo because I want to inspect serious repository work before it lands.

Install improve-codex