improve-codex

The plan comes first

I built improve-codex because an audit finding is not yet an implementation task. “Fix the billing” leaves an executor to invent the scope, failure cases and proof.

The orchestrating session turns findings into self-contained plans. Sol checks those plans against the repository. Terra implements accepted work. The main session and Sol review the resulting diffs.

Codex or Claude Code can orchestrate. The important split is between proposing work, implementing it and checking it.

Use the repository’s execution path

A repository with the full reviewed Symphony/Clanker contract can run its own serialized execution and integration workflow. Elsewhere, the portable runner creates an isolated git worktree for each plan.

That fallback confines writes, removes browser and MCP tools, applies a timeout and runs at lower CPU priority. It returns browser checks as skipped for the main session to finish. It does not try to install a fleet into a repository that only needs one worker.

A clean exit is not a review

The portable runner fails when an executor produces no usable status, even if the process exits zero. Critic verdicts are tied to a particular run, and each round has its own report path. A later attempt cannot erase the first review.

After two rounds, a surviving major finding means the plan needs to be split or reconsidered. Repeating the same request is not a repair for unclear intent.

Approved still means waiting

The skill returns approve, revise or block. It never merges or pushes approved implementation branches. The human merge decision and any remaining browser verification stay visible.

Read how the current workflow runs or install the skill.