Orchestrate a prose fix

Keep findings and retries out of the parent context. Parse the request into a target and complete lint command, then run:

bash .agents/skills/codex-fix-prose/scripts/preflight.sh <target> <lint command>

Stop on BLOCKED. On PROCEED, create a plan with update_plan: baseline, writer, mechanical verification, independent review, final lint.

Record the baseline before delegation:

bash .agents/skills/codex-fix-prose/scripts/check-suppressions.sh --baseline <target>

Save the absolute state: path that command prints. It identifies this workflow and records its target. Until cleanup, the script refuses a second workflow on the same target. Pass the state to every delegated writer and reviewer in this run.

Delegate codex-write-prose-fix with fork_turns: "none". Give it only the repository root, target, workflow state, lint command, and instruction to follow that skill. Wait for its terminal report. Never edit the target in the parent.

Run check-suppressions.sh --verify <state> <target>. Any finding means the run isn't clean. Delegate codex-review-prose-fix with fork_turns: "none", giving it the repository root, target, and workflow state. If it returns changes required, save those findings in <state>/notes, delegate the writer again with the same minimal prompt, and repeat verification and review. Bound this writer/reviewer loop at three review rounds.

Finally run the complete lint command once in the parent. Always run check-suppressions.sh --cleanup <state> before returning once no delegated agent needs the state, even when the outcome is unresolved or blocked. Report PROSE: CLEAN only when the writer reported clean, the suppression check is silent, the independent reviewer returned REVIEW: CLEAN, and the final lint command passes. Otherwise return the short unresolved or blocked report. Mark each plan item as it completes.

Safety rules:

• Only the delegated writer may edit the target, and the target is the only file it may edit.
• Never change lint configuration, dictionaries, hooks, or suppression comments.
• Don't infer clean from silent output unless preflight proved the target has prose rules.
• Don't expose full lint findings in the parent response unless the workflow blocks and the operator needs a decision.