Improvement

---
name: career-improvement
description: Prepare consented synthetic product proposals and evaluate optional isolated retrospective improvements without exporting client memory or automatically adopting changes.
---

# Shared improvement proposals

Check explicit enabled feedback consent, disclosure version and authorized GitHub identity before any issue/PR send. Unset or disabled means no send. Explain at setup that only generic descriptions and synthetic reproductions may be submitted; users can disable this later. No scanning of every upstream issue is required; maintainer deduplication owns that work.

Reconstruct a problem from scratch using invented people, IDs, URLs, paths and text. Do not redact a private log and assume it is safe. Inspect title/body/filenames/patch/commit metadata and any attachment for customer facts, source text, secret values and identifying environmental details. Uncertain content stays local for human inspection. Never attach raw journals, screenshots, memory, CVs or correspondence.

Use the user's own authorized account to create an issue or isolated tested PR. Do not merge, request merge powers, push to the default branch or alter shared instructions automatically. Report the proposal link and keep the private provenance separately. A client issue cannot disclose another client's existence or data.

For a patch, use a fresh product-only checkout and dedicated `codex/feedback/` branch. Commit only reviewed product files with deliberate public metadata. Run `feedback review-patch` against the explicit upstream base **before pushing**; it checks newly introduced history as well as the final diff. Then push only that reviewed branch to the user's authorized fork and submit the draft PR. Submission checks the remote head against the unchanged reviewed revision. A secret removed in a later commit is still present in history and blocks this route.

## Optional bounded retrospective study

A replay is a separate explicitly requested experiment, not background optimization. Freeze as-of task inputs, versions, referenced files and hashes. Keep future corrections and evaluator ground truth inaccessible to baseline/adaptation workers. Predeclare discovery and held-out groups to avoid near-duplicate leakage; label faithful, partial or unusable reconstructions.

Before execution prove no production credentials, authenticated browser/mail tools, private host files, unrestricted network or live send/publication route is reachable. If isolation cannot be enforced, mark execution blocked; inert analysis can be recorded as analysis, never a completed Harness² trial. Do not ship an unlicensed local adapter or assume its prompt is a sandbox.

Compare paired artifacts blindly on factual fidelity first, then fit, visual/writing quality and actual measured work. Record unavailable cost/token information as unavailable. Retain baseline on ties, regressions or unverifiable gain. A useful pattern needs held-out evidence and explicit maintainer adoption; a negative result is valid. Offline comparisons cannot prove delivery reliability, hiring benefit or market causality.