---
name: ideas
description: Offer evidence-backed improvement options without changing code or inventing a backlog.
category: analysis
---

# Improvement options

This is a proposal session, not implementation. Keep it write-free unless the user explicitly requests a scan artifact or chooses an action.

1. Query relevant capabilities/routes/decisions and inspect recent history for the requested area. Do not load every index or assume an unlisted package exists.
2. Inspect promising gaps directly: missing business tests, duplicate capabilities, weak access policy, user friction or measured performance trouble. Names, missing exports and TODOs are leads, not defects by themselves.
3. Present a short finite list with source/evidence pointers, benefit, rough effort, risk and recommended priority. Separate confirmed problems from hypotheses.
4. Change nothing until the user selects scope. Do not turn discussed alternatives into accepted decisions, findings or new skills.

Use `AGENTS.md` capture and package-recommendation policy. Effort estimates are uncertainty-bearing advice, not commitments.