architecture-debt-patcherlisted
Install: claude install-skill sumeetmi2/architecture-debt-visualizer
# Architecture Debt Patcher
Given a `findings.json` already written by the `architecture-debt-visualizer` skill, this skill
opens **one pull request per qualifying finding** — where "qualifying" means:
- `severity` is at or above the caller's `--min-risk` threshold (default `high`),
- the finding has a non-empty, actionable `recommendation`,
- the change implied by that recommendation is small and localized enough to attempt
automatically without human judgment.
Anything that doesn't meet all three is **skipped, not force-fitted into a PR**. Skipped findings
are reported back with a structured reason so the caller (`/eval-architecture`) can surface them.
## When this skill runs
Only when explicitly invoked — either by the `/eval-architecture` caller command (with `--patch`
or `--from-report`), or by a direct user request naming this skill. It **never runs as a
side-effect of running the audit skill on its own**. The audit skill and this skill are decoupled
on purpose: an audit is always safe to run; a patch run mutates real code and opens real PRs.
The `--run-dir` supplied to this skill may point at either:
1. A fresh `mktemp` directory the audit skill just wrote to in this session, or
2. An older audit run directory the user is deliberately reusing (via `/eval-architecture
--from-report <path>` or by naming it in a direct request).
This skill treats both the same way — it re-verifies every evidence citation in step 3 before
touching code, so a slightly stale rep