sort-unsorted

name	sort-unsorted
description	Re-classify rows in INDEX/sources.md and INDEX/tools.md that landed in the `unsorted` bucket — either with `category=unsorted` or `kind=unsorted`. Re-applies the categories.md vocabulary and the kind heuristic with a fresh attempt, using GitHub topics, description, README excerpts. Use on demand or on a slow weekly schedule to drain the unsorted backlog.

sort-unsorted

The classifying agent runs in add-source and star-sweep are intentionally conservative — when in doubt, a row lands with kind=unsorted or category=unsorted. This skill is the cleanup path: it takes a second, more careful look at those rows.

Inputs

input	type	notes
target	enum	tools | sources | both. Default both.
max_rows	int	Cap per run. Default 50. Prevents thrashing on a large backlog.
dry_run	bool	If true, write the proposal to /tmp/sort-unsorted-proposal.md and stop. Default false.

Procedure

1. Snapshot. Parse INDEX/sources.md and INDEX/tools.md. Collect rows where category=unsorted or kind=unsorted (active rows only — archived stays archived).

2. Enrich each candidate row. For repos, refetch gh api repos/<owner>/<repo> (description, topics, language, README excerpt — gh api repos/<owner>/<repo>/readme then base64-decode and keep the first 4000 chars). For sources, fetch the page title and meta description if not already present.

3. Re-classify. Apply the heuristic in .claude/skills/research-standard/TAXONOMY/categories.md:

   • Category: match against scope columns. Pick the most specific. Falls back to unsorted if still no fit.
   • Kind (tools only): topics → description verbs → README content (look for [[bin]] in Cargo.toml, main package in Go, bin/ directory, CLI entry-point in package.json, "Usage" sections that show shell commands). Falls back to unsorted if no clear signal.

4. Decide stay-or-move.

   • If new classification differs from current → write the row update.
   • If new classification is also unsorted after second pass → leave as-is and mark with notes: "second-pass uncertain" so it isn't re-tried until something changes.

5. Apply. Write updates back to INDEX/sources.md and INDEX/tools.md. Keep id ordering. Bump the row's last_seen to today.

6. Open a PR. Branch sort-unsorted/<YYYY-MM-DD>. Labels sort-unsorted,automated,agent-authored. PR body: per-category re-distribution counts, list of rows still uncertain.

Constraints

• Only operates on rows already in INDEX. Never adds new rows. Never archives.
• Never widens a category — once a row was sorted into a fine category, it stays there. sort-unsorted only narrows from unsorted.
• The heuristic is the same one in categories.md — read that file, do not invent rules.
• Respect notes: "second-pass uncertain". Skip those rows; humans can re-tag them by editing notes.

Verification checklist

• Total rows touched ≤ max_rows.
• No row moved from a fine category back to unsorted.
• Updates committed with a single branch and PR; the auto-review workflow gates the merge.

Related

• /sort-unsorted — slash command wrapper.
• add-source — applies the same heuristic to a single new row.
• star-sweep — bulk add that may produce unsorted rows.