Tristan-Hauber Runs against every ~/.agents/*-notes.md file and anything linked under ~/.agents/notes/. A new agent's file is picked up automatically. Local files only — via Codex directly or Cowork's device_bash.
scripts/notes_signals.py [~/.agents] runs first — read-only, no LLM calls: per-file line/word counts vs the size thresholds, dated-entry extraction, and difflib near-duplicate candidates (3+ dated entries within a file; matches across files). Reason from its output and the flagged bullets, not every file cold.
- Size: use the script's
size_flag; growth belongs in split-out topic docs, not the root. - Clarity: flag vague, unbounded, or now-contradicted statements — a later entry superseding an earlier one that isn't trimmed or marked stale.
- Repetition: for each script-flagged group, confirm it's a real recurrence before reporting.
- Recurring task-shaped group: report as a skill candidate,
weekly-skill-gap-review's shape — name, evidence, gap, proposed change, benefit, cautions. Do not create or edit a skill. - Recurring or oversized reference material: propose a split into
~/.agents/notes/<topic-slug>.md, shared if more than one file touches the topic. Show the diff; do not apply it. - A
cross_file_candidatesmatch: note it once as a shared-notes candidate, not per-file repetition. $workflow-analysis/$mistake-analysiscover workflow and git/store mistakes, not notes hygiene — don't fold their output in or query their store.
Output: one compact report, sectioned per file, plus cross-file findings from #6. Nothing here edits a notes file, its topic docs, or any skill; every action needs explicit confirmation to apply.