vault-compact

What this skill does

# Vault Compact

Reduce repeated thinking across the active vault. This skill is for semantic
overlap, not just exact duplicate text: compact passages when they do the same
job in the knowledge system.

Portent reference: `../references/portent-knowledge-base-spec.md`.

Prefer compaction when files represent the same Portent object or repeat the
same claim for the same object. Do not merge merely because objects share a
Topic. Preserve `belongs_to` and `related_to` metadata when absorbing files.

## Recency Priority

Work the active surface newest-first, ranked by each object's `updated`
frontmatter date (fall back to file modification time when `updated` is absent).
Concentrate scrutiny on the freshest material — recently-touched objects are
where active thinking lives and where new mismatches, overlaps, and gaps appear.
Long-untouched objects are assumed settled and rank lower, but are never
excluded: they remain valid link targets, merge destinations, and canonical
homes, and still get a structural check (broken links, orphaning).

Seed overlap-candidate groups from recent files. When a recent note overlaps an
older canonical one, keep the older established note as the canonical
destination, consistent with the Canonical Destination Rules below.

## Command

```
vault-compact [--scope notes|ideas|projects|all] [--mode report|apply-safe|apply] [--aggression conservative|normal|aggressive]
```

## Parameters

- `--scope notes|ideas|projects|all` (default: `all`)
- `--mode report|apply-safe|apply` (default: `report`)
  - `report`: identify overlap and propose canonical destinations only
  - `apply-safe`: apply high-confidence local compaction only
  - `apply`: also merge high-confidence cross-file overlap when provenance and
    links can be preserved; delete absorbed curated files when appropriate
- `--aggression conservative|normal|aggressive` (default: `conservative`)
  - `conservative`: exact/near-duplicate sections and obvious repeated claims
  - `normal`: same-claim overlap with unique examples preserved
  - `aggressive`: broader synthesis; appropriate for shipped/archived project
    compaction or when explicitly requested

## What Counts As Overlap

Compact when multiple passages are doing the same job:

- repeated thesis or framing
- duplicated audience/problem/value-prop language
- duplicated project status or next-action text
- same source summary in multiple curated notes
- same idea split across multiple idea notes
- concept explanation repeated instead of linked to the concept page
- stale restatement that is superseded by a clearer canonical section
- shipped project launch artifacts, drafts, and research that have served their
  temporary purpose

Do not compact just because notes share keywords.

## Unit Of Judgment

Compare passages as claim-like units:

- thesis
- audience
- problem
- product idea
- decision
- evidence
- example
- caveat
- open question
- next action
- implementation detail
- source summary

If two units make the same claim, collapse them. If one adds a distinct example,
decision, caveat, source, audience, or implementation detail, preserve that
detail.

## Workflow

1. Read `AGENTS.md`, then `index.md`.
2. Confirm the scoped directories:
   - `notes/`
   - `ideas/`
   - `projects/`
3. Exclude `raw/`, `raw/processed/`, `raw/assets/`, and `archive/` unless the
   user explicitly asks for an audit that includes archived curated material.
4. Use qmd first for semantic retrieval:
   - `qmd query` for themes and semantic overlap
   - `qmd search` for exact phrases, titles, URLs, and aliases
   - `rg` for migration residue, raw strings, and link rewrites
5. Build candidate groups by canonical topic, project, idea, source, or claim.
6. Classify each group by Portent object identity, confidence, and compaction
   action.
7. In `report` mode, stop after a compact report.
8. In `apply-safe` mode, apply only local, high-confidence edits inside one file
   or obvious link substitutions to canonical concept notes.
9. In `apply` mode, apply high-confidence cross-file compaction only when all
   unique content and provenance can be retained.
10. Delete absorbed curated files only after the canonical destination preserves
    their unique claims, examples, decisions, and provenance.
11. Update `updated:` frontmatter on edited curated notes where practical.
12. Run `qmd update` after any edits, moves, or deletions.
13. Append one concise line to `log.md` for every material vault change.

## Compaction Actions

- `dedupe`: remove exact or near-exact repeated paragraphs, bullets, or
  sections.
- `collapse`: replace equivalent claims with one sharper canonical statement and
  preserve unique examples, caveats, evidence, and decisions below it.
- `route`: replace repeated local explanation with a wikilink to the canonical
  note or section.
- `merge-plan`: propose a merge, canonical destination, unique content to
  preserve, and link rewrites; do not edit.
- `absorb-delete`: after a successful merge, remove curated source files whose
  useful content now lives in the canonical destination.
- `link-only`: add or propose a relationship link because notes are related but
  distinct.

## Confidence Rubric

### High Confidence

- Same source URL or same clipped source identity.
- Same title, slug, project, or idea identity.
- Same Portent object identity or same claim serving the same object.
- One note explicitly continues, replaces, or supersedes another.
- Passages make the same claim with compatible meaning.
- Differences are examples, wording, or ordering, not substance.
- Canonical destination is obvious.

High-confidence items may be compacted in `apply-safe` when local to one file,
or in `apply` when cross-file provenance and links can be preserved.

### Medium Confidence

- Same broad thesis with a meaningful caveat in one passage.
- Same product shape but slightly different audience or monetization.
- Same concept repeated across notes but no existing canonical destination.
- Canonical destination is plausible but not obvious.

Medium-confidence items become merge plans or concept-promotion candidates.

### Low Confidence

- Same broad domain only.
- Similar vocabulary but different use cases.
- Similar AI-generated structure but different actual content.
- Different lifecycle states or strategic decisions.

Low-confidence items should remain separate; propose links only when useful.

## Canonical Destination Rules

### Notes

- Prefer canonical concept pages under `notes/concepts/` for repeated theses.
- Prefer existing topic notes over creating new notes.
- External clipped notes are evidence, not owned synthesis. Preserve source
  identity and do not absorb complete external text into owned notes.
- Repeated source summaries should cite complete source records under
  `raw/processed/YYYY-MM-DD/` when used for synthesis.
- Preserve `belongs_to` and `related_to` metadata when compacting or absorbing
  notes.

### Ideas

- Keep the canonical idea in its lifecycle location under `ideas/`.
- Preserve distinctions in audience, wedge, pricing, validation status, and kill
  criteria.
- If two ideas share a product shape but differ in customer or strategy, do not
  merge; create a comparison or shared concept link instead.

### Projects

- Keep project PRDs, plans, decisions, research, and postmortems separate unless
  one is clearly obsolete or duplicative.
- Canonical project state should live in Portent frontmatter, the main project
  PRD, README, or active index navigation according to the local convention.
- Do not merge active, paused, shipped, and rejected rationales into one generic
  project summary; lifecycle context matters.
- For shipped projects, run heavier compaction after the shipping transition.
  Merge temporary launch materials, tweet drafts, old outlines, research
  snippets, and implementation leftovers into one canonical shipped project
  record when they no longer serve active execution.
- Shipped project comp