Wiki Status

Wiki Status — Audit & Delta

You are computing the current state of the wiki: what's been ingested, what's new since last ingest, and what the delta looks like. This helps the user decide whether to append (ingest the delta) or rebuild (archive and reprocess everything).

Before You Start

Writing profile: Before drafting or rewriting natural-language Markdown, read and apply the Writing Profile Resolution section in llm-wiki/SKILL.md. Framework schema, provenance, safety, and operation-specific requirements take precedence. Apply WRITING.md preferences only to generated _insights.md prose; keep the analyser snapshot verbatim.

1. Resolve config — follow the Config Resolution Protocol in llm-wiki/SKILL.md (inline @name override → walk up CWD for .env → global config → prompt setup). This gives OBSIDIAN_VAULT_PATH, OBSIDIAN_SOURCES_DIR, CLAUDE_HISTORY_PATH, and CODEX_HISTORY_PATH.
2. Read .manifest.json at the vault root — this is the ingest tracking ledger

The Manifest

The manifest lives at $OBSIDIAN_VAULT_PATH/.manifest.json. It tracks every source file that has been ingested. If it doesn't exist, this is a fresh vault with nothing ingested.

Source keys are portable — never bare absolute paths. The canonical key is vault-relative for in-vault sources (Raw/database/x.pdf), ~-relative for sources under $HOME (~/.claude/projects/.../abc.jsonl), or a pseudo-key (repo:/url:/agent:) for sources with no portable path form. Never mix forms — the same file would be tracked twice and re-ingested. See llm-wiki/SKILL.md → .manifest.json (Source key contract v2). Convert a legacy absolute-key manifest with scripts/manifest.py migrate <vault>.

{
  "version": 1,
  "last_updated": "2026-04-06T10:30:00Z",
  "sources": {
    "Raw/papers/attention.pdf": {
      "ingested_at": "2026-04-06T10:30:00Z",
      "size_bytes": 4523,
      "modified_at": "2026-04-05T08:00:00Z",
      "source_type": "document",
      "project": null,
      "pages_created": ["concepts/transformers.md"],
      "pages_updated": ["entities/vaswani.md"]
    },
    "~/.claude/projects/-Users-name-my-app/abc123.jsonl": {
      "ingested_at": "2026-04-06T11:00:00Z",
      "size_bytes": 128000,
      "modified_at": "2026-04-06T09:00:00Z",
      "source_type": "claude_conversation",
      "project": "my-app",
      "pages_created": ["entities/my-app.md"],
      "pages_updated": ["skills/react-debugging.md"]
    }
  },
  "projects": {
    "my-app": {
      "source_repo": "github.com/owner/my-app",
      "source_cwd_hint": "~/.claude/projects/-Users-name-my-app",
      "vault_path": "projects/my-app",
      "last_ingested": "2026-04-06T11:00:00Z",
      "conversations_ingested": 5,
      "conversations_total": 8,
      "memory_files_ingested": 3
    }
  },
  "stats": {
    "total_sources_ingested": 42,
    "total_pages": 87,
    "total_projects": 6,
    "last_full_rebuild": null
  }
}

Step 1: Scan Current Sources

Build an inventory of everything available to ingest right now:

Documents (from OBSIDIAN_SOURCES_DIR)

Glob each directory in OBSIDIAN_SOURCES_DIR for all text files
Record: path, size, modification time

Claude History (from CLAUDE_HISTORY_PATH)

Glob: ~/.claude/projects/*/          → project directories
Glob: ~/.claude/projects/*/*.jsonl   → conversation files
Glob: ~/.claude/projects/*/memory/*.md → memory files
Record: path, size, modification time, parent project

Codex History (from CODEX_HISTORY_PATH)

Glob: ~/.codex/session_index.jsonl            → session inventory index
Glob: ~/.codex/sessions/**/rollout-*.jsonl    → session rollout transcripts
Glob: ~/.codex/history.jsonl                  → optional local history log
Glob: ~/.codex/archived_sessions/**/rollout-*.jsonl → archived rollouts (if user wants archive coverage)
Record: path, size, modification time, inferred project from cwd when available

Any other sources the user has pointed at previously

Check the manifest for source paths outside the standard directories.

Step 2: Compute the Delta

Compare current sources against the manifest. Classify each source file:

When a manifest entry has no content_hash (older entry), fall back to mtime comparison only.

For Claude history specifically, also compute:

• New projects (directories in ~/.claude/projects/ not in manifest)
• New conversations within existing projects
• Updated memory files

For Codex history specifically, also compute:

• New rollout files under sessions/**
• Updated session_index.jsonl entries (session title/freshness changes)
• Archived rollout delta only when archive coverage is requested

Step 3: Report the Status

Visibility tally (before rendering the report): Grep frontmatter across all vault .md pages for visibility/internal and visibility/pii tag values. Count:

• public = pages with visibility/public tag or no visibility/ tag at all
• internal = pages with visibility/internal tag
• pii = pages with visibility/pii tag

Include this in the Overview section as Page visibility: N public · M internal · K pii. Skip the line if all pages are untagged (fully public vault).

Present a clear summary:

# Wiki Status

## Overview
- **Total wiki pages:** 87 across 6 categories
- **Page visibility:** 72 public · 11 internal · 4 pii
- **Total sources ingested:** 42
- **Projects tracked:** 6
- **Last ingest:** 2026-04-06T11:00:00Z
- **Staged writes pending:** 4 pages · 2 patches (oldest: 3 days ago)  ← only when WIKI_STAGED_WRITES=true

## Delta (what's changed since last ingest)

### New sources (never ingested): 12
| Source | Type | Size |
|---|---|---|
| ~/Documents/research/new-paper.pdf | document | 2.1 MB |
| ~/.claude/projects/-Users-.../session-xyz.jsonl | claude_conversation | 340 KB |
| ~/.codex/sessions/2026/04/12/rollout-...jsonl | codex_rollout | 220 KB |
| ... | | |

### Modified sources (need re-ingesting): 3
| Source | Last ingested | Last modified | Delta |
|---|---|---|---|
| ~/notes/architecture.md | 2026-04-01 | 2026-04-05 | 4 days newer |
| ... | | | |

### New projects (not yet in wiki): 2
- **tractorex** (3 conversations, 2 memory files)
- **papertech** (1 conversation, 0 memory files)

### Deleted sources (ingested but gone): 0

## Summary
- **Ready to ingest:** 12 new + 3 modified = 15 sources
- **Up to date:** 27 sources unchanged
- **Recommendation:** Append (delta is small relative to total)

## Token Footprint (estimated)

| Scope | Pages | ~Tokens |
|---|---|---|
| core tier | 12 | 18,400 |
| supporting tier | 87 | 94,200 |
| peripheral tier | 43 | 31,600 |
| **Full wiki (all)** | **142** | **144,200** |

Index-only pass (frontmatter + summaries): ~8,900 tokens
Typical query (index + 5 full pages):      ~14,200 tokens

⚠️  Full wiki exceeds 100K tokens. Consider:
  - Demoting peripheral pages (promote tier suggestions from wiki-status insights mode)
  - Running /wiki-lint --consolidate to merge near-duplicates
  - Using wiki-query fast mode for most queries

Step 3b: Compute Token Footprint

After building the status summary, compute the token footprint estimate:

1. Per-tier page sizes — Glob all .md pages. Read the tier: frontmatter field of each (cheap grep). Group pages by tier value (core, supporting, peripheral; unset → supporting).

2. Estimate tokens — For each page, estimate token count as file_size_bytes / 4 (4 chars/token heuristic — no actual tokenizer needed). Sum per tier and total.

3. Index-only estimate — Estimate the cost of an index-only pass: sum len(title) + len(summary) + len(tags) for each page frontmatter (~100 chars each on average), divided by 4.

4. Typical query estimate — Index-only estimate + average full-read cost of 5 pages (total_chars / total_pages * 5 / 4).

5. Threshold check — Read WIKI_TOKEN_WARN_THRESHOLD from config (default: 100000). If 0, skip the warning. If full-wiki token estimate exceeds the threshold, emit a ⚠️ warning with the three remediation suggestions shown in the template above.

6. Include in every standard status run — both normal and insights mode. The methodology note (4 chars/token heuristic) appears as a footnote below the table.

Step 4: What to Do Next

Replace the old single-line Recommendation with a ranked What to Do Next section. Gather these signals before rendering:

4a: Gather signals

0. Staged writes pending (only when WIKI_STAGED_WRITES=true) — Glob $OBSIDIAN_VAULT_PATH/_staging/**/*.md and **/*.patch.md. Count new pages and patches separately. Report the oldest file's age (mtime). This is always listed first if any staged files exist — it has the highest intent signal (the LLM already did the work; the human just needs to review).

1. _raw/ files — list every file at the top level of $OBSIDIAN_VAULT_PATH/_raw/ that isn't a .gitkeep, excluding the _archived/ subdirectory (already-promoted drafts kept for provenance, not pending work). Count and name them.

2. Stale core pages — scan all vault .md files. A page is "stale" when its updated frontmatter field is ≥90 days before today's date AND it has ≥5 incoming wikilinks (i.e., it's "core" — other pages depend on it). List them by name + last-updated date.

3. Orphan pages — pages with zero incoming wikilinks. To compute: glob all .md pages, extract every [[wikilink]], count references to each page, collect pages with incoming == 0. Show up to 5 names; report total count.

4. Synthesis opportunities — check hot.md for any recent /wiki-synthesize run summary. If the last synthesis run reported N opportunities, surface that count. If no synthesis has been run recently (not in hot.md or log.md within last 14 days), flag it as "synthesis scan overdue".

5. Source delta — from Step 2: count of new + modified sources ready to ingest.

6. Lint issues — check log.md for a recent /wiki-lint run (within last 30 days). If a recent run recorded broken links or missing frontmatter, surface the count. If no lint run appears in the log, flag "lint not run recently".

4b: Rank and render

Score each category and emit a ranked list, capped at 6 items. Always rank in this priority order (skip a category if its count is 0 or it has nothing to report):

Render as:

## What to Do Next

0. 📋  6 staged pages waiting for review (oldest: 3 days ago)
   → 4 new pages + 2 patches in _staging/
   run: /wiki-stage-commit

1. 📥  Ingest 3 files waiting in _raw/
   → architecture-notes.md, meeting-2026-05-10.md, paper-draft.pdf
   run: /wiki-ingest

2. 🔄  Refresh 2 stale core pages (not updated in 90+ days)
   → [[System Architecture]] (last updated 2026-02-10), [[API Design]] (2026-01-15)
   run: open these pages and re-run /wiki-update

3. 🔗  Link 7 orphan pages  →  run: /cross-linker
   Disconnected: [[Redis Caching]], [[JWT Tokens]], +5 more

4. 🧩  2 synthesis opportunities identified  →  run: /wiki-synthesize
   [[Redis Caching]] × [[Session Management]] (co-occur in 8 pages)

5. ✅  4 sources modified since last ingest  →  run: /wiki-ingest (append mode)

6. 🩺  Lint not run in 30+ days — run: /wiki-lint

Empty state: If all categories have nothing to report (no staged files, no _raw/ files, no orphans, no stale pages, no synthesis opportunities, no new sources, no lint issues), output instead:

## What to Do Next

✅  Wiki is healthy — nothing urgent.
    All sources up to date · no orphans · no stale core pages · no _raw/ files pending · no staged writes

Overflow: If more than 6 items would be shown, add a footer line: _(N more items available — run /wiki-status --full to see all)_. The --full flag is not yet implemented; this is forward-looking copy that sets expectations.

Insights Mode

Triggered when the user asks something like "wiki insights", "what's central in my wiki", "show me the hubs", "cross-domain bridges", "what pages are most important", or "wiki structure". This mode is additive — it doesn't replace the delta report, it analyzes the shape of the wiki itself.

Where the delta report tells the user what's pending, insights mode tells them what they've already built and where the interesting structure lives. Complements wiki-lint (which finds problems) by surfacing interesting structure.

What to compute

First, run the graph analyser. This replaces manual wikilink parsing — one command produces all the raw data you need:

obsidian-wiki graph-analyse "$OBSIDIAN_VAULT_PATH" --pretty --snapshot \
  --diff-against "$OBSIDIAN_VAULT_PATH/_insights.md"   # omit if no previous _insights.md

The analyser runs the same algorithm family graphify applies to code graphs, in pure Python: degree ranking, community detection (Leiden if obsidian-wiki[graph] is installed, else label propagation), community cohesion, Brandes betweenness centrality, cross-community surprise scoring, and snapshot diffing. Vault bookkeeping files (index, log, hot, _insights, _meta/, _readouts/) are excluded so they don't dominate every ranking.

Output fields used below:

• god_nodes — pages ranked by total degree (in + out), with in_degree/out_degree. Use for anchor pages and hub classification.
• bridges — pages ranked by betweenness centrality (share of shortest paths passing through them), with community and the list of other communities they connects. Use directly for the Bridge Pages section — no manual tag-pair search needed.
• communities — page clusters by link density, each with label, size, pages, and cohesion (intra-cluster edge density, 0–1). Use for cluster labelling and the cohesion section.
• surprising_connections — cross-community edges ranked by unexpectedness, one per community pair before any pair repeats, each with a note naming the two communities. Use directly in the Surprising Connections section.
• suggested_questions — {type, question, why} derived from bridges, sink hubs, isolates and low-cohesion clusters. Seed the Questions section with these.
• dead_ends — pages with zero outgoing links. Use for orphan-adjacent and cross-linker suggestions.
• isolated — pages with zero links in either direction. Use for stubs/orphan reporting.
• stats — total pages, edges, communities, graph density.
• diff (only with --diff-against) — added_pages, removed_pages, added_edges, removed_edges, newly_connected, lost_incoming, summary. Use for the Graph Delta section.
• snapshot (only with --snapshot) — write this JSON verbatim into the <!-- GRAPH_SNAPSHOT: ... --> comment at the end of _insights.md.

Query modes (useful when the user asks a follow-up): --path A B returns the shortest link chain between two pages; --around PAGE --depth N [--direction in|out|both] returns the N-hop neighbourhood (--direction in = the blast radius if PAGE were renamed or removed).

Fallback (if obsidian-wiki is not installed): glob all .md pages, extract every [[wikilink]], and build incoming, outgoing, and tags maps manually, then approximate the sections below by hand.

You'll reuse this data across all sections below.

1. Anchor pages (top hubs). Pages with the most incoming links — the load-bearing concepts.

   • Rank all pages by incoming count, take top 10
   • For each, note both incoming and outgoing counts: pages with high incoming and high outgoing are connector hubs (most valuable)
   • Pages with high incoming but zero outgoing are sink hubs — flag as cross-linker candidates

2. Bridge pages. Pages that connect otherwise-disconnected clusters — removing them would partition the graph. These are often more structurally important than raw hub count suggests.

   • Take the top 5 of bridges (already ranked by betweenness centrality)
   • Label each using the community labels: "P bridges [label of community] ↔ [labels of connects]". A bridge with an empty connects list is an intra-cluster chokepoint — note it as "internal hub of [label]"
   • Show the betweenness score so runs are comparable over time

3. Cluster cohesion. How tightly each community's pages are interlinked — read straight from communities[].cohesion (actual_links / (n × (n−1) / 2)).

   • Fragmented clusters (cohesion < 0.15, size ≥ 5): these pages share a topic but aren't woven together. Surface them as cross-linker targets.
   • Show the top 5 communities by cohesion (strongest clusters) and bottom 5 (most fragmented), each with its label and size
   • Optionally repeat the same formula per tag (for tags with ≥ 5 pages) if the user cares about tag-level cohesion

4. Surprising connections. Cross-community wikilinks that are non-obvious.

   • Start from surprising_connections (structural score = 1/√(cross-degree(A)·cross-degree(B)); one edge per community pair before any pair repeats, so a single hub can't fill the list)
   • Re-rank the candidates with content signals the analyser can't see:
     • +3 if the linking page or claim is marked ^[ambiguous] (uncertain connection, worth reviewing)
     • +2 if the linking page is marked ^[inferred] (synthesized, not directly stated)
     • +2 if the categories are in different knowledge layers (e.g., concepts ↔ entities more surprising than concepts ↔ concepts)
   • Show top 5 with a plain-language reason for each (use the note — "bridges [label A] → [label B]" — plus any content bonus)

5. Orphan-adjacent suggestions. Pages linked from a top-10 hub but with zero outgoing links of their own. Dead-ends in high-traffic areas — prime cross-linker candidates.

6. Rough clusters. Group anchor pages by dominant tag. (Simple tag intersection — just for orientation.)

7. Graph delta since last run. Read straight from the diff field (produced by --diff-against on the previous _insights.md, which the analyser reads from its <!-- GRAPH_SNAPSHOT: ... --> comment):

   • summary → "+N pages, +M wikilinks"; list added_pages / removed_pages
   • newly_connected → "newly connected: X, Y" (pages that had no incoming links last run)
   • lost_incoming → "link target may have been renamed: A, B"
   • If no previous snapshot exists (no diff field), skip this section

8. Tier assignment suggestions. After computing hubs and bridges, recommend tier: changes. Never write tier: to pages — only surface suggestions so the human can decide.

   • Promote to core: pages with ≥5 incoming links OR top-5 bridge position that currently have tier: supporting or no tier: field
   • Demote to peripheral: pages with ≤1 incoming link AND not updated in 90+ days that currently have tier: supporting or tier: core
   • Show up to 10 suggestions (promotions first, then demotions), formatted as:

     Tier Suggestions:
     ↑ core    [[concepts/attention-mechanism]] — 14 incoming links, currently tier=supporting
     ↑ core    [[entities/andrej-karpathy]]     — bridge (3 cluster pairs), currently unset
     ↓ peripheral [[concepts/old-concept]]       — 0 incoming, 120 days stale

   • If all high-link pages already have tier: core and all low-link pages have tier: peripheral, emit: "Tier assignments look healthy — no changes suggested."

9. Suggested questions. Questions this wiki structure is uniquely positioned to answer — or that reveal gaps:

   • From ^[ambiguous] claims (content scan — the analyser doesn't see these): "Resolve: What is the exact relationship between X and Y?"
   • From suggested_questions — already generated for you: bridge_node → "Explore: …", sink_hub → "Link: …", isolated_nodes → "Link: …", low_cohesion → "Audit: …". Rewrite each into the matching prefix form
   • Show up to 7, prioritizing AMBIGUOUS first, then bridge nodes, then sink hubs / isolates, then cohesion audits