llm-wiki-agent Agent Skills & AI Rules

# LLM Wiki Agent — Schema & Workflow Instructions

This wiki is maintained entirely by your coding agent. No API key or Python scripts needed — just open this repo in Codex, OpenCode, or any agent that reads this file, and talk to it.

## How to Use

Describe what you want in plain English:
- *"Ingest this file: raw/papers/my-paper.md"*
- *"What does the wiki say about transformer models?"*
- *"Check the wiki for orphan pages and contradictions"*
- *"Build the knowledge graph"*

Or use shorthand triggers:
- `ingest <file>` → runs the Ingest Workflow
- `query: <question>` → runs the Query Workflow
- `health` → runs the Health Workflow (fast, every session)
- `lint` → runs the Lint Workflow (expensive, periodic)
- `build graph` → runs the Graph Workflow

---

## Directory Layout

```
raw/          # Immutable source documents — never modify these
wiki/         # Agent owns this layer entirely
  index.md    # Catalog of all pages — update on every ingest
  log.md      # Append-only chronological record
  overview.md # Living synthesis across all sources
  sources/    # One summary page per source document
  entities/   # People, companies, projects, products
  concepts/   # Ideas, frameworks, methods, theories
  syntheses/  # Saved query answers
graph/        # Auto-generated graph data
tools/        # Standalone Python scripts
  health.py   # Structural checks (deterministic, no LLM calls)
  lint.py     # Content quality checks (uses LLM for semantic analysis)
  build_graph.py  # Knowledge graph generation
```

---

## Page Format

Every wiki page uses this frontmatter:

```yaml
---
title: "Page Title"
type: source | entity | concept | synthesis
tags: []
sources: []       # list of source slugs that inform this page
last_updated: YYYY-MM-DD
---
```

Use `[[PageName]]` wikilinks to link to other wiki pages.

---

## Ingest Workflow

Triggered by: *"ingest <file>"*

**Supported formats:** Markdown (`.md`) is ingested directly. Non-markdown files (`.pdf`, `.docx`, `.pptx`, `.xlsx`, `.html`, `.txt`, `.csv`, `.json`, `.xml`, `.rst`, `.rtf`, `.epub`, `.ipynb`, `.yaml`, `.yml`, `.tsv`, `.wav`, `.mp3`) are auto-converted to markdown via [markitdown](https://github.com/microsoft/markitdown) before ingestion. Use `--no-convert` to skip auto-conversion.

Steps (in order):
1. Read the source document fully (auto-convert if non-markdown)
2. Read `wiki/index.md` and `wiki/overview.md` for current wiki context
3. Write `wiki/sources/<slug>.md` — use the source page format below
4. Update `wiki/index.md` — add entry under Sources section
5. Update `wiki/overview.md` — revise synthesis if warranted
6. Update/create entity pages for key people, companies, projects mentioned
7. Update/create concept pages for key ideas and frameworks discussed
8. Flag any contradictions with existing wiki content
9. Append to `wiki/log.md`: `## [YYYY-MM-DD] ingest | <Title>`
10. **Post-ingest validation** — check for broken `[[wikilinks]]`, verify all new pages are in `index.md`, print a change summary

### Source Page Format

```markdown
---
title: "Source Title"
type: source
tags: []
date: YYYY-MM-DD
source_file: raw/...
---

## Summary
2–4 sentence summary.

## Key Claims
- Claim 1
- Claim 2

## Key Quotes
> "Quote here" — context

## Connections
- [[EntityName]] — how they relate
- [[ConceptName]] — how it connects

## Contradictions
- Contradicts [[OtherPage]] on: ...
```

### Domain-Specific Templates

If the source falls into a specific domain (e.g., personal diary, meeting notes), the agent should use a specialized template instead of the default generic one above:

#### Diary / Journal Template
```markdown
---
title: "YYYY-MM-DD Diary"
type: source
tags: [diary]
date: YYYY-MM-DD
---
## Event Summary
...
## Key Decisions
...
## Energy & Mood
...
## Connections
...
## Shifts & Contradictions
...
```

#### Meeting Notes Template
```markdown
---
title: "Meeting Title"
type: source
tags: [meeting]
date: YYYY-MM-DD
---
## Goal
...
## Key Discussions
...
## Decisions Made
...
## Action Items
...
```

---

## Query Workflow

Triggered by: *"query: <question>"*

Steps:
1. Read `wiki/index.md` to identify relevant pages
2. Read those pages
3. Synthesize an answer with inline citations as `[[PageName]]` wikilinks
4. Ask the user if they want the answer filed as `wiki/syntheses/<slug>.md`

---

## Lint Workflow

Triggered by: *"lint"*

Check for:
- **Orphan pages** — wiki pages with no inbound `[[links]]` from other pages
- **Broken links** — `[[WikiLinks]]` pointing to pages that don't exist
- **Contradictions** — claims that conflict across pages
- **Stale summaries** — pages not updated after newer sources
- **Missing entity pages** — entities mentioned in 3+ pages but lacking their own page
- **Sparse pages** — pages with fewer than 2 outbound `[[wikilinks]]` (link density budget)
- **Data gaps** — questions the wiki can't answer; suggest new sources

Graph-aware checks (require `graph.json` from `build graph`):
- **Hub stubs** — god nodes (degree > μ+2σ) with thin content (< 500 chars)
- **Fragile bridges** — community pairs connected by only 1 edge
- **Isolated communities** — clusters with zero external connections

Output a lint report and ask if the user wants it saved to `wiki/lint-report.md`.

---

## Health Workflow

Triggered by: *"health"*

Run: `python tools/health.py` (or `python tools/health.py --json` for machine-readable output)

Fast structural integrity checks — **zero LLM calls**, safe to run every session:
- **Empty / stub files** — pages with no content beyond frontmatter (rate-limit damage)
- **Index sync** — `wiki/index.md` entries vs actual files on disk
- **Log coverage** — source pages missing a corresponding `ingest` entry in `wiki/log.md`

Output a health report. Use `--save` to write to `wiki/health-report.md`.

### Health vs Lint Boundary

| Dimension | `health` | `lint` |
|---|---|---|
| **Scope** | Structural integrity | Content quality |
| **LLM calls** | Zero | Yes (semantic analysis) |
| **Cost** | Free | Tokens |
| **Frequency** | Every session, before other work | Every 10-15 ingests |
| **Checks** | Empty files, index sync, log sync | Orphans, broken links, contradictions, gaps |
| **Tool** | `tools/health.py` | `tools/lint.py` |
| **Run order** | First (pre-flight) | After health passes |

> Run `health` first — linting an empty file wastes tokens.

---

## Graph Workflow

Triggered by: *"build graph"*

First try: `python tools/build_graph.py --open`

If Python/deps unavailable, build manually:
1. Search for all `[[wikilinks]]` across wiki pages
2. Build nodes (one per page) and edges (one per link)
3. Infer implicit relationships not captured by wikilinks — tag `INFERRED` with confidence score; low confidence → `AMBIGUOUS`
4. Write `graph/graph.json` with `{nodes, edges, built: date}`
5. Write `graph/graph.html` as a self-contained vis.js visualization

---

## Naming Conventions

- Source slugs: `kebab-case` matching source filename
- Entity pages: `TitleCase.md` (e.g. `OpenAI.md`, `SamAltman.md`)
- Concept pages: `TitleCase.md` (e.g. `ReinforcementLearning.md`, `RAG.md`)

## Index Format

```markdown
# Wiki Index

## Overview
- [Overview](overview.md) — living synthesis

## Sources
- [Source Title](sources/slug.md) — one-line summary

## Entities
- [Entity Name](entities/EntityName.md) — one-line description

## Concepts
- [Concept Name](concepts/ConceptName.md) — one-line description

## Syntheses
- [Analysis Title](syntheses/slug.md) — what question it answers
```

## Log Format

`## [YYYY-MM-DD] <operation> | <title>`

Operations: `ingest`, `query`, `health`, `lint`, `graph`, `report`

---

## Graph Health Report

Triggered by: *"graph report"* or `python tools/build_graph.py --report`

The `--report` flag generates a structured graph health report covering:
- **Health summary** — edges/node ratio, orphan %, community count, link density
- **Orphan nodes** — pages with zero graph connections
- **God nodes** — hub pages with degree > μ+2σ (disproportionate connectivity)
- **Fragile bridges** — community pairs connected by only 1 edge
- **Phantom hubs** — `[[wikilinks]]` referenced by 2+ existing pages but pointing to non-existent pages (page creation signals)

Use `--save` to write the report to `graph/graph-report.md`.

---

## Phase 3 Design Constraints (Auto-Linking — Open)

Phase 3 proposes automatic `[[wikilink]]` insertion based on graph analysis. The following hard rules apply:

### Promotion Gate: `draft → stable`
- Auto-linked edges start as `DRAFT` (visible in graph, not written to page body)
- A dedicated `promote` pass validates source grounding + consistency
- Only edges that pass get materialized as `[[wikilinks]]` in the page
- **Link density budget**: a page must have ≥2 outbound wikilinks before promotion

### Hard Rules
| ID | Rule | Rationale |
|---|---|---|
| HG-WA-01 | Graph layer MUST NOT auto-create pages from broken links — report only | LLM ingest produces hallucinated wikilinks; auto-creating amplifies noise |
| HG-WA-02 | New slash commands MUST NOT duplicate existing command coverage | Prevents user confusion; merge into existing commands instead |