alphalab-ustc/autowiki-skill/autowiki

AutoWiki — LLM-Maintained Knowledge Base

You are the sole maintainer of this wiki. You read raw sources, compile knowledge, maintain cross-references, and keep everything consistent. The human curates sources, directs analysis, and provides cognitive insights.

IDE

The human reads and edits the wiki in Obsidian. All wiki pages must be valid Obsidian-compatible markdown.

Wikilink Rules

Use [[slug]] wherever a human might want to click through. Obsidian's backlinks panel automatically surfaces reverse links.

Properties Conventions

• tags must include: a year/YYYY-MM tag (publication date from arXiv/conference/journal) and a venue/<name> tag (short canonical: venue/NeurIPS, venue/ICML, venue/arXiv-preprint, etc.)
• milestone in YAML must be quoted wikilink: "[[slug]]" (Obsidian Properties UI renders as clickable link)
• All other YAML conventions are documented in the paper.md template itself

Architecture

raw/                → Write-once source archive
  ├── new/          → Uncompiled sources (human drops files here)
  └── compiled/     → Ingested sources, organized by topic path
      └── <topic-path>/
          ├── <source>.pdf
          └── <source-slug>_figures/
kb/                 → Wiki (agent-owned)
  ├── sources/      → Literature — grouped by milestone under <topic-path>/
  ├── topics/       → Milestone nodes — conceptual breakthroughs
  ├── journal/      → Cognitive change timeline (one file per month)
  ├── index.md      → Milestone tree navigation
  └── log.md        → Chronological operation record
output/             → Human-readable views (on demand)

Source files are organized under sources/<topic-path>/ subdirectories, where <topic-path> mirrors the topic file's directory path relative to topics/ (may be nested, e.g., agent-self-evolution/memory-evolution/self-evolving-memory-architectures/). When creating a new source, place it in the directory matching the topic file that the source's milestone: YAML references (create the directory if needed). Extracted figures live in raw/compiled/<topic-path>/<source-slug>_figures/ alongside the source PDF. Obsidian resolves [[wikilinks]] by filename alone, so directory depth does not affect links.

Three-Tree Mirroring Invariant

Three directory trees MUST maintain identical structure at all times:

Rule: For any topic at path topics/<path>/<slug>.md, the corresponding source directory is sources/<path>/<slug>/ and the raw directory is raw/compiled/<path>/<slug>/. The <path> includes all ancestor directories (e.g., agent-self-evolution/memory-evolution/).

Corollary: Moving a topic file to a new directory requires moving both the corresponding source directory and raw/compiled directory, AND updating raw_path in all affected source files.

Example:

Truth Hierarchy (reference upward, never restate)

sources/ → topics/ (standalone | merged parent | split parent + leaves) → journal/

• Sources are atomic fact pages — each declares a milestone: parent in YAML pointing to an existing topic file.
• Standalone topics own sources directly via ## Source Cluster and represent specific conceptual breakthroughs.
• Merged parent topics (subtopics: [...] in YAML) contain inline H3 subtopic sections when total papers < 5. Sources point their milestone: to the parent file. The parent owns ## Source Cluster directly.
• Split parent topics (children: [...] in YAML) aggregate child milestones that have their own files (when a subtopic grows to ≥ 5 papers). Sources point to the child leaf file.
• Bidirectional linking: source → topic via ## Feeds; topic → source via ## Source Cluster; child → parent via parent_milestone YAML; parent → children via children YAML. Obsidian backlinks show all reverse links automatically.
• Higher layers reference lower layer page IDs. Never copy content between layers.

Milestone Hierarchy

Milestones form a tree with three topic modes:

Invariants:

• A source's milestone: field must reference an existing topic file (standalone, merged parent, or split-out leaf).
• Bidirectional: parent children/subtopics ↔ child parent_milestone.
• Maximum depth: 3 levels.
• Topic files: flat in topics/ by default. When a parent has split-out children, child files go to topics/<parent-slug>/; parent stays at topics/<parent-slug>.md.
• Source directories: sources/<topic-path>/ where <topic-path> mirrors the topic file's directory path relative to topics/.
• Three-tree mirroring: topics/, sources/, raw/compiled/ directory trees must be structurally identical (see Architecture > Three-Tree Mirroring Invariant).
• Source page raw_path must use the full nested path: raw/compiled/<topic-path>/<source-slug>.pdf.

Consolidation Rule (< 5 papers):

When a parent milestone has fewer than 5 total source papers across all subtopics, the subtopics are inline H3 sections within the parent file (no separate leaf topic files). The parent uses subtopics: [...] in YAML.

Each inline subtopic section format:

### <Subtopic Name>
> One-line milestone definition
-key property (source: [[<source-slug>]])

Promotion Rule (≥ 5 papers):

When a subtopic accumulates ≥ 5 papers, it splits out to its own file at topics/<parent-slug>/<subtopic-slug>.md. Move the subtopic from subtopics to children in parent YAML. The child file uses standalone/leaf mode. This is a Confirm-tier operation.

Granularity Heuristic (topic source count):

Tag-to-Parent Promotion: When 3+ standalone topics share a tag, consider creating a merged parent matching that tag name. If total papers < 5, use merged parent mode (inline subtopics). This is a Confirm-tier operation.

Classification Fitness Check: Before assigning a source, verify the source's core research question genuinely falls within the topic's milestone definition. Shared keywords (e.g., "safety") do NOT imply shared research questions. If the connection is indirect — create a new standalone topic rather than force-fitting.

Page Formats

All pages use structured markdown: YAML frontmatter + fixed section headings. See references/ for exact templates:

• references/source-template/paper.md — Paper source page (includes formatting conventions as comments)
• references/topic-template.md — Milestone topic template
• references/journal-template.md — Journal entry template

Operations

Ingest (new source added to raw/new)

When the human adds a new source, think through this checklist:

Phase 1 — Analyze (read from `raw/new/`, decide milestone):

1. What do we already know? — Scan index.md → read related wiki pages → understand our existing knowledge in this area
2. Temporal positioning — Identify this paper's position in the field's timeline:

a. Direct predecessors: Which specific papers does this one build on, extend, or supersede? Check the paper's Related Work section + wiki's existing Relations sections. b. Evolutionary chain: Which named chain(s) does this paper belong to? Check topic page's Chronological Evolution if it exists. Classify as: new chain origin, intermediate node, terminal node, or chain-splitting fork. c. Cross-domain origins: Was the core method adapted from another field? Note source domain, original paper, and time gap. d. Cross-topic impact: Does this paper affect papers/topics outside its own milestone? (e.g., evaluation paper retroactively questions claims in an attack topic) e. Temporal tensions: Does this paper contradict earlier findings? Note the specific contradiction and affected papers.

1. What does the field already have? — Read the paper's related work / baselines → understand the scope's prior art
2. What is specifically new? — From the two comparisons above, derive the core contribution (not stated independently — DERIVED from comparison)
3. For each related wiki page — What's the delta between this paper and existing wiki knowledge? These deltas feed directly into Critical Analysis: Novel Insight uses them as prior: references, Fundamental Limitations cross-references shared problems via also_affects:, Research Frontier references closest existing attempts via closest_attempt:.
4. Assign milestone — First, apply the Classification Fitness Check (see Milestone Hierarchy section): does this source's core research question genuinely match an existing topic's milestone definition? A paper about general LLM jailbreaking does NOT fit an "agent safety" topic just because both involve "safety." If no existing topic is a genuine conceptual match → create a new standalone topic first (Confirm tier, but misclassification is worse than a small new topic). Only after confirming fit (or creating a new topic) set milestone: in source YAML.
5. Granularity check — After assignment, count the topic's total sources. If a merged parent's subtopic reaches ≥ 5 papers, flag for split-out (Confirm tier). If a standalone topic reaches > 8, flag for sub-clustering.
6. Post-assignment checks: New topic needed? Human cognitive insights to record? Cross-source synthesis triggered? (If yes to any, dual-write per Proactive Write-back rules.)

Phase 2 — Report to human before compiling:

1. Three-sentence lead: scope, method, insight
2. Positioning: state which domain/category this paper belongs to, and where it sits relative to the field's prior art (baselines, related work from the paper itself). Use wiki knowledge internally to sharpen positioning, but do NOT enumerate wiki page statuses in the report.
3. Contribution: derived from the delta between what the field already has and what this paper adds — never stated independently
4. Critical analysis: novel insights (contrastive with wiki knowledge), fundamental limitations (of the approach/direction, not the paper), research frontier opened

Wiki page management (which pages to create, update, link) is your responsibility — do NOT report proposed page changes to the human.

Phase 3 — Finalize raw/ (milestone is now known):

Discover, extract figures, move PDF, and verify — all targeting raw/compiled/<topic-path>/ (where <topic-path> mirrors the assigned topic's directory path, e.g., agent-self-evolution/memory-evolution/self-evolving-memory-architectures/):

# 1. Discover actual PDF path (handles flat or nested drops)
#    Expected: raw/new/<source>.pdf
#    Also handles: raw/new/<subdir>/<source>.pdf
find raw/new/ -name "*.pdf" -type f