Wiki Engine

Store saved articles, videos, repositories, and notes, then publish short reading notes that still point back to the source. Wiki Engine is a local command-line tool for people who save more material than they have time to review. A human or model writes the interpretation; Wiki Engine checks and records its link to the source.

🌐 Open the website · 📖 Read the master report · ⚙️ See the architecture · 🛡️ Read the safety model

Status: public alpha. Alpha.1 stores exact source bytes, publishes source-linked reading notes, can combine several sources into provisional wiki notes, recovers interrupted writes, and lists work that still needs review. Its 19 repository tests cover the local mechanisms; they do not prove that model-written summaries are accurate or that the product works well for external users.

What you get

📥 Capture  →  ⚡ Signal  →  🔗 Connect

Moment	What Wiki Engine does	Why it matters
📥 Capture	Admits one local file into hidden content-addressed storage.	The source remains exact and user-owned.
⚡ Signal	Publishes one concise reading note bound to that Source ID and digest.	You read what matters before building a graph.
🔗 Connect	Allows provisional Wiki synthesis across multiple distinct sources.	Recurring ideas earn durable life; they are not generated by default.
🧭 Review	Lists Capture, Signal, Wiki, and integrity debt without mutating the vault.	You get a read-only orientation queue; acceptance and promotion remain outside Alpha.1.

A Signal is a short reading note tied to one saved source. A Wiki note is a provisional summary built from more than one source. Wiki Engine keeps both linked to the exact material you admitted.

Save a source. Get a useful reading note. Keep the source attached.

Wiki Engine is not a hosted second brain, universal scraper, embedding database, automatic truth engine, or background collector.

Why the reading note comes first

Most research systems stop at one of two extremes:

• A backlog: everything is preserved, but you still have to read everything.
• A synthetic graph: the AI generates so many pages and links that verification becomes another backlog.

The reading note sits between capture and long-term knowledge. One source normally produces no more than one Signal. A useful Signal answers:

• ⚡ What is the bottom line?
• 💡 What is genuinely useful or novel?
• 🧾 What evidence or timestamp supports it?
• ✓ What should I do next?
• 🔍 Which linked resources are still uninspected?
• ? What remains uncertain?

Quick start

Requires Python 3.10+ on macOS or Linux.

git clone https://github.com/youkistudios/wiki-engine
cd wiki-engine

python3 -m venv .venv
. .venv/bin/activate
pip install -e .

wiki-engine --vault ~/Desktop/My-Wiki init
mkdir -p ~/Desktop/My-Wiki/Inbox/Web
printf '# Example\n\nA saved article to review.\n' \
  > ~/Desktop/My-Wiki/Inbox/Web/example.md

wiki-engine --vault ~/Desktop/My-Wiki scan
wiki-engine --vault ~/Desktop/My-Wiki ingest \
  ~/Desktop/My-Wiki/Inbox/Web/example.md \
  --url https://example.com \
  --kind article

The ingest command prints the new source_id and sha256. Put those exact values into a Signal draft that you or a model writes:

---
source_id: PASTE_SOURCE_ID
source_sha256: PASTE_SHA256
---
# Example signal

## Bottom line
The source's main point, written after reviewing it.

Save that draft as signal.md, replace the placeholders, then publish and check it:

wiki-engine --vault ~/Desktop/My-Wiki signal-publish PASTE_SOURCE_ID signal.md
wiki-engine --vault ~/Desktop/My-Wiki review
wiki-engine --vault ~/Desktop/My-Wiki lint

Import the templates in clipper/ into the official Obsidian Web Clipper for general web, YouTube, and X captures.

What makes it different

🧾 Provenance before polish

The kernel owns paths, identities, hashes, bounds, collision behavior, and receipts. A human or model writes fallible interpretation. Valid structure makes review possible; it does not prove the prose is true.

⚡ Signal before Wiki

The daily reading surface comes first. Cross-source synthesis is optional, provisional, and downstream from exact source identity.

🎬 Transcript-first video research

Alpha.1 produces one of three bounded route plans:

transcript-only      default for speech-led material
targeted-frames      exact cues for demos and “look at the screen” moments
full-visual-review   only for visual-dominant material

A URL found in a description remains uninspected until separately opened and admitted. Rights-safe acquisition and canonical frame Evidence remain under development.

🔒 Local ownership

• plain Markdown and JSON;
• one user-selected local vault;
• no hosted account;
• no daemon or runtime telemetry;
• no embeddings or vector database;
• no core network calls;
• manual invocation by default.

How the vault is organized

My-Wiki/
├── Inbox/Web/       📥 untrusted captures
├── Sources/Web/     📦 safe visible source receipts
├── Signal/          ⚡ one-source reading notes
├── Wiki/            🔗 provisional cross-source synthesis
└── .wiki-engine/    🧾 payloads, records, receipts, locks, recovery

Exact raw bytes live as hidden content-addressed payloads. Visible Markdown is a readable projection—not the canonical source authority.

Core commands

Command	Job	Source truth changed?
init	Create missing vault and engine paths	No existing notes
scan	Bounded, no-follow Inbox inventory	No
ingest	Admit one regular file	Creates Source
source / source-content	Inspect one selected record or payload	No
video-route / video-bind	Plan video treatment and bind transcript metadata	Adds evidence metadata
signal-publish	Publish one Source-bound reading note	No Source mutation
wiki-publish	Publish provisional multi-source synthesis	No Source mutation
review	List Capture, Signal, and Wiki debt	No
lint	Verify paths, records, lineage, and payload bytes	No

Trust boundary

📦 Source   proves what bytes were admitted
⚡ Signal   proposes what one source may mean
🔗 Wiki     proposes what distinct sources suggest together
🧭 Review   reports what needs attention next; it does not record acceptance

The alpha includes:

• SHA-256 source identities;
• hidden exact-byte payloads;
• regular-file and no-follow reads;
• 8 MiB/source and 5,000-file Inbox bounds;
• strict Source ID and digest publication fields;
• active Markdown rejection;
• a cooperative POSIX writer lock;
• journaled multi-file recovery;
• explicit replacement rather than silent overwrite;
• read-only Review.

See docs/SAFETY-MODEL.md for the complete boundary.

Verify the repository

No network access is required by the tests.

PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src \
  python3 -m unittest discover -s tests -v

python3 scripts/generate_manifest.py
python3 scripts/validate_repo.py

Current Alpha.1 evidence: 19 offline tests plus a public six-job macOS/Linux × Python CI witness. The suite covers initialization, deduplication, strict lineage, video manifests, path/control-plane tampering, source tamper detection, transaction recovery, bounded routing, Inbox acknowledgement, and read-only Review.

Research lineage

Wiki Engine builds on Andrej Karpathy’s influential LLM Wiki formulation while placing it inside the broader Memex, personal-knowledge, provenance, and human-review lineage.

Read From Compounding Wiki to Traceable Signal for the complete comparison, current market map, significance assessment, revision-bound prior-art ledger, and the next boundary into a second brain.

🧠 How it plugs into a second brain

Wiki Engine should feed a second brain—not become the whole brain.

Capture → Signal → Human review → Explicit promotion → Second-brain knowledge

Wiki Engine owns research custody and provisional interpretation. A second brain owns personal meaning, decisions, projects, commitments, and durable promotion. This boundary is a thesis, not an implemented integration.

Repository map

src/wiki_engine/       deterministic local kernel and CLI
skills/                Claude/Codex semantic workflows
clipper/               Obsidian Web Clipper templates
tests/                 Alpha.1 mechanism tests
docs/                  branded Pages site, research, architecture, safety
docs/assets/           banner, favicon, and social-preview assets
provenance/             observed prior-art revisions and authorship statement

Honest limits

• Model-authored Signal and Wiki prose can be incomplete or wrong.
• Structural citation validity does not guarantee semantic support.
• Runtime behavior across every Obsidian, browser, and agent version is not proven.
• Generalized YouTube downloading is not a stable first-party promise.
• Windows, multi-writer operation, network filesystems, and active cloud-sync control planes are unsupported.
• A hostile same-user process that coherently replaces all payload and control metadata is outside the alpha threat model.
• External users, semantic-quality evaluation, backup/restore, stable migration, accessibility witnesses, and long-duration soak evidence remain future gates.

Brand and prior art

The brand system is documented in docs/BRAND.md. Public projects were studied as attributed behavioral references; no donor code, templates, branding, or assets are represented as copied. See docs/PRIOR-ART.md and provenance/BUILD-PROVENANCE.json.

License

MIT © 2026 Youki Studios.