syntheca 0.2.0a3

CLI for scaffolding LLM-ready Syntheca wiki workspaces

Syntheca

English | 한국어

Schema-first framework for agent-maintained wikis

This guide adapts ideas from Andrej Karpathy's "LLM Wiki" pattern and the "LLM Wiki v2" / agentmemory production lessons. See REFERENCES.md and WIKI-ATTRIBUTION.md for full provenance.

Syntheca is a schema specification and starter framework for using LLM agents as disciplined wiki maintainers. It documents how agents can incrementally build and maintain a structured wiki that compounds over time.

---

What is Syntheca?

Syntheca is a framework with early alpha CLI tooling. It provides:

• Schema: Canonical templates, page types, and frontmatter specifications
• Workflows: Step-by-step procedures for ingest, query, lint, crystallization, and migration
• Runtime adapter notes: Instructions for applying the schema with Claude Code, Codex, OpenCode, and MCP-compatible agents
• Starter examples: Sample wiki demonstrating the pattern
• Migration workflow: Agent-guided preservation rules plus alpha dry-run reporting for existing knowledge bases
• Alpha CLI: syntheca init, syntheca doctor, syntheca inspect, syntheca lint, syntheca lint --fix, and syntheca migrate --mode dry-run for scaffolded workspace experiments

What Syntheca is NOT:

• ❌ A GUI application
• ❌ A vector database or search engine
• ❌ A replacement for Obsidian, Notion, or other note-taking apps
• ❌ A RAG system

What Syntheca IS:

• ✅ A schema and workflow specification
• ✅ A set of templates and conventions
• ✅ Documentation for agents to follow
• ✅ An agent-guided migration specification and deterministic dry-run report for unstructured knowledge bases
• ✅ An alpha CLI slice for workspace scaffolding, structural validation, deterministic inspection, deterministic lint, narrow mechanical lint fixes, and migration dry-run

Responsibility Boundary

Syntheca deliberately splits deterministic tooling from semantic wiki work.

Layer	Owns	Does Not Own
CLI	Workspace scaffold, structural validation, deterministic inspection, deterministic lint, narrow mechanical lint fixes, migration dry-run reports.	Source interpretation, page synthesis, contradiction resolution, broken-link intent, stale-claim judgment.
Agent	Ingest, query, crystallization, synthesis, link/index/log maintenance following schema/.	Destructive migration, unreviewed raw source changes, final authority on ambiguous claims.
Human	Review, approval, curation, destructive operation decisions, release decisions.	Repeating mechanical checks that the CLI can perform.

---

Core Concept: Compilation Over Retrieval

Retrieval-first workflow: Index sources → Retrieve relevant passages at query time → Generate an answer

Persistent wiki workflow: Ingest source → Extract and compile into structured pages → Query maintained wiki → Crystallize valuable answers back into wiki → Knowledge can compound

The wiki is a persistent, compounding artifact. The documented workflows guide agents through cross-reference maintenance, contradiction review, and synthesis updates while humans retain responsibility for curation and quality review.

---

Quick Start

Option A: Scaffold a workspace with the alpha CLI

From a checkout of this repository:

python -m pip install -e .
syntheca init my-wiki
syntheca doctor my-wiki
syntheca inspect my-wiki
syntheca lint my-wiki
syntheca lint my-wiki --fix
syntheca migrate --source ./my-vault --mode dry-run \
  --output-report ./migration-report.md \
  --output-manifest ./migration-manifest.json

Run the test suite with coverage:

python -m pip install -e '.[dev]'
python -m coverage run -m unittest
python -m coverage report

This creates an LLM-ready workspace with README.md, AGENTS.md, optional runtime entrypoints and provider-native skills, raw/, wiki/, and .syntheca/syntheca.yaml.

Project status: Alpha CLI available (init, doctor, inspect, lint, lint --fix, migrate --mode dry-run). lint --fix is limited to mechanical index/frontmatter fixes that do not require LLM judgment. Output/apply migration modes and MCP server work are planned (not implemented in the current alpha CLI).

Option B: Explore the schema-first starter

1. Clone the repository

git clone https://github.com/swj9707/syntheca.git
cd syntheca

Stay at the repository root so your agent can read schema/ and work with the example knowledge base under starter/.

Directory structure:

starter/
├── raw/                    # Your source material (immutable)
│   └── sample-source.md
└── wiki/                   # Agent-maintained wiki
    ├── sources/            # Processed source pages
    ├── entities/           # Named objects (people, projects, products)
    ├── concepts/           # Reusable ideas and patterns
    ├── syntheses/          # Cross-source analysis
    ├── decisions/          # Important choices (ADR pattern)
    ├── unclassified/       # Ambiguous or legacy content
    ├── index.md            # Content catalog
    └── log.md              # Chronological operation record

2. Load the schema into your agent

Claude Code:

Read schema/AGENTS.md and schema/adapters/claude/CLAUDE.md

Codex:

Read schema/AGENTS.md and schema/adapters/codex/AGENTS.md

OpenCode:

Read schema/AGENTS.md and schema/adapters/opencode/AGENTS.md

If you maintain a local OpenCode skill for Syntheca, you can load it before reading the schema. The v0.1 repository does not ship a skill package.

3. Ingest a source

Add your source to starter/raw/, then tell the agent:

Use starter/ as the wiki root.
Ingest starter/raw/my-article.md into starter/wiki/.

Ask the agent to follow the documented ingest workflow:

1. Read the source
2. Create a source page with summary and key points
3. Extract entities (people, projects, products) → create entity pages
4. Extract concepts (ideas, patterns) → create concept pages
5. Add evidence-grounded cross-links and review whether reciprocal links are appropriate
6. Update index.md and log.md

4. Query the wiki

What does the wiki say about [topic]?

The documented query workflow instructs the agent to search the wiki, synthesize an answer with citations, and evaluate whether the answer should be crystallized back into the wiki as a synthesis page.

5. Explore the starter

Open starter/wiki/index.md to see:

• 2 source pages demonstrating ingest and cross-source synthesis
• 1 entity page (Syntheca framework)
• 3 concept pages (persistent wiki, crystallization, and three-layer architecture)
• 1 synthesis page (retrieval, persistent wiki, and hybrid comparison)

All pages are cross-referenced and follow the schema. Run the read-only v0.1 checklist in docs/guides/starter-lint.md after changing the starter wiki. See docs/concepts/framework.md for the framework overview and docs/project/release-checklist.md for the public release gate.

---

Architecture

Three Layers

1. Raw Sources (raw/): Immutable source material. Agents read from here but never modify.

2. Wiki (wiki/): Agent-maintained structured pages. Agents create, update, and cross-link pages following the schema.

3. Schema (schema/): Instructions defining how the wiki is maintained. Templates, workflows, and runtime adapters.

Page Types

Type	Purpose	Example
Source	Summarize raw sources, identify extraction candidates	sources/article-2026-05-30.md
Entity	Named objects: people, projects, products, libraries	entities/syntheca.md
Concept	Reusable ideas, patterns, methods	concepts/persistent-wiki.md
Synthesis	Cross-source analysis, comparisons, crystallized insights	syntheses/wiki-vs-rag.md
Decision	Important choices following ADR pattern	decisions/use-markdown.md
Unclassified	Ambiguous or legacy content, safe preservation	unclassified/legacy-note.md

Core Workflows

Ingest (schema/workflows/ingest.md): Process raw sources → create/update pages → maintain cross-references

Query (schema/workflows/query.md): Search wiki → synthesize answer → evaluate crystallization

Lint (schema/workflows/lint.md): Health-check for orphans, broken links, contradictions, stale claims

Crystallization (schema/workflows/crystallization.md): File valuable query answers back into the wiki as synthesis/concept/decision pages

Migration (schema/workflows/migration.md): Guide an agent through safe imports with dry-run, classification heuristics, and unknown field preservation

---

Migration from Existing Wikis

The v0.1 agent-guided migration workflow is designed for:

• Obsidian vaults
• Notion exports
• Legacy markdown wikis
• Flat collections of notes

Process:

1. Dry-run first: Generate report without modifying files
2. Review classification: Check page type assignments (source/entity/concept/synthesis/decision/unclassified)
3. Apply with explicit approval: Ask an agent to output to a new directory (safer) or modify files in place (destructive)
4. Preserve unknowns: Custom frontmatter fields are never deleted

Alpha migration dry-run CLI:

syntheca migrate --source ./my-vault --mode dry-run \
  --output-report ./migration-report.md \
  --output-manifest ./migration-manifest.json

Output/apply/in-place migration modes are planned (not implemented in the current alpha CLI).

See docs/guides/migration.md for the full guide with examples, docs/guides/cli-reference.md for command details, and docs/guides/configuration.md for generated workspace settings.

---

Runtime Compatibility

Syntheca is designed to be runtime-neutral. The canonical schema (schema/AGENTS.md) and adapter notes describe how to apply the framework across different agent systems; v0.1 does not claim tested, automated support for every runtime.

Runtime	Adapter Focus	Relevant Runtime Capabilities
Claude Code	Conversational schema application	Runtime-specific capabilities vary by setup
Codex	Repository-oriented schema application	CLI, IDE, cloud/web, app; optional GitHub and SDK integration
OpenCode	Multi-agent schema application	Delegation, background execution, and skills depend on runtime configuration
MCP Generic	Optional external integration path	Capabilities depend on the selected MCP servers

See docs/concepts/capability-matrix.md for detailed comparison.

Adapter Documents

• Claude Code: schema/adapters/claude/CLAUDE.md
• Codex: schema/adapters/codex/AGENTS.md
• OpenCode: schema/adapters/opencode/AGENTS.md

Each adapter explains how to apply the canonical schema in that runtime without redefining policy.

---

Use Cases

Personal Research

Track papers, articles, and notes over weeks or months. Maintained pages can build a more comprehensive picture as sources accumulate and are reviewed.

Reading Companion

Use the documented workflow to file chapter summaries, character pages, and theme pages as you read. The wiki can grow into a cross-referenced reading companion with agent assistance and human review.

Team Knowledge Base

Use meeting transcripts, project documents, and customer notes as curated inputs. The documented workflow guides an agent to maintain project entities, recurring concepts, and reviewed decisions.

Competitive Analysis

Track competitors, products, and market trends. The documented workflow guides agents to consolidate sources and surface potential contradictions for review.

Course Notes, Trip Planning, Hobby Deep-Dives

Anything where you accumulate knowledge over time and want it organized, not scattered.

---

Schema Highlights

Frontmatter Standards

Every page has structured metadata:

---
title: "Page Title"
type: concept
status: active
created: "2026-05-30"
updated: "2026-05-30"
sources: ["sources/article.md"]
related: ["concepts/other-concept.md"]
tags: [tag1, tag2]
---

See schema/frontmatter.md for canonical field definitions per page type.

Template Structure

All templates follow consistent pattern:

1. Frontmatter: Structured metadata
2. Required sections: Definition, explanation, relationships, sources
3. Optional sections: Examples, counterpoints, open questions, contradictions
4. Authoring rules: Guidelines for agent behavior
5. Minimum completion criteria: Validation checklist

Quality Standards

Adopted from LLM Wiki v2:

• Confidence scoring groundwork: Synthesis pages may record a manually assigned page-level evidence score
• Contradiction handling: Explicit sections for conflicting claims
• Uncertainty preservation: Limitations and open questions are required
• Evidence grounding: Claims must cite sources
• Crystallization threshold: "Would I want to read this 6 months from now?"

---

Roadmap

v0.1 (Current)

• ✅ Core schema and templates
• ✅ Five workflows (ingest, query, lint, crystallization, migration)
• ✅ Runtime adapter notes (Claude Code, Codex, OpenCode)
• ✅ Starter examples
• ✅ Agent-guided migration specification with dry-run

v0.2 alpha (In progress)

• CLI package skeleton
• syntheca init workspace scaffold
• syntheca doctor structural validation
• syntheca inspect deterministic workspace summary
• syntheca lint deterministic wiki checks
• syntheca lint --fix narrow mechanical fixes
• syntheca migrate --mode dry-run report and manifest generation
• Machine-readable baseline checks for required type/status, links, raw source paths, and index coverage
• Extended frontmatter validation beyond the deterministic baseline
• Migration output/apply modes
• Initial MCP interface design

v0.3 (Future)

• MCP server implementation
• Vector and hybrid search integrations
• Link graph visualization
• Extended stale-page and supersession checks
• Claim-level confidence and contradiction review experiments
• Multi-wiki federation

See docs/project/roadmap.md for detailed feature plans.

---

Project Status

Current: Pre-release (v0.1 schema and documentation framework plus v0.2 alpha CLI init/doctor/inspect/lint/lint-fix/migrate dry-run commands)

Not ready for:

• Production team wikis without human review
• Large wikis without separately evaluated search and review infrastructure
• Automated workflows without supervision

Suitable for exploratory evaluation:

• Small personal research wiki experiments
• Trying the documented workflows with Claude Code, Codex, or OpenCode
• Creating an experimental workspace with syntheca init
• Agent-guided migration experiments (dry-run workflow)
• Schema evaluation and feedback

---

Contributing

See CONTRIBUTING.md for guidelines.

High-priority contributions:

• Runtime adapter notes and evaluation reports for other agents
• Production experience reports (what works, what breaks)
• Planned migration tooling design and implementation
• Template refinements based on usage

---

Attribution

Syntheca is derived from:

• Andrej Karpathy's LLM Wiki: Original persistent wiki pattern, three-layer architecture, core operations
• agentmemory (rohitg00 et al.): Production lessons, confidence scoring, quality standards

See REFERENCES.md and WIKI-ATTRIBUTION.md for detailed attribution.

Syntheca's contribution: formalization, cross-runtime compatibility, an agent-guided migration specification, and concrete templates/workflows.

---

License

MIT License. See LICENSE for details.

---

Links

• Documentation hub: docs/README.md
• 한국어 문서: README.ko.md and docs/ko/
• Schema: schema/
• Starter examples: starter/
• CLI reference: docs/guides/cli-reference.md
• Configuration reference: docs/guides/configuration.md
• Migration guide: docs/guides/migration.md
• Capability matrix: docs/concepts/capability-matrix.md
• References: REFERENCES.md

---

Questions?

• How is this different from RAG? See starter/wiki/syntheses/persistent-wiki-vs-rag.md
• How do I migrate my Obsidian vault? See docs/guides/migration.md and docs/ko/migration.md
• Which runtime should I use? See docs/concepts/capability-matrix.md
• What are the page types? See schema/page-types.md
• What are the templates? See schema/templates/*.md

---

Start small, let it compound.

The schema guides agents through bookkeeping tasks while you curate sources and review quality. Over time, the wiki can become a richer, cross-referenced knowledge base.