VaultForge v3.1.0

From generation tool to knowledge system

v3.0.0 is a system-level redesign of VaultForge — transforming it from a text generator into an engineering support system for cognitive processes. Three pillars: runtime reliability, incremental knowledge accumulation, and output precision control.

AI agent skill for Obsidian. Give it a PDF, get a complete vault — atomic notes, logical wikilinks, MOCs, and deep research. 58 automated tests. 7-phase pipeline. ⭐ Star on GitHub · Dev.to

Pillar 1

Runtime Reliability

State machine + atomic writes ensure crash recovery. Interrupted mid-work? Restart and it resumes where it stopped.

Pillar 2

Knowledge Accumulation

Phase 0 detects existing notes. Add new materials without touching what you've edited. Three-tier ACL protection.

Pillar 3

Output Precision

5-type logical wikilinks + 6-dimension quality review. Three-stage funnel balances precision and cost.

7‑Phase Pipeline 58 Automated Tests 7 Cross‑File Contracts
8 Degradation Paths 16/16 Audit Closure 19/19 Cross‑File Verification

Runtime Reliability

v1.2.0

Stateless Runtime

Agent crash or network interruption → all progress lost. Files may be corrupted. Restart from page one.

v3.0.0

Five-State State Machine + Atomic Write

Every note carries a status field. The filling state splits into two crash sub-scenarios, recovered separately. .md.tmp → verify → rename guarantees file integrity.

Learning Theory Connection
Spaced Learning produces significantly better retention than massed learning — but requires seamless recovery from interruption. A tool without checkpoint recovery penalizes the most effective study strategy. The state machine decouples learning from continuous sessions into interruptible ones — the spacing effect translated into an engineering constraint.

Knowledge Accumulation

v1.2.0

Linear Processing

Read all files → generate roadmap → create all notes → build all links. New files trigger full re-run. Existing notes overwritten or orphaned.

v3.0.0

Phase 0 · Three-Tier Protection

Phase 0 scans existing notes before any content generation. Pristine = refreshable. User-modified = protected. Locked = frozen. Incremental mode processes only new files; existing notes untouched.

StatusDetectionBehavior
PristineUnmodified since generationAuto-refreshable
User-modifiedmtime differs from dateProtected from auto-writes
LockedManually set by userCompletely ignored

Link Strategy in Incremental Mode

New ↔ New: links written directly. New → Existing: suggestions JSON only, existing files untouched. The link network in existing notes is the user's mental model — new links should be an optional data source, not an automatic change.

Learning Theory Connection
Bruner's Spiral Curriculum demonstrates that the most effective knowledge construction revisits topics at increasing complexity over time — provided existing structures remain intact on each return. Phase 0 + incremental processing + three-tier protection transforms VaultForge from a one-shot tool into a spiral learning system.

Output Precision

v1.2.0

Keyword Matching

Two notes sharing a keyword get linked. Stops at recognition memory — "I've seen this word" rather than "how do these concepts relate."

v3.0.0

Five-Type Logical Links + Six-Dimension Review

Derivation · Analogy · Contradiction · Application · Context — five types covering Bloom's taxonomy from comprehension to evaluation. note‑reviewer independently audits every note across six dimensions.

RelationMeaningBloom's Level
DerivationA is the logical prerequisite of BUnderstand → Apply
AnalogyA and B share deep structure but surface differentlyAnalyze
ContradictionA and B reach conflicting conclusionsEvaluate
ApplicationA's framework is used in B's contextApply
ContextA and B are phases in the same domain's evolutionSynthesize
Stage 1 · Structural Affinity Zero cost. Only note pairs within the same H2 proceed. Filters ~80%
Stage 2 · TF‑IDF Semantic Scoring Pure Python, zero dependencies. Computes cosine similarity for high-confidence candidates
Stage 3 · LLM Classification Controlled token cost. Only ~20% of candidates undergo five-type classification
Learning Theory Connection
Kluger & DeNisi (1996) meta-analysis: effective feedback requires three conditions — task-focused, specific direction, and opportunity to correct. Six-dimension review + auto-rewrite loop forms the engineering closure: precise to file path and line number, graded HIGH / MEDIUM, corrected within the same session.

Evaluation

VaultForge v3.1.0 is the most educationally rigorous and architecturally disciplined project in the current Agent Skill ecosystem. It has evolved from a one-shot generation tool into a continuous knowledge management system — with Phase 0 metacognitive awareness, vf_ frontmatter as a persistent state layer, ACL-style three-tier note protection, and engineering quality backed by 58 tests and 16/16 audit closure.

Pedagogical Designv1.2.0v3.1.0Driver
Cognitive Load Management3.55.0Atomic note proportion check + Context Pre-extraction
Advance Organizers4.05.0MOC wikilink completeness auto-audit
Semantic Network3.04.5Five-type classification replaces keyword matching
Retrieval Guidance3.04.0Structured guidance replaces simple lists
Critical Thinking3.05.0Three-layer Controversy Analysis framework
Spiral Curriculum5.0Phase 0 + Incremental + three-tier ACL
Learner Autonomy4.5vf_status three-tier ACL model
Engineering & Agent Designv1.2.0v3.1.0Driver
Fault Tolerance1.05.0State machine + Atomic Write + 8 degradation paths
Output Accuracy2.54.56-D Review + funnel wikilinks + file-based stats
Token Efficiency2.04.5Context Pre-extraction + --note-filter
Cross‑Platform2.04.5Tri-mode CLI + Phase 0 no sub-agent dependency
Maintainability2.05.058 tests + 7 cross-file contracts + 16/16 audit
Internal Consistency2.05.019 cross-checks + automated guardrails
Incremental Architecture5.0Three-tier ACL + add-only + suggestion isolation + refresh loop

Get Started

git clone https://github.com/Easonnotsing/VaultForge.git ~/.agents/skills/VaultForge
pip install pypdf

Trigger in Claude Code, Codex, or Cursor. Select output language first. Phase 0 scans automatically — if a knowledge base exists, you'll be asked: incremental update or fresh generation.