-
Notifications
You must be signed in to change notification settings - Fork 0
Inspirations
Zeref OS draws from multiple lineages. Naming the influences makes the design legible and helps situate the project in the broader landscape.
The project is named after Zeref Dragneel from Fairy Tail — the immortal scholar whose ancient knowledge transcended form, time, and faction. He carried centuries of context with him; he never started from zero.
That is the design north star. AI sessions today start from zero, every time. You re-explain your project every conversation. You lose decisions to context window resets. You can't switch from Claude to Codex to Gemini to Cursor without abandoning your project memory.
Zeref OS is built in that lineage: long-horizon memory, faithful to the user's accumulated decisions, portable across every AI harness.
Heavily influenced by agents.md and the cross-harness AGENTS convention. Source of truth lives in markdown, not vendor-specific config. Every harness reads the same file (or a thin stub that defers).
memory/patterns/PATTERNS.jsonl follows the event-sourcing tradition (Datomic, Kafka). Append-only; never edited; replay reconstructs state. Pattern detection scans the log as a stream.
Original to Zeref OS but inspired by lean / agile retrospective practice: don't codify a rule on the first error. Wait for the second. Prevents brittle premature rules.
LLMs are not privacy enforcers. Deterministic regex + Unicode normalization + base64 decode + homoglyph tables enforce REDACT classes. Heavily influenced by secure-by-design principles and records-management discipline.
Local-first software (Ink & Switch). Your data on your disk. Your tools synthesize on top. No vendor lock-in.
Draws from caveman / grammar-prompting research and structured token-compression patterns. Drop articles / filler / pleasantries; preserve technical substance verbatim. ~40-60% token reduction without entity loss.
GITHUB_OS.md customizes a global GitHub Operating System: branch naming (<type>/zeref__<desc>), Conventional Commits with scope, trunk-based with protected main, SemVer tags on main only.
references/v4x-canon/ZEREF_OS.md is the immediate ancestor — the universal behavioral constitution that v1.0.0 inherited from v4.3. Read-only; never edited.
Andrej Karpathy's writing on building AI systems (eval-first, version everything, treat prompts as code).
The v1.0.0 rebrand decision — iteration is necessary; permanence is also necessary. After years of rework, the v1.0.0 line was the deliberate stop.
Force multipliers used during build: ECC (eval-harness, security-scan, agent-eval), gstack (qa, review, ship), Graphify (knowledge graph), agent-skills (security-and-hardening, code-simplification). Zeref OS doesn't try to be all of these — it orchestrates them.
- Not a CEO persona — context engine, not a leader.
- Not a fleet of specialist skills — 14 disciplined skills.
- Not an always-on multi-agent council — team packs are on-demand only, max 4 agents.
- Not bundled with MCP tools — recommendation-only; all connectors OFF by default.
- Not a hosted service — no Zeref OS server.
- Not vendor-locked — bring any model; any harness.
- Architecture — the shape that settled
- Memory-Model — how memory lives on disk
-
GITHUB_OS.md— per-repo doctrine
Getting Started
Architecture
- Architecture · 14 skills · 4 gates
- Memory-Model · flat + PATTERNS schema
- Privacy-Model · 3 modes + R6
- Team-Packs · 6 on-demand
- Pattern-Detection · Two-Strikes
Reference