Lodestone Protocol — Markdown-Embedded DAG for conversational knowledge convergence / Alias: MD-DAG
Lodestone(磁石协议)为人机对话的知识脱水与收敛而生长:节点如同磁石吸附 有效信息,边作为磁力线登记依赖与分歧,draft → converged → aligned 标记观点成型阶段。
This repository is the authoritative home of the protocol: the normative specification text, the error-code registry, and the Golden Fixture corpus. It is language-agnostic — the protocol is not a library.
spec/v1.3.md # normative spec text (v1.3.0, frozen line)
registry/errors.md # the 14 error/warning codes (§11 extraction)
fixtures/ # 25 Golden Fixtures (.md input + .json expected) + MANIFEST
adr/ # protocol-level architecture decision records
- Protocol versions use plain SemVer: the current release is v1.3.0. "Final" is a status carried by Release notes, not a version suffix.
- An implementation's major.minor tracks the protocol version; its patch is independent.
- Implementations MUST declare in their README and package metadata:
"Implements Lodestone Protocol v1.3.0", and their conformance tests MUST
stay green against this repository's
fixtures/.
| Level | Content | Typical use |
|---|---|---|
| L1 skeleton | node table + normalized edges + diagnostics | rebuild global awareness |
| L2 targeted body | one node read by body_start/body_end |
answer a specific question |
| L3 full text | all bytes | recall-complete tasks |
AI agents are the protocol's native consumers (L1 + L2 minimal subset); humans audit alignment state, warnings, and dispute domains.
| Repository | Language | Status |
|---|---|---|
| lodestone-md | Rust | Implements Lodestone Protocol v1.3.0 · 23 tests green · submodule-imports this corpus |
See CONTRIBUTING.md — especially the ADR attribution rule (spec wording here, implementation choices there, cross-cutting issues start here).
Apache-2.0