Skip to content

public_release_notes_v0.1.md

Latest commit

 

History

History
94 lines (75 loc) · 4.34 KB

public_release_notes_v0.1.md

File metadata and controls

94 lines (75 loc) · 4.34 KB
Raw

LynkMesh Public Release Notes — v0.1

Release status: early validation / research preview. LynkMesh is not production-ready. Expect breaking changes.

What LynkMesh is

LynkMesh is a deterministic context protocol for AI-assisted code understanding. Instead of feeding an AI agent raw files or relying on probabilistic retrieval, LynkMesh turns a codebase into a structured, evidence-linked, reproducible context layer that an agent can reason over safely. The semantic contracts that make up the protocol perform no LLM inference — interpretation is left to the consuming agent, under explicit guardrails.

What is included in this release

  • Graph serialization / graph payload foundation — a deterministic serialized snapshot of the semantic code graph.
  • MeshContext Report — a deterministic, machine-readable projection of the latest successful build (facts and candidates only; no inference).
  • MeshContext AI Context Pack — a compact, token-conscious, evidence-linked projection of the report, in compact, balanced, and expanded profiles.
  • Token Benchmark with source_baselines — deterministic token measurement calibrated against two baselines (mesh_context_report and serialized_graph_payload) so reduction is never read from a single source.
  • Structural Validation — a semantic contract that checks consistency across the payload → report → AI pack → benchmark chain and asserts the privacy and determinism guarantees.
  • Open-core safety export pipeline — a curated, allow-listed export with an automated privacy/safety scan over the generated public output.

What is not included / current limitations

  • Not production-ready. This is an early-validation research preview.
  • Not benchmark proof yet. A passing run demonstrates the protocol works end-to-end deterministically; it is not a published performance benchmark.
  • No aggregate risk_score yet. Risk is surfaced only as deterministic candidates.
  • No LLM inference inside the semantic contracts. Every contract carries contains_llm_inference = false.
  • Token estimates are deterministic heuristics, not tokenizer-exact counts and not model-specific billing figures.
  • Language/runtime coverage is still evolving. PHP is partial / early alpha; framework magic and runtime behavior are not fully modeled; other languages fall back to structural-only results.

Validation snapshot

The most recent recorded validation for this release candidate:

  • Semantic contract tests: pass.
  • MCP / internal validation suite: pass.
  • Open-core dry-run safety scan: clean (warnings expected at the documented baseline; safety errors: 0).
  • Open-core release-candidate smoke: passed.

Dogfooding note

LynkMesh uses its own deterministic context outputs as evidence for release documentation, with human review. The figures below come from running LynkMesh on its own repository through the validated MCP capabilities; they are deterministic facts, not LLM interpretation, and not a performance claim.

Sanitized self-analysis snapshot (illustrative, single run):

  • capabilities: feature_stage 4.4.1; MeshContext Report, AI Context Pack, and Token Benchmark all enabled; token benchmark calibrated; AI pack privacy-safe; structural validation available.
  • MeshContext Report status: ok; node_count 26; edge_count 53.
  • AI Context Pack (compact) status: ok.
  • Token Benchmark status: ok; benchmark_source_kind: mesh_context_report; source_baselines: mesh_context_report, serialized_graph_payload; calibration notes: 4.
  • guardrails.contains_llm_inference: false across report, AI pack, and token benchmark.

These values describe the output shape on one small run; they are not a public benchmark or a guarantee.

Privacy guarantees

  • No mesh_context_report.json is written by default.
  • No private absolute paths or private project names are exposed in the protocol outputs.
  • contains_llm_inference = false across all semantic contracts.
  • The token benchmark embeds derived metrics only, never raw payloads.

Related documentation