Skip to content

Releases

Zeno v1.0: Evidence-First Codebase Analysis

Choose a tag to compare

@anth0nylawrence anth0nylawrence released this 05 Jan 06:10

Release Title

Zeno v1.0: Evidence-First Codebase Analysis


Release Notes

Zeno v1.0.0

The first stable release of Zeno—an evidence-first, read-only workflow for analyzing massive codebases without context rot.

Zeno implements the Recursive Language Model (RLM) paradigm, treating your codebase as an external environment that AI models can programmatically explore, decompose, and query recursively. Every claim is backed by file:line citations, every session is reproducible, and every analysis leaves receipts.


Highlights

🔍 Unbounded Codebase Analysis
Analyze repositories of any size. Zeno keeps the corpus external in a JSONL REPL server, pulling only the lines needed—no more stuffing files into prompts until context rot kicks in.

📋 Evidence-First by Design
No claim without a citation. The claim-evidence ledger ensures every answer traces back to specific file:line references, making audits reproducible and verifiable.

🔁 Recursive Decomposition
Complex queries automatically split into sub-queries over smaller slices. Summarize, consolidate, repeat—the Zeno loop handles architectural questions spanning hundreds of files.

⚡ Multi-Runtime Support
Works with both Codex and Claude Code out of the box. Drop-in skill packages with persistence hooks for always-on evidence tracking.

🛡️ Read-Only & Budget-Controlled
Zeno never modifies files. Strict retrieval budgets (≤30 ops/answer, ≤2,000 lines total) prevent runaway costs and keep analysis focused.


What's Included

  • JSONL REPL Server — list_files, peek, read_file, grep, extract_symbols, stat
  • Codex Skill Package — with notify-based persistence and OTEL telemetry
  • Claude Code Skill Package — with hooks-based persistence
  • Evidence & Claims Ledgers — JSONL persistence for full audit trails
  • Documentation — protocol specs, architecture guides, and quick-start instructions

Use Cases

  • Security audits requiring citation trails
  • Dependency wiring analysis across large monorepos
  • Architecture reviews spanning hundreds of files
  • Log/trace analysis where only specific lines matter
  • Any scenario where "trust but verify" matters

Acknowledgments

Built on the RLM research by Alex Zhang et al. (MIT CSAIL). Thanks to the researchers who demonstrated that recursive inspection beats bulk ingestion—achieving 2.7x accuracy improvements on long-context benchmarks.


Full documentation: [README.md](README.md)
Paper: [arXiv:2512.24601v1](https://arxiv.org/abs/2512.24601v1)