janicre: A Log-Scale, Reversible Semantic-Commit Manifest for Multi-Agent Software Reasoning

ACL ARR 2025 May Submission4302 Authors

Abstract:

Large language models (LLMs) promise repository-scale assistance, yet real projects routinely exceed 128k-token windows and fragment key logic across heterogeneous files. Existing techniques—RAG pipelines, chunked retrieval, or brute-force long-context prompting—either leak code, stumble on cross-file reasoning, or incur quadratic cost.

We realise .janicre via an arbitrary-depth abstraction pipeline (Stage 1...k), whose cumulative JSON snapshots plus an HTML semantic map form the final manifest; stopping at any depth trades tokens for detail while the worst-case growth remains Θ(log N).

Beyond human-to-LLM use, .janicre functions as a model-agnostic exchange layer: distinct agents (GPT-4o, Claude, Gemini, etc.) can inspect the same manifest, attach thinking_trace logs, and negotiate edits—enabling true agent-to-agent (A2A) collaboration without exposing raw source. Coupled with git-like semantic diffs, the manifest becomes a living memory of design rationale that is auditable by both machines and humans.

We formalise the schema, analyse its compression bounds, and outline an empirical plan—HumanEval, MBPP, and delta-stability benchmarks—to compare .janicre–augmented reasoning against standard retrieval and full-file prompting.

.janicre thus upgrades IR-level code summarisation into a universal, dialogue-ready substrate for large-scale, multi-agent software intelligence.

Paper Type: Long

Research Area: Generation

Research Area Keywords: large language models, code representation, intermediate representations, multi-agent reasoning, semantic abstraction, prompt engineering

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Theory

Languages Studied: Python, JavaScript, HTML, JSON, YAML

Submission Number: 4302