Context Inheritance: A Git-Native Architecture and Pre-registered Study of Repository Context for AI Coding Agents

Read the published paper at ACM · Presentation & PDF slides · Open study materials

The question

A repository preserves the code a team shipped. It often loses the knowledge behind that code: the rejected approaches, constraints, and lessons from failed attempts. Can we make that knowledge a reviewed part of the repository, and can a future coding agent use it?

The Repository Context Layer

The paper describes a Git-native architecture for persistent agent context. Agents consult repository knowledge before working, execute their task, and propose context updates for human review. Context branches, merges, and rolls back with the code. Context.md supplies the discovery contract; Metatron is a reference implementation.

What the study found

The pre-registered study used SWE-bench Verified to examine context authorship, the agent reading it, and delivery. On tasks sharing relevant project constraints, an 8B local model’s code-localization accuracy increased from 21.1% to 47.2% with frontier-authored context. Its own lessons brought no improvement. We call this pattern context inheritance.

The strongest result concerns locating the right code, rather than a confirmed improvement in successful bug fixes. A frontier agent’s failure-derived lessons showed a directional resolve increase, but it did not remain significant after the reported multiple-comparison correction. On a broad random sample, no context delivery improved the frontier agent’s resolve rate.

The experiment evaluates the underlying architecture with a minimal harness. It is not an end-to-end benchmark of Metatron’s extraction or serving tools. The protocol, data, scripts, and reported deviations are public.

Publication & presentation

Published October 8, 2026, pages 12–20, in the Proceedings of the 1st International Workshop on Agentic AI for Next-Generation Software Development (AgenticDev ’26), Association for Computing Machinery. DOI: 10.1145/3843282.3844425.

I presented the work in Munich on October 12, 2026 at AgenticDev, co-located with ASE. View the talk’s slides and download the presentation PDF.

Cite this work

@inproceedings{kerbel2026contextinheritance,
  author = {Kerbel, Pavel and Kerbel, Milana and
            Abramov, Vitali and Abramov, Liat},
  title = {Context Inheritance: A Git-Native Architecture and
           Pre-registered Study of Repository Context
           for AI Coding Agents},
  booktitle = {Proceedings of the 1st International Workshop on
               Agentic AI for Next-Generation Software Development},
  series = {AgenticDev ’26},
  year = {2026},
  pages = {12--20},
  publisher = {Association for Computing Machinery},
  doi = {10.1145/3843282.3844425},
  url = {https://doi.org/10.1145/3843282.3844425}
}

Earlier work

The July 2026 Repository Context Layer preprint introduced the architecture. It remains available as the earlier version; the ACM paper adds the pre-registered empirical study and four-author publication.