Reduce LLM token costs by relying on the root MIT licence

[thomhurst]

Problem

The contributor guidance currently requires the full MIT licence text in every source file, even though the repository already has a root-level LICENCE.txt that applies to the repository.

This repeated text has a growing cost for agentic development. LLM tools commonly read complete files, so every file read includes the same 23-line block before reaching the code. That consumes context-window capacity, input tokens, time, and money without adding new information to each read.

A measurement from the current repository checkout found:

• 1,982 tracked C# files containing The MIT License (MIT)
• about 1,161 characters in a typical header
• roughly 290 tokens per file, using a conservative four-characters-per-token estimate
• about 2.29 million repeated header characters across the repository
• roughly 573,000 tokens if those headers are read once across the corpus
• reading only ten affected files adds roughly 2,900 repeated input tokens

The exact token count varies by model/tokenizer, but the scale of duplicated context remains significant.

Per-file contributor names are also an incomplete authorship record because multiple people may change a file over time. Git history provides a more accurate attribution record.

Question

Is there a legal, packaging, or tooling requirement for retaining the complete MIT licence block in every source file when the repository-level LICENCE.txt is present and included in distributions?

The MIT condition requires the copyright and permission notice to be included in copies or substantial portions of the software, but does not explicitly require the complete text in every individual source file.

Proposed change

If the root-level licence is sufficient:

1. Remove the full MIT licence blocks from project-owned source files.
2. Remove the per-file licence requirement from CONTRIBUTING.md and .agent_instructions/documentation.md.
3. Ensure LICENCE.txt continues to be included in source and package distributions.
4. Rely on Git history for contributor attribution.

If a file-local marker is required by tooling, a short SPDX-License-Identifier: MIT line could be considered as a lower-cost alternative, but the preferred outcome would be no repeated per-file boilerplate where the root licence is sufficient.

Third-party or vendored files with their own licensing requirements would remain untouched.

[iancooper]

Because our model assigns copyright to the original author, this doesn't work well. We could rely on GH history, but part of our promise and difference is the lack of rug pull, and I would like to center that. I am not convinced that the token burden is significant with caching. Editing over time is less the point, as the copyright is with the author, not really the editor. We do switch attribution on a significant rewrite.

[iancooper]

I may move this to a discussion