v0.3.3 — correct MIT licence

This ships the project's actual licence

Every artifact published before this one — 0.3.0, 0.3.1 and 0.3.2 — bundled the upstream Model Context Protocol project's LICENSE instead of this project's. They carry the Apache-2.0 text and a licensing-transition notice about a different project, while the README says MIT.

It arrived by accident: the correct MIT licence was added in f78e00c and overwritten by the v0.3.0 release commit 6561a8e. This restores that exact file, so it is a correction, not a relicence.

PyPI releases are immutable, so a new version is the only way to get the correct licence to users. If you installed 0.3.2 or earlier, the LICENSE on your disk is wrong — upgrade.

Licence metadata

The package previously declared no licence at all, which is why PyPI showed none. It now ships PEP 639 metadata:

License-Expression: MIT
License-File: LICENSE

Documentation

The README is a landing page rather than the whole manual — 417 words and two diagrams, down from 853 words and thirteen. Detail moved to the page that already owned each subject:

Page	Covers
Architecture	Retrieval/reasoning split, what gets indexed, trust boundary
Authentication	Device Flow handshake, connection state machine
Indexing and storage	Job states, page caps, webhook, storage modes, limits
Tools	Tool surface and when to reach for each

Removed

The benchmarks/ directory and its replay harness. The benchmark found no statistically demonstrated effect across ten merged PRs — too little to conclude anything either way — so keeping the apparatus implied a claim the data did not support.

Upgrade

uvx github-pr-context-mcp

Nothing to reconfigure. If you are coming from 0.2.x, note that 0.3.2 was the first release in which PR indexing worked at all; earlier versions shipped a GraphQL query GitHub rejected outright.

Full changelog: https://github.com/paarths-collab/github-pr-context-mcp/blob/main/CHANGELOG.md

v0.3.2 — PR indexing actually works

PR indexing works for the first time

PR_QUERY asked GitHub for diffHunk on PullRequestReviewThread, a field GitHub defines only on PullRequestReviewComment. GitHub rejected the entire query with undefinedField, so every fetch failed before a single PR was indexed.

If you are on 0.3.0 or 0.3.1, you cannot index any repository. Upgrade.

Verified against the live API: 93 of 95 review comments in a 300-PR langchain-ai/langchain fetch now carry a diff hunk, where the request previously returned nothing at all. tests/test_queries_schema.py guards the field's placement so the query cannot silently drift from GitHub's schema again.

First release carrying the v3 work

v0.3.1 was tagged before the hardening branch merged, so none of this has ever appeared in a published artifact:

• GitHub App Device Flow with OS-vault-only credentials — no personal access tokens
• Retrieval-only boundary: the server returns evidence, the IDE agent reasons
• Namespace-scoped storage and migrate-storage
• Async index jobs, durable capped-refresh continuation
• Incremental indexing by webhook
• Comment-quality filtering

Breaking change

Numbered as a patch, but Device Flow replaces personal access tokens. Upgrading from 0.3.0 or 0.3.1 means reconnecting GitHub and running migrate-storage.

Also in this release

• Webhook indexing no longer advances the refresh watermark, which had let later refreshes skip PRs updated earlier but not yet indexed
• Review-summary bot metadata no longer leaks the previous loop's inline-comment flag
• PR replay benchmark scoring against real merged diffs. It reports no statistically demonstrated effect — the measured gap is smaller than the judge's own scatter. See benchmarks/README.md
• Diagram-first README

uvx github-pr-context-mcp

Full changelog: https://github.com/paarths-collab/github-pr-context-mcp/blob/main/CHANGELOG.md