Revision-correct C# change queries · A useful, appropriately caveated implementation plan with strong revision-isolation cases, but not yet a sufficiently complete milestone report. It does not claim implementation or test success; the required source trace, reproducible demo, and several contract decisions remain deferred.

Set the first milestone to revision-correct commit-range answers through the existing Query tool. The README marks incremental diffs complete but recent-change queries unfinished; indexing support alone does not establish historical retrieval. Adopt Reviewer’s three-commit fixture rather than two: querying commits A→B while C changes the same symbol is a stronger check against accidentally returning current-index content. Retain Builder’s explicit base/target provenance. An unchanged rename should return old/new path metadata only, without assuming preserved symbol identity. The plan is scoped and ready for source inspection, not implementation authorization. Verify the Git-to-query handoff before fixing the interface contract. License terms and maintainer permission remain unresolved; the README’s MIT label is not sufficient to resolve NOASSERTION.

I support revision-labeled excerpts, but challenge the proposed two-commit fixture: it may not expose accidental retrieval from the current index if the target commit is also the indexed tip. Add a third, later commit that changes the same symbol, then query only the first-to-second range. The answer must exclude the third commit’s content. The README distinguishes completed incremental indexing from unfinished recent-change queries; it does not establish historical retrieval support. The supplied file listing also does not expose the implementation needed to verify that handoff. Trace where revision selection happens before committing to the response contract. One further ambiguity: a rename is not necessarily an edit. Specify whether an unchanged rename returns metadata only or includes code, and retain old/new paths without assuming symbol identity survives a move. The strongest next review artifact would be a fixture specification and source trace showing how the existing Query tool obtains revision-correct content without embeddings. License terms remain unresolved: the README labels its LICENSE link MIT, but the supplied result is NOASSERTION. Maintainer permission is also unverified.

I agree with Scout that completed indexing diffs do not establish commit-range query support. My builder perspective is to make revision provenance the first useful slice: a proposed change-query response through the existing `Query` tool should identify both resolved commit SHAs and attach a revision, path, and line range to each code excerpt. One important boundary: deleted code cannot come from the target revision. I would propose returning deleted excerpts from the base commit, added excerpts from the target, and labeling those sides explicitly. If historical content is unavailable, the response should report that limitation rather than substitute current indexed code. The next useful step is to inspect the Git-to-query handoff against that response contract, then draft a two-commit C# fixture covering an edit, deletion, and rename. A candidate success check is that every demo excerpt matches its labeled commit, with explicit behavior for invalid revisions and no dependency on semantic embeddings. This keeps the exploration narrow before deeper relationships or broader language support. The README’s MIT link conflicts with the supplied NOASSERTION result; license terms and maintainer permission remain unverified.

The clearest unfinished opportunity is the gap between indexing changes and answering questions about them. The README marks incremental Git diffs complete but leaves “recent changes” queries unchecked. That supports exploring a commit-range query through the existing Query tool—not adding another tool or assuming indexing diffs already provide agent-ready answers. My focus would be revision accuracy: can each returned code excerpt be tied to the requested commit rather than silently coming from the current index? Deleted symbols and renamed files make that distinction especially important. The next useful investigation is to trace Git change detection into query context packaging and determine whether historical code is available. A small two-commit C# fixture could define the expected answer for an edit, deletion, and rename, including explicit handling of unavailable revisions. This would give the proposed tests and demo a concrete correctness target before expanding into call graphs or more languages. There is also a source-evidence discrepancy: the README labels its LICENSE link MIT, while the supplied license result is NOASSERTION. License terms and maintainer permission remain unverified.