Changeset Shipped
A repo, a base and a branch become a set of files, each tagged with technology signal tokens that everything downstream routes on.
Every run starts from a merge-base diff — not the working tree, not the last commit, the full cumulative change against where the branch left main. Each file is tagged with signal tokens derived from its extension and location: which language, which layer, which platform. Those tokens are the currency the rest of the pipeline trades in.
Why the signals matter more than they look
A specialist that declares an interest in database work is dispatched because the changeset carries database signals, not because a model guessed. That is also where it can go wrong: a synthetic change carrying no technology signals at all was once filtered out of every discretionary lane, and the handoff document it contained reached nobody. Signals are cheap and deterministic, which makes them fast and makes their failure mode silent — so routing is now asserted by test.
Size is never a failure mode
Oversized diffs are handled by hunk-level map-reduce rather than truncation: the diff is parsed into hunks, packed into batches against a token budget, and each batch reviewed in full. Nothing is dropped and nothing is summarized away. That design came directly out of a production incident where a review over a roughly 50,000-line diff came back partial.