context_core — the portable half of the data plane

A dependency-free C++17 library holding the pure decision rules that turn stored rows into an answer, plus a flat C ABI so both hosts can call it: the macOS Swift package links it into ContextCore, and the Windows CMake project (../windows) builds the identical sources.

It contains no I/O, no SQLite, no platform API, and no text tokenization. That is the whole point of where the seam was cut.

Why these three pieces, and nothing else

The app is macOS-only and stays that way — capture is ScreenCaptureKit, CoreAudio taps and Vision OCR, none of which have a Windows equivalent worth pretending about. What is portable is the layer underneath: the rules that decide when a conversation ended, which of a hundred matching rows a reader should actually see, and how many separate things a run of screenshots represents. Those rules are calibrated against real captured data (the constants below carry their measurements), and a second implementation of them would be a second set of answers to the same question.

So the seam is drawn at everything that is arithmetic over already-extracted facts:

Extracted	Why it is portable	Why it is worth extracting
ctx_should_open_new_session	one subtraction and one comparison	it is the entire session model; a host that gets it wrong shatters or merges conversations
ctx_recall_score / ctx_relevance_floor	the blend of two normalised signals and a decay	the weights are calibrated against a real ranking defect; a divergent copy silently re-introduces it
ctx_moment_groups	set arithmetic over opaque hashes and opaque keys	the churn-relative novelty bar is the subtlest rule in the product and the most expensive to re-derive

What deliberately stays in Swift is everything that needs Unicode: tokenizing a query into FTS terms, folding a window title to its words, hashing OCR into a word set. CharacterSet.alphanumerics is a full Unicode property, and reimplementing it in C++ without ICU would be a correctness regression wearing a portability badge. So the caller hands this library keys it has already normalised and word hashes it has already computed; the library never looks inside either.

The ABI is therefore free of Unicode, of allocation, and of string ownership: keys arrive as opaque byte ranges compared for equality only, words arrive as opaque 64-bit values compared for set membership only, and every buffer is owned by the caller.

Building and testing

Standalone, on macOS or Windows:

cmake -S desktop/context-for-claude/core -B /tmp/context-core-build
cmake --build /tmp/context-core-build
ctest --test-dir /tmp/context-core-build --output-on-failure

The tests are a single translation unit with a hand-rolled assertion macro — no GoogleTest, no FetchContent, nothing to resolve. A test dependency that has to be downloaded is a test suite that stops running the first time a machine is offline, and this library exists precisely to be built on machines nobody has configured yet.

On macOS the same sources are also compiled by SwiftPM as the ContextCoreCxx target and covered a second time, through ContextCore, by swift test --package-path desktop/context-for-claude. The two builds must agree; if they ever do not, the CMake build is the one that is honest, because it is the one Windows uses.