// Self-checks for everything that runs without a model server. The whole point // of ctxstream is that planning and reduction are deterministic code, so they // are testable with no GPU, no network and no tokens. #include "../src/ctxstream.hpp" #include #include #include #include #include #include namespace fs = std::filesystem; using namespace ctxstream; static int checks = 0; #define CHECK(cond, msg) \ do { \ ++checks; \ if (!(cond)) { \ std::fprintf(stderr, "FAIL %s:%d %s\n", __FILE__, __LINE__, msg); \ return 1; \ } \ } while (0) static int test_manifest_covers_everything() { std::string text; for (int i = 0; i < 5000; ++i) { text += "row " + std::to_string(i) + " || status=alpha\n"; } ManifestOptions opt; opt.segment_chars = 4000; opt.overlap_chars = 0; const auto segs = plan(text, opt); CHECK(!segs.empty(), "no segments planned"); // Every byte must land in some segment: a planner that silently drops the // tail produces a confident answer over partial data. std::size_t covered = 0; for (const auto& s : segs) covered += s.length; CHECK(covered >= text.size(), "segments do not cover the whole text"); CHECK(segs.front().offset == 0, "first segment must start at 0"); CHECK(segs.back().offset + segs.back().length == text.size(), "last segment must reach the end"); // No segment may cut a row in half. for (const auto& s : segs) { if (s.offset + s.length >= text.size()) continue; CHECK(text[s.offset + s.length - 1] == '\n', "segment ends mid-record"); } return 0; } static int test_manifest_pathological() { // A single record longer than a segment must still make progress, not loop. const std::string text(20000, 'x'); ManifestOptions opt; opt.segment_chars = 1000; opt.overlap_chars = 0; const auto segs = plan(text, opt); CHECK(segs.size() >= 20, "no progress on a record larger than a segment"); CHECK(plan("", opt).empty(), "empty text must yield no segments"); return 0; } static int test_parse_records() { // The output contract, and the shapes a small model actually emits. auto r = parse_records("alpha\t12\nbeta\t7\n"); CHECK(r.size() == 2, "tab-separated records not parsed"); CHECK(r[0].key == "alpha" && r[0].value == 12, "wrong tab parse"); // Keys with spaces must survive: "Spatial Relationship" truncated to // "Spatial" is a real failure this replaces. r = parse_records("Spatial Relationship\t109\n"); CHECK(r.size() == 1 && r[0].key == "Spatial Relationship", "key with spaces truncated"); r = parse_records("alpha: 4846\n"); CHECK(r.size() == 1 && r[0].value == 4846, "colon form not parsed"); r = parse_records("gamma\t1,905\n"); CHECK(r.size() == 1 && r[0].value == 1905, "thousands separator not handled"); CHECK(parse_records("NONE\n").empty(), "NONE must yield no records"); CHECK(parse_records("I analysed the fragment and found several statuses.\n").empty(), "prose must not parse as a record"); return 0; } static int test_tally_and_construct() { std::vector rs(3); rs[0] = {0, "alpha\t10\nbeta\t5\ngamma\t2\n", true, false, "", 0, 0, 0}; rs[1] = {1, "alpha\t20\nbeta\t5\ndelta\t1\n", true, false, "", 0, 0, 0}; // A failed segment must not be treated as empty-but-fine. rs[2] = {2, "", false, false, "boom", 0, 0, 0}; const Tally t = tally(rs); CHECK(t.sorted.front().first == "delta", "least-common wrong"); CHECK(t.sorted.back().first == "alpha", "most-common wrong"); CHECK(t.total == 43, "total wrong"); // This is the exact question the 4B model answered with a list of all four // labels instead of a selection. CHECK(construct(t, Answer::Least, "Status") == "Status: delta", "construct(Least) wrong"); CHECK(construct(t, Answer::Most, "Status") == "Status: alpha", "construct(Most) wrong"); return 0; } static int test_truncation_guard() { CHECK(looks_truncated(350000, 16387), "gross clipping not detected"); CHECK(!looks_truncated(100000, 30000), "normal ratio flagged as truncated"); CHECK(!looks_truncated(100000, 0), "unreported token count must not flag"); return 0; } static int test_codegraph() { const fs::path root = fs::temp_directory_path() / "ctxstream_gtest"; fs::remove_all(root); fs::create_directories(root / "sub"); fs::create_directories(root / ".git"); { std::ofstream a(root / "a.py"); a << "import helper\n\ndef alpha():\n return 1\n\nclass Beta:\n pass\n"; std::ofstream h(root / "sub" / "helper.py"); h << "def helper_fn():\n return 2\n"; std::ofstream junk(root / ".git" / "ignored.py"); junk << "def should_not_appear():\n pass\n"; } ScanOptions so; const CodeGraph g = scan_repo(root.string(), so); CHECK(g.files.size() == 2, "excluded dir was scanned, or files missed"); bool saw_alpha = false, saw_beta = false, saw_helper = false; for (const auto& s : g.symbols) { if (s.name == "alpha" && s.kind == NodeKind::Function) saw_alpha = true; if (s.name == "Beta" && s.kind == NodeKind::Class) saw_beta = true; if (s.name == "helper_fn") saw_helper = true; CHECK(s.name != "should_not_appear", "symbol from excluded dir"); } CHECK(saw_alpha && saw_beta && saw_helper, "symbols missed"); CHECK(!g.edges.empty(), "import edge a.py -> helper.py not found"); std::string packed; ManifestOptions mo; mo.segment_chars = 100000; const auto segs = plan_codebase(g, mo, &packed); CHECK(!segs.empty(), "no codebase segments"); CHECK(packed.find("===== FILE") != std::string::npos, "packed text missing file headers"); // A dependency must be packed before its dependent. CHECK(packed.find("helper.py") < packed.find("a.py"), "topological order not applied"); std::size_t covered = 0; for (const auto& s : segs) covered += s.length; CHECK(covered >= packed.size(), "codebase segments do not cover packed text"); fs::remove_all(root); return 0; } int main() { if (test_manifest_covers_everything()) return 1; if (test_manifest_pathological()) return 1; if (test_parse_records()) return 1; if (test_tally_and_construct()) return 1; if (test_truncation_guard()) return 1; if (test_codegraph()) return 1; std::printf("ok: %d checks passed\n", checks); return 0; }