@kenkaiiii/ggcoder 4.14.3 → 4.15.0

package/dist/core/hashline-edit-benchmark.test.js

@@ -0,0 +1,141 @@
+import { describe, it, expect } from "vitest";
+import { lineHash, anchorFile } from "./hashline.js";
+import { applyBaseline, applyHashline, checkAnchors, stripFence, genFile, buildTasks, } from "./hashline-edit-benchmark.js";
+/**
+ * These tests prove the DETERMINISTIC half of the hashline benchmark — the
+ * apply + grading logic that decides OK / FAIL / corruption. The live-model
+ * numbers are only trustworthy if this logic is correct, so we exercise it from
+ * every angle: anchor uniqueness, corruption-avoidance, multi-edit ordering,
+ * malformed input, and the string-match baseline's ambiguity handling.
+ */
+/** Build a minimal EditTask around a hand-written file for precise assertions. */
+function taskFor(file, mustContain, mustPreserve) {
+    return {
+        name: "t",
+        approxLines: file.split("\n").length,
+        file,
+        instruction: "",
+        mustContain,
+        mustPreserve,
+    };
+}
+const FN_FILE = [
+    "export function computeTimeout(cfg: Config): number {",
+    "  return cfg.timeoutMs > 0 ? cfg.timeoutMs : DEFAULT_TIMEOUT;",
+    "}",
+].join("\n");
+/** The anchor for a given 0-based line index of a file. */
+function anchorAt(file, index) {
+    return lineHash(file.split("\n")[index], index);
+}
+describe("hashline anchoring", () => {
+    it("1. folds line position into the hash so identical lines get distinct anchors", () => {
+        // Two identical lines at different positions must NOT collide.
+        expect(lineHash("  return x;", 3)).not.toBe(lineHash("  return x;", 9));
+        // Same content + same position is stable (deterministic).
+        expect(lineHash("  return x;", 3)).toBe(lineHash("  return x;", 3));
+    });
+    it("2. gives every line a unique resolvable anchor — even repeated blank lines", () => {
+        const file = ["a", "", "b", "", "c"].join("\n"); // two blank lines
+        const a = anchorFile(file);
+        expect(a.ambiguous.size).toBe(0);
+        expect(a.anchorToIndex.size).toBe(5);
+        expect([...a.anchorToIndex.values()].sort((x, y) => x - y)).toEqual([0, 1, 2, 3, 4]);
+    });
+    it("3. renders as `anchor│line` and the line column reconstructs the file", () => {
+        const a = anchorFile(FN_FILE);
+        const stripped = a.rendered
+            .split("\n")
+            .map((l) => l.slice(l.indexOf("│") + 1))
+            .join("\n");
+        expect(stripped).toBe(FN_FILE);
+    });
+});
+describe("applyHashline (proposed)", () => {
+    it("4. applies a valid single-line edit and grades it correct", () => {
+        const anchored = anchorFile(FN_FILE);
+        const anchor = anchorAt(FN_FILE, 1); // the `return ...` line
+        const raw = JSON.stringify({
+            edits: [
+                {
+                    from: anchor,
+                    to: anchor,
+                    lines: [
+                        "  return cfg.retries > 0 ? Math.min(cfg.timeoutMs * cfg.retries, 30000) : (cfg.timeoutMs > 0 ? cfg.timeoutMs : DEFAULT_TIMEOUT);",
+                    ],
+                },
+            ],
+        });
+        const task = taskFor(FN_FILE, ["cfg.timeoutMs * cfg.retries", "30000"], ["computeTimeout", "DEFAULT_TIMEOUT"]);
+        const out = applyHashline(raw, task, anchored);
+        expect(out.applied).toBe(true);
+        expect(out.correct).toBe(true);
+        expect(out.ambiguousEdits).toBe(0);
+    });
+    it("5. rejects an edit whose anchor is not in the file — never corrupts", () => {
+        const anchored = anchorFile(FN_FILE);
+        const raw = JSON.stringify({ edits: [{ from: "0000", to: "0000", lines: ["// hijacked"] }] });
+        const task = taskFor(FN_FILE, [], ["computeTimeout"]);
+        const out = applyHashline(raw, task, anchored);
+        expect(out.applied).toBe(false);
+        expect(out.correct).toBe(false);
+        expect(out.ambiguousEdits).toBe(1);
+    });
+    it("6. rejects a reversed range (from after to)", () => {
+        const anchored = anchorFile(FN_FILE);
+        const raw = JSON.stringify({
+            edits: [{ from: anchorAt(FN_FILE, 2), to: anchorAt(FN_FILE, 0), lines: ["x"] }],
+        });
+        const out = applyHashline(raw, taskFor(FN_FILE, [], []), anchored);
+        expect(out.applied).toBe(false);
+    });
+    it("7. applies multiple edits bottom-up so earlier indices stay valid", () => {
+        const file = ["const a = 1;", "const b = 2;", "const c = 3;"].join("\n");
+        const anchored = anchorFile(file);
+        const raw = JSON.stringify({
+            edits: [
+                { from: anchorAt(file, 0), to: anchorAt(file, 0), lines: ["const a = 10;"] },
+                { from: anchorAt(file, 2), to: anchorAt(file, 2), lines: ["const c = 30;"] },
+            ],
+        });
+        const out = applyHashline(raw, taskFor(file, ["const a = 10;", "const c = 30;"], ["const b = 2;"]), anchored);
+        expect(out.applied).toBe(true);
+        expect(out.correct).toBe(true);
+    });
+    it("8. returns a clean failure on malformed JSON instead of throwing", () => {
+        const anchored = anchorFile(FN_FILE);
+        expect(() => applyHashline("not json {", taskFor(FN_FILE, [], []), anchored)).not.toThrow();
+        const out = applyHashline("not json {", taskFor(FN_FILE, [], []), anchored);
+        expect(out.applied).toBe(false);
+        expect(out.parsedEdits).toBe(0);
+    });
+});
+describe("applyBaseline (current string-match) + helpers", () => {
+    it("9. applies a unique old_text but rejects a non-unique one", () => {
+        const unique = taskFor("const x = 1;\nconst y = 2;\n", ["const x = 100;"], ["const y = 2;"]);
+        const okRaw = JSON.stringify({
+            edits: [{ old_text: "const x = 1;", new_text: "const x = 100;" }],
+        });
+        expect(applyBaseline(okRaw, unique).applied).toBe(true);
+        // Same literal appears twice → the real edit tool refuses; so must we.
+        const dup = taskFor("const x = 1;\nconst y = 2;\nconst x = 1;\n", [], []);
+        const dupRaw = JSON.stringify({
+            edits: [{ old_text: "const x = 1;", new_text: "const x = 9;" }],
+        });
+        const out = applyBaseline(dupRaw, dup);
+        expect(out.applied).toBe(false);
+        expect(out.ambiguousEdits).toBe(1);
+    });
+    it("10. flags not-found old_text, strips code fences, and grades generated tasks", () => {
+        // Paraphrased / drifted old_text that isn't in the file → rejected.
+        const notFound = JSON.stringify({ edits: [{ old_text: "const z = 0;", new_text: "x" }] });
+        expect(applyBaseline(notFound, taskFor("const a = 1;\n", [], [])).applied).toBe(false);
+        // stripFence removes ```ts ... ``` wrappers the model sometimes adds.
+        expect(stripFence('```ts\n{"edits":[]}\n```')).toBe('{"edits":[]}');
+        // The generator + grader agree: a freshly generated file preserves its anchors.
+        const task = buildTasks()[0];
+        expect(genFile(task.approxLines)).toContain("computeTimeout");
+        expect(checkAnchors(task.file, taskFor(task.file, [], task.mustPreserve))).toBe(true);
+    });
+});
+//# sourceMappingURL=hashline-edit-benchmark.test.js.map

package/dist/core/hashline-edit-benchmark.test.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"hashline-edit-benchmark.test.js","sourceRoot":"","sources":["../../src/core/hashline-edit-benchmark.test.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,EAAE,EAAE,MAAM,EAAE,MAAM,QAAQ,CAAC;AAC9C,OAAO,EAAE,QAAQ,EAAE,UAAU,EAAE,MAAM,eAAe,CAAC;AACrD,OAAO,EACL,aAAa,EACb,aAAa,EACb,YAAY,EACZ,UAAU,EACV,OAAO,EACP,UAAU,GAEX,MAAM,8BAA8B,CAAC;AAEtC;;;;;;GAMG;AAEH,kFAAkF;AAClF,SAAS,OAAO,CAAC,IAAY,EAAE,WAAqB,EAAE,YAAsB;IAC1E,OAAO;QACL,IAAI,EAAE,GAAG;QACT,WAAW,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,MAAM;QACpC,IAAI;QACJ,WAAW,EAAE,EAAE;QACf,WAAW;QACX,YAAY;KACb,CAAC;AACJ,CAAC;AAED,MAAM,OAAO,GAAG;IACd,uDAAuD;IACvD,+DAA+D;IAC/D,GAAG;CACJ,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAEb,2DAA2D;AAC3D,SAAS,QAAQ,CAAC,IAAY,EAAE,KAAa;IAC3C,OAAO,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,KAAK,CAAE,EAAE,KAAK,CAAC,CAAC;AACnD,CAAC;AAED,QAAQ,CAAC,oBAAoB,EAAE,GAAG,EAAE;IAClC,EAAE,CAAC,8EAA8E,EAAE,GAAG,EAAE;QACtF,+DAA+D;QAC/D,MAAM,CAAC,QAAQ,CAAC,aAAa,EAAE,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,QAAQ,CAAC,aAAa,EAAE,CAAC,CAAC,CAAC,CAAC;QACxE,0DAA0D;QAC1D,MAAM,CAAC,QAAQ,CAAC,aAAa,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,aAAa,EAAE,CAAC,CAAC,CAAC,CAAC;IACtE,CAAC,CAAC,CAAC;IAEH,EAAE,CAAC,4EAA4E,EAAE,GAAG,EAAE;QACpF,MAAM,IAAI,GAAG,CAAC,GAAG,EAAE,EAAE,EAAE,GAAG,EAAE,EAAE,EAAE,GAAG,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,kBAAkB;QACnE,MAAM,CAAC,GAAG,UAAU,CAAC,IAAI,CAAC,CAAC;QAC3B,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QACjC,MAAM,CAAC,CAAC,CAAC,aAAa,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QACrC,MAAM,CAAC,CAAC,GAAG,CAAC,CAAC,aAAa,CAAC,MAAM,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;IACvF,CAAC,CAAC,CAAC;IAEH,EAAE,CAAC,uEAAuE,EAAE,GAAG,EAAE;QAC/E,MAAM,CAAC,GAAG,UAAU,CAAC,OAAO,CAAC,CAAC;QAC9B,MAAM,QAAQ,GAAG,CAAC,CAAC,QAAQ;aACxB,KAAK,CAAC,IAAI,CAAC;aACX,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;aACvC,IAAI,CAAC,IAAI,CAAC,CAAC;QACd,MAAM,CAAC,QAAQ,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;IACjC,CAAC,CAAC,CAAC;AACL,CAAC,CAAC,CAAC;AAEH,QAAQ,CAAC,0BAA0B,EAAE,GAAG,EAAE;IACxC,EAAE,CAAC,2DAA2D,EAAE,GAAG,EAAE;QACnE,MAAM,QAAQ,GAAG,UAAU,CAAC,OAAO,CAAC,CAAC;QACrC,MAAM,MAAM,GAAG,QAAQ,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,CAAC,wBAAwB;QAC7D,MAAM,GAAG,GAAG,IAAI,CAAC,SAAS,CAAC;YACzB,KAAK,EAAE;gBACL;oBACE,IAAI,EAAE,MAAM;oBACZ,EAAE,EAAE,MAAM;oBACV,KAAK,EAAE;wBACL,kIAAkI;qBACnI;iBACF;aACF;SACF,CAAC,CAAC;QACH,MAAM,IAAI,GAAG,OAAO,CAClB,OAAO,EACP,CAAC,6BAA6B,EAAE,OAAO,CAAC,EACxC,CAAC,gBAAgB,EAAE,iBAAiB,CAAC,CACtC,CAAC;QACF,MAAM,GAAG,GAAG,aAAa,CAAC,GAAG,EAAE,IAAI,EAAE,QAAQ,CAAC,CAAC;QAC/C,MAAM,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAC/B,MAAM,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAC/B,MAAM,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IACrC,CAAC,CAAC,CAAC;IAEH,EAAE,CAAC,qEAAqE,EAAE,GAAG,EAAE;QAC7E,MAAM,QAAQ,GAAG,UAAU,CAAC,OAAO,CAAC,CAAC;QACrC,MAAM,GAAG,GAAG,IAAI,CAAC,SAAS,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,CAAC,aAAa,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC;QAC9F,MAAM,IAAI,GAAG,OAAO,CAAC,OAAO,EAAE,EAAE,EAAE,CAAC,gBAAgB,CAAC,CAAC,CAAC;QACtD,MAAM,GAAG,GAAG,aAAa,CAAC,GAAG,EAAE,IAAI,EAAE,QAAQ,CAAC,CAAC;QAC/C,MAAM,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QAChC,MAAM,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QAChC,MAAM,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IACrC,CAAC,CAAC,CAAC;IAEH,EAAE,CAAC,6CAA6C,EAAE,GAAG,EAAE;QACrD,MAAM,QAAQ,GAAG,UAAU,CAAC,OAAO,CAAC,CAAC;QACrC,MAAM,GAAG,GAAG,IAAI,CAAC,SAAS,CAAC;YACzB,KAAK,EAAE,CAAC,EAAE,IAAI,EAAE,QAAQ,CAAC,OAAO,EAAE,CAAC,CAAC,EAAE,EAAE,EAAE,QAAQ,CAAC,OAAO,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC;SAChF,CAAC,CAAC;QACH,MAAM,GAAG,GAAG,aAAa,CAAC,GAAG,EAAE,OAAO,CAAC,OAAO,EAAE,EAAE,EAAE,EAAE,CAAC,EAAE,QAAQ,CAAC,CAAC;QACnE,MAAM,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;IAClC,CAAC,CAAC,CAAC;IAEH,EAAE,CAAC,mEAAmE,EAAE,GAAG,EAAE;QAC3E,MAAM,IAAI,GAAG,CAAC,cAAc,EAAE,cAAc,EAAE,cAAc,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACzE,MAAM,QAAQ,GAAG,UAAU,CAAC,IAAI,CAAC,CAAC;QAClC,MAAM,GAAG,GAAG,IAAI,CAAC,SAAS,CAAC;YACzB,KAAK,EAAE;gBACL,EAAE,IAAI,EAAE,QAAQ,CAAC,IAAI,EAAE,CAAC,CAAC,EAAE,EAAE,EAAE,QAAQ,CAAC,IAAI,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,eAAe,CAAC,EAAE;gBAC5E,EAAE,IAAI,EAAE,QAAQ,CAAC,IAAI,EAAE,CAAC,CAAC,EAAE,EAAE,EAAE,QAAQ,CAAC,IAAI,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,eAAe,CAAC,EAAE;aAC7E;SACF,CAAC,CAAC;QACH,MAAM,GAAG,GAAG,aAAa,CACvB,GAAG,EACH,OAAO,CAAC,IAAI,EAAE,CAAC,eAAe,EAAE,eAAe,CAAC,EAAE,CAAC,cAAc,CAAC,CAAC,EACnE,QAAQ,CACT,CAAC;QACF,MAAM,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAC/B,MAAM,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACjC,CAAC,CAAC,CAAC;IAEH,EAAE,CAAC,kEAAkE,EAAE,GAAG,EAAE;QAC1E,MAAM,QAAQ,GAAG,UAAU,CAAC,OAAO,CAAC,CAAC;QACrC,MAAM,CAAC,GAAG,EAAE,CAAC,aAAa,CAAC,YAAY,EAAE,OAAO,CAAC,OAAO,EAAE,EAAE,EAAE,EAAE,CAAC,EAAE,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QAC5F,MAAM,GAAG,GAAG,aAAa,CAAC,YAAY,EAAE,OAAO,CAAC,OAAO,EAAE,EAAE,EAAE,EAAE,CAAC,EAAE,QAAQ,CAAC,CAAC;QAC5E,MAAM,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QAChC,MAAM,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAClC,CAAC,CAAC,CAAC;AACL,CAAC,CAAC,CAAC;AAEH,QAAQ,CAAC,gDAAgD,EAAE,GAAG,EAAE;IAC9D,EAAE,CAAC,2DAA2D,EAAE,GAAG,EAAE;QACnE,MAAM,MAAM,GAAG,OAAO,CAAC,8BAA8B,EAAE,CAAC,gBAAgB,CAAC,EAAE,CAAC,cAAc,CAAC,CAAC,CAAC;QAC7F,MAAM,KAAK,GAAG,IAAI,CAAC,SAAS,CAAC;YAC3B,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,cAAc,EAAE,QAAQ,EAAE,gBAAgB,EAAE,CAAC;SAClE,CAAC,CAAC;QACH,MAAM,CAAC,aAAa,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAExD,uEAAuE;QACvE,MAAM,GAAG,GAAG,OAAO,CAAC,4CAA4C,EAAE,EAAE,EAAE,EAAE,CAAC,CAAC;QAC1E,MAAM,MAAM,GAAG,IAAI,CAAC,SAAS,CAAC;YAC5B,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,cAAc,EAAE,QAAQ,EAAE,cAAc,EAAE,CAAC;SAChE,CAAC,CAAC;QACH,MAAM,GAAG,GAAG,aAAa,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC;QACvC,MAAM,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QAChC,MAAM,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IACrC,CAAC,CAAC,CAAC;IAEH,EAAE,CAAC,8EAA8E,EAAE,GAAG,EAAE;QACtF,oEAAoE;QACpE,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,cAAc,EAAE,QAAQ,EAAE,GAAG,EAAE,CAAC,EAAE,CAAC,CAAC;QAC1F,MAAM,CAAC,aAAa,CAAC,QAAQ,EAAE,OAAO,CAAC,gBAAgB,EAAE,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QAEvF,sEAAsE;QACtE,MAAM,CAAC,UAAU,CAAC,0BAA0B,CAAC,CAAC,CAAC,IAAI,CAAC,cAAc,CAAC,CAAC;QAEpE,gFAAgF;QAChF,MAAM,IAAI,GAAG,UAAU,EAAE,CAAC,CAAC,CAAE,CAAC;QAC9B,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC,CAAC,SAAS,CAAC,gBAAgB,CAAC,CAAC;QAC9D,MAAM,CAAC,YAAY,CAAC,IAAI,CAAC,IAAI,EAAE,OAAO,CAAC,IAAI,CAAC,IAAI,EAAE,EAAE,EAAE,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACxF,CAAC,CAAC,CAAC;AACL,CAAC,CAAC,CAAC"}

package/dist/core/hashline.d.ts

@@ -0,0 +1,50 @@
+/**
+ * 4-hex-char anchor for a line. Position is folded into the hash so anchors are
+ * UNIQUE by construction (blank lines and repeated lines no longer collide).
+ * Resolution stays O(1) via a lookup map. `index` is the 0-based line index.
+ */
+export declare function lineHash(line: string, index: number): string;
+/** File rendered with `anchor│line` prefixes for the model to read. */
+export declare function renderWithAnchors(file: string): string;
+export interface AnchoredFile {
+    /** File rendered with `anchor│line` prefixes for the model to read. */
+    rendered: string;
+    /** anchor → line index (0-based). Only UNIQUE anchors are resolvable. */
+    anchorToIndex: Map<string, number>;
+    /** anchors that collided (ambiguous → unresolvable, like a stale-file reject). */
+    ambiguous: Set<string>;
+    lines: string[];
+}
+export declare function anchorFile(file: string): AnchoredFile;
+/**
+ * True when the line at `index` (0-based) still hashes to `hash`. Out-of-range
+ * indices return false. This is the staleness gate the edit tool uses.
+ */
+export declare function verifyAnchor(lines: string[], index: number, hash: string): boolean;
+/** Optional per-edit anchor guard: pin both endpoints of an edit by line+hash. */
+export interface EditAnchor {
+    /** 1-based line number of the first line of the edited span. */
+    start_line: number;
+    /** Content anchor of the first line (from a `read` with `anchors: true`). */
+    start_hash: string;
+    /** 1-based line number of the last line of the edited span. */
+    end_line: number;
+    /** Content anchor of the last line. */
+    end_hash: string;
+}
+export type AnchorFailure = "out_of_range" | "hash_mismatch" | "reversed";
+export interface AnchorResolution {
+    ok: boolean;
+    /** 0-based index of the first line (only when ok). */
+    startIndex?: number;
+    /** 0-based index of the last line (only when ok). */
+    endIndex?: number;
+    reason?: AnchorFailure;
+}
+/**
+ * Resolve an anchor against the current file lines (0-based). Rejects the edit
+ * if either endpoint is out of range, the range is reversed, or either hash no
+ * longer matches — the corruption-avoidance property.
+ */
+export declare function resolveAnchoredEdit(lines: string[], anchor: EditAnchor): AnchorResolution;
+//# sourceMappingURL=hashline.d.ts.map

package/dist/core/hashline.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"hashline.d.ts","sourceRoot":"","sources":["../../src/core/hashline.ts"],"names":[],"mappings":"AAYA;;;;GAIG;AACH,wBAAgB,QAAQ,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,MAAM,CAE5D;AAED,uEAAuE;AACvE,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAKtD;AAED,MAAM,WAAW,YAAY;IAC3B,uEAAuE;IACvE,QAAQ,EAAE,MAAM,CAAC;IACjB,yEAAyE;IACzE,aAAa,EAAE,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACnC,kFAAkF;IAClF,SAAS,EAAE,GAAG,CAAC,MAAM,CAAC,CAAC;IACvB,KAAK,EAAE,MAAM,EAAE,CAAC;CACjB;AAED,wBAAgB,UAAU,CAAC,IAAI,EAAE,MAAM,GAAG,YAAY,CAiBrD;AAED;;;GAGG;AACH,wBAAgB,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,OAAO,CAGlF;AAED,kFAAkF;AAClF,MAAM,WAAW,UAAU;IACzB,gEAAgE;IAChE,UAAU,EAAE,MAAM,CAAC;IACnB,6EAA6E;IAC7E,UAAU,EAAE,MAAM,CAAC;IACnB,+DAA+D;IAC/D,QAAQ,EAAE,MAAM,CAAC;IACjB,uCAAuC;IACvC,QAAQ,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,MAAM,aAAa,GAAG,cAAc,GAAG,eAAe,GAAG,UAAU,CAAC;AAE1E,MAAM,WAAW,gBAAgB;IAC/B,EAAE,EAAE,OAAO,CAAC;IACZ,sDAAsD;IACtD,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,qDAAqD;IACrD,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,MAAM,CAAC,EAAE,aAAa,CAAC;CACxB;AAED;;;;GAIG;AACH,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,MAAM,EAAE,EAAE,MAAM,EAAE,UAAU,GAAG,gBAAgB,CAgBzF"}

package/dist/core/hashline.js

@@ -0,0 +1,76 @@
+/**
+ * Hash-anchored line addressing — the pure, UI-free core shared by the hashline
+ * benchmark and the opt-in anchor guard in the read/edit tools.
+ *
+ * Every line gets a short content+position hash. Anchors are UNIQUE by
+ * construction (the line's index is folded into the hash, so blank/repeated
+ * lines never collide), which is what lets an edit either resolve to exactly one
+ * location or be rejected — never silently corrupt a file that drifted since the
+ * model last read it.
+ */
+import { createHash } from "node:crypto";
+/**
+ * 4-hex-char anchor for a line. Position is folded into the hash so anchors are
+ * UNIQUE by construction (blank lines and repeated lines no longer collide).
+ * Resolution stays O(1) via a lookup map. `index` is the 0-based line index.
+ */
+export function lineHash(line, index) {
+    return createHash("sha1").update(`${index}:${line.trim()}`).digest("hex").slice(0, 4);
+}
+/** File rendered with `anchor│line` prefixes for the model to read. */
+export function renderWithAnchors(file) {
+    return file
+        .split("\n")
+        .map((l, i) => `${lineHash(l, i)}│${l}`)
+        .join("\n");
+}
+export function anchorFile(file) {
+    const lines = file.split("\n");
+    const counts = new Map();
+    lines.forEach((l, i) => {
+        const h = lineHash(l, i);
+        const arr = counts.get(h) ?? [];
+        arr.push(i);
+        counts.set(h, arr);
+    });
+    const anchorToIndex = new Map();
+    const ambiguous = new Set();
+    for (const [h, idxs] of counts) {
+        // Position-folded anchors are unique unless sha1 itself collides in 16 bits.
+        if (idxs.length === 1)
+            anchorToIndex.set(h, idxs[0]);
+        else
+            ambiguous.add(h);
+    }
+    return { rendered: renderWithAnchors(file), anchorToIndex, ambiguous, lines };
+}
+/**
+ * True when the line at `index` (0-based) still hashes to `hash`. Out-of-range
+ * indices return false. This is the staleness gate the edit tool uses.
+ */
+export function verifyAnchor(lines, index, hash) {
+    if (index < 0 || index >= lines.length)
+        return false;
+    return lineHash(lines[index], index) === hash;
+}
+/**
+ * Resolve an anchor against the current file lines (0-based). Rejects the edit
+ * if either endpoint is out of range, the range is reversed, or either hash no
+ * longer matches — the corruption-avoidance property.
+ */
+export function resolveAnchoredEdit(lines, anchor) {
+    const startIndex = anchor.start_line - 1;
+    const endIndex = anchor.end_line - 1;
+    if (startIndex < 0 || endIndex < 0 || startIndex >= lines.length || endIndex >= lines.length) {
+        return { ok: false, reason: "out_of_range" };
+    }
+    if (startIndex > endIndex) {
+        return { ok: false, reason: "reversed" };
+    }
+    if (!verifyAnchor(lines, startIndex, anchor.start_hash) ||
+        !verifyAnchor(lines, endIndex, anchor.end_hash)) {
+        return { ok: false, reason: "hash_mismatch" };
+    }
+    return { ok: true, startIndex, endIndex };
+}
+//# sourceMappingURL=hashline.js.map

package/dist/core/hashline.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"hashline.js","sourceRoot":"","sources":["../../src/core/hashline.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AACH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAEzC;;;;GAIG;AACH,MAAM,UAAU,QAAQ,CAAC,IAAY,EAAE,KAAa;IAClD,OAAO,UAAU,CAAC,MAAM,CAAC,CAAC,MAAM,CAAC,GAAG,KAAK,IAAI,IAAI,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;AACxF,CAAC;AAED,uEAAuE;AACvE,MAAM,UAAU,iBAAiB,CAAC,IAAY;IAC5C,OAAO,IAAI;SACR,KAAK,CAAC,IAAI,CAAC;SACX,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,QAAQ,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;SACvC,IAAI,CAAC,IAAI,CAAC,CAAC;AAChB,CAAC;AAYD,MAAM,UAAU,UAAU,CAAC,IAAY;IACrC,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;IAC/B,MAAM,MAAM,GAAG,IAAI,GAAG,EAAoB,CAAC;IAC3C,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE;QACrB,MAAM,CAAC,GAAG,QAAQ,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;QACzB,MAAM,GAAG,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;QAChC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QACZ,MAAM,CAAC,GAAG,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC;IACrB,CAAC,CAAC,CAAC;IACH,MAAM,aAAa,GAAG,IAAI,GAAG,EAAkB,CAAC;IAChD,MAAM,SAAS,GAAG,IAAI,GAAG,EAAU,CAAC;IACpC,KAAK,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,IAAI,MAAM,EAAE,CAAC;QAC/B,6EAA6E;QAC7E,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;YAAE,aAAa,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,CAAE,CAAC,CAAC;;YACjD,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;IACxB,CAAC;IACD,OAAO,EAAE,QAAQ,EAAE,iBAAiB,CAAC,IAAI,CAAC,EAAE,aAAa,EAAE,SAAS,EAAE,KAAK,EAAE,CAAC;AAChF,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,YAAY,CAAC,KAAe,EAAE,KAAa,EAAE,IAAY;IACvE,IAAI,KAAK,GAAG,CAAC,IAAI,KAAK,IAAI,KAAK,CAAC,MAAM;QAAE,OAAO,KAAK,CAAC;IACrD,OAAO,QAAQ,CAAC,KAAK,CAAC,KAAK,CAAE,EAAE,KAAK,CAAC,KAAK,IAAI,CAAC;AACjD,CAAC;AAyBD;;;;GAIG;AACH,MAAM,UAAU,mBAAmB,CAAC,KAAe,EAAE,MAAkB;IACrE,MAAM,UAAU,GAAG,MAAM,CAAC,UAAU,GAAG,CAAC,CAAC;IACzC,MAAM,QAAQ,GAAG,MAAM,CAAC,QAAQ,GAAG,CAAC,CAAC;IACrC,IAAI,UAAU,GAAG,CAAC,IAAI,QAAQ,GAAG,CAAC,IAAI,UAAU,IAAI,KAAK,CAAC,MAAM,IAAI,QAAQ,IAAI,KAAK,CAAC,MAAM,EAAE,CAAC;QAC7F,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,cAAc,EAAE,CAAC;IAC/C,CAAC;IACD,IAAI,UAAU,GAAG,QAAQ,EAAE,CAAC;QAC1B,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,UAAU,EAAE,CAAC;IAC3C,CAAC;IACD,IACE,CAAC,YAAY,CAAC,KAAK,EAAE,UAAU,EAAE,MAAM,CAAC,UAAU,CAAC;QACnD,CAAC,YAAY,CAAC,KAAK,EAAE,QAAQ,EAAE,MAAM,CAAC,QAAQ,CAAC,EAC/C,CAAC;QACD,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,eAAe,EAAE,CAAC;IAChD,CAAC;IACD,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,QAAQ,EAAE,CAAC;AAC5C,CAAC"}

package/dist/core/semantic-search-benchmark.d.ts

@@ -0,0 +1,37 @@
+/**
+ * Semantic AST-chunk retrieval vs whole-file reads — real-API measurement of
+ * whether Feature #3 is worth building.
+ *
+ * The claim (cocoindex-code / oh-my-pi): replacing "grep then read whole files"
+ * with "retrieve only the relevant AST chunks" cuts ~70% of the tokens an agent
+ * spends locating code, with no loss of answer quality. We test that directly on
+ * OUR OWN repo against a live model.
+ *
+ * For a set of natural-language questions about real files in this repo, we build
+ * three context strategies and ask the model the same question with each:
+ *
+ *   BASELINE (whole-file): deliver the FULL text of the top files a lexical
+ *   grep would surface — this is what the agent reads today (read + grep).
+ *
+ *   SEMANTIC (AST chunks): parse every file into top-level declarations
+ *   (functions / classes / interfaces / consts), rank chunks with a real BM25
+ *   retriever, and deliver only the top-k chunks. No embedding model needed; a
+ *   learned embedding retriever would land between BM25 and ORACLE.
+ *
+ *   ORACLE (upper bound): deliver only the hand-labelled answer chunk(s) — the
+ *   best case any retriever could achieve.
+ *
+ * We measure, per question and strategy: INPUT tokens delivered (the headline
+ * cost) and whether the model's answer was correct (deterministic keyword grade).
+ * The verdict: does SEMANTIC reach BASELINE-level correctness at a fraction of
+ * the input tokens?
+ *
+ * Usage:
+ *   npx tsx src/core/semantic-search-benchmark.ts
+ *
+ * Env overrides:
+ *   GG_SS_PROVIDER / GG_SS_MODEL   (default openai / gpt-5.5)
+ *   GG_SS_TOPK                     (chunks/files delivered, default 3)
+ */
+export declare function grade(answer: string, mustInclude: string[]): boolean;
+//# sourceMappingURL=semantic-search-benchmark.d.ts.map

package/dist/core/semantic-search-benchmark.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"semantic-search-benchmark.d.ts","sourceRoot":"","sources":["../../src/core/semantic-search-benchmark.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AAwHH,wBAAgB,KAAK,CAAC,MAAM,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,GAAG,OAAO,CAGpE"}

package/dist/core/semantic-search-benchmark.js

@@ -0,0 +1,211 @@
+/**
+ * Semantic AST-chunk retrieval vs whole-file reads — real-API measurement of
+ * whether Feature #3 is worth building.
+ *
+ * The claim (cocoindex-code / oh-my-pi): replacing "grep then read whole files"
+ * with "retrieve only the relevant AST chunks" cuts ~70% of the tokens an agent
+ * spends locating code, with no loss of answer quality. We test that directly on
+ * OUR OWN repo against a live model.
+ *
+ * For a set of natural-language questions about real files in this repo, we build
+ * three context strategies and ask the model the same question with each:
+ *
+ *   BASELINE (whole-file): deliver the FULL text of the top files a lexical
+ *   grep would surface — this is what the agent reads today (read + grep).
+ *
+ *   SEMANTIC (AST chunks): parse every file into top-level declarations
+ *   (functions / classes / interfaces / consts), rank chunks with a real BM25
+ *   retriever, and deliver only the top-k chunks. No embedding model needed; a
+ *   learned embedding retriever would land between BM25 and ORACLE.
+ *
+ *   ORACLE (upper bound): deliver only the hand-labelled answer chunk(s) — the
+ *   best case any retriever could achieve.
+ *
+ * We measure, per question and strategy: INPUT tokens delivered (the headline
+ * cost) and whether the model's answer was correct (deterministic keyword grade).
+ * The verdict: does SEMANTIC reach BASELINE-level correctness at a fraction of
+ * the input tokens?
+ *
+ * Usage:
+ *   npx tsx src/core/semantic-search-benchmark.ts
+ *
+ * Env overrides:
+ *   GG_SS_PROVIDER / GG_SS_MODEL   (default openai / gpt-5.5)
+ *   GG_SS_TOPK                     (chunks/files delivered, default 3)
+ */
+import fs from "node:fs";
+import path from "node:path";
+import { fileURLToPath } from "node:url";
+import { stream } from "@kenkaiiii/gg-ai";
+import { AuthStorage } from "./auth-storage.js";
+import { chunkFile, bm25Rank, rankFiles } from "./code-retrieval.js";
+const HERE = path.dirname(fileURLToPath(import.meta.url));
+const SRC = path.resolve(HERE, ".."); // packages/ggcoder/src
+const QUESTIONS = [
+    {
+        q: "Which method resolves provider credentials and auto-refreshes expired OAuth tokens, and what happens if it is not logged in?",
+        files: ["core/auth-storage.ts", "core/loop-breaker.ts", "tools/edit-diff.ts"],
+        oracle: { file: "core/auth-storage.ts", symbol: "resolveCredentials" },
+        mustInclude: ["resolvecredentials", "refresh"],
+    },
+    {
+        q: "What function performs fuzzy text matching for the edit tool, and how does it tolerate indentation drift?",
+        files: ["tools/edit-diff.ts", "core/auth-storage.ts", "core/checkpoint-store.ts"],
+        oracle: { file: "tools/edit-diff.ts", symbol: "fuzzyFindText" },
+        mustInclude: ["fuzzyfindtext", "indent"],
+    },
+    {
+        q: "What restore modes does the checkpoint / rewind system support?",
+        files: ["core/checkpoint-store.ts", "core/loop-breaker.ts", "tools/edit-diff.ts"],
+        oracle: { file: "core/checkpoint-store.ts", symbol: "RestoreMode" },
+        mustInclude: ["code", "conversation", "both"],
+    },
+];
+function sleep(ms) {
+    return new Promise((r) => setTimeout(r, ms));
+}
+async function ask(provider, model, c, context, question) {
+    const messages = [
+        {
+            role: "system",
+            content: "You answer questions about a codebase using ONLY the provided context. " +
+                "Be specific: name the exact functions/types involved. If the context is insufficient, say so. " +
+                "Answer in 1-3 sentences.",
+        },
+        { role: "user", content: `CONTEXT:\n${context}\n\nQUESTION: ${question}` },
+    ];
+    let lastErr;
+    for (let attempt = 0; attempt < 4; attempt++) {
+        try {
+            let text = "";
+            const result = stream({
+                provider: provider,
+                model,
+                messages,
+                maxTokens: 512,
+                apiKey: c.apiKey,
+                baseUrl: c.baseUrl,
+                accountId: c.accountId,
+            });
+            for await (const event of result) {
+                if (event.type === "text_delta")
+                    text += event.text;
+            }
+            const response = await result.response;
+            const content = response.message.content;
+            const finalText = typeof content === "string"
+                ? content
+                : content
+                    .filter((p) => p.type === "text")
+                    .map((p) => p.text ?? "")
+                    .join("");
+            return {
+                text: finalText || text,
+                inputTokens: response.usage.inputTokens,
+                outputTokens: response.usage.outputTokens,
+            };
+        }
+        catch (err) {
+            lastErr = err;
+            await sleep(2000 * (attempt + 1));
+        }
+    }
+    throw lastErr;
+}
+export function grade(answer, mustInclude) {
+    const a = answer.toLowerCase();
+    return mustInclude.every((t) => a.includes(t));
+}
+async function main() {
+    const provider = process.env.GG_SS_PROVIDER ?? "openai";
+    const model = process.env.GG_SS_MODEL ?? "gpt-5.5";
+    const topK = Math.max(1, parseInt(process.env.GG_SS_TOPK ?? "3", 10));
+    const auth = new AuthStorage();
+    await auth.load();
+    const cr = await auth.resolveCredentials(provider);
+    const creds = { apiKey: cr.accessToken, baseUrl: cr.baseUrl, accountId: cr.accountId };
+    console.log(`\n🔎 Semantic-search benchmark — ${provider}/${model} (top-${topK})\n`);
+    const rows = [];
+    for (const q of QUESTIONS) {
+        // Load the real corpus files; skip the question if any are missing/moved.
+        const files = new Map();
+        let missing = false;
+        for (const rel of q.files) {
+            const abs = path.join(SRC, rel);
+            if (!fs.existsSync(abs)) {
+                console.log(`   ⚠ skipping question — missing ${rel}`);
+                missing = true;
+                break;
+            }
+            files.set(rel, fs.readFileSync(abs, "utf-8"));
+        }
+        if (missing)
+            continue;
+        // BASELINE context: full text of the top-k files a grep would surface.
+        const baseFiles = rankFiles(q.q, files, topK);
+        const baseContext = baseFiles.map((f) => `// FILE: ${f}\n${files.get(f)}`).join("\n\n");
+        // SEMANTIC context: top-k AST chunks across all corpus files.
+        const allChunks = [...files].flatMap(([rel, src]) => chunkFile(rel, src));
+        const semChunks = bm25Rank(q.q, allChunks, topK);
+        const semContext = semChunks.map((c) => `// ${c.file} → ${c.symbol}\n${c.text}`).join("\n\n");
+        // ORACLE context: just the labelled answer chunk.
+        const oracleChunk = allChunks.find((c) => c.file === q.oracle.file && c.symbol === q.oracle.symbol);
+        const oracleContext = oracleChunk
+            ? `// ${oracleChunk.file} → ${oracleChunk.symbol}\n${oracleChunk.text}`
+            : semContext;
+        process.stdout.write(`▶ ${q.q.slice(0, 64)}…\n`);
+        await sleep(1200);
+        const base = await ask(provider, model, creds, baseContext, q.q);
+        await sleep(1200);
+        const sem = await ask(provider, model, creds, semContext, q.q);
+        await sleep(1200);
+        const oracle = await ask(provider, model, creds, oracleContext, q.q);
+        const row = {
+            q: q.q.slice(0, 40),
+            baseInTok: base.inputTokens,
+            baseOk: grade(base.text, q.mustInclude),
+            semInTok: sem.inputTokens,
+            semOk: grade(sem.text, q.mustInclude),
+            oracleInTok: oracle.inputTokens,
+            oracleOk: grade(oracle.text, q.mustInclude),
+        };
+        rows.push(row);
+        process.stdout.write(`   baseline ${row.baseInTok} in tok ${row.baseOk ? "OK" : "FAIL"} | ` +
+            `semantic ${row.semInTok} in tok ${row.semOk ? "OK" : "FAIL"} | ` +
+            `oracle ${row.oracleInTok} in tok ${row.oracleOk ? "OK" : "FAIL"}\n\n`);
+    }
+    if (rows.length === 0) {
+        console.log("No questions ran (corpus files not found).");
+        return;
+    }
+    // ── Report ──
+    console.log("══════════════════════ RESULTS ══════════════════════\n");
+    console.log("Question                                 | base in-tok | sem in-tok | oracle | ok b/s/o");
+    for (const r of rows) {
+        console.log(`${r.q.padEnd(40)} | ${String(r.baseInTok).padStart(11)} | ${String(r.semInTok).padStart(10)} | ` +
+            `${String(r.oracleInTok).padStart(6)} | ${r.baseOk ? "1" : "0"}/${r.semOk ? "1" : "0"}/${r.oracleOk ? "1" : "0"}`);
+    }
+    const sum = (f) => rows.reduce((s, r) => s + f(r), 0);
+    const baseIn = sum((r) => r.baseInTok);
+    const semIn = sum((r) => r.semInTok);
+    const oracleIn = sum((r) => r.oracleInTok);
+    console.log(`\nInput tokens to answer: baseline ${baseIn} | semantic ${semIn} ` +
+        `(${(((baseIn - semIn) / baseIn) * 100).toFixed(0)}% fewer) | oracle ${oracleIn} ` +
+        `(${(((baseIn - oracleIn) / baseIn) * 100).toFixed(0)}% fewer)`);
+    console.log(`Correctness: baseline ${rows.filter((r) => r.baseOk).length}/${rows.length} | ` +
+        `semantic ${rows.filter((r) => r.semOk).length}/${rows.length} | ` +
+        `oracle ${rows.filter((r) => r.oracleOk).length}/${rows.length}`);
+    console.log(`\nVerdict: worth building if SEMANTIC keeps correctness ≈ baseline while cutting input tokens. ` +
+        `cocoindex claims ~70% fewer; ORACLE shows the ceiling a better (embedding) retriever could reach.\n`);
+}
+// Run when executed directly (not when imported by tests).
+const isDirectRun = process.argv[1]?.endsWith("semantic-search-benchmark.ts") ||
+    process.argv[1]?.endsWith("semantic-search-benchmark.js") ||
+    process.argv[1]?.endsWith("semantic-search-benchmark");
+if (isDirectRun) {
+    main().catch((err) => {
+        console.error("Benchmark failed:", err);
+        process.exit(1);
+    });
+}
+//# sourceMappingURL=semantic-search-benchmark.js.map

package/dist/core/semantic-search-benchmark.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"semantic-search-benchmark.js","sourceRoot":"","sources":["../../src/core/semantic-search-benchmark.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AAEH,OAAO,EAAE,MAAM,SAAS,CAAC;AACzB,OAAO,IAAI,MAAM,WAAW,CAAC;AAC7B,OAAO,EAAE,aAAa,EAAE,MAAM,UAAU,CAAC;AACzC,OAAO,EAAE,MAAM,EAA8C,MAAM,kBAAkB,CAAC;AACtF,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAChD,OAAO,EAAE,SAAS,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,qBAAqB,CAAC;AAErE,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC;AAC1D,MAAM,GAAG,GAAG,IAAI,CAAC,OAAO,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC,uBAAuB;AAc7D,MAAM,SAAS,GAAe;IAC5B;QACE,CAAC,EAAE,8HAA8H;QACjI,KAAK,EAAE,CAAC,sBAAsB,EAAE,sBAAsB,EAAE,oBAAoB,CAAC;QAC7E,MAAM,EAAE,EAAE,IAAI,EAAE,sBAAsB,EAAE,MAAM,EAAE,oBAAoB,EAAE;QACtE,WAAW,EAAE,CAAC,oBAAoB,EAAE,SAAS,CAAC;KAC/C;IACD;QACE,CAAC,EAAE,2GAA2G;QAC9G,KAAK,EAAE,CAAC,oBAAoB,EAAE,sBAAsB,EAAE,0BAA0B,CAAC;QACjF,MAAM,EAAE,EAAE,IAAI,EAAE,oBAAoB,EAAE,MAAM,EAAE,eAAe,EAAE;QAC/D,WAAW,EAAE,CAAC,eAAe,EAAE,QAAQ,CAAC;KACzC;IACD;QACE,CAAC,EAAE,iEAAiE;QACpE,KAAK,EAAE,CAAC,0BAA0B,EAAE,sBAAsB,EAAE,oBAAoB,CAAC;QACjF,MAAM,EAAE,EAAE,IAAI,EAAE,0BAA0B,EAAE,MAAM,EAAE,aAAa,EAAE;QACnE,WAAW,EAAE,CAAC,MAAM,EAAE,cAAc,EAAE,MAAM,CAAC;KAC9C;CACF,CAAC;AAYF,SAAS,KAAK,CAAC,EAAU;IACvB,OAAO,IAAI,OAAO,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,UAAU,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;AAC/C,CAAC;AAQD,KAAK,UAAU,GAAG,CAChB,QAAgB,EAChB,KAAa,EACb,CAAQ,EACR,OAAe,EACf,QAAgB;IAEhB,MAAM,QAAQ,GAAc;QAC1B;YACE,IAAI,EAAE,QAAQ;YACd,OAAO,EACL,yEAAyE;gBACzE,gGAAgG;gBAChG,0BAA0B;SAC7B;QACD,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,aAAa,OAAO,iBAAiB,QAAQ,EAAE,EAAE;KAC3E,CAAC;IACF,IAAI,OAAgB,CAAC;IACrB,KAAK,IAAI,OAAO,GAAG,CAAC,EAAE,OAAO,GAAG,CAAC,EAAE,OAAO,EAAE,EAAE,CAAC;QAC7C,IAAI,CAAC;YACH,IAAI,IAAI,GAAG,EAAE,CAAC;YACd,MAAM,MAAM,GAAG,MAAM,CAAC;gBACpB,QAAQ,EAAE,QAAiB;gBAC3B,KAAK;gBACL,QAAQ;gBACR,SAAS,EAAE,GAAG;gBACd,MAAM,EAAE,CAAC,CAAC,MAAM;gBAChB,OAAO,EAAE,CAAC,CAAC,OAAO;gBAClB,SAAS,EAAE,CAAC,CAAC,SAAS;aACvB,CAAC,CAAC;YACH,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,MAAoC,EAAE,CAAC;gBAC/D,IAAI,KAAK,CAAC,IAAI,KAAK,YAAY;oBAAE,IAAI,IAAI,KAAK,CAAC,IAAI,CAAC;YACtD,CAAC;YACD,MAAM,QAAQ,GAAuC,MAAM,MAAM,CAAC,QAAQ,CAAC;YAC3E,MAAM,OAAO,GAAG,QAAQ,CAAC,OAAO,CAAC,OAAO,CAAC;YACzC,MAAM,SAAS,GACb,OAAO,OAAO,KAAK,QAAQ;gBACzB,CAAC,CAAC,OAAO;gBACT,CAAC,CAAE,OAAkD;qBAChD,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC;qBAChC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,IAAI,EAAE,CAAC;qBACxB,IAAI,CAAC,EAAE,CAAC,CAAC;YAClB,OAAO;gBACL,IAAI,EAAE,SAAS,IAAI,IAAI;gBACvB,WAAW,EAAE,QAAQ,CAAC,KAAK,CAAC,WAAW;gBACvC,YAAY,EAAE,QAAQ,CAAC,KAAK,CAAC,YAAY;aAC1C,CAAC;QACJ,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,OAAO,GAAG,GAAG,CAAC;YACd,MAAM,KAAK,CAAC,IAAI,GAAG,CAAC,OAAO,GAAG,CAAC,CAAC,CAAC,CAAC;QACpC,CAAC;IACH,CAAC;IACD,MAAM,OAAO,CAAC;AAChB,CAAC;AAED,MAAM,UAAU,KAAK,CAAC,MAAc,EAAE,WAAqB;IACzD,MAAM,CAAC,GAAG,MAAM,CAAC,WAAW,EAAE,CAAC;IAC/B,OAAO,WAAW,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;AACjD,CAAC;AAcD,KAAK,UAAU,IAAI;IACjB,MAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,CAAC,cAAc,IAAI,QAAQ,CAAC;IACxD,MAAM,KAAK,GAAG,OAAO,CAAC,GAAG,CAAC,WAAW,IAAI,SAAS,CAAC;IACnD,MAAM,IAAI,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,QAAQ,CAAC,OAAO,CAAC,GAAG,CAAC,UAAU,IAAI,GAAG,EAAE,EAAE,CAAC,CAAC,CAAC;IAEtE,MAAM,IAAI,GAAG,IAAI,WAAW,EAAE,CAAC;IAC/B,MAAM,IAAI,CAAC,IAAI,EAAE,CAAC;IAClB,MAAM,EAAE,GAAG,MAAM,IAAI,CAAC,kBAAkB,CAAC,QAAQ,CAAC,CAAC;IACnD,MAAM,KAAK,GAAU,EAAE,MAAM,EAAE,EAAE,CAAC,WAAW,EAAE,OAAO,EAAE,EAAE,CAAC,OAAO,EAAE,SAAS,EAAE,EAAE,CAAC,SAAS,EAAE,CAAC;IAE9F,OAAO,CAAC,GAAG,CAAC,oCAAoC,QAAQ,IAAI,KAAK,SAAS,IAAI,KAAK,CAAC,CAAC;IAErF,MAAM,IAAI,GAAU,EAAE,CAAC;IAEvB,KAAK,MAAM,CAAC,IAAI,SAAS,EAAE,CAAC;QAC1B,0EAA0E;QAC1E,MAAM,KAAK,GAAG,IAAI,GAAG,EAAkB,CAAC;QACxC,IAAI,OAAO,GAAG,KAAK,CAAC;QACpB,KAAK,MAAM,GAAG,IAAI,CAAC,CAAC,KAAK,EAAE,CAAC;YAC1B,MAAM,GAAG,GAAG,IAAI,CAAC,IAAI,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC;YAChC,IAAI,CAAC,EAAE,CAAC,UAAU,CAAC,GAAG,CAAC,EAAE,CAAC;gBACxB,OAAO,CAAC,GAAG,CAAC,oCAAoC,GAAG,EAAE,CAAC,CAAC;gBACvD,OAAO,GAAG,IAAI,CAAC;gBACf,MAAM;YACR,CAAC;YACD,KAAK,CAAC,GAAG,CAAC,GAAG,EAAE,EAAE,CAAC,YAAY,CAAC,GAAG,EAAE,OAAO,CAAC,CAAC,CAAC;QAChD,CAAC;QACD,IAAI,OAAO;YAAE,SAAS;QAEtB,uEAAuE;QACvE,MAAM,SAAS,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,IAAI,CAAC,CAAC;QAC9C,MAAM,WAAW,GAAG,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,YAAY,CAAC,KAAK,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAExF,8DAA8D;QAC9D,MAAM,SAAS,GAAG,CAAC,GAAG,KAAK,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,EAAE,GAAG,CAAC,EAAE,EAAE,CAAC,SAAS,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC,CAAC;QAC1E,MAAM,SAAS,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,CAAC;QACjD,MAAM,UAAU,GAAG,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,MAAM,CAAC,CAAC,IAAI,MAAM,CAAC,CAAC,MAAM,KAAK,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAE9F,kDAAkD;QAClD,MAAM,WAAW,GAAG,SAAS,CAAC,IAAI,CAChC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,MAAM,CAAC,IAAI,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC,CAAC,MAAM,CAAC,MAAM,CAChE,CAAC;QACF,MAAM,aAAa,GAAG,WAAW;YAC/B,CAAC,CAAC,MAAM,WAAW,CAAC,IAAI,MAAM,WAAW,CAAC,MAAM,KAAK,WAAW,CAAC,IAAI,EAAE;YACvE,CAAC,CAAC,UAAU,CAAC;QAEf,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,KAAK,CAAC,CAAC;QACjD,MAAM,KAAK,CAAC,IAAI,CAAC,CAAC;QAClB,MAAM,IAAI,GAAG,MAAM,GAAG,CAAC,QAAQ,EAAE,KAAK,EAAE,KAAK,EAAE,WAAW,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QACjE,MAAM,KAAK,CAAC,IAAI,CAAC,CAAC;QAClB,MAAM,GAAG,GAAG,MAAM,GAAG,CAAC,QAAQ,EAAE,KAAK,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QAC/D,MAAM,KAAK,CAAC,IAAI,CAAC,CAAC;QAClB,MAAM,MAAM,GAAG,MAAM,GAAG,CAAC,QAAQ,EAAE,KAAK,EAAE,KAAK,EAAE,aAAa,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QAErE,MAAM,GAAG,GAAQ;YACf,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC;YACnB,SAAS,EAAE,IAAI,CAAC,WAAW;YAC3B,MAAM,EAAE,KAAK,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,WAAW,CAAC;YACvC,QAAQ,EAAE,GAAG,CAAC,WAAW;YACzB,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC,CAAC,WAAW,CAAC;YACrC,WAAW,EAAE,MAAM,CAAC,WAAW;YAC/B,QAAQ,EAAE,KAAK,CAAC,MAAM,CAAC,IAAI,EAAE,CAAC,CAAC,WAAW,CAAC;SAC5C,CAAC;QACF,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QACf,OAAO,CAAC,MAAM,CAAC,KAAK,CAClB,eAAe,GAAG,CAAC,SAAS,WAAW,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,KAAK;YACpE,YAAY,GAAG,CAAC,QAAQ,WAAW,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,KAAK;YACjE,UAAU,GAAG,CAAC,WAAW,WAAW,GAAG,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,MAAM,CACzE,CAAC;IACJ,CAAC;IAED,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACtB,OAAO,CAAC,GAAG,CAAC,4CAA4C,CAAC,CAAC;QAC1D,OAAO;IACT,CAAC;IAED,eAAe;IACf,OAAO,CAAC,GAAG,CAAC,yDAAyD,CAAC,CAAC;IACvE,OAAO,CAAC,GAAG,CACT,yFAAyF,CAC1F,CAAC;IACF,KAAK,MAAM,CAAC,IAAI,IAAI,EAAE,CAAC;QACrB,OAAO,CAAC,GAAG,CACT,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,QAAQ,CAAC,EAAE,CAAC,MAAM,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,QAAQ,CAAC,EAAE,CAAC,KAAK;YAC/F,GAAG,MAAM,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,EAAE,CACpH,CAAC;IACJ,CAAC;IACD,MAAM,GAAG,GAAG,CAAC,CAAqB,EAAE,EAAE,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IAC1E,MAAM,MAAM,GAAG,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC;IACvC,MAAM,KAAK,GAAG,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC;IACrC,MAAM,QAAQ,GAAG,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC;IAC3C,OAAO,CAAC,GAAG,CACT,sCAAsC,MAAM,eAAe,KAAK,GAAG;QACjE,IAAI,CAAC,CAAC,CAAC,MAAM,GAAG,KAAK,CAAC,GAAG,MAAM,CAAC,GAAG,GAAG,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,qBAAqB,QAAQ,GAAG;QAClF,IAAI,CAAC,CAAC,CAAC,MAAM,GAAG,QAAQ,CAAC,GAAG,MAAM,CAAC,GAAG,GAAG,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,UAAU,CAClE,CAAC;IACF,OAAO,CAAC,GAAG,CACT,yBAAyB,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,MAAM,IAAI,IAAI,CAAC,MAAM,KAAK;QAC9E,YAAY,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,MAAM,IAAI,IAAI,CAAC,MAAM,KAAK;QAClE,UAAU,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,MAAM,IAAI,IAAI,CAAC,MAAM,EAAE,CACnE,CAAC;IACF,OAAO,CAAC,GAAG,CACT,iGAAiG;QAC/F,qGAAqG,CACxG,CAAC;AACJ,CAAC;AAED,2DAA2D;AAC3D,MAAM,WAAW,GACf,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,QAAQ,CAAC,8BAA8B,CAAC;IACzD,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,QAAQ,CAAC,8BAA8B,CAAC;IACzD,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,QAAQ,CAAC,2BAA2B,CAAC,CAAC;AAEzD,IAAI,WAAW,EAAE,CAAC;IAChB,IAAI,EAAE,CAAC,KAAK,CAAC,CAAC,GAAG,EAAE,EAAE;QACnB,OAAO,CAAC,KAAK,CAAC,mBAAmB,EAAE,GAAG,CAAC,CAAC;QACxC,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAClB,CAAC,CAAC,CAAC;AACL,CAAC"}

package/dist/core/semantic-search-benchmark.test.d.ts

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=semantic-search-benchmark.test.d.ts.map

package/dist/core/semantic-search-benchmark.test.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"semantic-search-benchmark.test.d.ts","sourceRoot":"","sources":["../../src/core/semantic-search-benchmark.test.ts"],"names":[],"mappings":""}