@kolisachint/hoocode-agent 0.4.164 → 0.4.165

package/dist/core/search/rrf.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"rrf.d.ts","sourceRoot":"","sources":["../../../src/core/search/rrf.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,KAAK,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,YAAY,CAAC;AAEtD;;;;;;GAMG;AACH,eAAO,MAAM,aAAa,IAAI,CAAC;AAE/B;;;;;;;;;GASG;AACH,wBAAgB,OAAO,CAAC,KAAK,EAAE,SAAS,CAAC,SAAS,SAAS,EAAE,CAAC,EAAE,EAAE,CAAC,SAAgB,GAAG,QAAQ,EAAE,CA6C/F","sourcesContent":["/**\n * Reciprocal Rank Fusion — the transparent, rank-only recall layer of hybrid\n * search (docs/hybrid-retrieval-design.md, Decision 2).\n *\n * Rank-only on purpose: BM25 and cosine scores are not comparable across\n * retrievers, so raw scores are carried through as diagnostics but never\n * enter the fused score.\n */\n\nimport type { FusedHit, RankedHit } from \"./types.js\";\n\n/**\n * Default RRF constant. The literature folklore default is 60; the eval gate\n * (scripts/search-eval.mjs, 12-query gold set) measured k ∈ {0, 2} beating\n * k = 60 on every differing query, twice, with reranking on top of either —\n * small k keeps fusion top-heavy toward each retriever's best hits, and the\n * reranker corrects the tail. Small sample: re-sweep when the gold set grows.\n */\nexport const DEFAULT_RRF_K = 2;\n\n/**\n * Fuse ranked lists into one deterministic ordering by summed `1/(k + rank)`.\n *\n * Ties break by number of agreeing retrievers, then lexicographic id, so the\n * same inputs always produce the same context.\n *\n * Duplicate `source:id` pairs within one list are counted once (best rank\n * wins). The adapter dedupes upstream, so this guard should never fire in\n * practice — it exists so a misbehaving retriever cannot inflate its vote.\n */\nexport function rrfFuse(lists: readonly (readonly RankedHit[])[], k = DEFAULT_RRF_K): FusedHit[] {\n\tif (!Number.isFinite(k) || k < 0) {\n\t\tthrow new Error(`RRF k must be a finite non-negative number; got ${k}`);\n\t}\n\n\tconst acc = new Map<string, FusedHit>();\n\n\tfor (const list of lists) {\n\t\t// Collapse duplicates to their best rank first, so the single vote a\n\t\t// duplicated id gets is cast at the best rank regardless of emit order.\n\t\tconst collapsed = new Map<string, RankedHit>();\n\t\tfor (const hit of list) {\n\t\t\tif (!Number.isInteger(hit.rank) || hit.rank < 1) {\n\t\t\t\tthrow new Error(`RRF rank must be a positive integer; got ${hit.rank}`);\n\t\t\t}\n\t\t\tconst dedupeKey = `${hit.source}:${hit.id}`;\n\t\t\tconst existing = collapsed.get(dedupeKey);\n\t\t\tif (!existing || hit.rank < existing.rank) collapsed.set(dedupeKey, hit);\n\t\t}\n\n\t\tfor (const hit of collapsed.values()) {\n\t\t\tlet current = acc.get(hit.id);\n\t\t\tif (!current) {\n\t\t\t\tcurrent = { id: hit.id, rrfScore: 0, ranks: {}, rawScores: {} };\n\t\t\t\tacc.set(hit.id, current);\n\t\t\t}\n\n\t\t\tcurrent.rrfScore += 1 / (k + hit.rank);\n\n\t\t\tconst oldRank = current.ranks[hit.source];\n\t\t\tif (oldRank === undefined || hit.rank < oldRank) {\n\t\t\t\tcurrent.ranks[hit.source] = hit.rank;\n\t\t\t\t// rawScores follows the best rank; a best-ranked hit without a\n\t\t\t\t// score leaves any earlier score in place rather than erasing it.\n\t\t\t\tif (hit.score !== undefined) current.rawScores[hit.source] = hit.score;\n\t\t\t}\n\t\t}\n\t}\n\n\treturn [...acc.values()].sort(\n\t\t(a, b) =>\n\t\t\tb.rrfScore - a.rrfScore ||\n\t\t\tObject.keys(b.ranks).length - Object.keys(a.ranks).length ||\n\t\t\ta.id.localeCompare(b.id),\n\t);\n}\n"]}
+{"version":3,"file":"rrf.d.ts","sourceRoot":"","sources":["../../../src/core/search/rrf.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,KAAK,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,YAAY,CAAC;AAEtD;;;;;;;;;;;;;;;;;;;GAmBG;AACH,eAAO,MAAM,aAAa,KAAK,CAAC;AAEhC;;;;;;;;;GASG;AACH,wBAAgB,OAAO,CAAC,KAAK,EAAE,SAAS,CAAC,SAAS,SAAS,EAAE,CAAC,EAAE,EAAE,CAAC,SAAgB,GAAG,QAAQ,EAAE,CA6C/F","sourcesContent":["/**\n * Reciprocal Rank Fusion — the transparent, rank-only recall layer of hybrid\n * search (docs/hybrid-retrieval-design.md, Decision 2).\n *\n * Rank-only on purpose: BM25 and cosine scores are not comparable across\n * retrievers, so raw scores are carried through as diagnostics but never\n * enter the fused score.\n */\n\nimport type { FusedHit, RankedHit } from \"./types.js\";\n\n/**\n * Default RRF constant, re-swept on the 62-query harness.\n *\n * History matters here, because the answer changed twice. The original\n * 12-query gold set picked k = 2 over k = 60; that measurement ran on an\n * unpinned corpus with a file-level metric and was never reproducible. The\n * 62-query re-sweep then found the two *indistinguishable* — a 3pp R@10 gap\n * carried by two queries (p = 0.50), with MRR worse on more queries than it\n * was better.\n *\n * What made k decidable was fixing the reranker. Once it could tell a\n * declaration from a call site, the deeper, flatter candidate mix that k = 60\n * produces became worth having: MRR 0.403 -> 0.464, 20 queries better against\n * 7 worse (p <= 0.05), with R@10 unchanged. Small k keeps fusion top-heavy\n * toward each retriever's best hits, which only pays when the reranker cannot\n * exploit the tail — and it now can.\n *\n * Re-sweep again (`bun run search-eval:compare`) after any reranker change,\n * since that is what this value trades against.\n */\nexport const DEFAULT_RRF_K = 60;\n\n/**\n * Fuse ranked lists into one deterministic ordering by summed `1/(k + rank)`.\n *\n * Ties break by number of agreeing retrievers, then lexicographic id, so the\n * same inputs always produce the same context.\n *\n * Duplicate `source:id` pairs within one list are counted once (best rank\n * wins). The adapter dedupes upstream, so this guard should never fire in\n * practice — it exists so a misbehaving retriever cannot inflate its vote.\n */\nexport function rrfFuse(lists: readonly (readonly RankedHit[])[], k = DEFAULT_RRF_K): FusedHit[] {\n\tif (!Number.isFinite(k) || k < 0) {\n\t\tthrow new Error(`RRF k must be a finite non-negative number; got ${k}`);\n\t}\n\n\tconst acc = new Map<string, FusedHit>();\n\n\tfor (const list of lists) {\n\t\t// Collapse duplicates to their best rank first, so the single vote a\n\t\t// duplicated id gets is cast at the best rank regardless of emit order.\n\t\tconst collapsed = new Map<string, RankedHit>();\n\t\tfor (const hit of list) {\n\t\t\tif (!Number.isInteger(hit.rank) || hit.rank < 1) {\n\t\t\t\tthrow new Error(`RRF rank must be a positive integer; got ${hit.rank}`);\n\t\t\t}\n\t\t\tconst dedupeKey = `${hit.source}:${hit.id}`;\n\t\t\tconst existing = collapsed.get(dedupeKey);\n\t\t\tif (!existing || hit.rank < existing.rank) collapsed.set(dedupeKey, hit);\n\t\t}\n\n\t\tfor (const hit of collapsed.values()) {\n\t\t\tlet current = acc.get(hit.id);\n\t\t\tif (!current) {\n\t\t\t\tcurrent = { id: hit.id, rrfScore: 0, ranks: {}, rawScores: {} };\n\t\t\t\tacc.set(hit.id, current);\n\t\t\t}\n\n\t\t\tcurrent.rrfScore += 1 / (k + hit.rank);\n\n\t\t\tconst oldRank = current.ranks[hit.source];\n\t\t\tif (oldRank === undefined || hit.rank < oldRank) {\n\t\t\t\tcurrent.ranks[hit.source] = hit.rank;\n\t\t\t\t// rawScores follows the best rank; a best-ranked hit without a\n\t\t\t\t// score leaves any earlier score in place rather than erasing it.\n\t\t\t\tif (hit.score !== undefined) current.rawScores[hit.source] = hit.score;\n\t\t\t}\n\t\t}\n\t}\n\n\treturn [...acc.values()].sort(\n\t\t(a, b) =>\n\t\t\tb.rrfScore - a.rrfScore ||\n\t\t\tObject.keys(b.ranks).length - Object.keys(a.ranks).length ||\n\t\t\ta.id.localeCompare(b.id),\n\t);\n}\n"]}

package/dist/core/search/rrf.js

@@ -7,13 +7,26 @@
  * enter the fused score.
  */
 /**
- * Default RRF constant. The literature folklore default is 60; the eval gate
- * (scripts/search-eval.mjs, 12-query gold set) measured k ∈ {0, 2} beating
- * k = 60 on every differing query, twice, with reranking on top of either —
- * small k keeps fusion top-heavy toward each retriever's best hits, and the
- * reranker corrects the tail. Small sample: re-sweep when the gold set grows.
+ * Default RRF constant, re-swept on the 62-query harness.
+ *
+ * History matters here, because the answer changed twice. The original
+ * 12-query gold set picked k = 2 over k = 60; that measurement ran on an
+ * unpinned corpus with a file-level metric and was never reproducible. The
+ * 62-query re-sweep then found the two *indistinguishable* — a 3pp R@10 gap
+ * carried by two queries (p = 0.50), with MRR worse on more queries than it
+ * was better.
+ *
+ * What made k decidable was fixing the reranker. Once it could tell a
+ * declaration from a call site, the deeper, flatter candidate mix that k = 60
+ * produces became worth having: MRR 0.403 -> 0.464, 20 queries better against
+ * 7 worse (p <= 0.05), with R@10 unchanged. Small k keeps fusion top-heavy
+ * toward each retriever's best hits, which only pays when the reranker cannot
+ * exploit the tail — and it now can.
+ *
+ * Re-sweep again (`bun run search-eval:compare`) after any reranker change,
+ * since that is what this value trades against.
  */
-export const DEFAULT_RRF_K = 2;
+export const DEFAULT_RRF_K = 60;
 /**
  * Fuse ranked lists into one deterministic ordering by summed `1/(k + rank)`.
  *

package/dist/core/search/rrf.js.map

@@ -1 +1 @@
-{"version":3,"file":"rrf.js","sourceRoot":"","sources":["../../../src/core/search/rrf.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,aAAa,GAAG,CAAC,CAAC;AAE/B;;;;;;;;;GASG;AACH,MAAM,UAAU,OAAO,CAAC,KAAwC,EAAE,CAAC,GAAG,aAAa,EAAc;IAChG,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;QAClC,MAAM,IAAI,KAAK,CAAC,mDAAmD,CAAC,EAAE,CAAC,CAAC;IACzE,CAAC;IAED,MAAM,GAAG,GAAG,IAAI,GAAG,EAAoB,CAAC;IAExC,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QAC1B,qEAAqE;QACrE,wEAAwE;QACxE,MAAM,SAAS,GAAG,IAAI,GAAG,EAAqB,CAAC;QAC/C,KAAK,MAAM,GAAG,IAAI,IAAI,EAAE,CAAC;YACxB,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,IAAI,GAAG,CAAC,EAAE,CAAC;gBACjD,MAAM,IAAI,KAAK,CAAC,4CAA4C,GAAG,CAAC,IAAI,EAAE,CAAC,CAAC;YACzE,CAAC;YACD,MAAM,SAAS,GAAG,GAAG,GAAG,CAAC,MAAM,IAAI,GAAG,CAAC,EAAE,EAAE,CAAC;YAC5C,MAAM,QAAQ,GAAG,SAAS,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;YAC1C,IAAI,CAAC,QAAQ,IAAI,GAAG,CAAC,IAAI,GAAG,QAAQ,CAAC,IAAI;gBAAE,SAAS,CAAC,GAAG,CAAC,SAAS,EAAE,GAAG,CAAC,CAAC;QAC1E,CAAC;QAED,KAAK,MAAM,GAAG,IAAI,SAAS,CAAC,MAAM,EAAE,EAAE,CAAC;YACtC,IAAI,OAAO,GAAG,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YAC9B,IAAI,CAAC,OAAO,EAAE,CAAC;gBACd,OAAO,GAAG,EAAE,EAAE,EAAE,GAAG,CAAC,EAAE,EAAE,QAAQ,EAAE,CAAC,EAAE,KAAK,EAAE,EAAE,EAAE,SAAS,EAAE,EAAE,EAAE,CAAC;gBAChE,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,EAAE,OAAO,CAAC,CAAC;YAC1B,CAAC;YAED,OAAO,CAAC,QAAQ,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,IAAI,CAAC,CAAC;YAEvC,MAAM,OAAO,GAAG,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;YAC1C,IAAI,OAAO,KAAK,SAAS,IAAI,GAAG,CAAC,IAAI,GAAG,OAAO,EAAE,CAAC;gBACjD,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,MAAM,CAAC,GAAG,GAAG,CAAC,IAAI,CAAC;gBACrC,+DAA+D;gBAC/D,kEAAkE;gBAClE,IAAI,GAAG,CAAC,KAAK,KAAK,SAAS;oBAAE,OAAO,CAAC,SAAS,CAAC,GAAG,CAAC,MAAM,CAAC,GAAG,GAAG,CAAC,KAAK,CAAC;YACxE,CAAC;QACF,CAAC;IACF,CAAC;IAED,OAAO,CAAC,GAAG,GAAG,CAAC,MAAM,EAAE,CAAC,CAAC,IAAI,CAC5B,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CACR,CAAC,CAAC,QAAQ,GAAG,CAAC,CAAC,QAAQ;QACvB,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,MAAM;QACzD,CAAC,CAAC,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,CACzB,CAAC;AAAA,CACF","sourcesContent":["/**\n * Reciprocal Rank Fusion — the transparent, rank-only recall layer of hybrid\n * search (docs/hybrid-retrieval-design.md, Decision 2).\n *\n * Rank-only on purpose: BM25 and cosine scores are not comparable across\n * retrievers, so raw scores are carried through as diagnostics but never\n * enter the fused score.\n */\n\nimport type { FusedHit, RankedHit } from \"./types.js\";\n\n/**\n * Default RRF constant. The literature folklore default is 60; the eval gate\n * (scripts/search-eval.mjs, 12-query gold set) measured k ∈ {0, 2} beating\n * k = 60 on every differing query, twice, with reranking on top of either —\n * small k keeps fusion top-heavy toward each retriever's best hits, and the\n * reranker corrects the tail. Small sample: re-sweep when the gold set grows.\n */\nexport const DEFAULT_RRF_K = 2;\n\n/**\n * Fuse ranked lists into one deterministic ordering by summed `1/(k + rank)`.\n *\n * Ties break by number of agreeing retrievers, then lexicographic id, so the\n * same inputs always produce the same context.\n *\n * Duplicate `source:id` pairs within one list are counted once (best rank\n * wins). The adapter dedupes upstream, so this guard should never fire in\n * practice — it exists so a misbehaving retriever cannot inflate its vote.\n */\nexport function rrfFuse(lists: readonly (readonly RankedHit[])[], k = DEFAULT_RRF_K): FusedHit[] {\n\tif (!Number.isFinite(k) || k < 0) {\n\t\tthrow new Error(`RRF k must be a finite non-negative number; got ${k}`);\n\t}\n\n\tconst acc = new Map<string, FusedHit>();\n\n\tfor (const list of lists) {\n\t\t// Collapse duplicates to their best rank first, so the single vote a\n\t\t// duplicated id gets is cast at the best rank regardless of emit order.\n\t\tconst collapsed = new Map<string, RankedHit>();\n\t\tfor (const hit of list) {\n\t\t\tif (!Number.isInteger(hit.rank) || hit.rank < 1) {\n\t\t\t\tthrow new Error(`RRF rank must be a positive integer; got ${hit.rank}`);\n\t\t\t}\n\t\t\tconst dedupeKey = `${hit.source}:${hit.id}`;\n\t\t\tconst existing = collapsed.get(dedupeKey);\n\t\t\tif (!existing || hit.rank < existing.rank) collapsed.set(dedupeKey, hit);\n\t\t}\n\n\t\tfor (const hit of collapsed.values()) {\n\t\t\tlet current = acc.get(hit.id);\n\t\t\tif (!current) {\n\t\t\t\tcurrent = { id: hit.id, rrfScore: 0, ranks: {}, rawScores: {} };\n\t\t\t\tacc.set(hit.id, current);\n\t\t\t}\n\n\t\t\tcurrent.rrfScore += 1 / (k + hit.rank);\n\n\t\t\tconst oldRank = current.ranks[hit.source];\n\t\t\tif (oldRank === undefined || hit.rank < oldRank) {\n\t\t\t\tcurrent.ranks[hit.source] = hit.rank;\n\t\t\t\t// rawScores follows the best rank; a best-ranked hit without a\n\t\t\t\t// score leaves any earlier score in place rather than erasing it.\n\t\t\t\tif (hit.score !== undefined) current.rawScores[hit.source] = hit.score;\n\t\t\t}\n\t\t}\n\t}\n\n\treturn [...acc.values()].sort(\n\t\t(a, b) =>\n\t\t\tb.rrfScore - a.rrfScore ||\n\t\t\tObject.keys(b.ranks).length - Object.keys(a.ranks).length ||\n\t\t\ta.id.localeCompare(b.id),\n\t);\n}\n"]}
+{"version":3,"file":"rrf.js","sourceRoot":"","sources":["../../../src/core/search/rrf.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH;;;;;;;;;;;;;;;;;;;GAmBG;AACH,MAAM,CAAC,MAAM,aAAa,GAAG,EAAE,CAAC;AAEhC;;;;;;;;;GASG;AACH,MAAM,UAAU,OAAO,CAAC,KAAwC,EAAE,CAAC,GAAG,aAAa,EAAc;IAChG,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;QAClC,MAAM,IAAI,KAAK,CAAC,mDAAmD,CAAC,EAAE,CAAC,CAAC;IACzE,CAAC;IAED,MAAM,GAAG,GAAG,IAAI,GAAG,EAAoB,CAAC;IAExC,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QAC1B,qEAAqE;QACrE,wEAAwE;QACxE,MAAM,SAAS,GAAG,IAAI,GAAG,EAAqB,CAAC;QAC/C,KAAK,MAAM,GAAG,IAAI,IAAI,EAAE,CAAC;YACxB,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,IAAI,GAAG,CAAC,EAAE,CAAC;gBACjD,MAAM,IAAI,KAAK,CAAC,4CAA4C,GAAG,CAAC,IAAI,EAAE,CAAC,CAAC;YACzE,CAAC;YACD,MAAM,SAAS,GAAG,GAAG,GAAG,CAAC,MAAM,IAAI,GAAG,CAAC,EAAE,EAAE,CAAC;YAC5C,MAAM,QAAQ,GAAG,SAAS,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;YAC1C,IAAI,CAAC,QAAQ,IAAI,GAAG,CAAC,IAAI,GAAG,QAAQ,CAAC,IAAI;gBAAE,SAAS,CAAC,GAAG,CAAC,SAAS,EAAE,GAAG,CAAC,CAAC;QAC1E,CAAC;QAED,KAAK,MAAM,GAAG,IAAI,SAAS,CAAC,MAAM,EAAE,EAAE,CAAC;YACtC,IAAI,OAAO,GAAG,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YAC9B,IAAI,CAAC,OAAO,EAAE,CAAC;gBACd,OAAO,GAAG,EAAE,EAAE,EAAE,GAAG,CAAC,EAAE,EAAE,QAAQ,EAAE,CAAC,EAAE,KAAK,EAAE,EAAE,EAAE,SAAS,EAAE,EAAE,EAAE,CAAC;gBAChE,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,EAAE,OAAO,CAAC,CAAC;YAC1B,CAAC;YAED,OAAO,CAAC,QAAQ,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,IAAI,CAAC,CAAC;YAEvC,MAAM,OAAO,GAAG,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;YAC1C,IAAI,OAAO,KAAK,SAAS,IAAI,GAAG,CAAC,IAAI,GAAG,OAAO,EAAE,CAAC;gBACjD,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,MAAM,CAAC,GAAG,GAAG,CAAC,IAAI,CAAC;gBACrC,+DAA+D;gBAC/D,kEAAkE;gBAClE,IAAI,GAAG,CAAC,KAAK,KAAK,SAAS;oBAAE,OAAO,CAAC,SAAS,CAAC,GAAG,CAAC,MAAM,CAAC,GAAG,GAAG,CAAC,KAAK,CAAC;YACxE,CAAC;QACF,CAAC;IACF,CAAC;IAED,OAAO,CAAC,GAAG,GAAG,CAAC,MAAM,EAAE,CAAC,CAAC,IAAI,CAC5B,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CACR,CAAC,CAAC,QAAQ,GAAG,CAAC,CAAC,QAAQ;QACvB,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,MAAM;QACzD,CAAC,CAAC,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,CACzB,CAAC;AAAA,CACF","sourcesContent":["/**\n * Reciprocal Rank Fusion — the transparent, rank-only recall layer of hybrid\n * search (docs/hybrid-retrieval-design.md, Decision 2).\n *\n * Rank-only on purpose: BM25 and cosine scores are not comparable across\n * retrievers, so raw scores are carried through as diagnostics but never\n * enter the fused score.\n */\n\nimport type { FusedHit, RankedHit } from \"./types.js\";\n\n/**\n * Default RRF constant, re-swept on the 62-query harness.\n *\n * History matters here, because the answer changed twice. The original\n * 12-query gold set picked k = 2 over k = 60; that measurement ran on an\n * unpinned corpus with a file-level metric and was never reproducible. The\n * 62-query re-sweep then found the two *indistinguishable* — a 3pp R@10 gap\n * carried by two queries (p = 0.50), with MRR worse on more queries than it\n * was better.\n *\n * What made k decidable was fixing the reranker. Once it could tell a\n * declaration from a call site, the deeper, flatter candidate mix that k = 60\n * produces became worth having: MRR 0.403 -> 0.464, 20 queries better against\n * 7 worse (p <= 0.05), with R@10 unchanged. Small k keeps fusion top-heavy\n * toward each retriever's best hits, which only pays when the reranker cannot\n * exploit the tail — and it now can.\n *\n * Re-sweep again (`bun run search-eval:compare`) after any reranker change,\n * since that is what this value trades against.\n */\nexport const DEFAULT_RRF_K = 60;\n\n/**\n * Fuse ranked lists into one deterministic ordering by summed `1/(k + rank)`.\n *\n * Ties break by number of agreeing retrievers, then lexicographic id, so the\n * same inputs always produce the same context.\n *\n * Duplicate `source:id` pairs within one list are counted once (best rank\n * wins). The adapter dedupes upstream, so this guard should never fire in\n * practice — it exists so a misbehaving retriever cannot inflate its vote.\n */\nexport function rrfFuse(lists: readonly (readonly RankedHit[])[], k = DEFAULT_RRF_K): FusedHit[] {\n\tif (!Number.isFinite(k) || k < 0) {\n\t\tthrow new Error(`RRF k must be a finite non-negative number; got ${k}`);\n\t}\n\n\tconst acc = new Map<string, FusedHit>();\n\n\tfor (const list of lists) {\n\t\t// Collapse duplicates to their best rank first, so the single vote a\n\t\t// duplicated id gets is cast at the best rank regardless of emit order.\n\t\tconst collapsed = new Map<string, RankedHit>();\n\t\tfor (const hit of list) {\n\t\t\tif (!Number.isInteger(hit.rank) || hit.rank < 1) {\n\t\t\t\tthrow new Error(`RRF rank must be a positive integer; got ${hit.rank}`);\n\t\t\t}\n\t\t\tconst dedupeKey = `${hit.source}:${hit.id}`;\n\t\t\tconst existing = collapsed.get(dedupeKey);\n\t\t\tif (!existing || hit.rank < existing.rank) collapsed.set(dedupeKey, hit);\n\t\t}\n\n\t\tfor (const hit of collapsed.values()) {\n\t\t\tlet current = acc.get(hit.id);\n\t\t\tif (!current) {\n\t\t\t\tcurrent = { id: hit.id, rrfScore: 0, ranks: {}, rawScores: {} };\n\t\t\t\tacc.set(hit.id, current);\n\t\t\t}\n\n\t\t\tcurrent.rrfScore += 1 / (k + hit.rank);\n\n\t\t\tconst oldRank = current.ranks[hit.source];\n\t\t\tif (oldRank === undefined || hit.rank < oldRank) {\n\t\t\t\tcurrent.ranks[hit.source] = hit.rank;\n\t\t\t\t// rawScores follows the best rank; a best-ranked hit without a\n\t\t\t\t// score leaves any earlier score in place rather than erasing it.\n\t\t\t\tif (hit.score !== undefined) current.rawScores[hit.source] = hit.score;\n\t\t\t}\n\t\t}\n\t}\n\n\treturn [...acc.values()].sort(\n\t\t(a, b) =>\n\t\t\tb.rrfScore - a.rrfScore ||\n\t\t\tObject.keys(b.ranks).length - Object.keys(a.ranks).length ||\n\t\t\ta.id.localeCompare(b.id),\n\t);\n}\n"]}

package/dist/core/search/types.d.ts

@@ -7,7 +7,17 @@
  * query (fusion and span expansion consult the same sidecar snapshot) but not
  * across edits or rebuilds — never persist them as durable references.
  */
-type RetrieverSource = "grep" | "embed";
+/**
+ * A retriever whose ranked list can enter fusion.
+ *
+ * `bm25` is the daemon's Okapi lexical index, fetched as its own leg via the
+ * `retriever: "lexical"` op rather than pre-fused by `query_hybrid`. Keeping
+ * it separate from `grep` matters: they are both "lexical" but they fail
+ * differently — BM25 has IDF and misses identifiers written in another naming
+ * convention, ripgrep has neither IDF nor an index but sees the working tree,
+ * including edits made this session.
+ */
+type RetrieverSource = "grep" | "embed" | "bm25";
 export type SearchMode = "auto" | "lexical" | "semantic" | "hybrid";
 export type ResolvedSearchMode = Exclude<SearchMode, "auto">;
 export interface RankedHit {

package/dist/core/search/types.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../../src/core/search/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,KAAK,eAAe,GAAG,MAAM,GAAG,OAAO,CAAC;AAExC,MAAM,MAAM,UAAU,GAAG,MAAM,GAAG,SAAS,GAAG,UAAU,GAAG,QAAQ,CAAC;AACpE,MAAM,MAAM,kBAAkB,GAAG,OAAO,CAAC,UAAU,EAAE,MAAM,CAAC,CAAC;AAE7D,MAAM,WAAW,SAAS;IACzB,EAAE,EAAE,MAAM,CAAC;IACX,4DAA4D;IAC5D,IAAI,EAAE,MAAM,CAAC;IACb,uFAAmF;IACnF,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,MAAM,EAAE,eAAe,CAAC;CACxB;AAED,MAAM,WAAW,QAAQ;IACxB,EAAE,EAAE,MAAM,CAAC;IACX,QAAQ,EAAE,MAAM,CAAC;IACjB,4CAA4C;IAC5C,KAAK,EAAE,OAAO,CAAC,MAAM,CAAC,eAAe,EAAE,MAAM,CAAC,CAAC,CAAC;IAChD,kEAAkE;IAClE,SAAS,EAAE,OAAO,CAAC,MAAM,CAAC,eAAe,EAAE,MAAM,CAAC,CAAC,CAAC;CACpD;AAED,gFAAgF;AAChF,MAAM,WAAW,aAAa;IAC7B,IAAI,EAAE,MAAM,CAAC;IACb,yBAAyB;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,sEAAsE;IACtE,OAAO,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,cAAe,SAAQ,QAAQ,EAAE,aAAa;CAAG;AAElE;0BAC0B;AAC1B,MAAM,WAAW,WAAW;IAC3B,WAAW,EAAE,MAAM,CAAC;IACpB,KAAK,EAAE,MAAM,CAAC;IACd,aAAa,EAAE,UAAU,CAAC;IAC1B,YAAY,EAAE,kBAAkB,CAAC;IACjC,wEAAwE;IACxE,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,UAAU,EAAE,OAAO,GAAG,UAAU,GAAG,aAAa,CAAC;IACjD,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,UAAU,EAAE,OAAO,CAAC,MAAM,CAAC,eAAe,EAAE;QAAE,SAAS,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC,CAAC;IACtF,KAAK,EAAE,QAAQ,EAAE,CAAC;IAClB,MAAM,CAAC,EAAE;QAAE,OAAO,EAAE,OAAO,CAAC;QAAC,cAAc,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAA;KAAE,CAAC;CACzE","sourcesContent":["/**\n * Shared types for hybrid code retrieval (see docs/hybrid-retrieval-design.md).\n *\n * Identity note: candidate ids are per-index-build chunk ids (`relpath#index`)\n * from the embedding sidecar, or synthetic `relpath#L<line>` fallback ids for\n * grep hits in files the index does not cover. They are stable within one\n * query (fusion and span expansion consult the same sidecar snapshot) but not\n * across edits or rebuilds — never persist them as durable references.\n */\n\ntype RetrieverSource = \"grep\" | \"embed\";\n\nexport type SearchMode = \"auto\" | \"lexical\" | \"semantic\" | \"hybrid\";\nexport type ResolvedSearchMode = Exclude<SearchMode, \"auto\">;\n\nexport interface RankedHit {\n\tid: string;\n\t/** 1-indexed, gap-free rank within its retriever's list. */\n\trank: number;\n\t/** Retriever-local score (BM25-ish, cosine, …). Diagnostics only — never fused. */\n\tscore?: number;\n\tsource: RetrieverSource;\n}\n\nexport interface FusedHit {\n\tid: string;\n\trrfScore: number;\n\t/** Best rank per contributing retriever. */\n\tranks: Partial<Record<RetrieverSource, number>>;\n\t/** Raw score at the best rank per retriever. Diagnostics only. */\n\trawScores: Partial<Record<RetrieverSource, number>>;\n}\n\n/** Line-span identity a candidate id resolves to, for post-fusion expansion. */\nexport interface CandidateSpan {\n\tpath: string;\n\t/** 1-based inclusive. */\n\tstartLine: number;\n\t/** 1-based inclusive. May exceed the file's length; readers clamp. */\n\tendLine: number;\n}\n\nexport interface FusedCandidate extends FusedHit, CandidateSpan {}\n\n/** Per-call diagnostic record, written to the store-dir trace jsonl — never\n *  into model context. */\nexport interface SearchTrace {\n\ttimestampMs: number;\n\tquery: string;\n\trequestedMode: SearchMode;\n\tresolvedMode: ResolvedSearchMode;\n\t/** Set when the resolved mode is a degradation of the requested one. */\n\tdegradedReason?: string;\n\tindexPhase: \"ready\" | \"indexing\" | \"unavailable\";\n\trrfK?: number;\n\tretrievers: Partial<Record<RetrieverSource, { latencyMs: number; hitCount: number }>>;\n\tfused: FusedHit[];\n\trerank?: { applied: boolean; candidateCount: number; latencyMs: number };\n}\n"]}
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../../src/core/search/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH;;;;;;;;;GASG;AACH,KAAK,eAAe,GAAG,MAAM,GAAG,OAAO,GAAG,MAAM,CAAC;AAEjD,MAAM,MAAM,UAAU,GAAG,MAAM,GAAG,SAAS,GAAG,UAAU,GAAG,QAAQ,CAAC;AACpE,MAAM,MAAM,kBAAkB,GAAG,OAAO,CAAC,UAAU,EAAE,MAAM,CAAC,CAAC;AAE7D,MAAM,WAAW,SAAS;IACzB,EAAE,EAAE,MAAM,CAAC;IACX,4DAA4D;IAC5D,IAAI,EAAE,MAAM,CAAC;IACb,uFAAmF;IACnF,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,MAAM,EAAE,eAAe,CAAC;CACxB;AAED,MAAM,WAAW,QAAQ;IACxB,EAAE,EAAE,MAAM,CAAC;IACX,QAAQ,EAAE,MAAM,CAAC;IACjB,4CAA4C;IAC5C,KAAK,EAAE,OAAO,CAAC,MAAM,CAAC,eAAe,EAAE,MAAM,CAAC,CAAC,CAAC;IAChD,kEAAkE;IAClE,SAAS,EAAE,OAAO,CAAC,MAAM,CAAC,eAAe,EAAE,MAAM,CAAC,CAAC,CAAC;CACpD;AAED,gFAAgF;AAChF,MAAM,WAAW,aAAa;IAC7B,IAAI,EAAE,MAAM,CAAC;IACb,yBAAyB;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,sEAAsE;IACtE,OAAO,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,cAAe,SAAQ,QAAQ,EAAE,aAAa;CAAG;AAElE;0BAC0B;AAC1B,MAAM,WAAW,WAAW;IAC3B,WAAW,EAAE,MAAM,CAAC;IACpB,KAAK,EAAE,MAAM,CAAC;IACd,aAAa,EAAE,UAAU,CAAC;IAC1B,YAAY,EAAE,kBAAkB,CAAC;IACjC,wEAAwE;IACxE,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,UAAU,EAAE,OAAO,GAAG,UAAU,GAAG,aAAa,CAAC;IACjD,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,UAAU,EAAE,OAAO,CAAC,MAAM,CAAC,eAAe,EAAE;QAAE,SAAS,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC,CAAC;IACtF,KAAK,EAAE,QAAQ,EAAE,CAAC;IAClB,MAAM,CAAC,EAAE;QAAE,OAAO,EAAE,OAAO,CAAC;QAAC,cAAc,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAA;KAAE,CAAC;CACzE","sourcesContent":["/**\n * Shared types for hybrid code retrieval (see docs/hybrid-retrieval-design.md).\n *\n * Identity note: candidate ids are per-index-build chunk ids (`relpath#index`)\n * from the embedding sidecar, or synthetic `relpath#L<line>` fallback ids for\n * grep hits in files the index does not cover. They are stable within one\n * query (fusion and span expansion consult the same sidecar snapshot) but not\n * across edits or rebuilds — never persist them as durable references.\n */\n\n/**\n * A retriever whose ranked list can enter fusion.\n *\n * `bm25` is the daemon's Okapi lexical index, fetched as its own leg via the\n * `retriever: \"lexical\"` op rather than pre-fused by `query_hybrid`. Keeping\n * it separate from `grep` matters: they are both \"lexical\" but they fail\n * differently — BM25 has IDF and misses identifiers written in another naming\n * convention, ripgrep has neither IDF nor an index but sees the working tree,\n * including edits made this session.\n */\ntype RetrieverSource = \"grep\" | \"embed\" | \"bm25\";\n\nexport type SearchMode = \"auto\" | \"lexical\" | \"semantic\" | \"hybrid\";\nexport type ResolvedSearchMode = Exclude<SearchMode, \"auto\">;\n\nexport interface RankedHit {\n\tid: string;\n\t/** 1-indexed, gap-free rank within its retriever's list. */\n\trank: number;\n\t/** Retriever-local score (BM25-ish, cosine, …). Diagnostics only — never fused. */\n\tscore?: number;\n\tsource: RetrieverSource;\n}\n\nexport interface FusedHit {\n\tid: string;\n\trrfScore: number;\n\t/** Best rank per contributing retriever. */\n\tranks: Partial<Record<RetrieverSource, number>>;\n\t/** Raw score at the best rank per retriever. Diagnostics only. */\n\trawScores: Partial<Record<RetrieverSource, number>>;\n}\n\n/** Line-span identity a candidate id resolves to, for post-fusion expansion. */\nexport interface CandidateSpan {\n\tpath: string;\n\t/** 1-based inclusive. */\n\tstartLine: number;\n\t/** 1-based inclusive. May exceed the file's length; readers clamp. */\n\tendLine: number;\n}\n\nexport interface FusedCandidate extends FusedHit, CandidateSpan {}\n\n/** Per-call diagnostic record, written to the store-dir trace jsonl — never\n *  into model context. */\nexport interface SearchTrace {\n\ttimestampMs: number;\n\tquery: string;\n\trequestedMode: SearchMode;\n\tresolvedMode: ResolvedSearchMode;\n\t/** Set when the resolved mode is a degradation of the requested one. */\n\tdegradedReason?: string;\n\tindexPhase: \"ready\" | \"indexing\" | \"unavailable\";\n\trrfK?: number;\n\tretrievers: Partial<Record<RetrieverSource, { latencyMs: number; hitCount: number }>>;\n\tfused: FusedHit[];\n\trerank?: { applied: boolean; candidateCount: number; latencyMs: number };\n}\n"]}

package/dist/core/search/types.js.map

@@ -1 +1 @@
-{"version":3,"file":"types.js","sourceRoot":"","sources":["../../../src/core/search/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG","sourcesContent":["/**\n * Shared types for hybrid code retrieval (see docs/hybrid-retrieval-design.md).\n *\n * Identity note: candidate ids are per-index-build chunk ids (`relpath#index`)\n * from the embedding sidecar, or synthetic `relpath#L<line>` fallback ids for\n * grep hits in files the index does not cover. They are stable within one\n * query (fusion and span expansion consult the same sidecar snapshot) but not\n * across edits or rebuilds — never persist them as durable references.\n */\n\ntype RetrieverSource = \"grep\" | \"embed\";\n\nexport type SearchMode = \"auto\" | \"lexical\" | \"semantic\" | \"hybrid\";\nexport type ResolvedSearchMode = Exclude<SearchMode, \"auto\">;\n\nexport interface RankedHit {\n\tid: string;\n\t/** 1-indexed, gap-free rank within its retriever's list. */\n\trank: number;\n\t/** Retriever-local score (BM25-ish, cosine, …). Diagnostics only — never fused. */\n\tscore?: number;\n\tsource: RetrieverSource;\n}\n\nexport interface FusedHit {\n\tid: string;\n\trrfScore: number;\n\t/** Best rank per contributing retriever. */\n\tranks: Partial<Record<RetrieverSource, number>>;\n\t/** Raw score at the best rank per retriever. Diagnostics only. */\n\trawScores: Partial<Record<RetrieverSource, number>>;\n}\n\n/** Line-span identity a candidate id resolves to, for post-fusion expansion. */\nexport interface CandidateSpan {\n\tpath: string;\n\t/** 1-based inclusive. */\n\tstartLine: number;\n\t/** 1-based inclusive. May exceed the file's length; readers clamp. */\n\tendLine: number;\n}\n\nexport interface FusedCandidate extends FusedHit, CandidateSpan {}\n\n/** Per-call diagnostic record, written to the store-dir trace jsonl — never\n *  into model context. */\nexport interface SearchTrace {\n\ttimestampMs: number;\n\tquery: string;\n\trequestedMode: SearchMode;\n\tresolvedMode: ResolvedSearchMode;\n\t/** Set when the resolved mode is a degradation of the requested one. */\n\tdegradedReason?: string;\n\tindexPhase: \"ready\" | \"indexing\" | \"unavailable\";\n\trrfK?: number;\n\tretrievers: Partial<Record<RetrieverSource, { latencyMs: number; hitCount: number }>>;\n\tfused: FusedHit[];\n\trerank?: { applied: boolean; candidateCount: number; latencyMs: number };\n}\n"]}
+{"version":3,"file":"types.js","sourceRoot":"","sources":["../../../src/core/search/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG","sourcesContent":["/**\n * Shared types for hybrid code retrieval (see docs/hybrid-retrieval-design.md).\n *\n * Identity note: candidate ids are per-index-build chunk ids (`relpath#index`)\n * from the embedding sidecar, or synthetic `relpath#L<line>` fallback ids for\n * grep hits in files the index does not cover. They are stable within one\n * query (fusion and span expansion consult the same sidecar snapshot) but not\n * across edits or rebuilds — never persist them as durable references.\n */\n\n/**\n * A retriever whose ranked list can enter fusion.\n *\n * `bm25` is the daemon's Okapi lexical index, fetched as its own leg via the\n * `retriever: \"lexical\"` op rather than pre-fused by `query_hybrid`. Keeping\n * it separate from `grep` matters: they are both \"lexical\" but they fail\n * differently — BM25 has IDF and misses identifiers written in another naming\n * convention, ripgrep has neither IDF nor an index but sees the working tree,\n * including edits made this session.\n */\ntype RetrieverSource = \"grep\" | \"embed\" | \"bm25\";\n\nexport type SearchMode = \"auto\" | \"lexical\" | \"semantic\" | \"hybrid\";\nexport type ResolvedSearchMode = Exclude<SearchMode, \"auto\">;\n\nexport interface RankedHit {\n\tid: string;\n\t/** 1-indexed, gap-free rank within its retriever's list. */\n\trank: number;\n\t/** Retriever-local score (BM25-ish, cosine, …). Diagnostics only — never fused. */\n\tscore?: number;\n\tsource: RetrieverSource;\n}\n\nexport interface FusedHit {\n\tid: string;\n\trrfScore: number;\n\t/** Best rank per contributing retriever. */\n\tranks: Partial<Record<RetrieverSource, number>>;\n\t/** Raw score at the best rank per retriever. Diagnostics only. */\n\trawScores: Partial<Record<RetrieverSource, number>>;\n}\n\n/** Line-span identity a candidate id resolves to, for post-fusion expansion. */\nexport interface CandidateSpan {\n\tpath: string;\n\t/** 1-based inclusive. */\n\tstartLine: number;\n\t/** 1-based inclusive. May exceed the file's length; readers clamp. */\n\tendLine: number;\n}\n\nexport interface FusedCandidate extends FusedHit, CandidateSpan {}\n\n/** Per-call diagnostic record, written to the store-dir trace jsonl — never\n *  into model context. */\nexport interface SearchTrace {\n\ttimestampMs: number;\n\tquery: string;\n\trequestedMode: SearchMode;\n\tresolvedMode: ResolvedSearchMode;\n\t/** Set when the resolved mode is a degradation of the requested one. */\n\tdegradedReason?: string;\n\tindexPhase: \"ready\" | \"indexing\" | \"unavailable\";\n\trrfK?: number;\n\tretrievers: Partial<Record<RetrieverSource, { latencyMs: number; hitCount: number }>>;\n\tfused: FusedHit[];\n\trerank?: { applied: boolean; candidateCount: number; latencyMs: number };\n}\n"]}

package/examples/extensions/custom-provider-anthropic/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@kolisachint/hoocode-extension-custom-provider-anthropic",
 	"private": true,
-	"version": "0.2.161",
+	"version": "0.2.162",
 	"type": "module",
 	"engines": {
 		"bun": ">=1.0.0"

package/examples/extensions/custom-provider-gitlab-duo/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@kolisachint/hoocode-extension-custom-provider-gitlab-duo",
 	"private": true,
-	"version": "0.2.161",
+	"version": "0.2.162",
 	"type": "module",
 	"engines": {
 		"bun": ">=1.0.0"

package/examples/extensions/sandbox/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@kolisachint/hoocode-extension-sandbox",
 	"private": true,
-	"version": "0.2.161",
+	"version": "0.2.162",
 	"type": "module",
 	"engines": {
 		"bun": ">=1.0.0"

package/examples/extensions/with-deps/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@kolisachint/hoocode-extension-with-deps",
 	"private": true,
-	"version": "0.2.161",
+	"version": "0.2.162",
 	"type": "module",
 	"engines": {
 		"bun": ">=1.0.0"

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@kolisachint/hoocode-agent",
-	"version": "0.4.164",
+	"version": "0.4.165",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"type": "module",
 	"hoocodeConfig": {
@@ -42,12 +42,15 @@
 		"copy-assets": "shx mkdir -p dist/modes/interactive/theme && shx cp src/modes/interactive/theme/*.json dist/modes/interactive/theme/ && shx mkdir -p dist/core/export-html/vendor && shx cp src/core/export-html/template.html src/core/export-html/template.css src/core/export-html/template.js dist/core/export-html/ && shx cp src/core/export-html/vendor/*.js dist/core/export-html/vendor/ && shx mkdir -p dist/core/extensions/plugins && shx cp -r src/core/extensions/plugins/default-marketplace dist/core/extensions/plugins/",
 		"copy-binary-assets": "shx cp package.json dist/ && shx cp README.md dist/ && shx cp CHANGELOG.md dist/ && shx mkdir -p dist/theme && shx cp src/modes/interactive/theme/*.json dist/theme/ && shx mkdir -p dist/export-html/vendor && shx cp src/core/export-html/template.html dist/export-html/ && shx cp src/core/export-html/vendor/*.js dist/export-html/vendor/ && shx cp -r docs dist/ && shx cp -r examples dist/ && shx cp ../../node_modules/@silvia-odwyer/photon-node/photon_rs_bg.wasm dist/",
 		"test": "vitest --run",
+		"search-eval": "bun scripts/search-eval.ts",
+		"search-eval:gold": "bun scripts/search-eval-gold.ts",
+		"search-eval:compare": "bun scripts/search-eval-compare.ts",
 		"prepublishOnly": "npm run clean && npm run build"
 	},
 	"dependencies": {
-		"@kolisachint/hoocode-agent-core": "^0.4.164",
-		"@kolisachint/hoocode-ai": "^0.4.164",
-		"@kolisachint/hoocode-tui": "^0.4.164",
+		"@kolisachint/hoocode-agent-core": "^0.4.165",
+		"@kolisachint/hoocode-ai": "^0.4.165",
+		"@kolisachint/hoocode-tui": "^0.4.165",
 		"@silvia-odwyer/photon-node": "^0.3.4",
 		"chalk": "^5.5.0",
 		"cli-highlight": "^2.1.11",