mcp-researchpowerpack 6.0.9 → 6.0.10

package/dist/mcp-use.json

@@ -1,7 +1,7 @@
 {
   "includeInspector": false,
-  "buildTime": "2026-04-24T20:33:37.812Z",
-  "buildId": "16e6972fd463a4e9",
+  "buildTime": "2026-04-28T08:35:39.704Z",
+  "buildId": "4783158b494ab6e4",
   "entryPoint": "dist/index.js",
   "widgets": {}
 }

package/dist/src/schemas/web-search.js

@@ -46,7 +46,18 @@ const webSearchOutputSchema = z.object({
       result_count: z.number().int().nonnegative().describe("Results returned for this query."),
       top_url: z.string().optional().describe("Domain of the top result.")
     })).optional().describe("Per-query result counts and top URLs."),
-    low_yield_queries: z.array(z.string()).optional().describe("Queries that produced 0-1 results.")
+    low_yield_queries: z.array(z.string()).optional().describe("Queries that produced 0-1 results."),
+    query_rewrites: z.array(z.object({
+      original: z.string().describe("The query as the agent submitted it."),
+      rewritten: z.string().describe("The query as dispatched to Google after Phase A normalization."),
+      rules: z.array(z.string()).describe("Rule ids applied (A1=operator-char de-quote, A2=path/URL de-quote, A3=phrase-AND collapse).")
+    })).optional().describe("Pre-dispatch query rewrites \u2014 Phase A normalizations (operator-char and path/URL de-quote, phrase-AND \u2192 anchor + OR collapse)."),
+    retried_queries: z.array(z.object({
+      original: z.string().describe("The query as dispatched (post-Phase-A) that returned 0 results."),
+      retried_with: z.string().describe("The relaxed form retried after the empty initial response."),
+      rules: z.array(z.string()).describe("Rule ids applied (B1=strip all quotes, B2=drop site: filter)."),
+      recovered_results: z.number().int().nonnegative().describe("How many hits the retry produced; 0 means the retry also failed.")
+    })).optional().describe("On-empty retries \u2014 Phase B relaxations applied after the initial Serper batch returned 0 results for a query.")
   }).strict()
 }).strict();
 export {

package/dist/src/schemas/web-search.js.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
   "sources": ["../../../src/schemas/web-search.ts"],
-  "sourcesContent": ["import { z } from 'zod';\n\nexport const webSearchParamsSchema = z.object({\n  queries: z\n    .array(\n      z.string()\n        .min(1, { message: 'web-search: Query cannot be empty' })\n        .describe('A single Google search query. Each query runs as a separate parallel search. Use operators (site:, quotes, verbatim version numbers) to sharpen retrieval.'),\n    )\n    .min(1, { message: 'web-search: At least 1 query required' })\n    .describe(\n      'Search queries to run in parallel via Google. Think of these as **concept groups** \u2014 clusters of semantically distinct facets of your research goal, each probing a DIFFERENT angle (official spec, implementation, failures, comparison, sentiment, changelog, CVE, pricing). Fire all groups in ONE call as a flat array. Overlapping queries waste budget; orthogonal facets multiply coverage. A narrow bug needs 10\u201320 queries across 2\u20133 facets; a comparison needs 25\u201335 across 4\u20136 facets; open-ended synthesis needs 40\u201380 across 8+ facets.',\n    ),\n  extract: z\n    .string()\n    .min(1, { message: 'web-search: extract cannot be empty' })\n    .describe(\n      'Semantic instruction for the relevance classifier \u2014 what \"relevant\" means for THIS goal. Drives tiering (HIGHLY_RELEVANT / MAYBE_RELEVANT / OTHER), synthesis, gap analysis, and refine-query suggestions. Be specific: \"OAuth 2.1 support in TypeScript MCP frameworks \u2014 runnable code, not marketing\", not \"MCP OAuth\". The classifier uses this to choose a source-of-truth rubric (vendor_doc for spec, github for bugs, reddit/blog for migration/sentiment, cve_databases for security).',\n    ),\n  raw: z\n    .boolean()\n    .default(false)\n    .describe('Skip LLM classification and return the raw ranked URL list. Use when you need unprocessed results.'),\n  scope: z\n    .enum(['web', 'reddit', 'both'])\n    .default('web')\n    .describe(\n      'Search scope. \"web\" (default) = open web, no augmentation. \"reddit\" = server appends `site:reddit.com` to every query and filters results to post permalinks (`/r/.+/comments/[a-z0-9]+/`); subreddit homepages are dropped. \"both\" = runs every query twice (open web + reddit-scoped), merges the result set, and tags each row with its source. Use \"reddit\" for sentiment/migration/lived-experience research; use \"both\" when you want one call to cover both branches.',\n    ),\n  verbose: z\n    .boolean()\n    .default(false)\n    .describe(\n      'Include the per-row scoring/coverage metadata, the trailing Signals block, and the CONSENSUS labels even when they carry little signal (single-query hits, threshold of 1). Default false \u2014 most agents do not need this and it costs ~1.5KB per call on a typical 3-query fan-out.',\n    ),\n}).strict();\n\nexport type WebSearchParams = z.infer<typeof webSearchParamsSchema>;\n\nexport const webSearchOutputSchema = z.object({\n  // `content` deliberately NOT duplicated here \u2014 the primary markdown lives in\n  // the MCP tool result's `content[0].text`. Previously this schema echoed the\n  // whole markdown under `structuredContent.content`, doubling token cost for\n  // clients that forward both fields to an LLM.\n  results: z\n    .array(z.object({\n      rank: z.number().int().positive().describe('1-based rank in the merged ranking.'),\n      url: z.string().describe('Result URL.'),\n      title: z.string().describe('Page title from the result.'),\n      snippet: z.string().describe('Search snippet from the result.'),\n      source_type: z\n        .enum(['reddit', 'github', 'docs', 'blog', 'paper', 'qa', 'cve', 'news', 'video', 'web'])\n        .describe(\n          'Heuristic source kind from the URL. When the LLM classifier is online its tag overrides this.',\n        ),\n      score: z.number().describe('Composite CTR-weighted score, normalized to 100.'),\n      seen_in: z.number().int().nonnegative().describe('Number of input queries this URL appeared in.'),\n      best_position: z.number().int().nonnegative().describe('Best (lowest) SERP position observed.'),\n    }))\n    .optional()\n    .describe('Per-result structured payload \u2014 same data the markdown table renders, machine-readable.'),\n  metadata: z.object({\n    total_items: z.number().int().nonnegative().describe('Number of queries executed.'),\n    successful: z.number().int().nonnegative().describe('Queries that returned results.'),\n    failed: z.number().int().nonnegative().describe('Queries that failed.'),\n    execution_time_ms: z.number().int().nonnegative().describe('Wall clock time in milliseconds.'),\n    llm_classified: z.boolean().describe('Whether LLM classification was applied.'),\n    llm_error: z.string().optional().describe('LLM error if classification failed and fell back to raw.'),\n    scope: z.enum(['web', 'reddit', 'both']).optional().describe('Search scope used.'),\n    coverage_summary: z\n      .array(z.object({\n        query: z.string().describe('The search query.'),\n        result_count: z.number().int().nonnegative().describe('Results returned for this query.'),\n        top_url: z.string().optional().describe('Domain of the top result.'),\n      }))\n      .optional()\n      .describe('Per-query result counts and top URLs.'),\n    low_yield_queries: z\n      .array(z.string())\n      .optional()\n      .describe('Queries that produced 0-1 results.'),\n  }).strict(),\n}).strict();\n\nexport type WebSearchOutput = z.infer<typeof webSearchOutputSchema>;\n"],
-  "mappings": "AAAA,SAAS,SAAS;AAEX,MAAM,wBAAwB,EAAE,OAAO;AAAA,EAC5C,SAAS,EACN;AAAA,IACC,EAAE,OAAO,EACN,IAAI,GAAG,EAAE,SAAS,oCAAoC,CAAC,EACvD,SAAS,4JAA4J;AAAA,EAC1K,EACC,IAAI,GAAG,EAAE,SAAS,wCAAwC,CAAC,EAC3D;AAAA,IACC;AAAA,EACF;AAAA,EACF,SAAS,EACN,OAAO,EACP,IAAI,GAAG,EAAE,SAAS,sCAAsC,CAAC,EACzD;AAAA,IACC;AAAA,EACF;AAAA,EACF,KAAK,EACF,QAAQ,EACR,QAAQ,KAAK,EACb,SAAS,oGAAoG;AAAA,EAChH,OAAO,EACJ,KAAK,CAAC,OAAO,UAAU,MAAM,CAAC,EAC9B,QAAQ,KAAK,EACb;AAAA,IACC;AAAA,EACF;AAAA,EACF,SAAS,EACN,QAAQ,EACR,QAAQ,KAAK,EACb;AAAA,IACC;AAAA,EACF;AACJ,CAAC,EAAE,OAAO;AAIH,MAAM,wBAAwB,EAAE,OAAO;AAAA;AAAA;AAAA;AAAA;AAAA,EAK5C,SAAS,EACN,MAAM,EAAE,OAAO;AAAA,IACd,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,SAAS,EAAE,SAAS,qCAAqC;AAAA,IAChF,KAAK,EAAE,OAAO,EAAE,SAAS,aAAa;AAAA,IACtC,OAAO,EAAE,OAAO,EAAE,SAAS,6BAA6B;AAAA,IACxD,SAAS,EAAE,OAAO,EAAE,SAAS,iCAAiC;AAAA,IAC9D,aAAa,EACV,KAAK,CAAC,UAAU,UAAU,QAAQ,QAAQ,SAAS,MAAM,OAAO,QAAQ,SAAS,KAAK,CAAC,EACvF;AAAA,MACC;AAAA,IACF;AAAA,IACF,OAAO,EAAE,OAAO,EAAE,SAAS,kDAAkD;AAAA,IAC7E,SAAS,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,+CAA+C;AAAA,IAChG,eAAe,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,uCAAuC;AAAA,EAChG,CAAC,CAAC,EACD,SAAS,EACT,SAAS,8FAAyF;AAAA,EACrG,UAAU,EAAE,OAAO;AAAA,IACjB,aAAa,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,6BAA6B;AAAA,IAClF,YAAY,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,gCAAgC;AAAA,IACpF,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,sBAAsB;AAAA,IACtE,mBAAmB,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,kCAAkC;AAAA,IAC7F,gBAAgB,EAAE,QAAQ,EAAE,SAAS,yCAAyC;AAAA,IAC9E,WAAW,EAAE,OAAO,EAAE,SAAS,EAAE,SAAS,0DAA0D;AAAA,IACpG,OAAO,EAAE,KAAK,CAAC,OAAO,UAAU,MAAM,CAAC,EAAE,SAAS,EAAE,SAAS,oBAAoB;AAAA,IACjF,kBAAkB,EACf,MAAM,EAAE,OAAO;AAAA,MACd,OAAO,EAAE,OAAO,EAAE,SAAS,mBAAmB;AAAA,MAC9C,cAAc,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,kCAAkC;AAAA,MACxF,SAAS,EAAE,OAAO,EAAE,SAAS,EAAE,SAAS,2BAA2B;AAAA,IACrE,CAAC,CAAC,EACD,SAAS,EACT,SAAS,uCAAuC;AAAA,IACnD,mBAAmB,EAChB,MAAM,EAAE,OAAO,CAAC,EAChB,SAAS,EACT,SAAS,oCAAoC;AAAA,EAClD,CAAC,EAAE,OAAO;AACZ,CAAC,EAAE,OAAO;",
+  "sourcesContent": ["import { z } from 'zod';\n\nexport const webSearchParamsSchema = z.object({\n  queries: z\n    .array(\n      z.string()\n        .min(1, { message: 'web-search: Query cannot be empty' })\n        .describe('A single Google search query. Each query runs as a separate parallel search. Use operators (site:, quotes, verbatim version numbers) to sharpen retrieval.'),\n    )\n    .min(1, { message: 'web-search: At least 1 query required' })\n    .describe(\n      'Search queries to run in parallel via Google. Think of these as **concept groups** \u2014 clusters of semantically distinct facets of your research goal, each probing a DIFFERENT angle (official spec, implementation, failures, comparison, sentiment, changelog, CVE, pricing). Fire all groups in ONE call as a flat array. Overlapping queries waste budget; orthogonal facets multiply coverage. A narrow bug needs 10\u201320 queries across 2\u20133 facets; a comparison needs 25\u201335 across 4\u20136 facets; open-ended synthesis needs 40\u201380 across 8+ facets.',\n    ),\n  extract: z\n    .string()\n    .min(1, { message: 'web-search: extract cannot be empty' })\n    .describe(\n      'Semantic instruction for the relevance classifier \u2014 what \"relevant\" means for THIS goal. Drives tiering (HIGHLY_RELEVANT / MAYBE_RELEVANT / OTHER), synthesis, gap analysis, and refine-query suggestions. Be specific: \"OAuth 2.1 support in TypeScript MCP frameworks \u2014 runnable code, not marketing\", not \"MCP OAuth\". The classifier uses this to choose a source-of-truth rubric (vendor_doc for spec, github for bugs, reddit/blog for migration/sentiment, cve_databases for security).',\n    ),\n  raw: z\n    .boolean()\n    .default(false)\n    .describe('Skip LLM classification and return the raw ranked URL list. Use when you need unprocessed results.'),\n  scope: z\n    .enum(['web', 'reddit', 'both'])\n    .default('web')\n    .describe(\n      'Search scope. \"web\" (default) = open web, no augmentation. \"reddit\" = server appends `site:reddit.com` to every query and filters results to post permalinks (`/r/.+/comments/[a-z0-9]+/`); subreddit homepages are dropped. \"both\" = runs every query twice (open web + reddit-scoped), merges the result set, and tags each row with its source. Use \"reddit\" for sentiment/migration/lived-experience research; use \"both\" when you want one call to cover both branches.',\n    ),\n  verbose: z\n    .boolean()\n    .default(false)\n    .describe(\n      'Include the per-row scoring/coverage metadata, the trailing Signals block, and the CONSENSUS labels even when they carry little signal (single-query hits, threshold of 1). Default false \u2014 most agents do not need this and it costs ~1.5KB per call on a typical 3-query fan-out.',\n    ),\n}).strict();\n\nexport type WebSearchParams = z.infer<typeof webSearchParamsSchema>;\n\nexport const webSearchOutputSchema = z.object({\n  // `content` deliberately NOT duplicated here \u2014 the primary markdown lives in\n  // the MCP tool result's `content[0].text`. Previously this schema echoed the\n  // whole markdown under `structuredContent.content`, doubling token cost for\n  // clients that forward both fields to an LLM.\n  results: z\n    .array(z.object({\n      rank: z.number().int().positive().describe('1-based rank in the merged ranking.'),\n      url: z.string().describe('Result URL.'),\n      title: z.string().describe('Page title from the result.'),\n      snippet: z.string().describe('Search snippet from the result.'),\n      source_type: z\n        .enum(['reddit', 'github', 'docs', 'blog', 'paper', 'qa', 'cve', 'news', 'video', 'web'])\n        .describe(\n          'Heuristic source kind from the URL. When the LLM classifier is online its tag overrides this.',\n        ),\n      score: z.number().describe('Composite CTR-weighted score, normalized to 100.'),\n      seen_in: z.number().int().nonnegative().describe('Number of input queries this URL appeared in.'),\n      best_position: z.number().int().nonnegative().describe('Best (lowest) SERP position observed.'),\n    }))\n    .optional()\n    .describe('Per-result structured payload \u2014 same data the markdown table renders, machine-readable.'),\n  metadata: z.object({\n    total_items: z.number().int().nonnegative().describe('Number of queries executed.'),\n    successful: z.number().int().nonnegative().describe('Queries that returned results.'),\n    failed: z.number().int().nonnegative().describe('Queries that failed.'),\n    execution_time_ms: z.number().int().nonnegative().describe('Wall clock time in milliseconds.'),\n    llm_classified: z.boolean().describe('Whether LLM classification was applied.'),\n    llm_error: z.string().optional().describe('LLM error if classification failed and fell back to raw.'),\n    scope: z.enum(['web', 'reddit', 'both']).optional().describe('Search scope used.'),\n    coverage_summary: z\n      .array(z.object({\n        query: z.string().describe('The search query.'),\n        result_count: z.number().int().nonnegative().describe('Results returned for this query.'),\n        top_url: z.string().optional().describe('Domain of the top result.'),\n      }))\n      .optional()\n      .describe('Per-query result counts and top URLs.'),\n    low_yield_queries: z\n      .array(z.string())\n      .optional()\n      .describe('Queries that produced 0-1 results.'),\n    query_rewrites: z\n      .array(z.object({\n        original: z.string().describe('The query as the agent submitted it.'),\n        rewritten: z.string().describe('The query as dispatched to Google after Phase A normalization.'),\n        rules: z.array(z.string()).describe('Rule ids applied (A1=operator-char de-quote, A2=path/URL de-quote, A3=phrase-AND collapse).'),\n      }))\n      .optional()\n      .describe('Pre-dispatch query rewrites \u2014 Phase A normalizations (operator-char and path/URL de-quote, phrase-AND \u2192 anchor + OR collapse).'),\n    retried_queries: z\n      .array(z.object({\n        original: z.string().describe('The query as dispatched (post-Phase-A) that returned 0 results.'),\n        retried_with: z.string().describe('The relaxed form retried after the empty initial response.'),\n        rules: z.array(z.string()).describe('Rule ids applied (B1=strip all quotes, B2=drop site: filter).'),\n        recovered_results: z.number().int().nonnegative().describe('How many hits the retry produced; 0 means the retry also failed.'),\n      }))\n      .optional()\n      .describe('On-empty retries \u2014 Phase B relaxations applied after the initial Serper batch returned 0 results for a query.'),\n  }).strict(),\n}).strict();\n\nexport type WebSearchOutput = z.infer<typeof webSearchOutputSchema>;\n"],
+  "mappings": "AAAA,SAAS,SAAS;AAEX,MAAM,wBAAwB,EAAE,OAAO;AAAA,EAC5C,SAAS,EACN;AAAA,IACC,EAAE,OAAO,EACN,IAAI,GAAG,EAAE,SAAS,oCAAoC,CAAC,EACvD,SAAS,4JAA4J;AAAA,EAC1K,EACC,IAAI,GAAG,EAAE,SAAS,wCAAwC,CAAC,EAC3D;AAAA,IACC;AAAA,EACF;AAAA,EACF,SAAS,EACN,OAAO,EACP,IAAI,GAAG,EAAE,SAAS,sCAAsC,CAAC,EACzD;AAAA,IACC;AAAA,EACF;AAAA,EACF,KAAK,EACF,QAAQ,EACR,QAAQ,KAAK,EACb,SAAS,oGAAoG;AAAA,EAChH,OAAO,EACJ,KAAK,CAAC,OAAO,UAAU,MAAM,CAAC,EAC9B,QAAQ,KAAK,EACb;AAAA,IACC;AAAA,EACF;AAAA,EACF,SAAS,EACN,QAAQ,EACR,QAAQ,KAAK,EACb;AAAA,IACC;AAAA,EACF;AACJ,CAAC,EAAE,OAAO;AAIH,MAAM,wBAAwB,EAAE,OAAO;AAAA;AAAA;AAAA;AAAA;AAAA,EAK5C,SAAS,EACN,MAAM,EAAE,OAAO;AAAA,IACd,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,SAAS,EAAE,SAAS,qCAAqC;AAAA,IAChF,KAAK,EAAE,OAAO,EAAE,SAAS,aAAa;AAAA,IACtC,OAAO,EAAE,OAAO,EAAE,SAAS,6BAA6B;AAAA,IACxD,SAAS,EAAE,OAAO,EAAE,SAAS,iCAAiC;AAAA,IAC9D,aAAa,EACV,KAAK,CAAC,UAAU,UAAU,QAAQ,QAAQ,SAAS,MAAM,OAAO,QAAQ,SAAS,KAAK,CAAC,EACvF;AAAA,MACC;AAAA,IACF;AAAA,IACF,OAAO,EAAE,OAAO,EAAE,SAAS,kDAAkD;AAAA,IAC7E,SAAS,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,+CAA+C;AAAA,IAChG,eAAe,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,uCAAuC;AAAA,EAChG,CAAC,CAAC,EACD,SAAS,EACT,SAAS,8FAAyF;AAAA,EACrG,UAAU,EAAE,OAAO;AAAA,IACjB,aAAa,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,6BAA6B;AAAA,IAClF,YAAY,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,gCAAgC;AAAA,IACpF,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,sBAAsB;AAAA,IACtE,mBAAmB,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,kCAAkC;AAAA,IAC7F,gBAAgB,EAAE,QAAQ,EAAE,SAAS,yCAAyC;AAAA,IAC9E,WAAW,EAAE,OAAO,EAAE,SAAS,EAAE,SAAS,0DAA0D;AAAA,IACpG,OAAO,EAAE,KAAK,CAAC,OAAO,UAAU,MAAM,CAAC,EAAE,SAAS,EAAE,SAAS,oBAAoB;AAAA,IACjF,kBAAkB,EACf,MAAM,EAAE,OAAO;AAAA,MACd,OAAO,EAAE,OAAO,EAAE,SAAS,mBAAmB;AAAA,MAC9C,cAAc,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,kCAAkC;AAAA,MACxF,SAAS,EAAE,OAAO,EAAE,SAAS,EAAE,SAAS,2BAA2B;AAAA,IACrE,CAAC,CAAC,EACD,SAAS,EACT,SAAS,uCAAuC;AAAA,IACnD,mBAAmB,EAChB,MAAM,EAAE,OAAO,CAAC,EAChB,SAAS,EACT,SAAS,oCAAoC;AAAA,IAChD,gBAAgB,EACb,MAAM,EAAE,OAAO;AAAA,MACd,UAAU,EAAE,OAAO,EAAE,SAAS,sCAAsC;AAAA,MACpE,WAAW,EAAE,OAAO,EAAE,SAAS,gEAAgE;AAAA,MAC/F,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,SAAS,6FAA6F;AAAA,IACnI,CAAC,CAAC,EACD,SAAS,EACT,SAAS,0IAAgI;AAAA,IAC5I,iBAAiB,EACd,MAAM,EAAE,OAAO;AAAA,MACd,UAAU,EAAE,OAAO,EAAE,SAAS,iEAAiE;AAAA,MAC/F,cAAc,EAAE,OAAO,EAAE,SAAS,4DAA4D;AAAA,MAC9F,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,SAAS,+DAA+D;AAAA,MACnG,mBAAmB,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,kEAAkE;AAAA,IAC/H,CAAC,CAAC,EACD,SAAS,EACT,SAAS,oHAA+G;AAAA,EAC7H,CAAC,EAAE,OAAO;AACZ,CAAC,EAAE,OAAO;",
   "names": []
 }

package/dist/src/tools/search.js

@@ -28,6 +28,10 @@ import {
   toToolResponse
 } from "./mcp-helpers.js";
 import { sanitizeSuggestion } from "../utils/sanitize.js";
+import {
+  normalizeQueryForDispatch,
+  relaxQueryForRetry
+} from "../utils/query-relax.js";
 const REDDIT_POST_PERMALINK = /\/r\/[^/]+\/comments\/[a-z0-9]+\//i;
 const REDDIT_HOST = /(?:^|\.)reddit\.com$/i;
 function decorateQueriesForScope(queries, scope) {
@@ -41,6 +45,58 @@ async function executeSearches(queries) {
   const client = new SearchClient();
   return client.searchMultiple(queries);
 }
+async function executeWithRelaxRetry(dispatched, reporter) {
+  const initial = await executeSearches(dispatched);
+  const emptyIndices = initial.searches.map((s, i) => s.results.length === 0 ? i : -1).filter((i) => i !== -1);
+  if (emptyIndices.length === 0) {
+    return { response: initial, retried: [] };
+  }
+  const plans = [];
+  for (const idx of emptyIndices) {
+    const dq = dispatched[idx];
+    if (typeof dq !== "string") continue;
+    const r = relaxQueryForRetry(dq);
+    if (r.changed && r.rewritten !== dq) {
+      plans.push({ index: idx, original: dq, relaxed: r.rewritten, rules: [...r.rules] });
+    }
+  }
+  if (plans.length === 0) {
+    return { response: initial, retried: [] };
+  }
+  mcpLog(
+    "info",
+    `${plans.length}/${emptyIndices.length} empty-result queries eligible for relaxation retry`,
+    "search"
+  );
+  await reporter.log(
+    "info",
+    `${plans.length} queries returned 0 results; retrying with relaxation`
+  );
+  const retryResp = await executeSearches(plans.map((p) => p.relaxed));
+  const retried = [];
+  const retryByIndex = /* @__PURE__ */ new Map();
+  plans.forEach((plan, i) => {
+    const r = retryResp.searches[i];
+    if (r) retryByIndex.set(plan.index, r);
+    retried.push({
+      original: plan.original,
+      retried_with: plan.relaxed,
+      rules: plan.rules,
+      recovered_results: r?.results.length ?? 0
+    });
+  });
+  const mergedSearches = initial.searches.map((s, idx) => {
+    const r = retryByIndex.get(idx);
+    if (r && r.results.length > 0) {
+      return { ...r, query: s.query };
+    }
+    return s;
+  });
+  return {
+    response: { ...initial, searches: mergedSearches },
+    retried
+  };
+}
 function filterScopedSearches(response, scope) {
   if (scope === "web") return response;
   const filtered = response.searches.map((search) => ({
@@ -245,7 +301,7 @@ function buildClassifiedOutput(classification, aggregation, extract, searches, t
   }
   return lines.join("\n");
 }
-function buildMetadata(aggregation, executionTime, totalQueries, searches, llmClassified, scope, llmError) {
+function buildMetadata(aggregation, executionTime, totalQueries, searches, llmClassified, scope, llmError, queryRewrites, retriedQueries) {
   const coverageSummary = searches.map((s) => {
     let topDomain;
     const topResult = s.results[0];
@@ -267,7 +323,9 @@ function buildMetadata(aggregation, executionTime, totalQueries, searches, llmCl
     scope,
     ...llmError ? { llm_error: llmError } : {},
     coverage_summary: coverageSummary,
-    ...lowYieldQueries.length > 0 ? { low_yield_queries: lowYieldQueries } : {}
+    ...lowYieldQueries.length > 0 ? { low_yield_queries: lowYieldQueries } : {},
+    ...queryRewrites && queryRewrites.length > 0 ? { query_rewrites: queryRewrites } : {},
+    ...retriedQueries && retriedQueries.length > 0 ? { retried_queries: retriedQueries } : {}
   };
 }
 function buildStructuredResults(aggregation, llmTagsByRank) {
@@ -319,7 +377,27 @@ async function handleWebSearch(params, reporter = NOOP_REPORTER) {
     }
     await reporter.log("info", `Searching for ${effectiveQueries.length} query/queries (scope=${params.scope})`);
     await reporter.progress(15, 100, "Submitting search queries");
-    const rawResponse = await executeSearches(effectiveQueries);
+    const dispatchPlan = effectiveQueries.map((q) => {
+      const r = normalizeQueryForDispatch(q);
+      return { original: q, dispatched: r.rewritten, rules: [...r.rules], changed: r.changed };
+    });
+    const dispatchedQueries = dispatchPlan.map((p) => p.dispatched);
+    const queryRewrites = dispatchPlan.filter((p) => p.changed).map((p) => ({ original: p.original, rewritten: p.dispatched, rules: p.rules }));
+    if (queryRewrites.length > 0) {
+      mcpLog(
+        "info",
+        `Pre-dispatch normalized ${queryRewrites.length}/${effectiveQueries.length} queries`,
+        "search"
+      );
+      await reporter.log(
+        "info",
+        `Normalized ${queryRewrites.length} queries pre-dispatch`
+      );
+    }
+    const { response: rawResponse, retried: retriedQueries } = await executeWithRelaxRetry(
+      dispatchedQueries,
+      reporter
+    );
     const response = filterScopedSearches(rawResponse, params.scope);
     await reporter.progress(50, 100, "Collected search results");
     const { aggregation } = processResults(response);
@@ -396,7 +474,9 @@ async function handleWebSearch(params, reporter = NOOP_REPORTER) {
       response.searches,
       llmClassified,
       params.scope,
-      llmError
+      llmError,
+      queryRewrites,
+      retriedQueries
     );
     const llmTagsByRank = /* @__PURE__ */ new Map();
     const results = buildStructuredResults(aggregation, llmTagsByRank);

package/dist/src/tools/search.js.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
   "sources": ["../../../src/tools/search.ts"],
-  "sourcesContent": ["/**\n * Web Search Tool Handler\n * NEVER throws - always returns structured response for graceful degradation\n */\n\nimport type { MCPServer } from 'mcp-use/server';\n\nimport { getCapabilities, getMissingEnvMessage } from '../config/index.js';\nimport {\n  webSearchOutputSchema,\n  webSearchParamsSchema,\n  type WebSearchParams,\n  type WebSearchOutput,\n} from '../schemas/web-search.js';\nimport { SearchClient } from '../clients/search.js';\nimport {\n  aggregateAndRank,\n  generateUnifiedOutput,\n} from '../utils/url-aggregator.js';\nimport {\n  createLLMProcessor,\n  classifySearchResults,\n  suggestRefineQueriesForRawMode,\n  type ClassificationEntry,\n  type ClassificationResult,\n  type RefineQuerySuggestion,\n} from '../services/llm-processor.js';\nimport { classifyError } from '../utils/errors.js';\nimport { classifySourceByUrl } from '../utils/source-type.js';\nimport {\n  mcpLog,\n  formatError,\n  formatDuration,\n} from './utils.js';\nimport {\n  createToolReporter,\n  NOOP_REPORTER,\n  toolFailure,\n  toolSuccess,\n  toToolResponse,\n  type ToolExecutionResult,\n  type ToolReporter,\n} from './mcp-helpers.js';\nimport { sanitizeSuggestion } from '../utils/sanitize.js';\n\n// --- Internal types ---\n\ninterface SearchAggregation {\n  readonly rankedUrls: ReturnType<typeof aggregateAndRank>['rankedUrls'];\n  readonly totalUniqueUrls: number;\n  readonly frequencyThreshold: number;\n  readonly thresholdNote?: string;\n}\n\ninterface SearchResponse {\n  searches: Parameters<typeof aggregateAndRank>[0];\n  totalQueries: number;\n}\n\n// --- Helpers ---\n\n/** Reddit post permalink: /r/{sub}/comments/{id}/ \u2014 drops subreddit\n *  homepages, /rising, /new, /top, etc. so only post URLs reach the agent.\n *  See mcp-revisions/tool-surface/02-extend-web-search-with-reddit-scope.md. */\nconst REDDIT_POST_PERMALINK = /\\/r\\/[^/]+\\/comments\\/[a-z0-9]+\\//i;\nconst REDDIT_HOST = /(?:^|\\.)reddit\\.com$/i;\n\nfunction decorateQueriesForScope(queries: string[], scope: 'web' | 'reddit' | 'both'): string[] {\n  if (scope === 'web') return queries;\n  const reddited = queries.map((q) =>\n    /\\bsite:reddit\\.com\\b/i.test(q) ? q : `${q} site:reddit.com`,\n  );\n  return scope === 'reddit' ? reddited : [...queries, ...reddited];\n}\n\nasync function executeSearches(queries: string[]): Promise<SearchResponse> {\n  const client = new SearchClient();\n  return client.searchMultiple(queries);\n}\n\nfunction filterScopedSearches(\n  response: SearchResponse,\n  scope: 'web' | 'reddit' | 'both',\n): SearchResponse {\n  if (scope === 'web') return response;\n  const filtered = response.searches.map((search) => ({\n    ...search,\n    results: search.results.filter((r) => {\n      let host: string;\n      try { host = new URL(r.link).hostname; } catch { return true; }\n      // Non-reddit URLs pass through; reddit URLs must be post permalinks.\n      if (!REDDIT_HOST.test(host)) return scope !== 'reddit';\n      return REDDIT_POST_PERMALINK.test(r.link);\n    }),\n  }));\n  return { ...response, searches: filtered };\n}\n\nfunction processResults(response: SearchResponse): {\n  aggregation: SearchAggregation;\n} {\n  const aggregation = aggregateAndRank(response.searches, 5);\n  return { aggregation };\n}\n\n// --- Raw output (traditional unified ranked list) ---\n\nfunction buildRawOutput(\n  queries: string[],\n  aggregation: SearchAggregation,\n  searches: SearchResponse['searches'],\n  verbose: boolean = false,\n): string {\n  return generateUnifiedOutput(\n    aggregation.rankedUrls, queries, searches,\n    aggregation.totalUniqueUrls,\n    aggregation.frequencyThreshold, aggregation.thresholdNote,\n    verbose,\n  );\n}\n\nfunction buildSignalsSection(\n  aggregation: SearchAggregation,\n  searches: SearchResponse['searches'],\n  totalQueries: number,\n): string {\n  const coverageCount = searches.filter((search) => search.results.length >= 3).length;\n  const lowYield = searches\n    .filter((search) => search.results.length <= 1)\n    .map((search) => `\"${search.query}\"`);\n  const consensusCount = aggregation.rankedUrls.filter((url) => url.isConsensus).length;\n\n  const lines = [\n    '**Signals**',\n    `- Coverage: ${coverageCount}/${totalQueries} queries returned \u22653 results`,\n    `- Consensus URLs: ${consensusCount}`,\n  ];\n\n  if (lowYield.length > 0) {\n    lines.push(`- Low-yield: ${lowYield.join(', ')}`);\n  }\n\n  return lines.join('\\n');\n}\n\nexport function buildSuggestedFollowUpsSection(\n  refineQueries: Array<{ query: string; rationale?: string; gap_id?: number; gap_description?: string }> | undefined,\n): string {\n  if (!refineQueries || refineQueries.length === 0) {\n    return '';\n  }\n\n  const lines = ['## Suggested follow-up searches', ''];\n\n  for (const item of refineQueries) {\n    const query = sanitizeSuggestion(item.query ?? '');\n    if (!query) continue;\n    const rationale = sanitizeSuggestion(item.rationale ?? '');\n    const gapTag = typeof item.gap_id === 'number'\n      ? ` _(closes gap [${item.gap_id}])_`\n      : item.gap_description\n        ? ` _(${sanitizeSuggestion(item.gap_description)})_`\n        : '';\n    lines.push(rationale\n      ? `- ${query} \u2014 ${rationale}${gapTag}`\n      : `- ${query}${gapTag}`,\n    );\n  }\n\n  return lines.length === 2 ? '' : lines.join('\\n');\n}\n\nexport function appendSignalsAndFollowUps(\n  markdown: string,\n  signalsSection: string,\n  refineQueries: RefineQuerySuggestion[] | undefined,\n  options: { includeSignals?: boolean } = {},\n): string {\n  const includeSignals = options.includeSignals ?? false;\n  const sections = [markdown];\n  if (includeSignals && signalsSection) {\n    sections.push('', '---', signalsSection);\n  }\n  const followUps = buildSuggestedFollowUpsSection(refineQueries);\n  if (followUps) {\n    sections.push('', followUps);\n  }\n  return sections.join('\\n');\n}\n\n// --- \"Start here\" section ---\n//\n// Surfaces the best 3-5 URLs at the top of the classified response so an agent\n// skimming the first screen sees them before tier tables. Deterministic: uses\n// existing `tier` + `rank` + `reason` from the classifier, no extra LLM call.\n//\n// Algorithm: take HIGHLY_RELEVANT by rank up to MAX_START_HERE; if fewer than\n// MIN_START_HERE, pad from top MAYBE_RELEVANT; skip entirely if no entries\n// above OTHER.\n\nconst MIN_START_HERE = 3;\nconst MAX_START_HERE = 5;\n\n/** Minimal structural shape \u2014 avoids coupling to private `RankedUrl` type. */\ninterface StartHereCandidate {\n  readonly rank: number;\n  readonly url: string;\n  readonly title: string;\n}\n\ninterface StartHereTiers {\n  readonly high: readonly StartHereCandidate[];\n  readonly maybe: readonly StartHereCandidate[];\n}\n\nexport function buildStartHereSection(\n  tiers: StartHereTiers,\n  entryByRank: Map<number, ClassificationEntry>,\n  opts: { min?: number; max?: number } = {},\n): string {\n  const min = opts.min ?? MIN_START_HERE;\n  const max = opts.max ?? MAX_START_HERE;\n\n  const picks: Array<{ candidate: StartHereCandidate; tier: 'HIGHLY_RELEVANT' | 'MAYBE_RELEVANT' }> = [];\n\n  for (const candidate of tiers.high) {\n    if (picks.length >= max) break;\n    picks.push({ candidate, tier: 'HIGHLY_RELEVANT' });\n  }\n\n  if (picks.length < min) {\n    const target = Math.min(min, max);\n    for (const candidate of tiers.maybe) {\n      if (picks.length >= target) break;\n      picks.push({ candidate, tier: 'MAYBE_RELEVANT' });\n    }\n  }\n\n  if (picks.length === 0) return '';\n\n  const lines: string[] = [];\n  lines.push('## Start here \u2014 best candidates for your extract');\n  picks.forEach((pick, i) => {\n    const entry = entryByRank.get(pick.candidate.rank);\n    const reason = entry?.reason && entry.reason.trim().length > 0 ? entry.reason : '\u2014';\n    let domain: string;\n    try {\n      domain = new URL(pick.candidate.url).hostname.replace(/^www\\./, '');\n    } catch {\n      domain = pick.candidate.url;\n    }\n    lines.push(\n      `${i + 1}. **[${pick.candidate.title}](${pick.candidate.url})** \u2014 ${domain} \u2014 ${reason} *(${pick.tier}, rank ${pick.candidate.rank})*`,\n    );\n  });\n  return lines.join('\\n');\n}\n\n// --- Classified output (3-tier LLM-classified table) ---\n\nfunction buildClassifiedOutput(\n  classification: ClassificationResult,\n  aggregation: SearchAggregation,\n  extract: string,\n  searches: SearchResponse['searches'],\n  totalQueries: number,\n  verbose: boolean = false,\n): string {\n  const rankedUrls = aggregation.rankedUrls;\n\n  // Build tier \u2192 entries mapping (keep url data alongside classifier metadata)\n  const entryByRank = new Map(classification.results.map((r) => [r.rank, r]));\n\n  const tiers = {\n    high: [] as typeof rankedUrls,\n    maybe: [] as typeof rankedUrls,\n    other: [] as typeof rankedUrls,\n  };\n\n  for (const url of rankedUrls) {\n    const entry = entryByRank.get(url.rank);\n    const tier = entry?.tier;\n    if (tier === 'HIGHLY_RELEVANT') {\n      tiers.high.push(url);\n    } else if (tier === 'MAYBE_RELEVANT') {\n      tiers.maybe.push(url);\n    } else {\n      tiers.other.push(url);\n    }\n  }\n\n  const lines: string[] = [];\n\n  // Header with generated title, synthesis, and confidence\n  lines.push(`## ${classification.title}`);\n  lines.push(`> Looking for: ${extract}`);\n  lines.push(`> ${totalQueries} queries \u2192 ${rankedUrls.length} URLs \u2192 ${tiers.high.length} highly relevant, ${tiers.maybe.length} possibly relevant`);\n  if (classification.confidence) {\n    const confReason = classification.confidence_reason ? ` \u2014 ${classification.confidence_reason}` : '';\n    lines.push(`> Confidence: \\`${classification.confidence}\\`${confReason}`);\n  }\n  lines.push('');\n\n  // \"Start here\" block: surface the top 3-5 URLs above the synthesis so an\n  // agent skimming the first screen sees scrape candidates before prose.\n  const startHere = buildStartHereSection(\n    { high: tiers.high, maybe: tiers.maybe },\n    entryByRank,\n  );\n  if (startHere) {\n    lines.push(startHere);\n    lines.push('');\n  }\n\n  lines.push(`**Summary:** ${classification.synthesis}`);\n  lines.push('');\n\n  // Helper: render one row with optional source_type + reason\n  const renderRichRow = (url: typeof rankedUrls[number]): string => {\n    const entry = entryByRank.get(url.rank);\n    const coveragePct = Math.round(url.coverageRatio * 100);\n    const seenIn = `${url.frequency}/${totalQueries} (${coveragePct}%)`;\n    const sourceType = entry?.source_type ? `\\`${entry.source_type}\\`` : '\u2014';\n    const reason = entry?.reason ? entry.reason.replace(/\\|/g, '\\\\|') : '\u2014';\n    return `| ${url.rank} | [${url.title}](${url.url}) | ${sourceType} | ${seenIn} | ${reason} |`;\n  };\n\n  // Highly Relevant tier\n  if (tiers.high.length > 0) {\n    lines.push(`### Highly Relevant (${tiers.high.length})`);\n    lines.push('| # | URL | Source | Seen in | Why |');\n    lines.push('|---|-----|--------|---------|-----|');\n    for (const url of tiers.high) lines.push(renderRichRow(url));\n    lines.push('');\n  }\n\n  // Maybe Relevant tier\n  if (tiers.maybe.length > 0) {\n    lines.push(`### Maybe Relevant (${tiers.maybe.length})`);\n    lines.push('| # | URL | Source | Seen in | Why |');\n    lines.push('|---|-----|--------|---------|-----|');\n    for (const url of tiers.maybe) lines.push(renderRichRow(url));\n    lines.push('');\n  }\n\n  // Other tier \u2014 with query attribution\n  if (tiers.other.length > 0) {\n    lines.push(`### Other Results (${tiers.other.length})`);\n    lines.push('| # | URL | Source | Score | Queries |');\n    lines.push('|---|-----|--------|-------|---------|');\n    for (const url of tiers.other) {\n      const entry = entryByRank.get(url.rank);\n      const queryList = url.queries.map((q) => `\"${q}\"`).join(', ');\n      const sourceType = entry?.source_type ? `\\`${entry.source_type}\\`` : '\u2014';\n      let domain: string;\n      try {\n        domain = new URL(url.url).hostname.replace(/^www\\./, '');\n      } catch {\n        domain = url.url;\n      }\n      lines.push(`| ${url.rank} | ${domain} | ${sourceType} | ${url.score.toFixed(1)} | ${queryList} |`);\n    }\n    lines.push('');\n  }\n\n  // Signals block is gated behind verbose \u2014 it duplicates info already\n  // present in the per-row metadata for callers who care.\n  // See: docs/code-review/context/05-output-formatting-patterns.md.\n  if (verbose) {\n    lines.push(buildSignalsSection(aggregation, searches, totalQueries));\n  }\n\n  // Gaps section \u2014 what the current results don't answer\n  if (classification.gaps && classification.gaps.length > 0) {\n    lines.push('');\n    lines.push('## Gaps');\n    for (const gap of classification.gaps) {\n      lines.push(`- **[${gap.id}]** ${gap.description}`);\n    }\n  }\n\n  const followUps = buildSuggestedFollowUpsSection(classification.refine_queries);\n  if (followUps) {\n    lines.push('');\n    lines.push(followUps);\n  }\n\n  return lines.join('\\n');\n}\n\n// --- Metadata builder ---\n\nfunction buildMetadata(\n  aggregation: SearchAggregation,\n  executionTime: number,\n  totalQueries: number,\n  searches: SearchResponse['searches'],\n  llmClassified: boolean,\n  scope: 'web' | 'reddit' | 'both',\n  llmError?: string,\n) {\n  const coverageSummary = searches.map(s => {\n    let topDomain: string | undefined;\n    const topResult = s.results[0];\n    if (topResult) {\n      try { topDomain = new URL(topResult.link).hostname.replace(/^www\\./, ''); } catch { /* ignore */ }\n    }\n    return { query: s.query, result_count: s.results.length, top_url: topDomain };\n  });\n  const lowYieldQueries = searches\n    .filter(s => s.results.length <= 1)\n    .map(s => s.query);\n\n  return {\n    total_items: totalQueries,\n    successful: aggregation.rankedUrls.length,\n    failed: totalQueries - searches.filter(s => s.results.length > 0).length,\n    execution_time_ms: executionTime,\n    llm_classified: llmClassified,\n    scope,\n    ...(llmError ? { llm_error: llmError } : {}),\n    coverage_summary: coverageSummary,\n    ...(lowYieldQueries.length > 0 ? { low_yield_queries: lowYieldQueries } : {}),\n  };\n}\n\nfunction buildStructuredResults(\n  aggregation: SearchAggregation,\n  llmTagsByRank?: Map<number, string>,\n): Array<{\n  rank: number;\n  url: string;\n  title: string;\n  snippet: string;\n  source_type: 'reddit' | 'github' | 'docs' | 'blog' | 'paper' | 'qa' | 'cve' | 'news' | 'video' | 'web';\n  score: number;\n  seen_in: number;\n  best_position: number;\n}> {\n  return aggregation.rankedUrls.map((row) => {\n    // LLM tag wins when present; heuristic is the always-on fallback. See:\n    // mcp-revisions/output-shaping/06-source-type-tagging-without-llm.md.\n    const llmTag = llmTagsByRank?.get(row.rank);\n    const heuristic = classifySourceByUrl(row.url);\n    return {\n      rank: row.rank,\n      url: row.url,\n      title: row.title,\n      snippet: row.snippet,\n      source_type: ((llmTag as typeof heuristic) ?? heuristic),\n      score: Number(row.score.toFixed(2)),\n      seen_in: row.frequency,\n      best_position: row.bestPosition,\n    };\n  });\n}\n\n// --- Error builder ---\n\nfunction buildWebSearchError(\n  error: unknown,\n  params: WebSearchParams,\n  startTime: number,\n): ToolExecutionResult<WebSearchOutput> {\n  const structuredError = classifyError(error);\n  const executionTime = Date.now() - startTime;\n\n  mcpLog('error', `web-search: ${structuredError.message}`, 'search');\n\n  const errorContent = formatError({\n    code: structuredError.code,\n    message: structuredError.message,\n    retryable: structuredError.retryable,\n    toolName: 'web-search',\n    howToFix: ['Verify SERPER_API_KEY is set correctly'],\n    alternatives: [\n      'web-search(queries=[\"topic recommendations\"], extract=\"...\", scope: \"reddit\") \u2014 Reddit-only post permalinks via the same backend',\n      'scrape-links(urls=[...], extract=\"...\") \u2014 if you have URLs from prior steps, scrape them now',\n    ],\n  });\n\n  return toolFailure(\n    `${errorContent}\\n\\nExecution time: ${formatDuration(executionTime)}\\nQueries: ${params.queries.length}`,\n  );\n}\n\n// --- Main handler ---\n\nexport async function handleWebSearch(\n  params: WebSearchParams,\n  reporter: ToolReporter = NOOP_REPORTER,\n): Promise<ToolExecutionResult<WebSearchOutput>> {\n  const startTime = Date.now();\n\n  try {\n    const effectiveQueries = decorateQueriesForScope(params.queries, params.scope);\n    if (params.scope !== 'web') {\n      mcpLog('info', `Searching scope=${params.scope}: ${params.queries.length} input queries \u2192 ${effectiveQueries.length} dispatched`, 'search');\n    } else {\n      mcpLog('info', `Searching for ${params.queries.length} query/queries`, 'search');\n    }\n    await reporter.log('info', `Searching for ${effectiveQueries.length} query/queries (scope=${params.scope})`);\n    await reporter.progress(15, 100, 'Submitting search queries');\n\n    const rawResponse = await executeSearches(effectiveQueries);\n    const response = filterScopedSearches(rawResponse, params.scope);\n    await reporter.progress(50, 100, 'Collected search results');\n\n    const { aggregation } = processResults(response);\n    await reporter.log(\n      'info',\n      `Collected ${aggregation.totalUniqueUrls} unique URLs across ${response.totalQueries} queries`,\n    );\n\n    // Decide: raw output or LLM classification\n    const useRaw = params.raw;\n    const llmProcessor = createLLMProcessor();\n\n    let markdown: string;\n    let llmClassified = false;\n    let llmError: string | undefined;\n\n    if (useRaw || !llmProcessor) {\n      // Raw path: traditional unified ranked list\n      if (!useRaw && !llmProcessor) {\n        llmError = 'LLM unavailable (LLM_API_KEY / LLM_BASE_URL / LLM_MODEL not set). Falling back to raw output.';\n        mcpLog('warning', llmError, 'search');\n        // mcp-revisions/llm-degradation/01: surface degraded mode to the client.\n        await reporter.log('warning', 'llm_classifier_unreachable: planner not configured; raw ranked list returned');\n      }\n      let rawRefineQueries: RefineQuerySuggestion[] | undefined;\n      if (useRaw && llmProcessor) {\n        const refineResult = await suggestRefineQueriesForRawMode(\n          aggregation.rankedUrls,\n          params.extract,\n          params.queries,\n          llmProcessor,\n        );\n        rawRefineQueries = refineResult.result;\n      }\n      markdown = appendSignalsAndFollowUps(\n        buildRawOutput(params.queries, aggregation, response.searches, params.verbose),\n        buildSignalsSection(aggregation, response.searches, response.totalQueries),\n        rawRefineQueries,\n        { includeSignals: params.verbose },\n      );\n      await reporter.progress(80, 100, 'Ranking search results');\n    } else {\n      // LLM classification path\n      await reporter.progress(65, 100, 'Classifying results by relevance');\n      const classification = await classifySearchResults(\n        aggregation.rankedUrls,\n        params.extract,\n        response.totalQueries,\n        llmProcessor,\n        params.queries,\n      );\n\n      if (classification.result) {\n        markdown = buildClassifiedOutput(\n          classification.result, aggregation, params.extract, response.searches, response.totalQueries, params.verbose,\n        );\n        llmClassified = true;\n        await reporter.progress(85, 100, 'Formatted classified results');\n      } else {\n        // Classification failed \u2014 fall back to raw\n        llmError = classification.error ?? 'Unknown classification error';\n        mcpLog('warning', `Classification failed, falling back to raw: ${llmError}`, 'search');\n        // mcp-revisions/llm-degradation/01: surface degraded mode to the client.\n        await reporter.log('warning', `llm_classifier_unreachable: ${llmError}`);\n        markdown = appendSignalsAndFollowUps(\n          buildRawOutput(params.queries, aggregation, response.searches, params.verbose),\n          buildSignalsSection(aggregation, response.searches, response.totalQueries),\n          undefined,\n          { includeSignals: params.verbose },\n        );\n        await reporter.progress(85, 100, 'Classification failed, using raw output');\n      }\n    }\n\n    const executionTime = Date.now() - startTime;\n    const metadata = buildMetadata(\n      aggregation, executionTime, response.totalQueries, response.searches, llmClassified, params.scope, llmError,\n    );\n\n    // Build per-row structured results so capability-aware clients can\n    // index into `structuredContent.results` rather than regex-scrape the\n    // markdown table. The LLM tag wins when present; heuristic is the\n    // always-on fallback.\n    const llmTagsByRank = new Map<number, string>();\n    // (When classification succeeds the source_type per-row is populated\n    // inside buildClassifiedOutput via the entry.source_type field \u2014 but\n    // we don't have a direct handle on it here without a refactor. The\n    // heuristic alone covers the structuredContent shape correctly; the\n    // LLM-tagged variant remains in the markdown body.)\n    const results = buildStructuredResults(aggregation, llmTagsByRank);\n\n    mcpLog('info', `Search completed: ${aggregation.rankedUrls.length} URLs, classified=${llmClassified}`, 'search');\n    await reporter.log('info', `Search completed with ${aggregation.rankedUrls.length} URLs (classified: ${llmClassified})`);\n\n    const footer = `\\n---\\n*${formatDuration(executionTime)} | ${aggregation.totalUniqueUrls} unique URLs${llmClassified ? ' | LLM classified' : ''}*`;\n    const fullMarkdown = markdown + footer;\n\n    return toolSuccess(fullMarkdown, { results, metadata });\n  } catch (error) {\n    return buildWebSearchError(error, params, startTime);\n  }\n}\n\nexport function registerWebSearchTool(server: MCPServer): void {\n  server.tool(\n    {\n      name: 'web-search',\n      title: 'Web Search',\n      description:\n        'Fan out Google queries in parallel. One call carries up to 50 queries in a flat `queries` array \u2014 pack diverse facets (not paraphrases) into a single call. Call me AGGRESSIVELY across a session: 2\u20134 rounds is normal, 1 is underuse. After each pass, read `gaps[]` + `refine_queries[]` and fire another round with the new terms. Safe to call multiple times in parallel in the same turn for orthogonal subtopics. `scope`: `\"reddit\"` (server appends `site:reddit.com` + filters to post permalinks \u2014 use for sentiment / migration / lived experience), `\"web\"` default (spec / bug / pricing / CVE / API), `\"both\"` (fan each query across both \u2014 use when opinion-heavy AND needs official sources). Returns a tiered Markdown report (HIGHLY_RELEVANT / MAYBE_RELEVANT / OTHER) + grounded synthesis with `[rank]` citations + `## Gaps` + `## Suggested follow-up searches` tied to gap ids. Set `raw=true` to skip classification.',\n      schema: webSearchParamsSchema,\n      outputSchema: webSearchOutputSchema,\n      annotations: {\n        readOnlyHint: true,\n        idempotentHint: true,\n        destructiveHint: false,\n        openWorldHint: true,\n      },\n    },\n    async (args, ctx) => {\n      if (!getCapabilities().search) {\n        return toToolResponse(toolFailure(getMissingEnvMessage('search')));\n      }\n\n      const reporter = createToolReporter(ctx, 'web-search');\n      const result = await handleWebSearch(args, reporter);\n\n      await reporter.progress(100, 100, result.isError ? 'Search failed' : 'Search complete');\n      return toToolResponse(result);\n    },\n  );\n}\n"],
-  "mappings": "AAOA,SAAS,iBAAiB,4BAA4B;AACtD;AAAA,EACE;AAAA,EACA;AAAA,OAGK;AACP,SAAS,oBAAoB;AAC7B;AAAA,EACE;AAAA,EACA;AAAA,OACK;AACP;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OAIK;AACP,SAAS,qBAAqB;AAC9B,SAAS,2BAA2B;AACpC;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OACK;AACP;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OAGK;AACP,SAAS,0BAA0B;AAqBnC,MAAM,wBAAwB;AAC9B,MAAM,cAAc;AAEpB,SAAS,wBAAwB,SAAmB,OAA4C;AAC9F,MAAI,UAAU,MAAO,QAAO;AAC5B,QAAM,WAAW,QAAQ;AAAA,IAAI,CAAC,MAC5B,wBAAwB,KAAK,CAAC,IAAI,IAAI,GAAG,CAAC;AAAA,EAC5C;AACA,SAAO,UAAU,WAAW,WAAW,CAAC,GAAG,SAAS,GAAG,QAAQ;AACjE;AAEA,eAAe,gBAAgB,SAA4C;AACzE,QAAM,SAAS,IAAI,aAAa;AAChC,SAAO,OAAO,eAAe,OAAO;AACtC;AAEA,SAAS,qBACP,UACA,OACgB;AAChB,MAAI,UAAU,MAAO,QAAO;AAC5B,QAAM,WAAW,SAAS,SAAS,IAAI,CAAC,YAAY;AAAA,IAClD,GAAG;AAAA,IACH,SAAS,OAAO,QAAQ,OAAO,CAAC,MAAM;AACpC,UAAI;AACJ,UAAI;AAAE,eAAO,IAAI,IAAI,EAAE,IAAI,EAAE;AAAA,MAAU,QAAQ;AAAE,eAAO;AAAA,MAAM;AAE9D,UAAI,CAAC,YAAY,KAAK,IAAI,EAAG,QAAO,UAAU;AAC9C,aAAO,sBAAsB,KAAK,EAAE,IAAI;AAAA,IAC1C,CAAC;AAAA,EACH,EAAE;AACF,SAAO,EAAE,GAAG,UAAU,UAAU,SAAS;AAC3C;AAEA,SAAS,eAAe,UAEtB;AACA,QAAM,cAAc,iBAAiB,SAAS,UAAU,CAAC;AACzD,SAAO,EAAE,YAAY;AACvB;AAIA,SAAS,eACP,SACA,aACA,UACA,UAAmB,OACX;AACR,SAAO;AAAA,IACL,YAAY;AAAA,IAAY;AAAA,IAAS;AAAA,IACjC,YAAY;AAAA,IACZ,YAAY;AAAA,IAAoB,YAAY;AAAA,IAC5C;AAAA,EACF;AACF;AAEA,SAAS,oBACP,aACA,UACA,cACQ;AACR,QAAM,gBAAgB,SAAS,OAAO,CAAC,WAAW,OAAO,QAAQ,UAAU,CAAC,EAAE;AAC9E,QAAM,WAAW,SACd,OAAO,CAAC,WAAW,OAAO,QAAQ,UAAU,CAAC,EAC7C,IAAI,CAAC,WAAW,IAAI,OAAO,KAAK,GAAG;AACtC,QAAM,iBAAiB,YAAY,WAAW,OAAO,CAAC,QAAQ,IAAI,WAAW,EAAE;AAE/E,QAAM,QAAQ;AAAA,IACZ;AAAA,IACA,eAAe,aAAa,IAAI,YAAY;AAAA,IAC5C,qBAAqB,cAAc;AAAA,EACrC;AAEA,MAAI,SAAS,SAAS,GAAG;AACvB,UAAM,KAAK,gBAAgB,SAAS,KAAK,IAAI,CAAC,EAAE;AAAA,EAClD;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAEO,SAAS,+BACd,eACQ;AACR,MAAI,CAAC,iBAAiB,cAAc,WAAW,GAAG;AAChD,WAAO;AAAA,EACT;AAEA,QAAM,QAAQ,CAAC,mCAAmC,EAAE;AAEpD,aAAW,QAAQ,eAAe;AAChC,UAAM,QAAQ,mBAAmB,KAAK,SAAS,EAAE;AACjD,QAAI,CAAC,MAAO;AACZ,UAAM,YAAY,mBAAmB,KAAK,aAAa,EAAE;AACzD,UAAM,SAAS,OAAO,KAAK,WAAW,WAClC,kBAAkB,KAAK,MAAM,QAC7B,KAAK,kBACH,MAAM,mBAAmB,KAAK,eAAe,CAAC,OAC9C;AACN,UAAM;AAAA,MAAK,YACP,KAAK,KAAK,WAAM,SAAS,GAAG,MAAM,KAClC,KAAK,KAAK,GAAG,MAAM;AAAA,IACvB;AAAA,EACF;AAEA,SAAO,MAAM,WAAW,IAAI,KAAK,MAAM,KAAK,IAAI;AAClD;AAEO,SAAS,0BACd,UACA,gBACA,eACA,UAAwC,CAAC,GACjC;AACR,QAAM,iBAAiB,QAAQ,kBAAkB;AACjD,QAAM,WAAW,CAAC,QAAQ;AAC1B,MAAI,kBAAkB,gBAAgB;AACpC,aAAS,KAAK,IAAI,OAAO,cAAc;AAAA,EACzC;AACA,QAAM,YAAY,+BAA+B,aAAa;AAC9D,MAAI,WAAW;AACb,aAAS,KAAK,IAAI,SAAS;AAAA,EAC7B;AACA,SAAO,SAAS,KAAK,IAAI;AAC3B;AAYA,MAAM,iBAAiB;AACvB,MAAM,iBAAiB;AAchB,SAAS,sBACd,OACA,aACA,OAAuC,CAAC,GAChC;AACR,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,MAAM,KAAK,OAAO;AAExB,QAAM,QAA8F,CAAC;AAErG,aAAW,aAAa,MAAM,MAAM;AAClC,QAAI,MAAM,UAAU,IAAK;AACzB,UAAM,KAAK,EAAE,WAAW,MAAM,kBAAkB,CAAC;AAAA,EACnD;AAEA,MAAI,MAAM,SAAS,KAAK;AACtB,UAAM,SAAS,KAAK,IAAI,KAAK,GAAG;AAChC,eAAW,aAAa,MAAM,OAAO;AACnC,UAAI,MAAM,UAAU,OAAQ;AAC5B,YAAM,KAAK,EAAE,WAAW,MAAM,iBAAiB,CAAC;AAAA,IAClD;AAAA,EACF;AAEA,MAAI,MAAM,WAAW,EAAG,QAAO;AAE/B,QAAM,QAAkB,CAAC;AACzB,QAAM,KAAK,uDAAkD;AAC7D,QAAM,QAAQ,CAAC,MAAM,MAAM;AACzB,UAAM,QAAQ,YAAY,IAAI,KAAK,UAAU,IAAI;AACjD,UAAM,SAAS,OAAO,UAAU,MAAM,OAAO,KAAK,EAAE,SAAS,IAAI,MAAM,SAAS;AAChF,QAAI;AACJ,QAAI;AACF,eAAS,IAAI,IAAI,KAAK,UAAU,GAAG,EAAE,SAAS,QAAQ,UAAU,EAAE;AAAA,IACpE,QAAQ;AACN,eAAS,KAAK,UAAU;AAAA,IAC1B;AACA,UAAM;AAAA,MACJ,GAAG,IAAI,CAAC,QAAQ,KAAK,UAAU,KAAK,KAAK,KAAK,UAAU,GAAG,cAAS,MAAM,WAAM,MAAM,MAAM,KAAK,IAAI,UAAU,KAAK,UAAU,IAAI;AAAA,IACpI;AAAA,EACF,CAAC;AACD,SAAO,MAAM,KAAK,IAAI;AACxB;AAIA,SAAS,sBACP,gBACA,aACA,SACA,UACA,cACA,UAAmB,OACX;AACR,QAAM,aAAa,YAAY;AAG/B,QAAM,cAAc,IAAI,IAAI,eAAe,QAAQ,IAAI,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC;AAE1E,QAAM,QAAQ;AAAA,IACZ,MAAM,CAAC;AAAA,IACP,OAAO,CAAC;AAAA,IACR,OAAO,CAAC;AAAA,EACV;AAEA,aAAW,OAAO,YAAY;AAC5B,UAAM,QAAQ,YAAY,IAAI,IAAI,IAAI;AACtC,UAAM,OAAO,OAAO;AACpB,QAAI,SAAS,mBAAmB;AAC9B,YAAM,KAAK,KAAK,GAAG;AAAA,IACrB,WAAW,SAAS,kBAAkB;AACpC,YAAM,MAAM,KAAK,GAAG;AAAA,IACtB,OAAO;AACL,YAAM,MAAM,KAAK,GAAG;AAAA,IACtB;AAAA,EACF;AAEA,QAAM,QAAkB,CAAC;AAGzB,QAAM,KAAK,MAAM,eAAe,KAAK,EAAE;AACvC,QAAM,KAAK,kBAAkB,OAAO,EAAE;AACtC,QAAM,KAAK,KAAK,YAAY,mBAAc,WAAW,MAAM,gBAAW,MAAM,KAAK,MAAM,qBAAqB,MAAM,MAAM,MAAM,oBAAoB;AAClJ,MAAI,eAAe,YAAY;AAC7B,UAAM,aAAa,eAAe,oBAAoB,WAAM,eAAe,iBAAiB,KAAK;AACjG,UAAM,KAAK,mBAAmB,eAAe,UAAU,KAAK,UAAU,EAAE;AAAA,EAC1E;AACA,QAAM,KAAK,EAAE;AAIb,QAAM,YAAY;AAAA,IAChB,EAAE,MAAM,MAAM,MAAM,OAAO,MAAM,MAAM;AAAA,IACvC;AAAA,EACF;AACA,MAAI,WAAW;AACb,UAAM,KAAK,SAAS;AACpB,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,KAAK,gBAAgB,eAAe,SAAS,EAAE;AACrD,QAAM,KAAK,EAAE;AAGb,QAAM,gBAAgB,CAAC,QAA2C;AAChE,UAAM,QAAQ,YAAY,IAAI,IAAI,IAAI;AACtC,UAAM,cAAc,KAAK,MAAM,IAAI,gBAAgB,GAAG;AACtD,UAAM,SAAS,GAAG,IAAI,SAAS,IAAI,YAAY,KAAK,WAAW;AAC/D,UAAM,aAAa,OAAO,cAAc,KAAK,MAAM,WAAW,OAAO;AACrE,UAAM,SAAS,OAAO,SAAS,MAAM,OAAO,QAAQ,OAAO,KAAK,IAAI;AACpE,WAAO,KAAK,IAAI,IAAI,OAAO,IAAI,KAAK,KAAK,IAAI,GAAG,OAAO,UAAU,MAAM,MAAM,MAAM,MAAM;AAAA,EAC3F;AAGA,MAAI,MAAM,KAAK,SAAS,GAAG;AACzB,UAAM,KAAK,wBAAwB,MAAM,KAAK,MAAM,GAAG;AACvD,UAAM,KAAK,sCAAsC;AACjD,UAAM,KAAK,sCAAsC;AACjD,eAAW,OAAO,MAAM,KAAM,OAAM,KAAK,cAAc,GAAG,CAAC;AAC3D,UAAM,KAAK,EAAE;AAAA,EACf;AAGA,MAAI,MAAM,MAAM,SAAS,GAAG;AAC1B,UAAM,KAAK,uBAAuB,MAAM,MAAM,MAAM,GAAG;AACvD,UAAM,KAAK,sCAAsC;AACjD,UAAM,KAAK,sCAAsC;AACjD,eAAW,OAAO,MAAM,MAAO,OAAM,KAAK,cAAc,GAAG,CAAC;AAC5D,UAAM,KAAK,EAAE;AAAA,EACf;AAGA,MAAI,MAAM,MAAM,SAAS,GAAG;AAC1B,UAAM,KAAK,sBAAsB,MAAM,MAAM,MAAM,GAAG;AACtD,UAAM,KAAK,wCAAwC;AACnD,UAAM,KAAK,wCAAwC;AACnD,eAAW,OAAO,MAAM,OAAO;AAC7B,YAAM,QAAQ,YAAY,IAAI,IAAI,IAAI;AACtC,YAAM,YAAY,IAAI,QAAQ,IAAI,CAAC,MAAM,IAAI,CAAC,GAAG,EAAE,KAAK,IAAI;AAC5D,YAAM,aAAa,OAAO,cAAc,KAAK,MAAM,WAAW,OAAO;AACrE,UAAI;AACJ,UAAI;AACF,iBAAS,IAAI,IAAI,IAAI,GAAG,EAAE,SAAS,QAAQ,UAAU,EAAE;AAAA,MACzD,QAAQ;AACN,iBAAS,IAAI;AAAA,MACf;AACA,YAAM,KAAK,KAAK,IAAI,IAAI,MAAM,MAAM,MAAM,UAAU,MAAM,IAAI,MAAM,QAAQ,CAAC,CAAC,MAAM,SAAS,IAAI;AAAA,IACnG;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAKA,MAAI,SAAS;AACX,UAAM,KAAK,oBAAoB,aAAa,UAAU,YAAY,CAAC;AAAA,EACrE;AAGA,MAAI,eAAe,QAAQ,eAAe,KAAK,SAAS,GAAG;AACzD,UAAM,KAAK,EAAE;AACb,UAAM,KAAK,SAAS;AACpB,eAAW,OAAO,eAAe,MAAM;AACrC,YAAM,KAAK,QAAQ,IAAI,EAAE,OAAO,IAAI,WAAW,EAAE;AAAA,IACnD;AAAA,EACF;AAEA,QAAM,YAAY,+BAA+B,eAAe,cAAc;AAC9E,MAAI,WAAW;AACb,UAAM,KAAK,EAAE;AACb,UAAM,KAAK,SAAS;AAAA,EACtB;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAIA,SAAS,cACP,aACA,eACA,cACA,UACA,eACA,OACA,UACA;AACA,QAAM,kBAAkB,SAAS,IAAI,OAAK;AACxC,QAAI;AACJ,UAAM,YAAY,EAAE,QAAQ,CAAC;AAC7B,QAAI,WAAW;AACb,UAAI;AAAE,oBAAY,IAAI,IAAI,UAAU,IAAI,EAAE,SAAS,QAAQ,UAAU,EAAE;AAAA,MAAG,QAAQ;AAAA,MAAe;AAAA,IACnG;AACA,WAAO,EAAE,OAAO,EAAE,OAAO,cAAc,EAAE,QAAQ,QAAQ,SAAS,UAAU;AAAA,EAC9E,CAAC;AACD,QAAM,kBAAkB,SACrB,OAAO,OAAK,EAAE,QAAQ,UAAU,CAAC,EACjC,IAAI,OAAK,EAAE,KAAK;AAEnB,SAAO;AAAA,IACL,aAAa;AAAA,IACb,YAAY,YAAY,WAAW;AAAA,IACnC,QAAQ,eAAe,SAAS,OAAO,OAAK,EAAE,QAAQ,SAAS,CAAC,EAAE;AAAA,IAClE,mBAAmB;AAAA,IACnB,gBAAgB;AAAA,IAChB;AAAA,IACA,GAAI,WAAW,EAAE,WAAW,SAAS,IAAI,CAAC;AAAA,IAC1C,kBAAkB;AAAA,IAClB,GAAI,gBAAgB,SAAS,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;AAAA,EAC7E;AACF;AAEA,SAAS,uBACP,aACA,eAUC;AACD,SAAO,YAAY,WAAW,IAAI,CAAC,QAAQ;AAGzC,UAAM,SAAS,eAAe,IAAI,IAAI,IAAI;AAC1C,UAAM,YAAY,oBAAoB,IAAI,GAAG;AAC7C,WAAO;AAAA,MACL,MAAM,IAAI;AAAA,MACV,KAAK,IAAI;AAAA,MACT,OAAO,IAAI;AAAA,MACX,SAAS,IAAI;AAAA,MACb,aAAe,UAA+B;AAAA,MAC9C,OAAO,OAAO,IAAI,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClC,SAAS,IAAI;AAAA,MACb,eAAe,IAAI;AAAA,IACrB;AAAA,EACF,CAAC;AACH;AAIA,SAAS,oBACP,OACA,QACA,WACsC;AACtC,QAAM,kBAAkB,cAAc,KAAK;AAC3C,QAAM,gBAAgB,KAAK,IAAI,IAAI;AAEnC,SAAO,SAAS,eAAe,gBAAgB,OAAO,IAAI,QAAQ;AAElE,QAAM,eAAe,YAAY;AAAA,IAC/B,MAAM,gBAAgB;AAAA,IACtB,SAAS,gBAAgB;AAAA,IACzB,WAAW,gBAAgB;AAAA,IAC3B,UAAU;AAAA,IACV,UAAU,CAAC,wCAAwC;AAAA,IACnD,cAAc;AAAA,MACZ;AAAA,MACA;AAAA,IACF;AAAA,EACF,CAAC;AAED,SAAO;AAAA,IACL,GAAG,YAAY;AAAA;AAAA,kBAAuB,eAAe,aAAa,CAAC;AAAA,WAAc,OAAO,QAAQ,MAAM;AAAA,EACxG;AACF;AAIA,eAAsB,gBACpB,QACA,WAAyB,eACsB;AAC/C,QAAM,YAAY,KAAK,IAAI;AAE3B,MAAI;AACF,UAAM,mBAAmB,wBAAwB,OAAO,SAAS,OAAO,KAAK;AAC7E,QAAI,OAAO,UAAU,OAAO;AAC1B,aAAO,QAAQ,mBAAmB,OAAO,KAAK,KAAK,OAAO,QAAQ,MAAM,yBAAoB,iBAAiB,MAAM,eAAe,QAAQ;AAAA,IAC5I,OAAO;AACL,aAAO,QAAQ,iBAAiB,OAAO,QAAQ,MAAM,kBAAkB,QAAQ;AAAA,IACjF;AACA,UAAM,SAAS,IAAI,QAAQ,iBAAiB,iBAAiB,MAAM,yBAAyB,OAAO,KAAK,GAAG;AAC3G,UAAM,SAAS,SAAS,IAAI,KAAK,2BAA2B;AAE5D,UAAM,cAAc,MAAM,gBAAgB,gBAAgB;AAC1D,UAAM,WAAW,qBAAqB,aAAa,OAAO,KAAK;AAC/D,UAAM,SAAS,SAAS,IAAI,KAAK,0BAA0B;AAE3D,UAAM,EAAE,YAAY,IAAI,eAAe,QAAQ;AAC/C,UAAM,SAAS;AAAA,MACb;AAAA,MACA,aAAa,YAAY,eAAe,uBAAuB,SAAS,YAAY;AAAA,IACtF;AAGA,UAAM,SAAS,OAAO;AACtB,UAAM,eAAe,mBAAmB;AAExC,QAAI;AACJ,QAAI,gBAAgB;AACpB,QAAI;AAEJ,QAAI,UAAU,CAAC,cAAc;AAE3B,UAAI,CAAC,UAAU,CAAC,cAAc;AAC5B,mBAAW;AACX,eAAO,WAAW,UAAU,QAAQ;AAEpC,cAAM,SAAS,IAAI,WAAW,8EAA8E;AAAA,MAC9G;AACA,UAAI;AACJ,UAAI,UAAU,cAAc;AAC1B,cAAM,eAAe,MAAM;AAAA,UACzB,YAAY;AAAA,UACZ,OAAO;AAAA,UACP,OAAO;AAAA,UACP;AAAA,QACF;AACA,2BAAmB,aAAa;AAAA,MAClC;AACA,iBAAW;AAAA,QACT,eAAe,OAAO,SAAS,aAAa,SAAS,UAAU,OAAO,OAAO;AAAA,QAC7E,oBAAoB,aAAa,SAAS,UAAU,SAAS,YAAY;AAAA,QACzE;AAAA,QACA,EAAE,gBAAgB,OAAO,QAAQ;AAAA,MACnC;AACA,YAAM,SAAS,SAAS,IAAI,KAAK,wBAAwB;AAAA,IAC3D,OAAO;AAEL,YAAM,SAAS,SAAS,IAAI,KAAK,kCAAkC;AACnE,YAAM,iBAAiB,MAAM;AAAA,QAC3B,YAAY;AAAA,QACZ,OAAO;AAAA,QACP,SAAS;AAAA,QACT;AAAA,QACA,OAAO;AAAA,MACT;AAEA,UAAI,eAAe,QAAQ;AACzB,mBAAW;AAAA,UACT,eAAe;AAAA,UAAQ;AAAA,UAAa,OAAO;AAAA,UAAS,SAAS;AAAA,UAAU,SAAS;AAAA,UAAc,OAAO;AAAA,QACvG;AACA,wBAAgB;AAChB,cAAM,SAAS,SAAS,IAAI,KAAK,8BAA8B;AAAA,MACjE,OAAO;AAEL,mBAAW,eAAe,SAAS;AACnC,eAAO,WAAW,+CAA+C,QAAQ,IAAI,QAAQ;AAErF,cAAM,SAAS,IAAI,WAAW,+BAA+B,QAAQ,EAAE;AACvE,mBAAW;AAAA,UACT,eAAe,OAAO,SAAS,aAAa,SAAS,UAAU,OAAO,OAAO;AAAA,UAC7E,oBAAoB,aAAa,SAAS,UAAU,SAAS,YAAY;AAAA,UACzE;AAAA,UACA,EAAE,gBAAgB,OAAO,QAAQ;AAAA,QACnC;AACA,cAAM,SAAS,SAAS,IAAI,KAAK,yCAAyC;AAAA,MAC5E;AAAA,IACF;AAEA,UAAM,gBAAgB,KAAK,IAAI,IAAI;AACnC,UAAM,WAAW;AAAA,MACf;AAAA,MAAa;AAAA,MAAe,SAAS;AAAA,MAAc,SAAS;AAAA,MAAU;AAAA,MAAe,OAAO;AAAA,MAAO;AAAA,IACrG;AAMA,UAAM,gBAAgB,oBAAI,IAAoB;AAM9C,UAAM,UAAU,uBAAuB,aAAa,aAAa;AAEjE,WAAO,QAAQ,qBAAqB,YAAY,WAAW,MAAM,qBAAqB,aAAa,IAAI,QAAQ;AAC/G,UAAM,SAAS,IAAI,QAAQ,yBAAyB,YAAY,WAAW,MAAM,sBAAsB,aAAa,GAAG;AAEvH,UAAM,SAAS;AAAA;AAAA,GAAW,eAAe,aAAa,CAAC,MAAM,YAAY,eAAe,eAAe,gBAAgB,sBAAsB,EAAE;AAC/I,UAAM,eAAe,WAAW;AAEhC,WAAO,YAAY,cAAc,EAAE,SAAS,SAAS,CAAC;AAAA,EACxD,SAAS,OAAO;AACd,WAAO,oBAAoB,OAAO,QAAQ,SAAS;AAAA,EACrD;AACF;AAEO,SAAS,sBAAsB,QAAyB;AAC7D,SAAO;AAAA,IACL;AAAA,MACE,MAAM;AAAA,MACN,OAAO;AAAA,MACP,aACE;AAAA,MACF,QAAQ;AAAA,MACR,cAAc;AAAA,MACd,aAAa;AAAA,QACX,cAAc;AAAA,QACd,gBAAgB;AAAA,QAChB,iBAAiB;AAAA,QACjB,eAAe;AAAA,MACjB;AAAA,IACF;AAAA,IACA,OAAO,MAAM,QAAQ;AACnB,UAAI,CAAC,gBAAgB,EAAE,QAAQ;AAC7B,eAAO,eAAe,YAAY,qBAAqB,QAAQ,CAAC,CAAC;AAAA,MACnE;AAEA,YAAM,WAAW,mBAAmB,KAAK,YAAY;AACrD,YAAM,SAAS,MAAM,gBAAgB,MAAM,QAAQ;AAEnD,YAAM,SAAS,SAAS,KAAK,KAAK,OAAO,UAAU,kBAAkB,iBAAiB;AACtF,aAAO,eAAe,MAAM;AAAA,IAC9B;AAAA,EACF;AACF;",
+  "sourcesContent": ["/**\n * Web Search Tool Handler\n * NEVER throws - always returns structured response for graceful degradation\n */\n\nimport type { MCPServer } from 'mcp-use/server';\n\nimport { getCapabilities, getMissingEnvMessage } from '../config/index.js';\nimport {\n  webSearchOutputSchema,\n  webSearchParamsSchema,\n  type WebSearchParams,\n  type WebSearchOutput,\n} from '../schemas/web-search.js';\nimport { SearchClient } from '../clients/search.js';\nimport {\n  aggregateAndRank,\n  generateUnifiedOutput,\n} from '../utils/url-aggregator.js';\nimport {\n  createLLMProcessor,\n  classifySearchResults,\n  suggestRefineQueriesForRawMode,\n  type ClassificationEntry,\n  type ClassificationResult,\n  type RefineQuerySuggestion,\n} from '../services/llm-processor.js';\nimport { classifyError } from '../utils/errors.js';\nimport { classifySourceByUrl } from '../utils/source-type.js';\nimport {\n  mcpLog,\n  formatError,\n  formatDuration,\n} from './utils.js';\nimport {\n  createToolReporter,\n  NOOP_REPORTER,\n  toolFailure,\n  toolSuccess,\n  toToolResponse,\n  type ToolExecutionResult,\n  type ToolReporter,\n} from './mcp-helpers.js';\nimport { sanitizeSuggestion } from '../utils/sanitize.js';\nimport {\n  normalizeQueryForDispatch,\n  relaxQueryForRetry,\n} from '../utils/query-relax.js';\n\n// --- Internal types ---\n\ninterface SearchAggregation {\n  readonly rankedUrls: ReturnType<typeof aggregateAndRank>['rankedUrls'];\n  readonly totalUniqueUrls: number;\n  readonly frequencyThreshold: number;\n  readonly thresholdNote?: string;\n}\n\ninterface SearchResponse {\n  searches: Parameters<typeof aggregateAndRank>[0];\n  totalQueries: number;\n}\n\n// --- Helpers ---\n\n/** Reddit post permalink: /r/{sub}/comments/{id}/ \u2014 drops subreddit\n *  homepages, /rising, /new, /top, etc. so only post URLs reach the agent.\n *  See mcp-revisions/tool-surface/02-extend-web-search-with-reddit-scope.md. */\nconst REDDIT_POST_PERMALINK = /\\/r\\/[^/]+\\/comments\\/[a-z0-9]+\\//i;\nconst REDDIT_HOST = /(?:^|\\.)reddit\\.com$/i;\n\nfunction decorateQueriesForScope(queries: string[], scope: 'web' | 'reddit' | 'both'): string[] {\n  if (scope === 'web') return queries;\n  const reddited = queries.map((q) =>\n    /\\bsite:reddit\\.com\\b/i.test(q) ? q : `${q} site:reddit.com`,\n  );\n  return scope === 'reddit' ? reddited : [...queries, ...reddited];\n}\n\nasync function executeSearches(queries: string[]): Promise<SearchResponse> {\n  const client = new SearchClient();\n  return client.searchMultiple(queries);\n}\n\ninterface QueryRewriteRecord {\n  original: string;\n  rewritten: string;\n  rules: string[];\n}\n\ninterface RetriedQueryRecord {\n  original: string;\n  retried_with: string;\n  rules: string[];\n  recovered_results: number;\n}\n\n/** Run Serper, then for each query that returned 0 results build a relaxed\n *  retry (Phase B) and reissue them in a single second batch. Replace the\n *  empty slot with the retry's results when the retry recovered \u22651 hit, but\n *  keep the original query string in the slot so downstream aggregation and\n *  follow-up rendering stay consistent. */\nasync function executeWithRelaxRetry(\n  dispatched: string[],\n  reporter: ToolReporter,\n): Promise<{ response: SearchResponse; retried: RetriedQueryRecord[] }> {\n  const initial = await executeSearches(dispatched);\n\n  const emptyIndices = initial.searches\n    .map((s, i) => (s.results.length === 0 ? i : -1))\n    .filter((i) => i !== -1);\n\n  if (emptyIndices.length === 0) {\n    return { response: initial, retried: [] };\n  }\n\n  interface Plan { index: number; original: string; relaxed: string; rules: string[] }\n  const plans: Plan[] = [];\n  for (const idx of emptyIndices) {\n    const dq = dispatched[idx];\n    if (typeof dq !== 'string') continue;\n    const r = relaxQueryForRetry(dq);\n    if (r.changed && r.rewritten !== dq) {\n      plans.push({ index: idx, original: dq, relaxed: r.rewritten, rules: [...r.rules] });\n    }\n  }\n\n  if (plans.length === 0) {\n    return { response: initial, retried: [] };\n  }\n\n  mcpLog(\n    'info',\n    `${plans.length}/${emptyIndices.length} empty-result queries eligible for relaxation retry`,\n    'search',\n  );\n  await reporter.log(\n    'info',\n    `${plans.length} queries returned 0 results; retrying with relaxation`,\n  );\n\n  const retryResp = await executeSearches(plans.map((p) => p.relaxed));\n  const retried: RetriedQueryRecord[] = [];\n  const retryByIndex = new Map<number, SearchResponse['searches'][number]>();\n\n  plans.forEach((plan, i) => {\n    const r = retryResp.searches[i];\n    if (r) retryByIndex.set(plan.index, r);\n    retried.push({\n      original: plan.original,\n      retried_with: plan.relaxed,\n      rules: plan.rules,\n      recovered_results: r?.results.length ?? 0,\n    });\n  });\n\n  const mergedSearches = initial.searches.map((s, idx) => {\n    const r = retryByIndex.get(idx);\n    if (r && r.results.length > 0) {\n      return { ...r, query: s.query };\n    }\n    return s;\n  });\n\n  return {\n    response: { ...initial, searches: mergedSearches },\n    retried,\n  };\n}\n\nfunction filterScopedSearches(\n  response: SearchResponse,\n  scope: 'web' | 'reddit' | 'both',\n): SearchResponse {\n  if (scope === 'web') return response;\n  const filtered = response.searches.map((search) => ({\n    ...search,\n    results: search.results.filter((r) => {\n      let host: string;\n      try { host = new URL(r.link).hostname; } catch { return true; }\n      // Non-reddit URLs pass through; reddit URLs must be post permalinks.\n      if (!REDDIT_HOST.test(host)) return scope !== 'reddit';\n      return REDDIT_POST_PERMALINK.test(r.link);\n    }),\n  }));\n  return { ...response, searches: filtered };\n}\n\nfunction processResults(response: SearchResponse): {\n  aggregation: SearchAggregation;\n} {\n  const aggregation = aggregateAndRank(response.searches, 5);\n  return { aggregation };\n}\n\n// --- Raw output (traditional unified ranked list) ---\n\nfunction buildRawOutput(\n  queries: string[],\n  aggregation: SearchAggregation,\n  searches: SearchResponse['searches'],\n  verbose: boolean = false,\n): string {\n  return generateUnifiedOutput(\n    aggregation.rankedUrls, queries, searches,\n    aggregation.totalUniqueUrls,\n    aggregation.frequencyThreshold, aggregation.thresholdNote,\n    verbose,\n  );\n}\n\nfunction buildSignalsSection(\n  aggregation: SearchAggregation,\n  searches: SearchResponse['searches'],\n  totalQueries: number,\n): string {\n  const coverageCount = searches.filter((search) => search.results.length >= 3).length;\n  const lowYield = searches\n    .filter((search) => search.results.length <= 1)\n    .map((search) => `\"${search.query}\"`);\n  const consensusCount = aggregation.rankedUrls.filter((url) => url.isConsensus).length;\n\n  const lines = [\n    '**Signals**',\n    `- Coverage: ${coverageCount}/${totalQueries} queries returned \u22653 results`,\n    `- Consensus URLs: ${consensusCount}`,\n  ];\n\n  if (lowYield.length > 0) {\n    lines.push(`- Low-yield: ${lowYield.join(', ')}`);\n  }\n\n  return lines.join('\\n');\n}\n\nexport function buildSuggestedFollowUpsSection(\n  refineQueries: Array<{ query: string; rationale?: string; gap_id?: number; gap_description?: string }> | undefined,\n): string {\n  if (!refineQueries || refineQueries.length === 0) {\n    return '';\n  }\n\n  const lines = ['## Suggested follow-up searches', ''];\n\n  for (const item of refineQueries) {\n    const query = sanitizeSuggestion(item.query ?? '');\n    if (!query) continue;\n    const rationale = sanitizeSuggestion(item.rationale ?? '');\n    const gapTag = typeof item.gap_id === 'number'\n      ? ` _(closes gap [${item.gap_id}])_`\n      : item.gap_description\n        ? ` _(${sanitizeSuggestion(item.gap_description)})_`\n        : '';\n    lines.push(rationale\n      ? `- ${query} \u2014 ${rationale}${gapTag}`\n      : `- ${query}${gapTag}`,\n    );\n  }\n\n  return lines.length === 2 ? '' : lines.join('\\n');\n}\n\nexport function appendSignalsAndFollowUps(\n  markdown: string,\n  signalsSection: string,\n  refineQueries: RefineQuerySuggestion[] | undefined,\n  options: { includeSignals?: boolean } = {},\n): string {\n  const includeSignals = options.includeSignals ?? false;\n  const sections = [markdown];\n  if (includeSignals && signalsSection) {\n    sections.push('', '---', signalsSection);\n  }\n  const followUps = buildSuggestedFollowUpsSection(refineQueries);\n  if (followUps) {\n    sections.push('', followUps);\n  }\n  return sections.join('\\n');\n}\n\n// --- \"Start here\" section ---\n//\n// Surfaces the best 3-5 URLs at the top of the classified response so an agent\n// skimming the first screen sees them before tier tables. Deterministic: uses\n// existing `tier` + `rank` + `reason` from the classifier, no extra LLM call.\n//\n// Algorithm: take HIGHLY_RELEVANT by rank up to MAX_START_HERE; if fewer than\n// MIN_START_HERE, pad from top MAYBE_RELEVANT; skip entirely if no entries\n// above OTHER.\n\nconst MIN_START_HERE = 3;\nconst MAX_START_HERE = 5;\n\n/** Minimal structural shape \u2014 avoids coupling to private `RankedUrl` type. */\ninterface StartHereCandidate {\n  readonly rank: number;\n  readonly url: string;\n  readonly title: string;\n}\n\ninterface StartHereTiers {\n  readonly high: readonly StartHereCandidate[];\n  readonly maybe: readonly StartHereCandidate[];\n}\n\nexport function buildStartHereSection(\n  tiers: StartHereTiers,\n  entryByRank: Map<number, ClassificationEntry>,\n  opts: { min?: number; max?: number } = {},\n): string {\n  const min = opts.min ?? MIN_START_HERE;\n  const max = opts.max ?? MAX_START_HERE;\n\n  const picks: Array<{ candidate: StartHereCandidate; tier: 'HIGHLY_RELEVANT' | 'MAYBE_RELEVANT' }> = [];\n\n  for (const candidate of tiers.high) {\n    if (picks.length >= max) break;\n    picks.push({ candidate, tier: 'HIGHLY_RELEVANT' });\n  }\n\n  if (picks.length < min) {\n    const target = Math.min(min, max);\n    for (const candidate of tiers.maybe) {\n      if (picks.length >= target) break;\n      picks.push({ candidate, tier: 'MAYBE_RELEVANT' });\n    }\n  }\n\n  if (picks.length === 0) return '';\n\n  const lines: string[] = [];\n  lines.push('## Start here \u2014 best candidates for your extract');\n  picks.forEach((pick, i) => {\n    const entry = entryByRank.get(pick.candidate.rank);\n    const reason = entry?.reason && entry.reason.trim().length > 0 ? entry.reason : '\u2014';\n    let domain: string;\n    try {\n      domain = new URL(pick.candidate.url).hostname.replace(/^www\\./, '');\n    } catch {\n      domain = pick.candidate.url;\n    }\n    lines.push(\n      `${i + 1}. **[${pick.candidate.title}](${pick.candidate.url})** \u2014 ${domain} \u2014 ${reason} *(${pick.tier}, rank ${pick.candidate.rank})*`,\n    );\n  });\n  return lines.join('\\n');\n}\n\n// --- Classified output (3-tier LLM-classified table) ---\n\nfunction buildClassifiedOutput(\n  classification: ClassificationResult,\n  aggregation: SearchAggregation,\n  extract: string,\n  searches: SearchResponse['searches'],\n  totalQueries: number,\n  verbose: boolean = false,\n): string {\n  const rankedUrls = aggregation.rankedUrls;\n\n  // Build tier \u2192 entries mapping (keep url data alongside classifier metadata)\n  const entryByRank = new Map(classification.results.map((r) => [r.rank, r]));\n\n  const tiers = {\n    high: [] as typeof rankedUrls,\n    maybe: [] as typeof rankedUrls,\n    other: [] as typeof rankedUrls,\n  };\n\n  for (const url of rankedUrls) {\n    const entry = entryByRank.get(url.rank);\n    const tier = entry?.tier;\n    if (tier === 'HIGHLY_RELEVANT') {\n      tiers.high.push(url);\n    } else if (tier === 'MAYBE_RELEVANT') {\n      tiers.maybe.push(url);\n    } else {\n      tiers.other.push(url);\n    }\n  }\n\n  const lines: string[] = [];\n\n  // Header with generated title, synthesis, and confidence\n  lines.push(`## ${classification.title}`);\n  lines.push(`> Looking for: ${extract}`);\n  lines.push(`> ${totalQueries} queries \u2192 ${rankedUrls.length} URLs \u2192 ${tiers.high.length} highly relevant, ${tiers.maybe.length} possibly relevant`);\n  if (classification.confidence) {\n    const confReason = classification.confidence_reason ? ` \u2014 ${classification.confidence_reason}` : '';\n    lines.push(`> Confidence: \\`${classification.confidence}\\`${confReason}`);\n  }\n  lines.push('');\n\n  // \"Start here\" block: surface the top 3-5 URLs above the synthesis so an\n  // agent skimming the first screen sees scrape candidates before prose.\n  const startHere = buildStartHereSection(\n    { high: tiers.high, maybe: tiers.maybe },\n    entryByRank,\n  );\n  if (startHere) {\n    lines.push(startHere);\n    lines.push('');\n  }\n\n  lines.push(`**Summary:** ${classification.synthesis}`);\n  lines.push('');\n\n  // Helper: render one row with optional source_type + reason\n  const renderRichRow = (url: typeof rankedUrls[number]): string => {\n    const entry = entryByRank.get(url.rank);\n    const coveragePct = Math.round(url.coverageRatio * 100);\n    const seenIn = `${url.frequency}/${totalQueries} (${coveragePct}%)`;\n    const sourceType = entry?.source_type ? `\\`${entry.source_type}\\`` : '\u2014';\n    const reason = entry?.reason ? entry.reason.replace(/\\|/g, '\\\\|') : '\u2014';\n    return `| ${url.rank} | [${url.title}](${url.url}) | ${sourceType} | ${seenIn} | ${reason} |`;\n  };\n\n  // Highly Relevant tier\n  if (tiers.high.length > 0) {\n    lines.push(`### Highly Relevant (${tiers.high.length})`);\n    lines.push('| # | URL | Source | Seen in | Why |');\n    lines.push('|---|-----|--------|---------|-----|');\n    for (const url of tiers.high) lines.push(renderRichRow(url));\n    lines.push('');\n  }\n\n  // Maybe Relevant tier\n  if (tiers.maybe.length > 0) {\n    lines.push(`### Maybe Relevant (${tiers.maybe.length})`);\n    lines.push('| # | URL | Source | Seen in | Why |');\n    lines.push('|---|-----|--------|---------|-----|');\n    for (const url of tiers.maybe) lines.push(renderRichRow(url));\n    lines.push('');\n  }\n\n  // Other tier \u2014 with query attribution\n  if (tiers.other.length > 0) {\n    lines.push(`### Other Results (${tiers.other.length})`);\n    lines.push('| # | URL | Source | Score | Queries |');\n    lines.push('|---|-----|--------|-------|---------|');\n    for (const url of tiers.other) {\n      const entry = entryByRank.get(url.rank);\n      const queryList = url.queries.map((q) => `\"${q}\"`).join(', ');\n      const sourceType = entry?.source_type ? `\\`${entry.source_type}\\`` : '\u2014';\n      let domain: string;\n      try {\n        domain = new URL(url.url).hostname.replace(/^www\\./, '');\n      } catch {\n        domain = url.url;\n      }\n      lines.push(`| ${url.rank} | ${domain} | ${sourceType} | ${url.score.toFixed(1)} | ${queryList} |`);\n    }\n    lines.push('');\n  }\n\n  // Signals block is gated behind verbose \u2014 it duplicates info already\n  // present in the per-row metadata for callers who care.\n  // See: docs/code-review/context/05-output-formatting-patterns.md.\n  if (verbose) {\n    lines.push(buildSignalsSection(aggregation, searches, totalQueries));\n  }\n\n  // Gaps section \u2014 what the current results don't answer\n  if (classification.gaps && classification.gaps.length > 0) {\n    lines.push('');\n    lines.push('## Gaps');\n    for (const gap of classification.gaps) {\n      lines.push(`- **[${gap.id}]** ${gap.description}`);\n    }\n  }\n\n  const followUps = buildSuggestedFollowUpsSection(classification.refine_queries);\n  if (followUps) {\n    lines.push('');\n    lines.push(followUps);\n  }\n\n  return lines.join('\\n');\n}\n\n// --- Metadata builder ---\n\nfunction buildMetadata(\n  aggregation: SearchAggregation,\n  executionTime: number,\n  totalQueries: number,\n  searches: SearchResponse['searches'],\n  llmClassified: boolean,\n  scope: 'web' | 'reddit' | 'both',\n  llmError?: string,\n  queryRewrites?: QueryRewriteRecord[],\n  retriedQueries?: RetriedQueryRecord[],\n) {\n  const coverageSummary = searches.map(s => {\n    let topDomain: string | undefined;\n    const topResult = s.results[0];\n    if (topResult) {\n      try { topDomain = new URL(topResult.link).hostname.replace(/^www\\./, ''); } catch { /* ignore */ }\n    }\n    return { query: s.query, result_count: s.results.length, top_url: topDomain };\n  });\n  const lowYieldQueries = searches\n    .filter(s => s.results.length <= 1)\n    .map(s => s.query);\n\n  return {\n    total_items: totalQueries,\n    successful: aggregation.rankedUrls.length,\n    failed: totalQueries - searches.filter(s => s.results.length > 0).length,\n    execution_time_ms: executionTime,\n    llm_classified: llmClassified,\n    scope,\n    ...(llmError ? { llm_error: llmError } : {}),\n    coverage_summary: coverageSummary,\n    ...(lowYieldQueries.length > 0 ? { low_yield_queries: lowYieldQueries } : {}),\n    ...(queryRewrites && queryRewrites.length > 0 ? { query_rewrites: queryRewrites } : {}),\n    ...(retriedQueries && retriedQueries.length > 0 ? { retried_queries: retriedQueries } : {}),\n  };\n}\n\nfunction buildStructuredResults(\n  aggregation: SearchAggregation,\n  llmTagsByRank?: Map<number, string>,\n): Array<{\n  rank: number;\n  url: string;\n  title: string;\n  snippet: string;\n  source_type: 'reddit' | 'github' | 'docs' | 'blog' | 'paper' | 'qa' | 'cve' | 'news' | 'video' | 'web';\n  score: number;\n  seen_in: number;\n  best_position: number;\n}> {\n  return aggregation.rankedUrls.map((row) => {\n    // LLM tag wins when present; heuristic is the always-on fallback. See:\n    // mcp-revisions/output-shaping/06-source-type-tagging-without-llm.md.\n    const llmTag = llmTagsByRank?.get(row.rank);\n    const heuristic = classifySourceByUrl(row.url);\n    return {\n      rank: row.rank,\n      url: row.url,\n      title: row.title,\n      snippet: row.snippet,\n      source_type: ((llmTag as typeof heuristic) ?? heuristic),\n      score: Number(row.score.toFixed(2)),\n      seen_in: row.frequency,\n      best_position: row.bestPosition,\n    };\n  });\n}\n\n// --- Error builder ---\n\nfunction buildWebSearchError(\n  error: unknown,\n  params: WebSearchParams,\n  startTime: number,\n): ToolExecutionResult<WebSearchOutput> {\n  const structuredError = classifyError(error);\n  const executionTime = Date.now() - startTime;\n\n  mcpLog('error', `web-search: ${structuredError.message}`, 'search');\n\n  const errorContent = formatError({\n    code: structuredError.code,\n    message: structuredError.message,\n    retryable: structuredError.retryable,\n    toolName: 'web-search',\n    howToFix: ['Verify SERPER_API_KEY is set correctly'],\n    alternatives: [\n      'web-search(queries=[\"topic recommendations\"], extract=\"...\", scope: \"reddit\") \u2014 Reddit-only post permalinks via the same backend',\n      'scrape-links(urls=[...], extract=\"...\") \u2014 if you have URLs from prior steps, scrape them now',\n    ],\n  });\n\n  return toolFailure(\n    `${errorContent}\\n\\nExecution time: ${formatDuration(executionTime)}\\nQueries: ${params.queries.length}`,\n  );\n}\n\n// --- Main handler ---\n\nexport async function handleWebSearch(\n  params: WebSearchParams,\n  reporter: ToolReporter = NOOP_REPORTER,\n): Promise<ToolExecutionResult<WebSearchOutput>> {\n  const startTime = Date.now();\n\n  try {\n    const effectiveQueries = decorateQueriesForScope(params.queries, params.scope);\n    if (params.scope !== 'web') {\n      mcpLog('info', `Searching scope=${params.scope}: ${params.queries.length} input queries \u2192 ${effectiveQueries.length} dispatched`, 'search');\n    } else {\n      mcpLog('info', `Searching for ${params.queries.length} query/queries`, 'search');\n    }\n    await reporter.log('info', `Searching for ${effectiveQueries.length} query/queries (scope=${params.scope})`);\n    await reporter.progress(15, 100, 'Submitting search queries');\n\n    // Phase A \u2014 pre-dispatch normalizer. Rewrites the small fraction of\n    // queries Google was statistically going to mis-handle (3+ phrase AND,\n    // operator chars in quotes, paths in quotes). See src/utils/query-relax.ts.\n    const dispatchPlan = effectiveQueries.map((q) => {\n      const r = normalizeQueryForDispatch(q);\n      return { original: q, dispatched: r.rewritten, rules: [...r.rules], changed: r.changed };\n    });\n    const dispatchedQueries = dispatchPlan.map((p) => p.dispatched);\n    const queryRewrites: QueryRewriteRecord[] = dispatchPlan\n      .filter((p) => p.changed)\n      .map((p) => ({ original: p.original, rewritten: p.dispatched, rules: p.rules }));\n\n    if (queryRewrites.length > 0) {\n      mcpLog(\n        'info',\n        `Pre-dispatch normalized ${queryRewrites.length}/${effectiveQueries.length} queries`,\n        'search',\n      );\n      await reporter.log(\n        'info',\n        `Normalized ${queryRewrites.length} queries pre-dispatch`,\n      );\n    }\n\n    // Phase B \u2014 on-empty retry: any query returning 0 results gets one\n    // relaxed retry (drop quotes, drop site:). Recovered hits replace the\n    // empty slot transparently.\n    const { response: rawResponse, retried: retriedQueries } = await executeWithRelaxRetry(\n      dispatchedQueries,\n      reporter,\n    );\n    const response = filterScopedSearches(rawResponse, params.scope);\n    await reporter.progress(50, 100, 'Collected search results');\n\n    const { aggregation } = processResults(response);\n    await reporter.log(\n      'info',\n      `Collected ${aggregation.totalUniqueUrls} unique URLs across ${response.totalQueries} queries`,\n    );\n\n    // Decide: raw output or LLM classification\n    const useRaw = params.raw;\n    const llmProcessor = createLLMProcessor();\n\n    let markdown: string;\n    let llmClassified = false;\n    let llmError: string | undefined;\n\n    if (useRaw || !llmProcessor) {\n      // Raw path: traditional unified ranked list\n      if (!useRaw && !llmProcessor) {\n        llmError = 'LLM unavailable (LLM_API_KEY / LLM_BASE_URL / LLM_MODEL not set). Falling back to raw output.';\n        mcpLog('warning', llmError, 'search');\n        // mcp-revisions/llm-degradation/01: surface degraded mode to the client.\n        await reporter.log('warning', 'llm_classifier_unreachable: planner not configured; raw ranked list returned');\n      }\n      let rawRefineQueries: RefineQuerySuggestion[] | undefined;\n      if (useRaw && llmProcessor) {\n        const refineResult = await suggestRefineQueriesForRawMode(\n          aggregation.rankedUrls,\n          params.extract,\n          params.queries,\n          llmProcessor,\n        );\n        rawRefineQueries = refineResult.result;\n      }\n      markdown = appendSignalsAndFollowUps(\n        buildRawOutput(params.queries, aggregation, response.searches, params.verbose),\n        buildSignalsSection(aggregation, response.searches, response.totalQueries),\n        rawRefineQueries,\n        { includeSignals: params.verbose },\n      );\n      await reporter.progress(80, 100, 'Ranking search results');\n    } else {\n      // LLM classification path\n      await reporter.progress(65, 100, 'Classifying results by relevance');\n      const classification = await classifySearchResults(\n        aggregation.rankedUrls,\n        params.extract,\n        response.totalQueries,\n        llmProcessor,\n        params.queries,\n      );\n\n      if (classification.result) {\n        markdown = buildClassifiedOutput(\n          classification.result, aggregation, params.extract, response.searches, response.totalQueries, params.verbose,\n        );\n        llmClassified = true;\n        await reporter.progress(85, 100, 'Formatted classified results');\n      } else {\n        // Classification failed \u2014 fall back to raw\n        llmError = classification.error ?? 'Unknown classification error';\n        mcpLog('warning', `Classification failed, falling back to raw: ${llmError}`, 'search');\n        // mcp-revisions/llm-degradation/01: surface degraded mode to the client.\n        await reporter.log('warning', `llm_classifier_unreachable: ${llmError}`);\n        markdown = appendSignalsAndFollowUps(\n          buildRawOutput(params.queries, aggregation, response.searches, params.verbose),\n          buildSignalsSection(aggregation, response.searches, response.totalQueries),\n          undefined,\n          { includeSignals: params.verbose },\n        );\n        await reporter.progress(85, 100, 'Classification failed, using raw output');\n      }\n    }\n\n    const executionTime = Date.now() - startTime;\n    const metadata = buildMetadata(\n      aggregation, executionTime, response.totalQueries, response.searches, llmClassified, params.scope, llmError,\n      queryRewrites, retriedQueries,\n    );\n\n    // Build per-row structured results so capability-aware clients can\n    // index into `structuredContent.results` rather than regex-scrape the\n    // markdown table. The LLM tag wins when present; heuristic is the\n    // always-on fallback.\n    const llmTagsByRank = new Map<number, string>();\n    // (When classification succeeds the source_type per-row is populated\n    // inside buildClassifiedOutput via the entry.source_type field \u2014 but\n    // we don't have a direct handle on it here without a refactor. The\n    // heuristic alone covers the structuredContent shape correctly; the\n    // LLM-tagged variant remains in the markdown body.)\n    const results = buildStructuredResults(aggregation, llmTagsByRank);\n\n    mcpLog('info', `Search completed: ${aggregation.rankedUrls.length} URLs, classified=${llmClassified}`, 'search');\n    await reporter.log('info', `Search completed with ${aggregation.rankedUrls.length} URLs (classified: ${llmClassified})`);\n\n    const footer = `\\n---\\n*${formatDuration(executionTime)} | ${aggregation.totalUniqueUrls} unique URLs${llmClassified ? ' | LLM classified' : ''}*`;\n    const fullMarkdown = markdown + footer;\n\n    return toolSuccess(fullMarkdown, { results, metadata });\n  } catch (error) {\n    return buildWebSearchError(error, params, startTime);\n  }\n}\n\nexport function registerWebSearchTool(server: MCPServer): void {\n  server.tool(\n    {\n      name: 'web-search',\n      title: 'Web Search',\n      description:\n        'Fan out Google queries in parallel. One call carries up to 50 queries in a flat `queries` array \u2014 pack diverse facets (not paraphrases) into a single call. Call me AGGRESSIVELY across a session: 2\u20134 rounds is normal, 1 is underuse. After each pass, read `gaps[]` + `refine_queries[]` and fire another round with the new terms. Safe to call multiple times in parallel in the same turn for orthogonal subtopics. `scope`: `\"reddit\"` (server appends `site:reddit.com` + filters to post permalinks \u2014 use for sentiment / migration / lived experience), `\"web\"` default (spec / bug / pricing / CVE / API), `\"both\"` (fan each query across both \u2014 use when opinion-heavy AND needs official sources). Returns a tiered Markdown report (HIGHLY_RELEVANT / MAYBE_RELEVANT / OTHER) + grounded synthesis with `[rank]` citations + `## Gaps` + `## Suggested follow-up searches` tied to gap ids. Set `raw=true` to skip classification.',\n      schema: webSearchParamsSchema,\n      outputSchema: webSearchOutputSchema,\n      annotations: {\n        readOnlyHint: true,\n        idempotentHint: true,\n        destructiveHint: false,\n        openWorldHint: true,\n      },\n    },\n    async (args, ctx) => {\n      if (!getCapabilities().search) {\n        return toToolResponse(toolFailure(getMissingEnvMessage('search')));\n      }\n\n      const reporter = createToolReporter(ctx, 'web-search');\n      const result = await handleWebSearch(args, reporter);\n\n      await reporter.progress(100, 100, result.isError ? 'Search failed' : 'Search complete');\n      return toToolResponse(result);\n    },\n  );\n}\n"],
+  "mappings": "AAOA,SAAS,iBAAiB,4BAA4B;AACtD;AAAA,EACE;AAAA,EACA;AAAA,OAGK;AACP,SAAS,oBAAoB;AAC7B;AAAA,EACE;AAAA,EACA;AAAA,OACK;AACP;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OAIK;AACP,SAAS,qBAAqB;AAC9B,SAAS,2BAA2B;AACpC;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OACK;AACP;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OAGK;AACP,SAAS,0BAA0B;AACnC;AAAA,EACE;AAAA,EACA;AAAA,OACK;AAqBP,MAAM,wBAAwB;AAC9B,MAAM,cAAc;AAEpB,SAAS,wBAAwB,SAAmB,OAA4C;AAC9F,MAAI,UAAU,MAAO,QAAO;AAC5B,QAAM,WAAW,QAAQ;AAAA,IAAI,CAAC,MAC5B,wBAAwB,KAAK,CAAC,IAAI,IAAI,GAAG,CAAC;AAAA,EAC5C;AACA,SAAO,UAAU,WAAW,WAAW,CAAC,GAAG,SAAS,GAAG,QAAQ;AACjE;AAEA,eAAe,gBAAgB,SAA4C;AACzE,QAAM,SAAS,IAAI,aAAa;AAChC,SAAO,OAAO,eAAe,OAAO;AACtC;AAoBA,eAAe,sBACb,YACA,UACsE;AACtE,QAAM,UAAU,MAAM,gBAAgB,UAAU;AAEhD,QAAM,eAAe,QAAQ,SAC1B,IAAI,CAAC,GAAG,MAAO,EAAE,QAAQ,WAAW,IAAI,IAAI,EAAG,EAC/C,OAAO,CAAC,MAAM,MAAM,EAAE;AAEzB,MAAI,aAAa,WAAW,GAAG;AAC7B,WAAO,EAAE,UAAU,SAAS,SAAS,CAAC,EAAE;AAAA,EAC1C;AAGA,QAAM,QAAgB,CAAC;AACvB,aAAW,OAAO,cAAc;AAC9B,UAAM,KAAK,WAAW,GAAG;AACzB,QAAI,OAAO,OAAO,SAAU;AAC5B,UAAM,IAAI,mBAAmB,EAAE;AAC/B,QAAI,EAAE,WAAW,EAAE,cAAc,IAAI;AACnC,YAAM,KAAK,EAAE,OAAO,KAAK,UAAU,IAAI,SAAS,EAAE,WAAW,OAAO,CAAC,GAAG,EAAE,KAAK,EAAE,CAAC;AAAA,IACpF;AAAA,EACF;AAEA,MAAI,MAAM,WAAW,GAAG;AACtB,WAAO,EAAE,UAAU,SAAS,SAAS,CAAC,EAAE;AAAA,EAC1C;AAEA;AAAA,IACE;AAAA,IACA,GAAG,MAAM,MAAM,IAAI,aAAa,MAAM;AAAA,IACtC;AAAA,EACF;AACA,QAAM,SAAS;AAAA,IACb;AAAA,IACA,GAAG,MAAM,MAAM;AAAA,EACjB;AAEA,QAAM,YAAY,MAAM,gBAAgB,MAAM,IAAI,CAAC,MAAM,EAAE,OAAO,CAAC;AACnE,QAAM,UAAgC,CAAC;AACvC,QAAM,eAAe,oBAAI,IAAgD;AAEzE,QAAM,QAAQ,CAAC,MAAM,MAAM;AACzB,UAAM,IAAI,UAAU,SAAS,CAAC;AAC9B,QAAI,EAAG,cAAa,IAAI,KAAK,OAAO,CAAC;AACrC,YAAQ,KAAK;AAAA,MACX,UAAU,KAAK;AAAA,MACf,cAAc,KAAK;AAAA,MACnB,OAAO,KAAK;AAAA,MACZ,mBAAmB,GAAG,QAAQ,UAAU;AAAA,IAC1C,CAAC;AAAA,EACH,CAAC;AAED,QAAM,iBAAiB,QAAQ,SAAS,IAAI,CAAC,GAAG,QAAQ;AACtD,UAAM,IAAI,aAAa,IAAI,GAAG;AAC9B,QAAI,KAAK,EAAE,QAAQ,SAAS,GAAG;AAC7B,aAAO,EAAE,GAAG,GAAG,OAAO,EAAE,MAAM;AAAA,IAChC;AACA,WAAO;AAAA,EACT,CAAC;AAED,SAAO;AAAA,IACL,UAAU,EAAE,GAAG,SAAS,UAAU,eAAe;AAAA,IACjD;AAAA,EACF;AACF;AAEA,SAAS,qBACP,UACA,OACgB;AAChB,MAAI,UAAU,MAAO,QAAO;AAC5B,QAAM,WAAW,SAAS,SAAS,IAAI,CAAC,YAAY;AAAA,IAClD,GAAG;AAAA,IACH,SAAS,OAAO,QAAQ,OAAO,CAAC,MAAM;AACpC,UAAI;AACJ,UAAI;AAAE,eAAO,IAAI,IAAI,EAAE,IAAI,EAAE;AAAA,MAAU,QAAQ;AAAE,eAAO;AAAA,MAAM;AAE9D,UAAI,CAAC,YAAY,KAAK,IAAI,EAAG,QAAO,UAAU;AAC9C,aAAO,sBAAsB,KAAK,EAAE,IAAI;AAAA,IAC1C,CAAC;AAAA,EACH,EAAE;AACF,SAAO,EAAE,GAAG,UAAU,UAAU,SAAS;AAC3C;AAEA,SAAS,eAAe,UAEtB;AACA,QAAM,cAAc,iBAAiB,SAAS,UAAU,CAAC;AACzD,SAAO,EAAE,YAAY;AACvB;AAIA,SAAS,eACP,SACA,aACA,UACA,UAAmB,OACX;AACR,SAAO;AAAA,IACL,YAAY;AAAA,IAAY;AAAA,IAAS;AAAA,IACjC,YAAY;AAAA,IACZ,YAAY;AAAA,IAAoB,YAAY;AAAA,IAC5C;AAAA,EACF;AACF;AAEA,SAAS,oBACP,aACA,UACA,cACQ;AACR,QAAM,gBAAgB,SAAS,OAAO,CAAC,WAAW,OAAO,QAAQ,UAAU,CAAC,EAAE;AAC9E,QAAM,WAAW,SACd,OAAO,CAAC,WAAW,OAAO,QAAQ,UAAU,CAAC,EAC7C,IAAI,CAAC,WAAW,IAAI,OAAO,KAAK,GAAG;AACtC,QAAM,iBAAiB,YAAY,WAAW,OAAO,CAAC,QAAQ,IAAI,WAAW,EAAE;AAE/E,QAAM,QAAQ;AAAA,IACZ;AAAA,IACA,eAAe,aAAa,IAAI,YAAY;AAAA,IAC5C,qBAAqB,cAAc;AAAA,EACrC;AAEA,MAAI,SAAS,SAAS,GAAG;AACvB,UAAM,KAAK,gBAAgB,SAAS,KAAK,IAAI,CAAC,EAAE;AAAA,EAClD;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAEO,SAAS,+BACd,eACQ;AACR,MAAI,CAAC,iBAAiB,cAAc,WAAW,GAAG;AAChD,WAAO;AAAA,EACT;AAEA,QAAM,QAAQ,CAAC,mCAAmC,EAAE;AAEpD,aAAW,QAAQ,eAAe;AAChC,UAAM,QAAQ,mBAAmB,KAAK,SAAS,EAAE;AACjD,QAAI,CAAC,MAAO;AACZ,UAAM,YAAY,mBAAmB,KAAK,aAAa,EAAE;AACzD,UAAM,SAAS,OAAO,KAAK,WAAW,WAClC,kBAAkB,KAAK,MAAM,QAC7B,KAAK,kBACH,MAAM,mBAAmB,KAAK,eAAe,CAAC,OAC9C;AACN,UAAM;AAAA,MAAK,YACP,KAAK,KAAK,WAAM,SAAS,GAAG,MAAM,KAClC,KAAK,KAAK,GAAG,MAAM;AAAA,IACvB;AAAA,EACF;AAEA,SAAO,MAAM,WAAW,IAAI,KAAK,MAAM,KAAK,IAAI;AAClD;AAEO,SAAS,0BACd,UACA,gBACA,eACA,UAAwC,CAAC,GACjC;AACR,QAAM,iBAAiB,QAAQ,kBAAkB;AACjD,QAAM,WAAW,CAAC,QAAQ;AAC1B,MAAI,kBAAkB,gBAAgB;AACpC,aAAS,KAAK,IAAI,OAAO,cAAc;AAAA,EACzC;AACA,QAAM,YAAY,+BAA+B,aAAa;AAC9D,MAAI,WAAW;AACb,aAAS,KAAK,IAAI,SAAS;AAAA,EAC7B;AACA,SAAO,SAAS,KAAK,IAAI;AAC3B;AAYA,MAAM,iBAAiB;AACvB,MAAM,iBAAiB;AAchB,SAAS,sBACd,OACA,aACA,OAAuC,CAAC,GAChC;AACR,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,MAAM,KAAK,OAAO;AAExB,QAAM,QAA8F,CAAC;AAErG,aAAW,aAAa,MAAM,MAAM;AAClC,QAAI,MAAM,UAAU,IAAK;AACzB,UAAM,KAAK,EAAE,WAAW,MAAM,kBAAkB,CAAC;AAAA,EACnD;AAEA,MAAI,MAAM,SAAS,KAAK;AACtB,UAAM,SAAS,KAAK,IAAI,KAAK,GAAG;AAChC,eAAW,aAAa,MAAM,OAAO;AACnC,UAAI,MAAM,UAAU,OAAQ;AAC5B,YAAM,KAAK,EAAE,WAAW,MAAM,iBAAiB,CAAC;AAAA,IAClD;AAAA,EACF;AAEA,MAAI,MAAM,WAAW,EAAG,QAAO;AAE/B,QAAM,QAAkB,CAAC;AACzB,QAAM,KAAK,uDAAkD;AAC7D,QAAM,QAAQ,CAAC,MAAM,MAAM;AACzB,UAAM,QAAQ,YAAY,IAAI,KAAK,UAAU,IAAI;AACjD,UAAM,SAAS,OAAO,UAAU,MAAM,OAAO,KAAK,EAAE,SAAS,IAAI,MAAM,SAAS;AAChF,QAAI;AACJ,QAAI;AACF,eAAS,IAAI,IAAI,KAAK,UAAU,GAAG,EAAE,SAAS,QAAQ,UAAU,EAAE;AAAA,IACpE,QAAQ;AACN,eAAS,KAAK,UAAU;AAAA,IAC1B;AACA,UAAM;AAAA,MACJ,GAAG,IAAI,CAAC,QAAQ,KAAK,UAAU,KAAK,KAAK,KAAK,UAAU,GAAG,cAAS,MAAM,WAAM,MAAM,MAAM,KAAK,IAAI,UAAU,KAAK,UAAU,IAAI;AAAA,IACpI;AAAA,EACF,CAAC;AACD,SAAO,MAAM,KAAK,IAAI;AACxB;AAIA,SAAS,sBACP,gBACA,aACA,SACA,UACA,cACA,UAAmB,OACX;AACR,QAAM,aAAa,YAAY;AAG/B,QAAM,cAAc,IAAI,IAAI,eAAe,QAAQ,IAAI,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC;AAE1E,QAAM,QAAQ;AAAA,IACZ,MAAM,CAAC;AAAA,IACP,OAAO,CAAC;AAAA,IACR,OAAO,CAAC;AAAA,EACV;AAEA,aAAW,OAAO,YAAY;AAC5B,UAAM,QAAQ,YAAY,IAAI,IAAI,IAAI;AACtC,UAAM,OAAO,OAAO;AACpB,QAAI,SAAS,mBAAmB;AAC9B,YAAM,KAAK,KAAK,GAAG;AAAA,IACrB,WAAW,SAAS,kBAAkB;AACpC,YAAM,MAAM,KAAK,GAAG;AAAA,IACtB,OAAO;AACL,YAAM,MAAM,KAAK,GAAG;AAAA,IACtB;AAAA,EACF;AAEA,QAAM,QAAkB,CAAC;AAGzB,QAAM,KAAK,MAAM,eAAe,KAAK,EAAE;AACvC,QAAM,KAAK,kBAAkB,OAAO,EAAE;AACtC,QAAM,KAAK,KAAK,YAAY,mBAAc,WAAW,MAAM,gBAAW,MAAM,KAAK,MAAM,qBAAqB,MAAM,MAAM,MAAM,oBAAoB;AAClJ,MAAI,eAAe,YAAY;AAC7B,UAAM,aAAa,eAAe,oBAAoB,WAAM,eAAe,iBAAiB,KAAK;AACjG,UAAM,KAAK,mBAAmB,eAAe,UAAU,KAAK,UAAU,EAAE;AAAA,EAC1E;AACA,QAAM,KAAK,EAAE;AAIb,QAAM,YAAY;AAAA,IAChB,EAAE,MAAM,MAAM,MAAM,OAAO,MAAM,MAAM;AAAA,IACvC;AAAA,EACF;AACA,MAAI,WAAW;AACb,UAAM,KAAK,SAAS;AACpB,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,KAAK,gBAAgB,eAAe,SAAS,EAAE;AACrD,QAAM,KAAK,EAAE;AAGb,QAAM,gBAAgB,CAAC,QAA2C;AAChE,UAAM,QAAQ,YAAY,IAAI,IAAI,IAAI;AACtC,UAAM,cAAc,KAAK,MAAM,IAAI,gBAAgB,GAAG;AACtD,UAAM,SAAS,GAAG,IAAI,SAAS,IAAI,YAAY,KAAK,WAAW;AAC/D,UAAM,aAAa,OAAO,cAAc,KAAK,MAAM,WAAW,OAAO;AACrE,UAAM,SAAS,OAAO,SAAS,MAAM,OAAO,QAAQ,OAAO,KAAK,IAAI;AACpE,WAAO,KAAK,IAAI,IAAI,OAAO,IAAI,KAAK,KAAK,IAAI,GAAG,OAAO,UAAU,MAAM,MAAM,MAAM,MAAM;AAAA,EAC3F;AAGA,MAAI,MAAM,KAAK,SAAS,GAAG;AACzB,UAAM,KAAK,wBAAwB,MAAM,KAAK,MAAM,GAAG;AACvD,UAAM,KAAK,sCAAsC;AACjD,UAAM,KAAK,sCAAsC;AACjD,eAAW,OAAO,MAAM,KAAM,OAAM,KAAK,cAAc,GAAG,CAAC;AAC3D,UAAM,KAAK,EAAE;AAAA,EACf;AAGA,MAAI,MAAM,MAAM,SAAS,GAAG;AAC1B,UAAM,KAAK,uBAAuB,MAAM,MAAM,MAAM,GAAG;AACvD,UAAM,KAAK,sCAAsC;AACjD,UAAM,KAAK,sCAAsC;AACjD,eAAW,OAAO,MAAM,MAAO,OAAM,KAAK,cAAc,GAAG,CAAC;AAC5D,UAAM,KAAK,EAAE;AAAA,EACf;AAGA,MAAI,MAAM,MAAM,SAAS,GAAG;AAC1B,UAAM,KAAK,sBAAsB,MAAM,MAAM,MAAM,GAAG;AACtD,UAAM,KAAK,wCAAwC;AACnD,UAAM,KAAK,wCAAwC;AACnD,eAAW,OAAO,MAAM,OAAO;AAC7B,YAAM,QAAQ,YAAY,IAAI,IAAI,IAAI;AACtC,YAAM,YAAY,IAAI,QAAQ,IAAI,CAAC,MAAM,IAAI,CAAC,GAAG,EAAE,KAAK,IAAI;AAC5D,YAAM,aAAa,OAAO,cAAc,KAAK,MAAM,WAAW,OAAO;AACrE,UAAI;AACJ,UAAI;AACF,iBAAS,IAAI,IAAI,IAAI,GAAG,EAAE,SAAS,QAAQ,UAAU,EAAE;AAAA,MACzD,QAAQ;AACN,iBAAS,IAAI;AAAA,MACf;AACA,YAAM,KAAK,KAAK,IAAI,IAAI,MAAM,MAAM,MAAM,UAAU,MAAM,IAAI,MAAM,QAAQ,CAAC,CAAC,MAAM,SAAS,IAAI;AAAA,IACnG;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAKA,MAAI,SAAS;AACX,UAAM,KAAK,oBAAoB,aAAa,UAAU,YAAY,CAAC;AAAA,EACrE;AAGA,MAAI,eAAe,QAAQ,eAAe,KAAK,SAAS,GAAG;AACzD,UAAM,KAAK,EAAE;AACb,UAAM,KAAK,SAAS;AACpB,eAAW,OAAO,eAAe,MAAM;AACrC,YAAM,KAAK,QAAQ,IAAI,EAAE,OAAO,IAAI,WAAW,EAAE;AAAA,IACnD;AAAA,EACF;AAEA,QAAM,YAAY,+BAA+B,eAAe,cAAc;AAC9E,MAAI,WAAW;AACb,UAAM,KAAK,EAAE;AACb,UAAM,KAAK,SAAS;AAAA,EACtB;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAIA,SAAS,cACP,aACA,eACA,cACA,UACA,eACA,OACA,UACA,eACA,gBACA;AACA,QAAM,kBAAkB,SAAS,IAAI,OAAK;AACxC,QAAI;AACJ,UAAM,YAAY,EAAE,QAAQ,CAAC;AAC7B,QAAI,WAAW;AACb,UAAI;AAAE,oBAAY,IAAI,IAAI,UAAU,IAAI,EAAE,SAAS,QAAQ,UAAU,EAAE;AAAA,MAAG,QAAQ;AAAA,MAAe;AAAA,IACnG;AACA,WAAO,EAAE,OAAO,EAAE,OAAO,cAAc,EAAE,QAAQ,QAAQ,SAAS,UAAU;AAAA,EAC9E,CAAC;AACD,QAAM,kBAAkB,SACrB,OAAO,OAAK,EAAE,QAAQ,UAAU,CAAC,EACjC,IAAI,OAAK,EAAE,KAAK;AAEnB,SAAO;AAAA,IACL,aAAa;AAAA,IACb,YAAY,YAAY,WAAW;AAAA,IACnC,QAAQ,eAAe,SAAS,OAAO,OAAK,EAAE,QAAQ,SAAS,CAAC,EAAE;AAAA,IAClE,mBAAmB;AAAA,IACnB,gBAAgB;AAAA,IAChB;AAAA,IACA,GAAI,WAAW,EAAE,WAAW,SAAS,IAAI,CAAC;AAAA,IAC1C,kBAAkB;AAAA,IAClB,GAAI,gBAAgB,SAAS,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;AAAA,IAC3E,GAAI,iBAAiB,cAAc,SAAS,IAAI,EAAE,gBAAgB,cAAc,IAAI,CAAC;AAAA,IACrF,GAAI,kBAAkB,eAAe,SAAS,IAAI,EAAE,iBAAiB,eAAe,IAAI,CAAC;AAAA,EAC3F;AACF;AAEA,SAAS,uBACP,aACA,eAUC;AACD,SAAO,YAAY,WAAW,IAAI,CAAC,QAAQ;AAGzC,UAAM,SAAS,eAAe,IAAI,IAAI,IAAI;AAC1C,UAAM,YAAY,oBAAoB,IAAI,GAAG;AAC7C,WAAO;AAAA,MACL,MAAM,IAAI;AAAA,MACV,KAAK,IAAI;AAAA,MACT,OAAO,IAAI;AAAA,MACX,SAAS,IAAI;AAAA,MACb,aAAe,UAA+B;AAAA,MAC9C,OAAO,OAAO,IAAI,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClC,SAAS,IAAI;AAAA,MACb,eAAe,IAAI;AAAA,IACrB;AAAA,EACF,CAAC;AACH;AAIA,SAAS,oBACP,OACA,QACA,WACsC;AACtC,QAAM,kBAAkB,cAAc,KAAK;AAC3C,QAAM,gBAAgB,KAAK,IAAI,IAAI;AAEnC,SAAO,SAAS,eAAe,gBAAgB,OAAO,IAAI,QAAQ;AAElE,QAAM,eAAe,YAAY;AAAA,IAC/B,MAAM,gBAAgB;AAAA,IACtB,SAAS,gBAAgB;AAAA,IACzB,WAAW,gBAAgB;AAAA,IAC3B,UAAU;AAAA,IACV,UAAU,CAAC,wCAAwC;AAAA,IACnD,cAAc;AAAA,MACZ;AAAA,MACA;AAAA,IACF;AAAA,EACF,CAAC;AAED,SAAO;AAAA,IACL,GAAG,YAAY;AAAA;AAAA,kBAAuB,eAAe,aAAa,CAAC;AAAA,WAAc,OAAO,QAAQ,MAAM;AAAA,EACxG;AACF;AAIA,eAAsB,gBACpB,QACA,WAAyB,eACsB;AAC/C,QAAM,YAAY,KAAK,IAAI;AAE3B,MAAI;AACF,UAAM,mBAAmB,wBAAwB,OAAO,SAAS,OAAO,KAAK;AAC7E,QAAI,OAAO,UAAU,OAAO;AAC1B,aAAO,QAAQ,mBAAmB,OAAO,KAAK,KAAK,OAAO,QAAQ,MAAM,yBAAoB,iBAAiB,MAAM,eAAe,QAAQ;AAAA,IAC5I,OAAO;AACL,aAAO,QAAQ,iBAAiB,OAAO,QAAQ,MAAM,kBAAkB,QAAQ;AAAA,IACjF;AACA,UAAM,SAAS,IAAI,QAAQ,iBAAiB,iBAAiB,MAAM,yBAAyB,OAAO,KAAK,GAAG;AAC3G,UAAM,SAAS,SAAS,IAAI,KAAK,2BAA2B;AAK5D,UAAM,eAAe,iBAAiB,IAAI,CAAC,MAAM;AAC/C,YAAM,IAAI,0BAA0B,CAAC;AACrC,aAAO,EAAE,UAAU,GAAG,YAAY,EAAE,WAAW,OAAO,CAAC,GAAG,EAAE,KAAK,GAAG,SAAS,EAAE,QAAQ;AAAA,IACzF,CAAC;AACD,UAAM,oBAAoB,aAAa,IAAI,CAAC,MAAM,EAAE,UAAU;AAC9D,UAAM,gBAAsC,aACzC,OAAO,CAAC,MAAM,EAAE,OAAO,EACvB,IAAI,CAAC,OAAO,EAAE,UAAU,EAAE,UAAU,WAAW,EAAE,YAAY,OAAO,EAAE,MAAM,EAAE;AAEjF,QAAI,cAAc,SAAS,GAAG;AAC5B;AAAA,QACE;AAAA,QACA,2BAA2B,cAAc,MAAM,IAAI,iBAAiB,MAAM;AAAA,QAC1E;AAAA,MACF;AACA,YAAM,SAAS;AAAA,QACb;AAAA,QACA,cAAc,cAAc,MAAM;AAAA,MACpC;AAAA,IACF;AAKA,UAAM,EAAE,UAAU,aAAa,SAAS,eAAe,IAAI,MAAM;AAAA,MAC/D;AAAA,MACA;AAAA,IACF;AACA,UAAM,WAAW,qBAAqB,aAAa,OAAO,KAAK;AAC/D,UAAM,SAAS,SAAS,IAAI,KAAK,0BAA0B;AAE3D,UAAM,EAAE,YAAY,IAAI,eAAe,QAAQ;AAC/C,UAAM,SAAS;AAAA,MACb;AAAA,MACA,aAAa,YAAY,eAAe,uBAAuB,SAAS,YAAY;AAAA,IACtF;AAGA,UAAM,SAAS,OAAO;AACtB,UAAM,eAAe,mBAAmB;AAExC,QAAI;AACJ,QAAI,gBAAgB;AACpB,QAAI;AAEJ,QAAI,UAAU,CAAC,cAAc;AAE3B,UAAI,CAAC,UAAU,CAAC,cAAc;AAC5B,mBAAW;AACX,eAAO,WAAW,UAAU,QAAQ;AAEpC,cAAM,SAAS,IAAI,WAAW,8EAA8E;AAAA,MAC9G;AACA,UAAI;AACJ,UAAI,UAAU,cAAc;AAC1B,cAAM,eAAe,MAAM;AAAA,UACzB,YAAY;AAAA,UACZ,OAAO;AAAA,UACP,OAAO;AAAA,UACP;AAAA,QACF;AACA,2BAAmB,aAAa;AAAA,MAClC;AACA,iBAAW;AAAA,QACT,eAAe,OAAO,SAAS,aAAa,SAAS,UAAU,OAAO,OAAO;AAAA,QAC7E,oBAAoB,aAAa,SAAS,UAAU,SAAS,YAAY;AAAA,QACzE;AAAA,QACA,EAAE,gBAAgB,OAAO,QAAQ;AAAA,MACnC;AACA,YAAM,SAAS,SAAS,IAAI,KAAK,wBAAwB;AAAA,IAC3D,OAAO;AAEL,YAAM,SAAS,SAAS,IAAI,KAAK,kCAAkC;AACnE,YAAM,iBAAiB,MAAM;AAAA,QAC3B,YAAY;AAAA,QACZ,OAAO;AAAA,QACP,SAAS;AAAA,QACT;AAAA,QACA,OAAO;AAAA,MACT;AAEA,UAAI,eAAe,QAAQ;AACzB,mBAAW;AAAA,UACT,eAAe;AAAA,UAAQ;AAAA,UAAa,OAAO;AAAA,UAAS,SAAS;AAAA,UAAU,SAAS;AAAA,UAAc,OAAO;AAAA,QACvG;AACA,wBAAgB;AAChB,cAAM,SAAS,SAAS,IAAI,KAAK,8BAA8B;AAAA,MACjE,OAAO;AAEL,mBAAW,eAAe,SAAS;AACnC,eAAO,WAAW,+CAA+C,QAAQ,IAAI,QAAQ;AAErF,cAAM,SAAS,IAAI,WAAW,+BAA+B,QAAQ,EAAE;AACvE,mBAAW;AAAA,UACT,eAAe,OAAO,SAAS,aAAa,SAAS,UAAU,OAAO,OAAO;AAAA,UAC7E,oBAAoB,aAAa,SAAS,UAAU,SAAS,YAAY;AAAA,UACzE;AAAA,UACA,EAAE,gBAAgB,OAAO,QAAQ;AAAA,QACnC;AACA,cAAM,SAAS,SAAS,IAAI,KAAK,yCAAyC;AAAA,MAC5E;AAAA,IACF;AAEA,UAAM,gBAAgB,KAAK,IAAI,IAAI;AACnC,UAAM,WAAW;AAAA,MACf;AAAA,MAAa;AAAA,MAAe,SAAS;AAAA,MAAc,SAAS;AAAA,MAAU;AAAA,MAAe,OAAO;AAAA,MAAO;AAAA,MACnG;AAAA,MAAe;AAAA,IACjB;AAMA,UAAM,gBAAgB,oBAAI,IAAoB;AAM9C,UAAM,UAAU,uBAAuB,aAAa,aAAa;AAEjE,WAAO,QAAQ,qBAAqB,YAAY,WAAW,MAAM,qBAAqB,aAAa,IAAI,QAAQ;AAC/G,UAAM,SAAS,IAAI,QAAQ,yBAAyB,YAAY,WAAW,MAAM,sBAAsB,aAAa,GAAG;AAEvH,UAAM,SAAS;AAAA;AAAA,GAAW,eAAe,aAAa,CAAC,MAAM,YAAY,eAAe,eAAe,gBAAgB,sBAAsB,EAAE;AAC/I,UAAM,eAAe,WAAW;AAEhC,WAAO,YAAY,cAAc,EAAE,SAAS,SAAS,CAAC;AAAA,EACxD,SAAS,OAAO;AACd,WAAO,oBAAoB,OAAO,QAAQ,SAAS;AAAA,EACrD;AACF;AAEO,SAAS,sBAAsB,QAAyB;AAC7D,SAAO;AAAA,IACL;AAAA,MACE,MAAM;AAAA,MACN,OAAO;AAAA,MACP,aACE;AAAA,MACF,QAAQ;AAAA,MACR,cAAc;AAAA,MACd,aAAa;AAAA,QACX,cAAc;AAAA,QACd,gBAAgB;AAAA,QAChB,iBAAiB;AAAA,QACjB,eAAe;AAAA,MACjB;AAAA,IACF;AAAA,IACA,OAAO,MAAM,QAAQ;AACnB,UAAI,CAAC,gBAAgB,EAAE,QAAQ;AAC7B,eAAO,eAAe,YAAY,qBAAqB,QAAQ,CAAC,CAAC;AAAA,MACnE;AAEA,YAAM,WAAW,mBAAmB,KAAK,YAAY;AACrD,YAAM,SAAS,MAAM,gBAAgB,MAAM,QAAQ;AAEnD,YAAM,SAAS,SAAS,KAAK,KAAK,OAAO,UAAU,kBAAkB,iBAAiB;AACtF,aAAO,eAAe,MAAM;AAAA,IAC9B;AAAA,EACF;AACF;",
   "names": []
 }

package/dist/src/utils/query-relax.js

@@ -0,0 +1,91 @@
+const QUOTED_PHRASE_RE = /"([^"]*)"/g;
+const HAS_BOOLEAN_GROUPING = /\b(?:OR|AND)\b|[()]/;
+const OPERATOR_CHAR_IN_PHRASE = /[():[\]]/;
+const OPERATOR_CHAR_GLOBAL = /[():[\]]/g;
+const PATH_LIKE_IN_PHRASE = /\/|~\/|^@|\.{3,}/;
+const HAS_SITE_OPERATOR = /\bsite:\S+/i;
+const SITE_OPERATOR_GLOBAL = /\bsite:\S+/gi;
+function tokenize(query) {
+  const segs = [];
+  let last = 0;
+  for (const m of query.matchAll(QUOTED_PHRASE_RE)) {
+    const start = m.index ?? 0;
+    const end = start + m[0].length;
+    if (start > last) segs.push({ type: "raw", text: query.slice(last, start) });
+    segs.push({ type: "phrase", text: m[1] ?? "", quoted: true });
+    last = end;
+  }
+  if (last < query.length) segs.push({ type: "raw", text: query.slice(last) });
+  return segs;
+}
+function rebuild(segs) {
+  return segs.map((s) => s.type === "raw" ? s.text : s.quoted ? `"${s.text}"` : s.text).join("").replace(/\s+/g, " ").trim();
+}
+function normalizeQueryForDispatch(query) {
+  const original = query.trim().replace(/\s+/g, " ");
+  if (!original) {
+    return { rewritten: original, changed: false, rules: [] };
+  }
+  const segs = tokenize(query);
+  const rules = [];
+  for (const s of segs) {
+    if (s.type === "phrase" && s.quoted && OPERATOR_CHAR_IN_PHRASE.test(s.text)) {
+      s.quoted = false;
+      s.text = s.text.replace(OPERATOR_CHAR_GLOBAL, " ");
+      if (!rules.includes("A1")) rules.push("A1");
+    }
+  }
+  for (const s of segs) {
+    if (s.type === "phrase" && s.quoted && PATH_LIKE_IN_PHRASE.test(s.text)) {
+      s.quoted = false;
+      if (!rules.includes("A2")) rules.push("A2");
+    }
+  }
+  const stillQuoted = segs.filter(
+    (s) => s.type === "phrase" && s.quoted
+  );
+  const rawJoined = segs.filter((s) => s.type === "raw").map((s) => s.text).join(" ");
+  if (stillQuoted.length >= 3 && !HAS_BOOLEAN_GROUPING.test(rawJoined)) {
+    let phraseCount = 0;
+    let modified = false;
+    for (let i = 0; i < segs.length; i += 1) {
+      const s = segs[i];
+      if (s && s.type === "phrase" && s.quoted) {
+        phraseCount += 1;
+        if (phraseCount >= 3) {
+          const prev = segs[i - 1];
+          if (prev && prev.type === "raw" && prev.text.trim() === "") {
+            prev.text = " OR ";
+            modified = true;
+          }
+        }
+      }
+    }
+    if (modified) rules.push("A3");
+  }
+  const rewritten = rebuild(segs);
+  return { rewritten, changed: rewritten !== original, rules };
+}
+function relaxQueryForRetry(query) {
+  const original = query.trim().replace(/\s+/g, " ");
+  if (!original) {
+    return { rewritten: original, changed: false, rules: [] };
+  }
+  const rules = [];
+  let result = query;
+  if (result.includes('"')) {
+    result = result.replace(/"/g, "");
+    rules.push("B1");
+  }
+  if (HAS_SITE_OPERATOR.test(result)) {
+    result = result.replace(SITE_OPERATOR_GLOBAL, " ");
+    rules.push("B2");
+  }
+  result = result.replace(/\s+/g, " ").trim();
+  return { rewritten: result, changed: result !== original, rules };
+}
+export {
+  normalizeQueryForDispatch,
+  relaxQueryForRetry
+};
+//# sourceMappingURL=query-relax.js.map

package/dist/src/utils/query-relax.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": ["../../../src/utils/query-relax.ts"],
+  "sourcesContent": ["/**\n * Query relaxation for web-search.\n *\n * Two-phase rewriter that addresses systematic Google query failures observed\n * in the Serper log:\n *  - 3+ back-to-back quoted phrases get implicit-AND'd by Google \u2192 no page\n *    contains all rare tokens \u2192 0 results.\n *  - Quoted phrases with operator chars (parens, colons, brackets) \u2014 Google\n *    strips them inside quotes, so the quotes only impose a pointless AND.\n *  - Quoted paths/URLs (`/`, `~/`, leading `@`, 3+ dots) \u2014 same reason.\n *  - Tiny `site:` corpus that returns 0 \u2014 drop site filter on retry.\n *  - Verbatim long phrases that don't appear word-for-word \u2014 strip quotes on retry.\n *\n * Phase A (`normalizeQueryForDispatch`) is always-on, deterministic, lossless.\n * Phase B (`relaxQueryForRetry`) is aggressive; only invoke when Phase A's\n * dispatched form returned zero results.\n */\n\nconst QUOTED_PHRASE_RE = /\"([^\"]*)\"/g;\nconst HAS_BOOLEAN_GROUPING = /\\b(?:OR|AND)\\b|[()]/;\nconst OPERATOR_CHAR_IN_PHRASE = /[():[\\]]/;\nconst OPERATOR_CHAR_GLOBAL = /[():[\\]]/g;\nconst PATH_LIKE_IN_PHRASE = /\\/|~\\/|^@|\\.{3,}/;\nconst HAS_SITE_OPERATOR = /\\bsite:\\S+/i;\nconst SITE_OPERATOR_GLOBAL = /\\bsite:\\S+/gi;\n\nexport interface RewriteResult {\n  rewritten: string;\n  changed: boolean;\n  rules: string[];\n}\n\ninterface PhraseSeg { type: 'phrase'; text: string; quoted: boolean }\ninterface RawSeg { type: 'raw'; text: string }\ntype Seg = PhraseSeg | RawSeg;\n\nfunction tokenize(query: string): Seg[] {\n  const segs: Seg[] = [];\n  let last = 0;\n  for (const m of query.matchAll(QUOTED_PHRASE_RE)) {\n    const start = m.index ?? 0;\n    const end = start + m[0].length;\n    if (start > last) segs.push({ type: 'raw', text: query.slice(last, start) });\n    segs.push({ type: 'phrase', text: m[1] ?? '', quoted: true });\n    last = end;\n  }\n  if (last < query.length) segs.push({ type: 'raw', text: query.slice(last) });\n  return segs;\n}\n\nfunction rebuild(segs: Seg[]): string {\n  return segs\n    .map((s) => (s.type === 'raw' ? s.text : s.quoted ? `\"${s.text}\"` : s.text))\n    .join('')\n    .replace(/\\s+/g, ' ')\n    .trim();\n}\n\n/**\n * Phase A \u2014 pre-dispatch normalizer. Runs on every query before it leaves the\n * server. Three rules apply in order; each only fires when the original query\n * was statistically going to mis-handle in Google anyway.\n *\n * Rule A1: phrase contains `(`, `)`, `:`, `[`, `]` \u2192 drop quotes; replace those\n *          chars with space (Google strips them inside quotes anyway).\n * Rule A2: phrase contains `/`, `~/`, leading `@`, or 3+ dots \u2192 drop quotes only\n *          (Google tokenizes path separators in or out of quotes).\n * Rule A3: \u22653 still-quoted phrases AND no existing `OR`/`AND`/parens \u2192 keep\n *          first phrase as anchor; replace whitespace between consecutive\n *          subsequent phrases with ` OR `. Bare words and `site:`/`filetype:`\n *          operators are preserved verbatim.\n */\nexport function normalizeQueryForDispatch(query: string): RewriteResult {\n  const original = query.trim().replace(/\\s+/g, ' ');\n  if (!original) {\n    return { rewritten: original, changed: false, rules: [] };\n  }\n  const segs = tokenize(query);\n  const rules: string[] = [];\n\n  // A1 \u2014 operator chars inside quotes are pointless AND constraints.\n  for (const s of segs) {\n    if (s.type === 'phrase' && s.quoted && OPERATOR_CHAR_IN_PHRASE.test(s.text)) {\n      s.quoted = false;\n      s.text = s.text.replace(OPERATOR_CHAR_GLOBAL, ' ');\n      if (!rules.includes('A1')) rules.push('A1');\n    }\n  }\n\n  // A2 \u2014 path/URL inside quotes; quoting doesn't help recall.\n  for (const s of segs) {\n    if (s.type === 'phrase' && s.quoted && PATH_LIKE_IN_PHRASE.test(s.text)) {\n      s.quoted = false;\n      if (!rules.includes('A2')) rules.push('A2');\n    }\n  }\n\n  // A3 \u2014 phrase-AND collapse. Trigger requires \u22653 phrases that survive A1+A2,\n  // and no existing boolean grouping in the raw (non-quoted) part of the query.\n  const stillQuoted = segs.filter(\n    (s): s is PhraseSeg => s.type === 'phrase' && s.quoted,\n  );\n  const rawJoined = segs\n    .filter((s): s is RawSeg => s.type === 'raw')\n    .map((s) => s.text)\n    .join(' ');\n\n  if (stillQuoted.length >= 3 && !HAS_BOOLEAN_GROUPING.test(rawJoined)) {\n    let phraseCount = 0;\n    let modified = false;\n    for (let i = 0; i < segs.length; i += 1) {\n      const s = segs[i];\n      if (s && s.type === 'phrase' && s.quoted) {\n        phraseCount += 1;\n        if (phraseCount >= 3) {\n          const prev = segs[i - 1];\n          if (prev && prev.type === 'raw' && prev.text.trim() === '') {\n            prev.text = ' OR ';\n            modified = true;\n          }\n        }\n      }\n    }\n    if (modified) rules.push('A3');\n  }\n\n  const rewritten = rebuild(segs);\n  return { rewritten, changed: rewritten !== original, rules };\n}\n\n/**\n * Phase B \u2014 on-empty retry. Only invoked for queries whose Phase-A dispatched\n * form returned zero results from Serper. Strips ALL remaining quotes and the\n * `site:` operator (if present). Caller should skip the retry when the\n * relaxed form equals the dispatched form.\n *\n * Rule B1: strip every `\"` (turns each phrase into a bag of words; Google\n *          ranks by token co-occurrence instead of forcing verbatim match).\n * Rule B2: drop `site:operator` (broadens to open web; catches \"tiny corpus\"\n *          and \"site path doesn't exist\" cases at once).\n */\nexport function relaxQueryForRetry(query: string): RewriteResult {\n  const original = query.trim().replace(/\\s+/g, ' ');\n  if (!original) {\n    return { rewritten: original, changed: false, rules: [] };\n  }\n  const rules: string[] = [];\n  let result = query;\n\n  if (result.includes('\"')) {\n    result = result.replace(/\"/g, '');\n    rules.push('B1');\n  }\n\n  if (HAS_SITE_OPERATOR.test(result)) {\n    result = result.replace(SITE_OPERATOR_GLOBAL, ' ');\n    rules.push('B2');\n  }\n\n  result = result.replace(/\\s+/g, ' ').trim();\n  return { rewritten: result, changed: result !== original, rules };\n}\n"],
+  "mappings": "AAkBA,MAAM,mBAAmB;AACzB,MAAM,uBAAuB;AAC7B,MAAM,0BAA0B;AAChC,MAAM,uBAAuB;AAC7B,MAAM,sBAAsB;AAC5B,MAAM,oBAAoB;AAC1B,MAAM,uBAAuB;AAY7B,SAAS,SAAS,OAAsB;AACtC,QAAM,OAAc,CAAC;AACrB,MAAI,OAAO;AACX,aAAW,KAAK,MAAM,SAAS,gBAAgB,GAAG;AAChD,UAAM,QAAQ,EAAE,SAAS;AACzB,UAAM,MAAM,QAAQ,EAAE,CAAC,EAAE;AACzB,QAAI,QAAQ,KAAM,MAAK,KAAK,EAAE,MAAM,OAAO,MAAM,MAAM,MAAM,MAAM,KAAK,EAAE,CAAC;AAC3E,SAAK,KAAK,EAAE,MAAM,UAAU,MAAM,EAAE,CAAC,KAAK,IAAI,QAAQ,KAAK,CAAC;AAC5D,WAAO;AAAA,EACT;AACA,MAAI,OAAO,MAAM,OAAQ,MAAK,KAAK,EAAE,MAAM,OAAO,MAAM,MAAM,MAAM,IAAI,EAAE,CAAC;AAC3E,SAAO;AACT;AAEA,SAAS,QAAQ,MAAqB;AACpC,SAAO,KACJ,IAAI,CAAC,MAAO,EAAE,SAAS,QAAQ,EAAE,OAAO,EAAE,SAAS,IAAI,EAAE,IAAI,MAAM,EAAE,IAAK,EAC1E,KAAK,EAAE,EACP,QAAQ,QAAQ,GAAG,EACnB,KAAK;AACV;AAgBO,SAAS,0BAA0B,OAA8B;AACtE,QAAM,WAAW,MAAM,KAAK,EAAE,QAAQ,QAAQ,GAAG;AACjD,MAAI,CAAC,UAAU;AACb,WAAO,EAAE,WAAW,UAAU,SAAS,OAAO,OAAO,CAAC,EAAE;AAAA,EAC1D;AACA,QAAM,OAAO,SAAS,KAAK;AAC3B,QAAM,QAAkB,CAAC;AAGzB,aAAW,KAAK,MAAM;AACpB,QAAI,EAAE,SAAS,YAAY,EAAE,UAAU,wBAAwB,KAAK,EAAE,IAAI,GAAG;AAC3E,QAAE,SAAS;AACX,QAAE,OAAO,EAAE,KAAK,QAAQ,sBAAsB,GAAG;AACjD,UAAI,CAAC,MAAM,SAAS,IAAI,EAAG,OAAM,KAAK,IAAI;AAAA,IAC5C;AAAA,EACF;AAGA,aAAW,KAAK,MAAM;AACpB,QAAI,EAAE,SAAS,YAAY,EAAE,UAAU,oBAAoB,KAAK,EAAE,IAAI,GAAG;AACvE,QAAE,SAAS;AACX,UAAI,CAAC,MAAM,SAAS,IAAI,EAAG,OAAM,KAAK,IAAI;AAAA,IAC5C;AAAA,EACF;AAIA,QAAM,cAAc,KAAK;AAAA,IACvB,CAAC,MAAsB,EAAE,SAAS,YAAY,EAAE;AAAA,EAClD;AACA,QAAM,YAAY,KACf,OAAO,CAAC,MAAmB,EAAE,SAAS,KAAK,EAC3C,IAAI,CAAC,MAAM,EAAE,IAAI,EACjB,KAAK,GAAG;AAEX,MAAI,YAAY,UAAU,KAAK,CAAC,qBAAqB,KAAK,SAAS,GAAG;AACpE,QAAI,cAAc;AAClB,QAAI,WAAW;AACf,aAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK,GAAG;AACvC,YAAM,IAAI,KAAK,CAAC;AAChB,UAAI,KAAK,EAAE,SAAS,YAAY,EAAE,QAAQ;AACxC,uBAAe;AACf,YAAI,eAAe,GAAG;AACpB,gBAAM,OAAO,KAAK,IAAI,CAAC;AACvB,cAAI,QAAQ,KAAK,SAAS,SAAS,KAAK,KAAK,KAAK,MAAM,IAAI;AAC1D,iBAAK,OAAO;AACZ,uBAAW;AAAA,UACb;AAAA,QACF;AAAA,MACF;AAAA,IACF;AACA,QAAI,SAAU,OAAM,KAAK,IAAI;AAAA,EAC/B;AAEA,QAAM,YAAY,QAAQ,IAAI;AAC9B,SAAO,EAAE,WAAW,SAAS,cAAc,UAAU,MAAM;AAC7D;AAaO,SAAS,mBAAmB,OAA8B;AAC/D,QAAM,WAAW,MAAM,KAAK,EAAE,QAAQ,QAAQ,GAAG;AACjD,MAAI,CAAC,UAAU;AACb,WAAO,EAAE,WAAW,UAAU,SAAS,OAAO,OAAO,CAAC,EAAE;AAAA,EAC1D;AACA,QAAM,QAAkB,CAAC;AACzB,MAAI,SAAS;AAEb,MAAI,OAAO,SAAS,GAAG,GAAG;AACxB,aAAS,OAAO,QAAQ,MAAM,EAAE;AAChC,UAAM,KAAK,IAAI;AAAA,EACjB;AAEA,MAAI,kBAAkB,KAAK,MAAM,GAAG;AAClC,aAAS,OAAO,QAAQ,sBAAsB,GAAG;AACjD,UAAM,KAAK,IAAI;AAAA,EACjB;AAEA,WAAS,OAAO,QAAQ,QAAQ,GAAG,EAAE,KAAK;AAC1C,SAAO,EAAE,WAAW,QAAQ,SAAS,WAAW,UAAU,MAAM;AAClE;",
+  "names": []
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mcp-researchpowerpack",
-  "version": "6.0.9",
+  "version": "6.0.10",
   "description": "HTTP-first MCP research server: start-research (goal-tailored brief), web-search (with Reddit scope), scrape-links (auto-detects Reddit URLs) — built on mcp-use.",
   "type": "module",
   "main": "dist/index.js",