Update benchmark prompts and README for post-rebase correctness

- Add isPrimaryKey: true to the natural PK column in each of the 4
  benchmark prompts so the orchestrator gets explicit pk guidance and
  deduplication works correctly with the run_subagent tool's .refine()
  enforcement
- Fix investigateCalls description: investigate_row → run_subagent
- Add note that enumeration-step tokens are not captured in RunMetrics
- Update custom-prompt example to show isPrimaryKey usage
- Update cost table model name: Kimi K2 → DeepSeek V4 Pro

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: MMeteorL

benchmarks/README.md

@@ -14,7 +14,7 @@ When a user clicks "Populate" in the app, or when the benchmark runner triggers
 |---|---|
 | `searchCalls` | Calls to `search_web` (TinyFish search API) |
 | `fetchCalls` | Calls to `fetch_page` (TinyFish fetch API) |
-| `investigateCalls` | `investigate_row` dispatches from the orchestrator |
+| `investigateCalls` | `run_subagent` dispatches from the orchestrator |
 | `rowsInserted` | Rows successfully inserted into the dataset |
 | `tokensInput` / `tokensOutput` | Total LLM tokens across all agents |
 | `orchestratorTokens*` / `investigateTokens*` | Token breakdown per agent tier |
@@ -24,6 +24,8 @@ When a user clicks "Populate" in the app, or when the benchmark runner triggers
 
 Each run also records `status` (`success` / `error`), any error message, and an `isBenchmark` flag so you can filter benchmark runs from real sessions.
 
+> **Note:** The workflow includes an enumeration classification step that calls an LLM directly (not through an agent) to decide whether to use a scraper or search strategy. The tokens used by that step are **not** captured in the metrics above — they're a small fixed cost per run (~100–200 input tokens, ~5 output tokens) but worth knowing if you're doing precise cost accounting.
+
 ---
 
 ## Prerequisites
@@ -225,13 +227,16 @@ Edit [`prompts.json`](./prompts.json) to add your own benchmark prompts. Each en
   "datasetName": "Human-readable dataset name",
   "description": "What the dataset is about — shown to the agent",
   "columns": [
-    { "name": "column_name", "type": "text", "description": "What this field is" }
+    { "name": "entity_name", "type": "text", "description": "The entity name", "isPrimaryKey": true },
+    { "name": "other_field", "type": "text", "description": "What this field is" }
   ]
 }
 ```
 
 Column types: `text`, `number`, `boolean`, `url`, `date`.
 
+Mark at least one column as `"isPrimaryKey": true` — the orchestrator uses this to tell subagents which field is the unique identifier, and the workflow uses it to reject duplicate rows automatically.
+
 Then run:
 ```bash
 make benchmark ARGS="--prompt my-prompt-id"
@@ -241,7 +246,7 @@ make benchmark ARGS="--prompt my-prompt-id"
 
 ## Cost estimation
 
-Rough estimates based on Kimi K2 pricing (as of writing):
+Rough estimates based on DeepSeek V4 Pro pricing (as of writing):
 
 | Per run (20 rows target) | Approximate cost |
 |---|---|

benchmarks/prompts.json

@@ -4,7 +4,7 @@
     "datasetName": "YC Recent Batch Companies",
     "description": "Companies from recent Y Combinator batches (W24, S24). Include name, website, one-line description, and batch.",
     "columns": [
-      { "name": "company_name", "type": "text", "description": "Company name" },
+      { "name": "company_name", "type": "text", "description": "Company name", "isPrimaryKey": true },
       { "name": "website", "type": "url", "description": "Company website URL" },
       { "name": "description", "type": "text", "description": "One-line description of what the company does" },
       { "name": "batch", "type": "text", "description": "YC batch, e.g. W24 or S24" }
@@ -15,7 +15,7 @@
     "datasetName": "B2B SaaS with Free Tiers",
     "description": "Popular B2B SaaS tools that offer a free tier or freemium plan. Include tool name, category, free-tier summary, and pricing page URL.",
     "columns": [
-      { "name": "tool_name", "type": "text", "description": "Name of the SaaS tool" },
+      { "name": "tool_name", "type": "text", "description": "Name of the SaaS tool", "isPrimaryKey": true },
       { "name": "category", "type": "text", "description": "Category, e.g. CRM, Analytics, DevOps" },
       { "name": "free_tier_summary", "type": "text", "description": "What the free tier includes" },
       { "name": "pricing_page_url", "type": "url", "description": "URL to the pricing page" }
@@ -26,7 +26,7 @@
     "datasetName": "US National Parks",
     "description": "All US National Parks with state, year established, and official NPS page URL.",
     "columns": [
-      { "name": "park_name", "type": "text", "description": "Official park name" },
+      { "name": "park_name", "type": "text", "description": "Official park name", "isPrimaryKey": true },
       { "name": "state", "type": "text", "description": "US state(s) the park is in" },
       { "name": "established_year", "type": "number", "description": "Year the park was established" },
       { "name": "official_page_url", "type": "url", "description": "URL on nps.gov" }
@@ -37,7 +37,7 @@
     "datasetName": "University AI Research Labs",
     "description": "Academic AI research labs at major universities. Include lab name, university, research focus, and lab website URL.",
     "columns": [
-      { "name": "lab_name", "type": "text", "description": "Name of the research lab" },
+      { "name": "lab_name", "type": "text", "description": "Name of the research lab", "isPrimaryKey": true },
       { "name": "university", "type": "text", "description": "Host university" },
       { "name": "research_focus", "type": "text", "description": "Primary research area, e.g. NLP, robotics, RL" },
       { "name": "lab_website_url", "type": "url", "description": "Lab homepage URL" }