From aaf7916ffff9e326b797cb5bf46b100036304327 Mon Sep 17 00:00:00 2001
From: Mert Koseoglu <bm.ksglu@gmail.com>
Date: Fri, 26 Jun 2026 00:46:53 +0300
Subject: [PATCH] refactor(pricing): single catalog under src/session;
 structured cost event

Consolidates the 5 vendor price files into one src/session/model-prices.json (61 models), moves the module to src/session/pricing.ts beside its only consumer (extract.ts), and drops the dev-only litellm NOTES from tracking (litellm base stays gitignored). Wave 2b: extractAgentUsage emits model_id + input/output/cache_read/cache_creation tokens + cost_usd as structured top-level event fields, replacing the colon-string, so the platform persists them as columns. 45 tests green, tsc clean.
---
 .gitignore                                    |   3 +-
 src/pricing/sources/anthropic.json            |  82 ----
 src/pricing/sources/chinese.json              | 155 -------
 src/pricing/sources/google.json               |  74 ---
 src/pricing/sources/openai.json               | 106 -----
 src/pricing/sources/others.json               | 137 ------
 src/session/extract.ts                        |  41 +-
 src/session/model-prices.json                 | 429 ++++++++++++++++++
 .../catalog.ts => session/pricing.ts}         |  62 +--
 tests/session/extract-agent-usage.test.ts     |  87 ++++
 .../pricing.test.ts}                          |  10 +-
 tools/pricing/litellm-NOTES.md                |  80 ----
 12 files changed, 583 insertions(+), 683 deletions(-)
 delete mode 100644 src/pricing/sources/anthropic.json
 delete mode 100644 src/pricing/sources/chinese.json
 delete mode 100644 src/pricing/sources/google.json
 delete mode 100644 src/pricing/sources/openai.json
 delete mode 100644 src/pricing/sources/others.json
 create mode 100644 src/session/model-prices.json
 rename src/{pricing/catalog.ts => session/pricing.ts} (75%)
 rename tests/{pricing/catalog.test.ts => session/pricing.test.ts} (95%)
 delete mode 100644 tools/pricing/litellm-NOTES.md

diff --git a/.gitignore b/.gitignore
index 2e2ea416..b1c6453d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -40,8 +40,7 @@ context-mode-guidance-*/
 .vibetree/
 .cw/
 # Dev-only pricing source: the full litellm catalog (~1.5MB) is the manual
-# refresh base for src/pricing/sources/*.json — never bundled into the plugin.
-# Refresh: see tools/pricing/litellm-NOTES.md. Keep NOTES tracked, ignore the blob.
+# refresh base for src/session/model-prices.json — never bundled into the plugin.
 tools/pricing/litellm-catalog.json
 /.cocoindex_code/
 /.kilo/
diff --git a/src/pricing/sources/anthropic.json b/src/pricing/sources/anthropic.json
deleted file mode 100644
index a15cad69..00000000
--- a/src/pricing/sources/anthropic.json
+++ /dev/null
@@ -1,82 +0,0 @@
-{
-  "claude-opus-4-8": {
-    "input_per_mtok": 5.00,
-    "output_per_mtok": 25.00,
-    "cache_read_per_mtok": 0.50,
-    "cache_write_per_mtok": 6.25,
-    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
-    "as_of": "2026-06"
-  },
-  "claude-opus-4-7": {
-    "input_per_mtok": 5.00,
-    "output_per_mtok": 25.00,
-    "cache_read_per_mtok": 0.50,
-    "cache_write_per_mtok": 6.25,
-    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
-    "as_of": "2026-06"
-  },
-  "claude-opus-4-6": {
-    "input_per_mtok": 5.00,
-    "output_per_mtok": 25.00,
-    "cache_read_per_mtok": 0.50,
-    "cache_write_per_mtok": 6.25,
-    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
-    "as_of": "2026-06"
-  },
-  "claude-opus-4-5": {
-    "input_per_mtok": 5.00,
-    "output_per_mtok": 25.00,
-    "cache_read_per_mtok": 0.50,
-    "cache_write_per_mtok": 6.25,
-    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
-    "as_of": "2026-06"
-  },
-  "claude-sonnet-4-6": {
-    "input_per_mtok": 3.00,
-    "output_per_mtok": 15.00,
-    "cache_read_per_mtok": 0.30,
-    "cache_write_per_mtok": 3.75,
-    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
-    "as_of": "2026-06"
-  },
-  "claude-sonnet-4-5": {
-    "input_per_mtok": 3.00,
-    "output_per_mtok": 15.00,
-    "cache_read_per_mtok": 0.30,
-    "cache_write_per_mtok": 3.75,
-    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
-    "as_of": "2026-06"
-  },
-  "claude-haiku-4-5": {
-    "input_per_mtok": 1.00,
-    "output_per_mtok": 5.00,
-    "cache_read_per_mtok": 0.10,
-    "cache_write_per_mtok": 1.25,
-    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
-    "as_of": "2026-06"
-  },
-  "claude-3-7-sonnet": {
-    "input_per_mtok": 3.00,
-    "output_per_mtok": 15.00,
-    "cache_read_per_mtok": 0.30,
-    "cache_write_per_mtok": 3.75,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "claude-3-5-haiku": {
-    "input_per_mtok": 0.80,
-    "output_per_mtok": 4.00,
-    "cache_read_per_mtok": 0.08,
-    "cache_write_per_mtok": 1.00,
-    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
-    "as_of": "2026-06"
-  },
-  "claude-fable-5": {
-    "input_per_mtok": 10.00,
-    "output_per_mtok": 50.00,
-    "cache_read_per_mtok": 1.00,
-    "cache_write_per_mtok": 12.50,
-    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
-    "as_of": "2026-06"
-  }
-}
diff --git a/src/pricing/sources/chinese.json b/src/pricing/sources/chinese.json
deleted file mode 100644
index 15e78b09..00000000
--- a/src/pricing/sources/chinese.json
+++ /dev/null
@@ -1,155 +0,0 @@
-{
-  "qwen3-coder": {
-    "input_per_mtok": 1.0,
-    "output_per_mtok": 5.0,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://www.alibabacloud.com/help/en/model-studio/models",
-    "as_of": "2026-06",
-    "_note": "Billing id maps to qwen3-coder-plus. Tiered by input length; headline is the 0-32K tier. Higher tiers: 32K-128K in 1.8/out 9, 128K-256K in 3/out 15, 256K-1M in 6/out 60. Cross-checked vs LiteLLM tiered_pricing."
-  },
-  "qwen-max": {
-    "input_per_mtok": 1.6,
-    "output_per_mtok": 6.4,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://www.alibabacloud.com/help/en/model-studio/models",
-    "as_of": "2026-06",
-    "_note": "Alibaba DashScope international (USD-denominated). Cross-checked vs LiteLLM dashscope/qwen-max."
-  },
-  "qwen-plus": {
-    "input_per_mtok": 0.4,
-    "output_per_mtok": 1.2,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://www.alibabacloud.com/help/en/model-studio/models",
-    "as_of": "2026-06",
-    "_note": "Alibaba DashScope international (USD). Cross-checked vs LiteLLM dashscope/qwen-plus."
-  },
-  "qwen-turbo": {
-    "input_per_mtok": 0.05,
-    "output_per_mtok": 0.2,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://www.alibabacloud.com/help/en/model-studio/models",
-    "as_of": "2026-06",
-    "_note": "Alibaba DashScope international (USD). Cross-checked vs LiteLLM dashscope/qwen-turbo."
-  },
-  "qwen3-max": {
-    "input_per_mtok": 1.2,
-    "output_per_mtok": 6.0,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://www.alibabacloud.com/help/en/model-studio/models",
-    "as_of": "2026-06",
-    "_note": "Tiered by input length; headline is the 0-32K tier. Higher tiers: 32K-128K in 2.4/out 12, 128K-252K in 3/out 15. Cross-checked vs LiteLLM dashscope/qwen3-max tiered_pricing."
-  },
-  "kimi-k2": {
-    "input_per_mtok": 0.6,
-    "output_per_mtok": 2.5,
-    "cache_read_per_mtok": 0.15,
-    "cache_write_per_mtok": null,
-    "source": "https://platform.moonshot.ai/docs/pricing/chat",
-    "as_of": "2026-06",
-    "_note": "Moonshot international platform (USD). Resolves to kimi-k2-0905-preview. Cache-hit input 0.15. Cross-checked vs LiteLLM moonshot/kimi-k2-0905-preview."
-  },
-  "kimi-k2-turbo": {
-    "input_per_mtok": 1.15,
-    "output_per_mtok": 8.0,
-    "cache_read_per_mtok": 0.15,
-    "cache_write_per_mtok": null,
-    "source": "https://platform.moonshot.ai/docs/pricing/chat",
-    "as_of": "2026-06",
-    "_note": "Moonshot international (USD). Resolves to kimi-k2-turbo-preview. Cache-hit input 0.15. Cross-checked vs LiteLLM moonshot/kimi-k2-turbo-preview."
-  },
-  "moonshot-v1-8k": {
-    "input_per_mtok": 0.2,
-    "output_per_mtok": 2.0,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://platform.moonshot.ai/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Moonshot international (USD). Cross-checked vs LiteLLM moonshot/moonshot-v1-8k."
-  },
-  "moonshot-v1-32k": {
-    "input_per_mtok": 1.0,
-    "output_per_mtok": 3.0,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://platform.moonshot.ai/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Moonshot international (USD). Cross-checked vs LiteLLM moonshot/moonshot-v1-32k."
-  },
-  "moonshot-v1-128k": {
-    "input_per_mtok": 2.0,
-    "output_per_mtok": 5.0,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://platform.moonshot.ai/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Moonshot international (USD). Cross-checked vs LiteLLM moonshot/moonshot-v1-128k."
-  },
-  "deepseek-v3": {
-    "input_per_mtok": 0.27,
-    "output_per_mtok": 1.1,
-    "cache_read_per_mtok": 0.07,
-    "cache_write_per_mtok": 0.0,
-    "source": "https://api-docs.deepseek.com/quick_start/pricing",
-    "as_of": "2026-06",
-    "_note": "Legacy DeepSeek-V3 era pricing. No longer shown on the official current pricing page (which now lists deepseek-v4-flash/pro). Value from LiteLLM deepseek/deepseek-v3, last sourced from the official page. cache-hit input 0.07."
-  },
-  "deepseek-r1": {
-    "input_per_mtok": 0.55,
-    "output_per_mtok": 2.19,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://api-docs.deepseek.com/quick_start/pricing",
-    "as_of": "2026-06",
-    "_note": "Legacy DeepSeek-R1 pricing. Not on the official current pricing page (now deepseek-v4-flash/pro). Value from LiteLLM deepseek/deepseek-r1."
-  },
-  "deepseek-chat": {
-    "input_per_mtok": 0.14,
-    "output_per_mtok": 0.28,
-    "cache_read_per_mtok": 0.0028,
-    "cache_write_per_mtok": null,
-    "source": "https://api-docs.deepseek.com/quick_start/pricing",
-    "as_of": "2026-06",
-    "_note": "Official page (fetched 2026-06): deepseek-chat is deprecated 2026/07/24 and now maps to deepseek-v4-flash non-thinking mode. Live billing = v4-flash: input(cache-miss) 0.14, output 0.28, input(cache-hit) 0.0028."
-  },
-  "deepseek-reasoner": {
-    "input_per_mtok": 0.14,
-    "output_per_mtok": 0.28,
-    "cache_read_per_mtok": 0.0028,
-    "cache_write_per_mtok": null,
-    "source": "https://api-docs.deepseek.com/quick_start/pricing",
-    "as_of": "2026-06",
-    "_note": "Official page (fetched 2026-06): deepseek-reasoner is deprecated 2026/07/24 and now maps to deepseek-v4-flash thinking mode. Live billing = v4-flash: input(cache-miss) 0.14, output 0.28, input(cache-hit) 0.0028."
-  },
-  "glm-4.6": {
-    "input_per_mtok": 0.6,
-    "output_per_mtok": 2.2,
-    "cache_read_per_mtok": 0.11,
-    "cache_write_per_mtok": null,
-    "source": "https://docs.z.ai/guides/overview/pricing",
-    "as_of": "2026-06",
-    "_note": "Z.AI international (USD). Cached-input 0.11. Cache-input storage currently limited-time free. Cross-checked vs LiteLLM zai/glm-4.6."
-  },
-  "glm-4-plus": {
-    "input_per_mtok": null,
-    "output_per_mtok": null,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": null,
-    "as_of": "2026-06",
-    "_note": "No authoritative current price found this session. glm-4-plus is a legacy Zhipu/BigModel-native model not listed on the Z.AI international catalog; open.bigmodel.cn pricing is SPA-rendered (no static price) and BigModel English docs returned HTTP 552/404. Not in LiteLLM. Left null per anti-hallucination rule."
-  },
-  "glm-4-air": {
-    "input_per_mtok": 0.2,
-    "output_per_mtok": 1.1,
-    "cache_read_per_mtok": 0.03,
-    "cache_write_per_mtok": null,
-    "source": "https://docs.z.ai/guides/overview/pricing",
-    "as_of": "2026-06",
-    "_note": "Mission billing id glm-4-air; current Z.AI international catalog lists this as GLM-4.5-Air (USD): input 0.2, cached-input 0.03, output 1.1. Cross-checked vs LiteLLM zai/glm-4.5-air (in 0.2/out 1.1)."
-  }
-}
diff --git a/src/pricing/sources/google.json b/src/pricing/sources/google.json
deleted file mode 100644
index 09399361..00000000
--- a/src/pricing/sources/google.json
+++ /dev/null
@@ -1,74 +0,0 @@
-{
-  "gemini-2.5-pro": {
-    "input_per_mtok": 1.25,
-    "output_per_mtok": 10.0,
-    "cache_read_per_mtok": 0.125,
-    "cache_write_per_mtok": null,
-    "source": "https://ai.google.dev/gemini-api/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Tiered: prompts >200k tokens cost input $2.50, output $15.00, cache_read $0.25 per Mtok. Cache storage $4.50/Mtok/hr."
-  },
-  "gemini-2.5-flash": {
-    "input_per_mtok": 0.3,
-    "output_per_mtok": 2.5,
-    "cache_read_per_mtok": 0.03,
-    "cache_write_per_mtok": null,
-    "source": "https://ai.google.dev/gemini-api/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Text/image/video input; audio input $1.00. Output includes thinking tokens. Cache storage $1.00/Mtok/hr. No >200k context tier."
-  },
-  "gemini-2.5-flash-lite": {
-    "input_per_mtok": 0.1,
-    "output_per_mtok": 0.4,
-    "cache_read_per_mtok": 0.01,
-    "cache_write_per_mtok": null,
-    "source": "https://ai.google.dev/gemini-api/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Text/image/video input; audio input $0.30. Cache storage $1.00/Mtok/hr. No >200k context tier."
-  },
-  "gemini-2.0-flash": {
-    "input_per_mtok": 0.1,
-    "output_per_mtok": 0.4,
-    "cache_read_per_mtok": 0.025,
-    "cache_write_per_mtok": null,
-    "source": "https://ai.google.dev/gemini-api/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Text/image/video input; audio input $0.70, audio cache_read $0.175. Cache storage $1.00/Mtok/hr. No >200k context tier."
-  },
-  "gemini-2.0-flash-lite": {
-    "input_per_mtok": 0.075,
-    "output_per_mtok": 0.3,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://ai.google.dev/gemini-api/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "DEPRECATED — shut down June 1, 2026. Context caching not available for this model. Prices retained for historical billing."
-  },
-  "gemini-2.0-pro": {
-    "input_per_mtok": null,
-    "output_per_mtok": null,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://ai.google.dev/gemini-api/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "No GA billing id. Only shipped as gemini-2.0-pro-exp (experimental, free of charge); never had paid pricing. Not listed on official pricing page or LiteLLM catalog."
-  },
-  "gemini-3-pro-preview": {
-    "input_per_mtok": 2.0,
-    "output_per_mtok": 12.0,
-    "cache_read_per_mtok": 0.2,
-    "cache_write_per_mtok": null,
-    "source": "https://ai.google.dev/gemini-api/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Current 3.x flagship (page labels it Gemini 3.1 Pro Preview). Tiered: prompts >200k cost input $4.00, output $18.00, cache_read $0.40 per Mtok. Cache storage $4.50/Mtok/hr."
-  },
-  "gemini-3-flash-preview": {
-    "input_per_mtok": 0.5,
-    "output_per_mtok": 3.0,
-    "cache_read_per_mtok": 0.05,
-    "cache_write_per_mtok": null,
-    "source": "https://ai.google.dev/gemini-api/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Text/image/video input; audio input $1.00, audio cache_read $0.10. Cache storage $1.00/Mtok/hr. No >200k context tier on standard."
-  }
-}
diff --git a/src/pricing/sources/openai.json b/src/pricing/sources/openai.json
deleted file mode 100644
index 7b6b7d51..00000000
--- a/src/pricing/sources/openai.json
+++ /dev/null
@@ -1,106 +0,0 @@
-{
-  "gpt-5": {
-    "input_per_mtok": 1.25,
-    "output_per_mtok": 10,
-    "cache_read_per_mtok": 0.125,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "gpt-5-mini": {
-    "input_per_mtok": 0.25,
-    "output_per_mtok": 2,
-    "cache_read_per_mtok": 0.025,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "gpt-5-nano": {
-    "input_per_mtok": 0.05,
-    "output_per_mtok": 0.4,
-    "cache_read_per_mtok": 0.005,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "gpt-5-codex": {
-    "input_per_mtok": 1.25,
-    "output_per_mtok": 10,
-    "cache_read_per_mtok": 0.125,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "gpt-4.1": {
-    "input_per_mtok": 2,
-    "output_per_mtok": 8,
-    "cache_read_per_mtok": 0.5,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "gpt-4.1-mini": {
-    "input_per_mtok": 0.4,
-    "output_per_mtok": 1.6,
-    "cache_read_per_mtok": 0.1,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "gpt-4.1-nano": {
-    "input_per_mtok": 0.1,
-    "output_per_mtok": 0.4,
-    "cache_read_per_mtok": 0.025,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "gpt-4o": {
-    "input_per_mtok": 2.5,
-    "output_per_mtok": 10,
-    "cache_read_per_mtok": 1.25,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "gpt-4o-mini": {
-    "input_per_mtok": 0.15,
-    "output_per_mtok": 0.6,
-    "cache_read_per_mtok": 0.075,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "o3": {
-    "input_per_mtok": 2,
-    "output_per_mtok": 8,
-    "cache_read_per_mtok": 0.5,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "o4-mini": {
-    "input_per_mtok": 1.1,
-    "output_per_mtok": 4.4,
-    "cache_read_per_mtok": 0.275,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "o3-mini": {
-    "input_per_mtok": 1.1,
-    "output_per_mtok": 4.4,
-    "cache_read_per_mtok": 0.55,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  },
-  "codex-mini-latest": {
-    "input_per_mtok": 1.5,
-    "output_per_mtok": 6,
-    "cache_read_per_mtok": 0.375,
-    "cache_write_per_mtok": null,
-    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
-    "as_of": "2026-06"
-  }
-}
diff --git a/src/pricing/sources/others.json b/src/pricing/sources/others.json
deleted file mode 100644
index 1cdeb5e6..00000000
--- a/src/pricing/sources/others.json
+++ /dev/null
@@ -1,137 +0,0 @@
-{
-  "grok-4": {
-    "input_per_mtok": 3,
-    "output_per_mtok": 15,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://docs.x.ai/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Legacy id; xAI public pricing page now lists newer Grok versions (4.3/4.20/build-0.1). Rate verified via LiteLLM xai/ catalog (BerriAI/litellm) which sources directly from xAI."
-  },
-  "grok-3": {
-    "input_per_mtok": 3,
-    "output_per_mtok": 15,
-    "cache_read_per_mtok": 0.75,
-    "cache_write_per_mtok": null,
-    "source": "https://docs.x.ai/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Legacy id; rate from LiteLLM xai/grok-3 catalog (sourced from xAI)."
-  },
-  "grok-code-fast-1": {
-    "input_per_mtok": 0.2,
-    "output_per_mtok": 1.5,
-    "cache_read_per_mtok": 0.02,
-    "cache_write_per_mtok": null,
-    "source": "https://docs.x.ai/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Legacy id; rate from LiteLLM xai/grok-code-fast-1 catalog (sourced from xAI)."
-  },
-  "grok-2": {
-    "input_per_mtok": 2,
-    "output_per_mtok": 10,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://docs.x.ai/docs/pricing",
-    "as_of": "2026-06",
-    "_note": "Legacy id (grok-2-1212); rate from LiteLLM xai/grok-2 catalog (sourced from xAI)."
-  },
-  "mistral-large-latest": {
-    "input_per_mtok": 0.5,
-    "output_per_mtok": 1.5,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://mistral.ai/pricing",
-    "as_of": "2026-06",
-    "_note": "mistral-large-latest = Mistral Large 3."
-  },
-  "codestral-latest": {
-    "input_per_mtok": 0.3,
-    "output_per_mtok": 0.9,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://mistral.ai/pricing",
-    "as_of": "2026-06",
-    "_note": "Codestral 2; price cross-confirmed via Google Vertex partner pricing ($0.30/$0.90)."
-  },
-  "devstral": {
-    "input_per_mtok": 0.4,
-    "output_per_mtok": 2,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://mistral.ai/pricing",
-    "as_of": "2026-06",
-    "_note": "Maps to devstral-medium-latest (Mistral hosted API). devstral-small is $0.10/$0.30."
-  },
-  "mistral-medium": {
-    "input_per_mtok": 0.4,
-    "output_per_mtok": 2,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://mistral.ai/pricing",
-    "as_of": "2026-06",
-    "_note": "Mistral Medium 3; cross-confirmed via Google Vertex partner pricing ($0.40/$2.00)."
-  },
-  "llama-4-maverick": {
-    "input_per_mtok": 0.27,
-    "output_per_mtok": 0.85,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://www.together.ai/pricing",
-    "as_of": "2026-06",
-    "_note": "Host-dependent. Priced via Together AI serverless. Other hosts differ (e.g. AWS Bedrock, Vertex $0.35/$1.15)."
-  },
-  "llama-4-scout": {
-    "input_per_mtok": 0.08,
-    "output_per_mtok": 0.3,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://www.together.ai/pricing",
-    "as_of": "2026-06",
-    "_note": "Host-dependent. Priced via Together AI serverless. Vertex lists $0.25/$0.70."
-  },
-  "llama-3.3-70b": {
-    "input_per_mtok": 0.88,
-    "output_per_mtok": 0.88,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://www.together.ai/pricing",
-    "as_of": "2026-06",
-    "_note": "Host-dependent. Together Llama-3.3-70B-Instruct-Turbo ($0.88/$0.88, from LiteLLM together_ai source). Cheaper on DeepInfra ($0.13/$0.39); Vertex $0.72/$0.72."
-  },
-  "command-a": {
-    "input_per_mtok": 2.5,
-    "output_per_mtok": 10,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://cohere.com/pricing",
-    "as_of": "2026-06",
-    "_note": "command-a-03-2025."
-  },
-  "command-r-plus": {
-    "input_per_mtok": 2.5,
-    "output_per_mtok": 10,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://cohere.com/pricing",
-    "as_of": "2026-06",
-    "_note": "command-r-plus-08-2024 (legacy)."
-  },
-  "amazon-nova-pro": {
-    "input_per_mtok": 0.8,
-    "output_per_mtok": 3.2,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://aws.amazon.com/bedrock/pricing/",
-    "as_of": "2026-06",
-    "_note": "AWS Bedrock on-demand ($0.0008/$0.0032 per 1K tokens)."
-  },
-  "amazon-nova-lite": {
-    "input_per_mtok": 0.06,
-    "output_per_mtok": 0.24,
-    "cache_read_per_mtok": null,
-    "cache_write_per_mtok": null,
-    "source": "https://aws.amazon.com/bedrock/pricing/",
-    "as_of": "2026-06",
-    "_note": "AWS Bedrock on-demand ($0.00006/$0.00024 per 1K tokens)."
-  }
-}
diff --git a/src/session/extract.ts b/src/session/extract.ts
index d6b4e857..1b0b0180 100644
--- a/src/session/extract.ts
+++ b/src/session/extract.ts
@@ -8,7 +8,7 @@
 import {
   lookupPrice as catalogLookupPrice,
   computeCostUsd as catalogComputeCostUsd,
-} from "../pricing/catalog.js";
+} from "./pricing.js";
 
 // ── Public interfaces ──────────────────────────────────────────────────────
 
@@ -30,6 +30,19 @@ export interface SessionEvent {
    * `Fetched and indexed N sections (XKB)` preamble.
    */
   bytes_avoided?: number;
+  /**
+   * Optional structured cost/usage fields (Wave 2b). Emitted by
+   * extractAgentUsage alongside the colon-string `data` so the forward
+   * envelope can spread them to the platform as typed columns instead of an
+   * opaque blob. Present only when the source signal is present; cost_usd is
+   * omitted on a price miss or a zero-token turn.
+   */
+  model_id?: string;
+  input_tokens?: number;
+  output_tokens?: number;
+  cache_read_tokens?: number;
+  cache_creation_tokens?: number;
+  cost_usd?: number;
 }
 
 export interface ToolCall {
@@ -1487,21 +1500,39 @@ function extractAgentUsage(input: HookInput): SessionEvent[] {
   const cacheRead = typeof usage.cache_read_input_tokens === "number"
     ? usage.cache_read_input_tokens
     : 0;
+  const modelId = resolveModelId(input, out);
   const anyTokens = inputTokens > 0 || outputTokens > 0 || cacheCreate > 0 || cacheRead > 0;
+  let cost: number | null = null;
   if (anyTokens) {
-    const modelId = resolveModelId(input, out);
     // null ⇒ unmatched model id (catalog warned once) — skip the cost token
     // rather than blend a wrong Claude rate (the old non-Claude bug).
-    const cost = computeTurnCostUsd(modelId, inputTokens, outputTokens, cacheCreate, cacheRead);
+    cost = computeTurnCostUsd(modelId, inputTokens, outputTokens, cacheCreate, cacheRead);
     if (cost !== null) parts.push(`cost_usd:${formatCostUsd(cost)}`);
   }
 
-  return [{
+  // Wave 2b — emit structured top-level fields alongside the colon-string so
+  // the forward envelope (which spreads `...event`) hands the platform typed
+  // columns. Each field is set only when its source signal is present, so the
+  // forward payload stays minimal; cost_usd is omitted on a price miss or a
+  // zero-token turn. The colon-string `data` stays for human/debug + back-compat.
+  const event: SessionEvent = {
     type: "agent_usage",
     category: "cost",
     data: safeString(parts.join(" ")),
     priority: 2,
-  }];
+  };
+  if (modelId.length > 0) event.model_id = modelId;
+  if (typeof usage.input_tokens === "number") event.input_tokens = usage.input_tokens;
+  if (typeof usage.output_tokens === "number") event.output_tokens = usage.output_tokens;
+  if (typeof usage.cache_read_input_tokens === "number") {
+    event.cache_read_tokens = usage.cache_read_input_tokens;
+  }
+  if (typeof usage.cache_creation_input_tokens === "number") {
+    event.cache_creation_tokens = usage.cache_creation_input_tokens;
+  }
+  if (cost !== null) event.cost_usd = cost;
+
+  return [event];
 }
 
 // ── User-message extractors ────────────────────────────────────────────────
diff --git a/src/session/model-prices.json b/src/session/model-prices.json
new file mode 100644
index 00000000..1d70fc0f
--- /dev/null
+++ b/src/session/model-prices.json
@@ -0,0 +1,429 @@
+{
+  "claude-opus-4-8": {
+    "input_per_mtok": 5,
+    "output_per_mtok": 25,
+    "cache_read_per_mtok": 0.5,
+    "cache_write_per_mtok": 6.25,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing"
+  },
+  "claude-opus-4-7": {
+    "input_per_mtok": 5,
+    "output_per_mtok": 25,
+    "cache_read_per_mtok": 0.5,
+    "cache_write_per_mtok": 6.25,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing"
+  },
+  "claude-opus-4-6": {
+    "input_per_mtok": 5,
+    "output_per_mtok": 25,
+    "cache_read_per_mtok": 0.5,
+    "cache_write_per_mtok": 6.25,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing"
+  },
+  "claude-opus-4-5": {
+    "input_per_mtok": 5,
+    "output_per_mtok": 25,
+    "cache_read_per_mtok": 0.5,
+    "cache_write_per_mtok": 6.25,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing"
+  },
+  "claude-sonnet-4-6": {
+    "input_per_mtok": 3,
+    "output_per_mtok": 15,
+    "cache_read_per_mtok": 0.3,
+    "cache_write_per_mtok": 3.75,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing"
+  },
+  "claude-sonnet-4-5": {
+    "input_per_mtok": 3,
+    "output_per_mtok": 15,
+    "cache_read_per_mtok": 0.3,
+    "cache_write_per_mtok": 3.75,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing"
+  },
+  "claude-haiku-4-5": {
+    "input_per_mtok": 1,
+    "output_per_mtok": 5,
+    "cache_read_per_mtok": 0.1,
+    "cache_write_per_mtok": 1.25,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing"
+  },
+  "claude-3-7-sonnet": {
+    "input_per_mtok": 3,
+    "output_per_mtok": 15,
+    "cache_read_per_mtok": 0.3,
+    "cache_write_per_mtok": 3.75,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "claude-3-5-haiku": {
+    "input_per_mtok": 0.8,
+    "output_per_mtok": 4,
+    "cache_read_per_mtok": 0.08,
+    "cache_write_per_mtok": 1,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing"
+  },
+  "claude-fable-5": {
+    "input_per_mtok": 10,
+    "output_per_mtok": 50,
+    "cache_read_per_mtok": 1,
+    "cache_write_per_mtok": 12.5,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing"
+  },
+  "gpt-5": {
+    "input_per_mtok": 1.25,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": 0.125,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "gpt-5-mini": {
+    "input_per_mtok": 0.25,
+    "output_per_mtok": 2,
+    "cache_read_per_mtok": 0.025,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "gpt-5-nano": {
+    "input_per_mtok": 0.05,
+    "output_per_mtok": 0.4,
+    "cache_read_per_mtok": 0.005,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "gpt-5-codex": {
+    "input_per_mtok": 1.25,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": 0.125,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "gpt-4.1": {
+    "input_per_mtok": 2,
+    "output_per_mtok": 8,
+    "cache_read_per_mtok": 0.5,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "gpt-4.1-mini": {
+    "input_per_mtok": 0.4,
+    "output_per_mtok": 1.6,
+    "cache_read_per_mtok": 0.1,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "gpt-4.1-nano": {
+    "input_per_mtok": 0.1,
+    "output_per_mtok": 0.4,
+    "cache_read_per_mtok": 0.025,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "gpt-4o": {
+    "input_per_mtok": 2.5,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": 1.25,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "gpt-4o-mini": {
+    "input_per_mtok": 0.15,
+    "output_per_mtok": 0.6,
+    "cache_read_per_mtok": 0.075,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "o3": {
+    "input_per_mtok": 2,
+    "output_per_mtok": 8,
+    "cache_read_per_mtok": 0.5,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "o4-mini": {
+    "input_per_mtok": 1.1,
+    "output_per_mtok": 4.4,
+    "cache_read_per_mtok": 0.275,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "o3-mini": {
+    "input_per_mtok": 1.1,
+    "output_per_mtok": 4.4,
+    "cache_read_per_mtok": 0.55,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "codex-mini-latest": {
+    "input_per_mtok": 1.5,
+    "output_per_mtok": 6,
+    "cache_read_per_mtok": 0.375,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
+  },
+  "gemini-2.5-pro": {
+    "input_per_mtok": 1.25,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": 0.125,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing"
+  },
+  "gemini-2.5-flash": {
+    "input_per_mtok": 0.3,
+    "output_per_mtok": 2.5,
+    "cache_read_per_mtok": 0.03,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing"
+  },
+  "gemini-2.5-flash-lite": {
+    "input_per_mtok": 0.1,
+    "output_per_mtok": 0.4,
+    "cache_read_per_mtok": 0.01,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing"
+  },
+  "gemini-2.0-flash": {
+    "input_per_mtok": 0.1,
+    "output_per_mtok": 0.4,
+    "cache_read_per_mtok": 0.025,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing"
+  },
+  "gemini-2.0-flash-lite": {
+    "input_per_mtok": 0.075,
+    "output_per_mtok": 0.3,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing"
+  },
+  "gemini-3-pro-preview": {
+    "input_per_mtok": 2,
+    "output_per_mtok": 12,
+    "cache_read_per_mtok": 0.2,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing"
+  },
+  "gemini-3-flash-preview": {
+    "input_per_mtok": 0.5,
+    "output_per_mtok": 3,
+    "cache_read_per_mtok": 0.05,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing"
+  },
+  "qwen3-coder": {
+    "input_per_mtok": 1,
+    "output_per_mtok": 5,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.alibabacloud.com/help/en/model-studio/models"
+  },
+  "qwen-max": {
+    "input_per_mtok": 1.6,
+    "output_per_mtok": 6.4,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.alibabacloud.com/help/en/model-studio/models"
+  },
+  "qwen-plus": {
+    "input_per_mtok": 0.4,
+    "output_per_mtok": 1.2,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.alibabacloud.com/help/en/model-studio/models"
+  },
+  "qwen-turbo": {
+    "input_per_mtok": 0.05,
+    "output_per_mtok": 0.2,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.alibabacloud.com/help/en/model-studio/models"
+  },
+  "qwen3-max": {
+    "input_per_mtok": 1.2,
+    "output_per_mtok": 6,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.alibabacloud.com/help/en/model-studio/models"
+  },
+  "kimi-k2": {
+    "input_per_mtok": 0.6,
+    "output_per_mtok": 2.5,
+    "cache_read_per_mtok": 0.15,
+    "cache_write_per_mtok": null,
+    "source": "https://platform.moonshot.ai/docs/pricing/chat"
+  },
+  "kimi-k2-turbo": {
+    "input_per_mtok": 1.15,
+    "output_per_mtok": 8,
+    "cache_read_per_mtok": 0.15,
+    "cache_write_per_mtok": null,
+    "source": "https://platform.moonshot.ai/docs/pricing/chat"
+  },
+  "moonshot-v1-8k": {
+    "input_per_mtok": 0.2,
+    "output_per_mtok": 2,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://platform.moonshot.ai/docs/pricing"
+  },
+  "moonshot-v1-32k": {
+    "input_per_mtok": 1,
+    "output_per_mtok": 3,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://platform.moonshot.ai/docs/pricing"
+  },
+  "moonshot-v1-128k": {
+    "input_per_mtok": 2,
+    "output_per_mtok": 5,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://platform.moonshot.ai/docs/pricing"
+  },
+  "deepseek-v3": {
+    "input_per_mtok": 0.27,
+    "output_per_mtok": 1.1,
+    "cache_read_per_mtok": 0.07,
+    "cache_write_per_mtok": 0,
+    "source": "https://api-docs.deepseek.com/quick_start/pricing"
+  },
+  "deepseek-r1": {
+    "input_per_mtok": 0.55,
+    "output_per_mtok": 2.19,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://api-docs.deepseek.com/quick_start/pricing"
+  },
+  "deepseek-chat": {
+    "input_per_mtok": 0.14,
+    "output_per_mtok": 0.28,
+    "cache_read_per_mtok": 0.0028,
+    "cache_write_per_mtok": null,
+    "source": "https://api-docs.deepseek.com/quick_start/pricing"
+  },
+  "deepseek-reasoner": {
+    "input_per_mtok": 0.14,
+    "output_per_mtok": 0.28,
+    "cache_read_per_mtok": 0.0028,
+    "cache_write_per_mtok": null,
+    "source": "https://api-docs.deepseek.com/quick_start/pricing"
+  },
+  "glm-4.6": {
+    "input_per_mtok": 0.6,
+    "output_per_mtok": 2.2,
+    "cache_read_per_mtok": 0.11,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.z.ai/guides/overview/pricing"
+  },
+  "glm-4-air": {
+    "input_per_mtok": 0.2,
+    "output_per_mtok": 1.1,
+    "cache_read_per_mtok": 0.03,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.z.ai/guides/overview/pricing"
+  },
+  "grok-4": {
+    "input_per_mtok": 3,
+    "output_per_mtok": 15,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.x.ai/docs/pricing"
+  },
+  "grok-3": {
+    "input_per_mtok": 3,
+    "output_per_mtok": 15,
+    "cache_read_per_mtok": 0.75,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.x.ai/docs/pricing"
+  },
+  "grok-code-fast-1": {
+    "input_per_mtok": 0.2,
+    "output_per_mtok": 1.5,
+    "cache_read_per_mtok": 0.02,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.x.ai/docs/pricing"
+  },
+  "grok-2": {
+    "input_per_mtok": 2,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.x.ai/docs/pricing"
+  },
+  "mistral-large-latest": {
+    "input_per_mtok": 0.5,
+    "output_per_mtok": 1.5,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://mistral.ai/pricing"
+  },
+  "codestral-latest": {
+    "input_per_mtok": 0.3,
+    "output_per_mtok": 0.9,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://mistral.ai/pricing"
+  },
+  "devstral": {
+    "input_per_mtok": 0.4,
+    "output_per_mtok": 2,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://mistral.ai/pricing"
+  },
+  "mistral-medium": {
+    "input_per_mtok": 0.4,
+    "output_per_mtok": 2,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://mistral.ai/pricing"
+  },
+  "llama-4-maverick": {
+    "input_per_mtok": 0.27,
+    "output_per_mtok": 0.85,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.together.ai/pricing"
+  },
+  "llama-4-scout": {
+    "input_per_mtok": 0.08,
+    "output_per_mtok": 0.3,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.together.ai/pricing"
+  },
+  "llama-3.3-70b": {
+    "input_per_mtok": 0.88,
+    "output_per_mtok": 0.88,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.together.ai/pricing"
+  },
+  "command-a": {
+    "input_per_mtok": 2.5,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://cohere.com/pricing"
+  },
+  "command-r-plus": {
+    "input_per_mtok": 2.5,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://cohere.com/pricing"
+  },
+  "amazon-nova-pro": {
+    "input_per_mtok": 0.8,
+    "output_per_mtok": 3.2,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://aws.amazon.com/bedrock/pricing/"
+  },
+  "amazon-nova-lite": {
+    "input_per_mtok": 0.06,
+    "output_per_mtok": 0.24,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://aws.amazon.com/bedrock/pricing/"
+  }
+}
diff --git a/src/pricing/catalog.ts b/src/session/pricing.ts
similarity index 75%
rename from src/pricing/catalog.ts
rename to src/session/pricing.ts
index fff0275a..cb9faa8e 100644
--- a/src/pricing/catalog.ts
+++ b/src/session/pricing.ts
@@ -1,9 +1,9 @@
 /**
  * Pricing catalog — single source of truth for per-model USD cost.
  *
- * Deep module, tiny interface. At load it merges the 5 curated vendor JSONs
- * (src/pricing/sources/{anthropic,openai,google,chinese,others}.json) into one
- * Map<modelId, Price> in per-Mtok units, then exposes three pure-ish functions:
+ * Deep module, tiny interface. At load it reads one curated multi-vendor JSON
+ * (src/session/model-prices.json) into a Map<modelId, Price> in per-Mtok units,
+ * then exposes three pure-ish functions:
  *
  *   lookupPrice(modelId)            → Price | null
  *   computeCostUsd(modelId, tokens) → number  | null
@@ -17,18 +17,14 @@
  * an unknown id resolves to `null` (no price) instead of a wrong Claude rate.
  *
  * The large litellm catalog (~1.5MB, ~2900 models) is NOT bundled — it lives at
- * tools/pricing/litellm-catalog.json as the dev-only refresh base for these
- * curated JSONs. See tools/pricing/litellm-NOTES.md for the refresh recipe.
+ * tools/pricing/litellm-catalog.json as the dev-only refresh base for this
+ * curated JSON.
  *
- * The 5 curated JSONs are small (~25KB total) and esbuild inlines them into the
+ * The curated JSON is small (~13KB, 61 models) and esbuild inlines it into the
  * hook/server bundles at build time (no runtime fs read, no external file).
  */
 
-import anthropic from "./sources/anthropic.json" with { type: "json" };
-import openai from "./sources/openai.json" with { type: "json" };
-import google from "./sources/google.json" with { type: "json" };
-import chinese from "./sources/chinese.json" with { type: "json" };
-import others from "./sources/others.json" with { type: "json" };
+import catalog from "./model-prices.json" with { type: "json" };
 
 /** Per-Mtok price for one model. Any of the four rates may be null ("unknown"). */
 export interface Price {
@@ -46,7 +42,7 @@ export interface TokenCounts {
   cache_creation_tokens?: number;
 }
 
-/** Raw shape of a curated source row (carries provenance fields we drop). */
+/** Raw shape of a curated source row (carries a provenance `source` we drop). */
 interface RawRow {
   input_per_mtok: number | null;
   output_per_mtok: number | null;
@@ -56,36 +52,28 @@ interface RawRow {
 }
 
 /**
- * Merge the 5 vendor JSONs into one Map. A row with a null *input* price is
+ * Read the curated JSON into one Map. A row with a null *input* price is
  * unusable for cost (the primary bucket has no rate) and is dropped at load so
  * lookupPrice returns null for it — matching "null-priced entries → no price".
- * Curated ids are globally unique across the five files (verified), so merge
- * order is irrelevant; later files would otherwise win on collision.
+ * (The two null-input ids are already pruned from the JSON itself; this guard
+ * keeps the loader robust if one is ever re-added.)
  */
 function buildCatalog(): Map<string, Price> {
   const map = new Map<string, Price>();
-  const sources: Record<string, RawRow>[] = [
-    anthropic as Record<string, RawRow>,
-    openai as Record<string, RawRow>,
-    google as Record<string, RawRow>,
-    chinese as Record<string, RawRow>,
-    others as Record<string, RawRow>,
-  ];
-  for (const src of sources) {
-    for (const id of Object.keys(src)) {
-      const row = src[id];
-      if (row == null || typeof row !== "object") continue;
-      // No input rate ⇒ no usable price for this model.
-      if (typeof row.input_per_mtok !== "number") continue;
-      map.set(id, {
-        input_per_mtok: row.input_per_mtok,
-        output_per_mtok: typeof row.output_per_mtok === "number" ? row.output_per_mtok : null,
-        cache_read_per_mtok:
-          typeof row.cache_read_per_mtok === "number" ? row.cache_read_per_mtok : null,
-        cache_write_per_mtok:
-          typeof row.cache_write_per_mtok === "number" ? row.cache_write_per_mtok : null,
-      });
-    }
+  const src = catalog as Record<string, RawRow>;
+  for (const id of Object.keys(src)) {
+    const row = src[id];
+    if (row == null || typeof row !== "object") continue;
+    // No input rate ⇒ no usable price for this model.
+    if (typeof row.input_per_mtok !== "number") continue;
+    map.set(id, {
+      input_per_mtok: row.input_per_mtok,
+      output_per_mtok: typeof row.output_per_mtok === "number" ? row.output_per_mtok : null,
+      cache_read_per_mtok:
+        typeof row.cache_read_per_mtok === "number" ? row.cache_read_per_mtok : null,
+      cache_write_per_mtok:
+        typeof row.cache_write_per_mtok === "number" ? row.cache_write_per_mtok : null,
+    });
   }
   return map;
 }
diff --git a/tests/session/extract-agent-usage.test.ts b/tests/session/extract-agent-usage.test.ts
index 5ae161da..85ac0b4d 100644
--- a/tests/session/extract-agent-usage.test.ts
+++ b/tests/session/extract-agent-usage.test.ts
@@ -292,3 +292,90 @@ describe("extractAgentUsage — Issue #4 AgentOutput.usage capture", () => {
     }
   });
 });
+
+/**
+ * Wave 2b — structured cost event.
+ *
+ * The colon-string `data` is opaque to the platform (it cannot column-ize a
+ * "tokens_in:123 cost_usd:0.02" blob). extractAgentUsage now also emits the
+ * cost/token signals as top-level SessionEvent fields, which the forward
+ * envelope spreads straight to the platform as typed columns:
+ *
+ *   model_id, input_tokens, output_tokens,
+ *   cache_read_tokens, cache_creation_tokens, cost_usd
+ *
+ * The colon-string `data` stays for human/debug + back-compat.
+ */
+describe("extractAgentUsage — Wave 2b structured cost fields", () => {
+  function usageEvent(toolInput: Record<string, unknown>, usage: Record<string, unknown>, extra: Record<string, unknown> = {}) {
+    return extractEvents({
+      tool_name: "Task",
+      tool_input: toolInput,
+      tool_response: JSON.stringify({ ...extra, usage }),
+    }).filter((e) => e.type === "agent_usage")[0];
+  }
+
+  // (a) the 6 structured fields ride the event with correct values
+  test("(a) Task usage yields event carrying the 6 structured fields with correct values", () => {
+    const ev = usageEvent(
+      { model: "claude-sonnet-4-6" },
+      {
+        input_tokens: 1000,
+        output_tokens: 500,
+        cache_creation_input_tokens: 1000,
+        cache_read_input_tokens: 1500,
+      },
+    );
+    expect(ev.model_id).toBe("claude-sonnet-4-6");
+    expect(ev.input_tokens).toBe(1000);
+    expect(ev.output_tokens).toBe(500);
+    expect(ev.cache_creation_tokens).toBe(1000);
+    expect(ev.cache_read_tokens).toBe(1500);
+    // 1000*3 + 500*15 + 1000*3.75 + 1500*0.30 = 14700 / 1e6 = 0.0147
+    expect(ev.cost_usd).toBeCloseTo(0.0147, 8);
+  });
+
+  // (b) cost_usd matches the catalog for a known model
+  test("(b) cost_usd matches the catalog for a known model (gpt-5)", () => {
+    const ev = usageEvent(
+      { model: "gpt-5" },
+      { input_tokens: 1000, output_tokens: 500 },
+    );
+    // gpt-5: 1000*1.25 + 500*10 = 6250 / 1e6 = 0.00625
+    expect(ev.cost_usd).toBeCloseTo(0.00625, 8);
+    expect(ev.model_id).toBe("gpt-5");
+  });
+
+  // (c) unknown model → tokens present, cost_usd omitted (no Claude fallback)
+  test("(c) unknown model → tokens present, cost_usd omitted/null", () => {
+    const warn = vi.spyOn(console, "warn").mockImplementation(() => {});
+    const ev = usageEvent(
+      { model: "claude-future-model-99" },
+      { input_tokens: 1000, output_tokens: 500 },
+    );
+    warn.mockRestore();
+    expect(ev.input_tokens).toBe(1000);
+    expect(ev.output_tokens).toBe(500);
+    expect(ev.cost_usd == null).toBe(true);
+  });
+
+  // (d) zero-token response → no cost_usd
+  test("(d) zero-token response → no cost_usd", () => {
+    const ev = usageEvent(
+      { model: "claude-sonnet-4-6" },
+      { input_tokens: 0, output_tokens: 0 },
+    );
+    expect(ev.cost_usd == null).toBe(true);
+  });
+
+  // (e) the existing colon-string `data` still present for back-compat
+  test("(e) colon-string data still present for back-compat", () => {
+    const ev = usageEvent(
+      { model: "claude-sonnet-4-6" },
+      { input_tokens: 1000, output_tokens: 500 },
+    );
+    expect(ev.data).toMatch(/tokens_in:1000/);
+    expect(ev.data).toMatch(/tokens_out:500/);
+    expect(ev.data).toMatch(/cost_usd:0\.0105/);
+  });
+});
diff --git a/tests/pricing/catalog.test.ts b/tests/session/pricing.test.ts
similarity index 95%
rename from tests/pricing/catalog.test.ts
rename to tests/session/pricing.test.ts
index dc5d522f..84fd48e9 100644
--- a/tests/pricing/catalog.test.ts
+++ b/tests/session/pricing.test.ts
@@ -6,7 +6,7 @@
  * a Claude rate (non-Claude models silently inherited Sonnet's default),
  * which over-/under-charged every OpenAI / Gemini / Qwen / DeepSeek turn.
  *
- * The catalog merges the 5 curated vendor JSONs (src/pricing/sources/*.json)
+ * The catalog reads the curated multi-vendor JSON (src/session/model-prices.json)
  * into one per-Mtok price map and prices each model from ITS OWN row.
  * Behaviours under test:
  *   (a) curated lookup hits across all five vendors
@@ -23,9 +23,9 @@ import {
   lookupPrice,
   computeCostUsd,
   nativeOrComputed,
-} from "../../src/pricing/catalog.js";
+} from "../../src/session/pricing.js";
 
-describe("pricing/catalog — lookupPrice", () => {
+describe("session/pricing — lookupPrice", () => {
   // (a) curated lookup hits
   test("(a) Anthropic curated hit returns per-Mtok price", () => {
     const p = lookupPrice("claude-opus-4-8");
@@ -79,7 +79,7 @@ describe("pricing/catalog — lookupPrice", () => {
   });
 });
 
-describe("pricing/catalog — computeCostUsd", () => {
+describe("session/pricing — computeCostUsd", () => {
   // (c) THE BUG: non-Claude model must use its own price, not Claude's.
   test("(c) gpt-5 priced from its own row, NOT Claude default", () => {
     const tokens = { input_tokens: 1000, output_tokens: 500 };
@@ -157,7 +157,7 @@ describe("pricing/catalog — computeCostUsd", () => {
   });
 });
 
-describe("pricing/catalog — nativeOrComputed", () => {
+describe("session/pricing — nativeOrComputed", () => {
   // (g) native-cost passthrough
   test("(g) provider native cost wins over computed", () => {
     const native = nativeOrComputed("gpt-5", { input_tokens: 1000 }, 0.42);
diff --git a/tools/pricing/litellm-NOTES.md b/tools/pricing/litellm-NOTES.md
deleted file mode 100644
index 954e216f..00000000
--- a/tools/pricing/litellm-NOTES.md
+++ /dev/null
@@ -1,80 +0,0 @@
-# LiteLLM Catalog — Adapter Notes
-
-Vendored from BerriAI/litellm as context-mode's comprehensive pricing base, so
-unknown/unseen models still resolve a price instead of failing.
-
-- **Source:** `https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json`
-- **Local file:** `litellm-catalog.json`
-- **Size:** ~1.53 MB (1,570,159 bytes)
-- **Top-level keys:** 2,901 (2,900 model entries + 1 `sample_spec` template — skip `sample_spec`)
-- **Entries with numeric `input_cost_per_token`:** 2,432
-- **Distinct `litellm_provider` values:** 120 (top: fireworks_ai, bedrock, openai, azure, gemini, mistral, openrouter …)
-
-## Shape
-
-The file is a flat JSON object. **Each key is a model id**; each value is a metadata object.
-There is no wrapper array. One special key, `sample_spec`, is a documentation template
-(not a real model) and must be excluded.
-
-```jsonc
-{
-  "sample_spec": { /* template — ignore */ },
-  "gpt-4o": { "input_cost_per_token": 0.0000025, ... },
-  "claude-sonnet-4-20250514": { ... }
-}
-```
-
-## Cost fields — verified present in the fetched JSON
-
-> All `*_cost_per_token` values are **PER TOKEN** (USD), not per-million.
-> To convert to per-Mtok (context-mode's internal unit): **multiply by `1e6` (×1,000,000).**
-> e.g. `gpt-4o` `input_cost_per_token = 0.0000025` → `0.0000025 × 1e6 = $2.50 / Mtok`.
-
-Core fields (use these; coverage count in parentheses):
-
-| Field | Meaning | Count |
-|-------|---------|-------|
-| `input_cost_per_token` | prompt cost per token | 2,433 |
-| `output_cost_per_token` | completion cost per token | 2,430 |
-| `cache_read_input_token_cost` | cached-prompt read cost per token | 645 |
-| `cache_creation_input_token_cost` | cache-write/creation cost per token | 210 |
-
-Real example (`claude-sonnet-4-20250514`, provider `anthropic`):
-`input_cost_per_token: 0.000003` (= $3.00/Mtok), `output_cost_per_token: 0.000015` (= $15.00/Mtok),
-`cache_read_input_token_cost: 3e-7` (= $0.30/Mtok), `cache_creation_input_token_cost: 0.00000375` (= $3.75/Mtok).
-
-### Schema surprises / gotchas
-
-- **Tiered & variant cost fields exist** — e.g. `input_cost_per_token_above_200k_tokens`,
-  `cache_read_input_token_cost_above_200k_tokens`, `_priority`, `_batches`, `_flex`,
-  `_above_1hr`, `_above_272k_tokens`. Treat these as optional overrides; fall back to the
-  base `input_cost_per_token` / `output_cost_per_token`.
-- **Non-token cost units also appear** and are NOT per-token: `input_cost_per_second`,
-  `input_cost_per_character`, `input_cost_per_image`, `input_cost_per_pixel`,
-  `output_cost_per_second`, `output_cost_per_reasoning_token`,
-  `search_context_cost_per_query` (the last can be an object keyed by search-context size).
-  Do not blindly ×1e6 these — only the `*_per_token` family is per-token.
-- **2,366 of 2,900 keys contain `/`** — namespaced ids like `bedrock/...`,
-  `1024-x-1024/dall-e-2`, image/resolution-prefixed entries. Match on the full key.
-- Metadata fields used for context limits: `max_tokens`, `max_input_tokens`,
-  `max_output_tokens` (and `mode` distinguishes `chat`, `embedding`, image, etc.).
-- Some entries carry `deprecation_date` (81 entries).
-
-## Lookup strategy (for the adapter)
-
-Given an incoming `model_id`:
-
-1. **Exact match** — `catalog[model_id]`. Fastest; covers the common case.
-2. **Provider-stripped / namespaced fallback** — many ids are `provider/model`.
-   Try stripping or adding a known provider prefix:
-   - if `model_id` has no `/`, try `catalog[provider + "/" + model_id]`;
-   - if `model_id` is `provider/model`, also try the bare `model` segment.
-3. **Provider + model match** — scan entries whose `litellm_provider` matches the resolved
-   provider and whose key endsWith the model segment.
-4. **Skip `sample_spec`** in every scan.
-5. From the matched entry, read `input_cost_per_token` / `output_cost_per_token`
-   (+ optional `cache_read_input_token_cost`, `cache_creation_input_token_cost`),
-   then **× 1e6** to get per-Mtok rates.
-6. If nothing matches or `input_cost_per_token` is absent (some entries price only by
-   second/character/image), the model has no usable per-token price — fall through to
-   context-mode's own default.