docs(llm): fix stale references in protocols/shared.ts and gemini.ts

- subtractTokens JSDoc said the raw payload lives on Usage.native, but that field was renamed to providerMetadata earlier in this PR. - totalTokens JSDoc still described the abandoned "additive" first-pass contract where inputTokens/outputTokens were non-cached / visible only. We landed on inclusive totals; the fallback already covers cache and reasoning. - Removed a duplicate inline comment in Gemini's mapUsage — the function-level comment already explains the visible/reasoning sum and the undefined-when-incomplete rule.
test(llm): re-record golden scenarios against live providers
2026-08-10 03:19:59 -04:00 · 2026-05-10 22:08:12 -04:00 · 2026-05-10 22:03:02 -04:00 · 2026-05-10 21:42:09 -04:00 · 2026-05-10 20:39:22 -04:00 · 2026-05-10 19:29:41 -04:00
85 changed files with 2461 additions and 3245 deletions
@@ -29,7 +29,7 @@
    },
    "packages/app": {
      "name": "@opencode-ai/app",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@kobalte/core": "catalog:",
        "@opencode-ai/core": "workspace:*",
@@ -85,7 +85,7 @@
    },
    "packages/console/app": {
      "name": "@opencode-ai/console-app",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@cloudflare/vite-plugin": "1.15.2",
        "@ibm/plex": "6.4.1",
@@ -120,7 +120,7 @@
    },
    "packages/console/core": {
      "name": "@opencode-ai/console-core",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@aws-sdk/client-sts": "3.782.0",
        "@jsx-email/render": "1.1.1",
@@ -147,7 +147,7 @@
    },
    "packages/console/function": {
      "name": "@opencode-ai/console-function",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@ai-sdk/anthropic": "3.0.64",
        "@ai-sdk/openai": "3.0.48",
@@ -171,7 +171,7 @@
    },
    "packages/console/mail": {
      "name": "@opencode-ai/console-mail",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@jsx-email/all": "2.2.3",
        "@jsx-email/cli": "1.4.3",
@@ -195,7 +195,7 @@
    },
    "packages/core": {
      "name": "@opencode-ai/core",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "bin": {
        "opencode": "./bin/opencode",
      },
@@ -229,7 +229,7 @@
    },
    "packages/desktop": {
      "name": "@opencode-ai/desktop",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "drizzle-orm": "catalog:",
        "effect": "catalog:",
@@ -283,7 +283,7 @@
    },
    "packages/enterprise": {
      "name": "@opencode-ai/enterprise",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@opencode-ai/core": "workspace:*",
        "@opencode-ai/ui": "workspace:*",
@@ -313,7 +313,7 @@
    },
    "packages/function": {
      "name": "@opencode-ai/function",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@octokit/auth-app": "8.0.1",
        "@octokit/rest": "catalog:",
@@ -329,7 +329,7 @@
    },
    "packages/http-recorder": {
      "name": "@opencode-ai/http-recorder",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@effect/platform-node": "catalog:",
        "effect": "catalog:",
@@ -342,7 +342,7 @@
    },
    "packages/llm": {
      "name": "@opencode-ai/llm",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@smithy/eventstream-codec": "4.2.14",
        "@smithy/util-utf8": "4.2.2",
@@ -360,7 +360,7 @@
    },
    "packages/opencode": {
      "name": "opencode",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "bin": {
        "opencode": "./bin/opencode",
      },
@@ -496,7 +496,7 @@
    },
    "packages/plugin": {
      "name": "@opencode-ai/plugin",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@opencode-ai/sdk": "workspace:*",
        "effect": "catalog:",
@@ -534,7 +534,7 @@
    },
    "packages/sdk/js": {
      "name": "@opencode-ai/sdk",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "cross-spawn": "catalog:",
      },
@@ -549,7 +549,7 @@
    },
    "packages/slack": {
      "name": "@opencode-ai/slack",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@opencode-ai/sdk": "workspace:*",
        "@slack/bolt": "^3.17.1",
@@ -584,7 +584,7 @@
    },
    "packages/ui": {
      "name": "@opencode-ai/ui",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@kobalte/core": "catalog:",
        "@opencode-ai/core": "workspace:*",
@@ -633,7 +633,7 @@
    },
    "packages/web": {
      "name": "@opencode-ai/web",
-      "version": "1.14.48",
+      "version": "1.14.46",
      "dependencies": {
        "@astrojs/cloudflare": "12.6.3",
        "@astrojs/markdown-remark": "6.3.1",
@@ -1,6 +1,6 @@
 {
  "name": "@opencode-ai/app",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "description": "",
  "type": "module",
  "exports": {
@@ -1,6 +1,6 @@
 {
  "name": "@opencode-ai/console-app",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "type": "module",
  "license": "MIT",
  "scripts": {
@@ -53,7 +53,7 @@ export function UsageSection() {
  }

  const calculateTotalOutputTokens = (u: Awaited<ReturnType<typeof getUsageInfo>>[0]) => {
-    return u.outputTokens
+    return u.outputTokens + (u.reasoningTokens ?? 0)
  }

  const goPrev = async () => {
@@ -889,6 +889,10 @@ export async function handler(

    const inputCost = modelCost.input * inputTokens * 100
    const outputCost = modelCost.output * outputTokens * 100
+    const reasoningCost = (() => {
+      if (!reasoningTokens) return undefined
+      return modelCost.output * reasoningTokens * 100
+    })()
    const cacheReadCost = (() => {
      if (!cacheReadTokens) return undefined
      if (!modelCost.cacheRead) return undefined
@@ -905,11 +909,17 @@ export async function handler(
      return modelCost.cacheWrite1h * cacheWrite1hTokens * 100
    })()
    const totalCostInCent =
-      inputCost + outputCost + (cacheReadCost ?? 0) + (cacheWrite5mCost ?? 0) + (cacheWrite1hCost ?? 0)
+      inputCost +
+      outputCost +
+      (reasoningCost ?? 0) +
+      (cacheReadCost ?? 0) +
+      (cacheWrite5mCost ?? 0) +
+      (cacheWrite1hCost ?? 0)
    return {
      totalCostInCent,
      inputCost,
      outputCost,
+      reasoningCost,
      cacheReadCost,
      cacheWrite5mCost,
      cacheWrite1hCost,
@@ -931,7 +941,8 @@ export async function handler(
  ) {
    const { inputTokens, outputTokens, reasoningTokens, cacheReadTokens, cacheWrite5mTokens, cacheWrite1hTokens } =
      usageInfo
-    const { totalCostInCent, inputCost, outputCost, cacheReadCost, cacheWrite5mCost, cacheWrite1hCost } = costInfo
+    const { totalCostInCent, inputCost, outputCost, reasoningCost, cacheReadCost, cacheWrite5mCost, cacheWrite1hCost } =
+      costInfo

    logger.metric({
      "tokens.input": inputTokens,
@@ -942,12 +953,14 @@ export async function handler(
      "tokens.cache_write_1h": cacheWrite1hTokens,
      "cost.input.microcents": centsToMicroCents(inputCost),
      "cost.output.microcents": centsToMicroCents(outputCost),
+      "cost.reasoning.microcents": reasoningCost ? centsToMicroCents(reasoningCost) : undefined,
      "cost.cache_read.microcents": cacheReadCost ? centsToMicroCents(cacheReadCost) : undefined,
      "cost.cache_write.microcents": cacheWrite5mCost ? centsToMicroCents(cacheWrite5mCost) : undefined,
      "cost.total.microcents": centsToMicroCents(totalCostInCent),
      // deprecated - remove after May 20, 2026
      "cost.input": Math.round(inputCost),
      "cost.output": Math.round(outputCost),
+      "cost.reasoning": reasoningCost ? Math.round(reasoningCost) : undefined,
      "cost.cache_read": cacheReadCost ? Math.round(cacheReadCost) : undefined,
      "cost.cache_write_5m": cacheWrite5mCost ? Math.round(cacheWrite5mCost) : undefined,
      "cost.cache_write_1h": cacheWrite1hCost ? Math.round(cacheWrite1hCost) : undefined,
@@ -50,7 +50,7 @@ export const openaiHelper: ProviderHelper = ({ workspaceID }) => ({
    const cacheReadTokens = usage.input_tokens_details?.cached_tokens ?? undefined
    return {
      inputTokens: inputTokens - (cacheReadTokens ?? 0),
-      outputTokens,
+      outputTokens: outputTokens - (reasoningTokens ?? 0),
      reasoningTokens,
      cacheReadTokens,
      cacheWrite5mTokens: undefined,
@@ -1,7 +1,7 @@
 {
  "$schema": "https://json.schemastore.org/package.json",
  "name": "@opencode-ai/console-core",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "private": true,
  "type": "module",
  "license": "MIT",
@@ -1,6 +1,6 @@
 {
  "name": "@opencode-ai/console-function",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "$schema": "https://json.schemastore.org/package.json",
  "private": true,
  "type": "module",
@@ -1,6 +1,6 @@
 {
  "name": "@opencode-ai/console-mail",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "dependencies": {
    "@jsx-email/all": "2.2.3",
    "@jsx-email/cli": "1.4.3",
@@ -1,6 +1,6 @@
 {
  "$schema": "https://json.schemastore.org/package.json",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "name": "@opencode-ai/core",
  "type": "module",
  "license": "MIT",
@@ -1,7 +1,7 @@
 {
  "name": "@opencode-ai/desktop",
  "private": true,
-  "version": "1.14.48",
+  "version": "1.14.46",
  "type": "module",
  "license": "MIT",
  "homepage": "https://opencode.ai",
@@ -1,6 +1,6 @@
 {
  "name": "@opencode-ai/enterprise",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "private": true,
  "type": "module",
  "license": "MIT",
@@ -1,7 +1,7 @@
 id = "opencode"
 name = "OpenCode"
 description = "The open source coding agent."
-version = "1.14.48"
+version = "1.14.46"
 schema_version = 1
 authors = ["Anomaly"]
 repository = "https://github.com/anomalyco/opencode"
@@ -11,26 +11,26 @@ name = "OpenCode"
 icon = "./icons/opencode.svg"

 [agent_servers.opencode.targets.darwin-aarch64]
-archive = "https://github.com/anomalyco/opencode/releases/download/v1.14.48/opencode-darwin-arm64.zip"
+archive = "https://github.com/anomalyco/opencode/releases/download/v1.14.46/opencode-darwin-arm64.zip"
 cmd = "./opencode"
 args = ["acp"]

 [agent_servers.opencode.targets.darwin-x86_64]
-archive = "https://github.com/anomalyco/opencode/releases/download/v1.14.48/opencode-darwin-x64.zip"
+archive = "https://github.com/anomalyco/opencode/releases/download/v1.14.46/opencode-darwin-x64.zip"
 cmd = "./opencode"
 args = ["acp"]

 [agent_servers.opencode.targets.linux-aarch64]
-archive = "https://github.com/anomalyco/opencode/releases/download/v1.14.48/opencode-linux-arm64.tar.gz"
+archive = "https://github.com/anomalyco/opencode/releases/download/v1.14.46/opencode-linux-arm64.tar.gz"
 cmd = "./opencode"
 args = ["acp"]

 [agent_servers.opencode.targets.linux-x86_64]
-archive = "https://github.com/anomalyco/opencode/releases/download/v1.14.48/opencode-linux-x64.tar.gz"
+archive = "https://github.com/anomalyco/opencode/releases/download/v1.14.46/opencode-linux-x64.tar.gz"
 cmd = "./opencode"
 args = ["acp"]

 [agent_servers.opencode.targets.windows-x86_64]
-archive = "https://github.com/anomalyco/opencode/releases/download/v1.14.48/opencode-windows-x64.zip"
+archive = "https://github.com/anomalyco/opencode/releases/download/v1.14.46/opencode-windows-x64.zip"
 cmd = "./opencode.exe"
 args = ["acp"]
@@ -1,6 +1,6 @@
 {
  "name": "@opencode-ai/function",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "$schema": "https://json.schemastore.org/package.json",
  "private": true,
  "type": "module",
@@ -1,6 +1,6 @@
 {
  "$schema": "https://json.schemastore.org/package.json",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "name": "@opencode-ai/http-recorder",
  "type": "module",
  "license": "MIT",
@@ -1,6 +1,6 @@
 {
  "$schema": "https://json.schemastore.org/package.json",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "name": "@opencode-ai/llm",
  "type": "module",
  "license": "MIT",
@@ -16,7 +16,6 @@ import {
  type ToolResultPart,
 } from "../schema"
 import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared"
-import * as Cache from "./utils/cache"
 import { ToolStream } from "./utils/tool-stream"

 const ADAPTER = "anthropic-messages"
@@ -26,10 +25,7 @@ export const PATH = "/messages"
 // =============================================================================
 // Request Body Schema
 // =============================================================================
-const AnthropicCacheControl = Schema.Struct({
-  type: Schema.tag("ephemeral"),
-  ttl: Schema.optional(Schema.Literals(["5m", "1h"])),
-})
+const AnthropicCacheControl = Schema.Struct({ type: Schema.tag("ephemeral") })

 const AnthropicTextBlock = Schema.Struct({
  type: Schema.tag("text"),
@@ -197,24 +193,8 @@ const invalid = ProviderShared.invalidRequest
 // =============================================================================
 // Request Lowering
 // =============================================================================
-// Anthropic accepts at most 4 explicit cache_control breakpoints per request,
-// across `tools`, `system`, and `messages`. Beyond the cap the API returns a
-// 400 — so the lowering layer counts emitted markers and silently drops any
-// that exceed it.
-const ANTHROPIC_BREAKPOINT_CAP = 4
-
-const EPHEMERAL_5M = { type: "ephemeral" as const }
-const EPHEMERAL_1H = { type: "ephemeral" as const, ttl: "1h" as const }
-
-const cacheControl = (breakpoints: Cache.Breakpoints, cache: CacheHint | undefined) => {
-  if (cache?.type !== "ephemeral" && cache?.type !== "persistent") return undefined
-  if (breakpoints.remaining <= 0) {
-    breakpoints.dropped += 1
-    return undefined
-  }
-  breakpoints.remaining -= 1
-  return Cache.ttlBucket(cache.ttlSeconds) === "1h" ? EPHEMERAL_1H : EPHEMERAL_5M
-}
+const cacheControl = (cache: CacheHint | undefined) =>
+  cache?.type === "ephemeral" ? { type: "ephemeral" as const } : undefined

 const anthropicMetadata = (metadata: Record<string, unknown>): ProviderMetadata => ({ anthropic: metadata })

@@ -224,11 +204,10 @@ const signatureFromMetadata = (metadata: ProviderMetadata | undefined): string |
  return typeof anthropic.signature === "string" ? anthropic.signature : undefined
 }

-const lowerTool = (breakpoints: Cache.Breakpoints, tool: ToolDefinition): AnthropicTool => ({
+const lowerTool = (tool: ToolDefinition): AnthropicTool => ({
  name: tool.name,
  description: tool.description,
  input_schema: tool.inputSchema,
-  cache_control: cacheControl(breakpoints, tool.cache),
 })

 const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
@@ -270,10 +249,7 @@ const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult
  return { type: wireType, tool_use_id: part.id, content: part.result.value } satisfies AnthropicServerToolResultBlock
 })

-const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
-  request: LLMRequest,
-  breakpoints: Cache.Breakpoints,
-) {
+const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (request: LLMRequest) {
  const messages: AnthropicMessage[] = []

  for (const message of request.messages) {
@@ -282,7 +258,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
      for (const part of message.content) {
        if (!ProviderShared.supportsContent(part, ["text"]))
          return yield* ProviderShared.unsupportedContent("Anthropic Messages", "user", ["text"])
-        content.push({ type: "text", text: part.text, cache_control: cacheControl(breakpoints, part.cache) })
+        content.push({ type: "text", text: part.text, cache_control: cacheControl(part.cache) })
      }
      messages.push({ role: "user", content })
      continue
@@ -292,7 +268,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
      const content: AnthropicAssistantBlock[] = []
      for (const part of message.content) {
        if (part.type === "text") {
-          content.push({ type: "text", text: part.text, cache_control: cacheControl(breakpoints, part.cache) })
+          content.push({ type: "text", text: part.text, cache_control: cacheControl(part.cache) })
          continue
        }
        if (part.type === "reasoning") {
@@ -328,7 +304,6 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
        tool_use_id: part.id,
        content: ProviderShared.toolResultText(part),
        is_error: part.result.type === "error" ? true : undefined,
-        cache_control: cacheControl(breakpoints, part.cache),
      })
    }
    messages.push({ role: "user", content })
@@ -355,33 +330,18 @@ const lowerThinking = Effect.fn("AnthropicMessages.lowerThinking")(function* (re
 const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (request: LLMRequest) {
  const toolChoice = request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined
  const generation = request.generation
-  // Allocate the 4-breakpoint budget in invalidation order: tools → system →
-  // messages. Tools live highest in the cache hierarchy, so when callers
-  // over-mark we keep their tool hints and shed the message-tail ones first.
-  const breakpoints = Cache.newBreakpoints(ANTHROPIC_BREAKPOINT_CAP)
-  const tools =
-    request.tools.length === 0 || request.toolChoice?.type === "none"
-      ? undefined
-      : request.tools.map((tool) => lowerTool(breakpoints, tool))
-  const system =
-    request.system.length === 0
-      ? undefined
-      : request.system.map((part) => ({
-          type: "text" as const,
-          text: part.text,
-          cache_control: cacheControl(breakpoints, part.cache),
-        }))
-  const messages = yield* lowerMessages(request, breakpoints)
-  if (breakpoints.dropped > 0) {
-    yield* Effect.logWarning(
-      `Anthropic Messages: dropped ${breakpoints.dropped} cache breakpoint(s); the API allows at most ${ANTHROPIC_BREAKPOINT_CAP} per request.`,
-    )
-  }
  return {
    model: request.model.id,
-    system,
-    messages,
-    tools,
+    system:
+      request.system.length === 0
+        ? undefined
+        : request.system.map((part) => ({
+            type: "text" as const,
+            text: part.text,
+            cache_control: cacheControl(part.cache),
+          })),
+    messages: yield* lowerMessages(request),
+    tools: request.tools.length === 0 || request.toolChoice?.type === "none" ? undefined : request.tools.map(lowerTool),
    tool_choice: toolChoice,
    stream: true as const,
    max_tokens: generation?.maxTokens ?? request.model.limits.output ?? 4096,
@@ -404,34 +364,56 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
  return "unknown"
 }

+// Anthropic reports the non-overlapping breakdown natively — its
+// `input_tokens` is the *non-cached* count per the Messages API docs, with
+// cache reads and writes as separate fields. We sum them to derive the
+// inclusive `inputTokens` the rest of the contract expects. Extended
+// thinking tokens are *not* broken out by Anthropic — they're billed as
+// part of `output_tokens`, so `reasoningTokens` stays `undefined` and
+// `outputTokens` carries the combined total.
 const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
  if (!usage) return undefined
+  const nonCached = usage.input_tokens
+  const cacheRead = usage.cache_read_input_tokens ?? undefined
+  const cacheWrite = usage.cache_creation_input_tokens ?? undefined
+  const inputTokens = ProviderShared.sumTokens(nonCached, cacheRead, cacheWrite)
  return new Usage({
-    inputTokens: usage.input_tokens,
+    inputTokens,
    outputTokens: usage.output_tokens,
-    cacheReadInputTokens: usage.cache_read_input_tokens ?? undefined,
-    cacheWriteInputTokens: usage.cache_creation_input_tokens ?? undefined,
-    totalTokens: ProviderShared.totalTokens(usage.input_tokens, usage.output_tokens, undefined),
-    native: usage,
+    nonCachedInputTokens: nonCached,
+    cacheReadInputTokens: cacheRead,
+    cacheWriteInputTokens: cacheWrite,
+    totalTokens: ProviderShared.totalTokens(inputTokens, usage.output_tokens, undefined),
+    providerMetadata: { anthropic: usage },
  })
 }

 // Anthropic emits usage on `message_start` and again on `message_delta` — the
 // final delta carries the authoritative totals. Right-biased merge: each
-// field prefers `right` when defined, falls back to `left`. `totalTokens` is
-// recomputed from the merged input/output to stay consistent.
+// field prefers `right` when defined, falls back to `left`. `inputTokens` is
+// recomputed from the merged breakdown so the inclusive total stays
+// consistent with `nonCached + cacheRead + cacheWrite`.
 const mergeUsage = (left: Usage | undefined, right: Usage | undefined) => {
  if (!left) return right
  if (!right) return left
-  const inputTokens = right.inputTokens ?? left.inputTokens
+  const nonCachedInputTokens = right.nonCachedInputTokens ?? left.nonCachedInputTokens
+  const cacheReadInputTokens = right.cacheReadInputTokens ?? left.cacheReadInputTokens
+  const cacheWriteInputTokens = right.cacheWriteInputTokens ?? left.cacheWriteInputTokens
+  const inputTokens = ProviderShared.sumTokens(nonCachedInputTokens, cacheReadInputTokens, cacheWriteInputTokens)
  const outputTokens = right.outputTokens ?? left.outputTokens
  return new Usage({
    inputTokens,
    outputTokens,
-    cacheReadInputTokens: right.cacheReadInputTokens ?? left.cacheReadInputTokens,
-    cacheWriteInputTokens: right.cacheWriteInputTokens ?? left.cacheWriteInputTokens,
+    nonCachedInputTokens,
+    cacheReadInputTokens,
+    cacheWriteInputTokens,
    totalTokens: ProviderShared.totalTokens(inputTokens, outputTokens, undefined),
-    native: { ...left.native, ...right.native },
+    providerMetadata: {
+      anthropic: {
+        ...(left.providerMetadata?.["anthropic"] ?? {}),
+        ...(right.providerMetadata?.["anthropic"] ?? {}),
+      },
+    },
  })
 }

@@ -108,7 +108,7 @@ type BedrockMessage = Schema.Schema.Type<typeof BedrockMessage>
 const BedrockSystemBlock = Schema.Union([BedrockTextBlock, BedrockCache.CachePointBlock])
 type BedrockSystemBlock = Schema.Schema.Type<typeof BedrockSystemBlock>

-const BedrockToolSpec = Schema.Struct({
+const BedrockTool = Schema.Struct({
  toolSpec: Schema.Struct({
    name: Schema.String,
    description: Schema.String,
@@ -117,9 +117,6 @@ const BedrockToolSpec = Schema.Struct({
    }),
  }),
 })
-type BedrockToolSpec = Schema.Schema.Type<typeof BedrockToolSpec>
-
-const BedrockTool = Schema.Union([BedrockToolSpec, BedrockCache.CachePointBlock])
 type BedrockTool = Schema.Schema.Type<typeof BedrockTool>

 const BedrockToolChoice = Schema.Union([
@@ -217,7 +214,7 @@ type BedrockEvent = Schema.Schema.Type<typeof BedrockEvent>
 // =============================================================================
 // Request Lowering
 // =============================================================================
-const lowerToolSpec = (tool: ToolDefinition): BedrockToolSpec => ({
+const lowerTool = (tool: ToolDefinition): BedrockTool => ({
  toolSpec: {
    name: tool.name,
    description: tool.description,
@@ -225,22 +222,11 @@ const lowerToolSpec = (tool: ToolDefinition): BedrockToolSpec => ({
  },
 })

-const lowerTools = (breakpoints: BedrockCache.Breakpoints, tools: ReadonlyArray<ToolDefinition>): BedrockTool[] => {
-  const result: BedrockTool[] = []
-  for (const tool of tools) {
-    result.push(lowerToolSpec(tool))
-    const cachePoint = BedrockCache.block(breakpoints, tool.cache)
-    if (cachePoint) result.push(cachePoint)
-  }
-  return result
-}
-
 const textWithCache = (
-  breakpoints: BedrockCache.Breakpoints,
  text: string,
  cache: CacheHint | undefined,
 ): Array<BedrockTextBlock | BedrockCache.CachePointBlock> => {
-  const cachePoint = BedrockCache.block(breakpoints, cache)
+  const cachePoint = BedrockCache.block(cache)
  return cachePoint ? [{ text }, cachePoint] : [{ text }]
 }

@@ -271,10 +257,7 @@ const lowerToolResult = (part: ToolResultPart): BedrockToolResultBlock => ({
  },
 })

-const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
-  request: LLMRequest,
-  breakpoints: BedrockCache.Breakpoints,
-) {
+const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (request: LLMRequest) {
  const messages: BedrockMessage[] = []

  for (const message of request.messages) {
@@ -284,7 +267,7 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
        if (!ProviderShared.supportsContent(part, ["text", "media"]))
          return yield* ProviderShared.unsupportedContent("Bedrock Converse", "user", ["text", "media"])
        if (part.type === "text") {
-          content.push(...textWithCache(breakpoints, part.text, part.cache))
+          content.push(...textWithCache(part.text, part.cache))
          continue
        }
        if (part.type === "media") {
@@ -306,7 +289,7 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
            "tool-call",
          ])
        if (part.type === "text") {
-          content.push(...textWithCache(breakpoints, part.text, part.cache))
+          content.push(...textWithCache(part.text, part.cache))
          continue
        }
        if (part.type === "reasoning") {
@@ -326,13 +309,11 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
      continue
    }

-    const content: BedrockUserBlock[] = []
+    const content: BedrockToolResultBlock[] = []
    for (const part of message.content) {
      if (!ProviderShared.supportsContent(part, ["tool-result"]))
        return yield* ProviderShared.unsupportedContent("Bedrock Converse", "tool", ["tool-result"])
      content.push(lowerToolResult(part))
-      const cachePoint = BedrockCache.block(breakpoints, part.cache)
-      if (cachePoint) content.push(cachePoint)
    }
    messages.push({ role: "user", content })
  }
@@ -342,32 +323,16 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (

 // System prompts share the cache-point convention: emit the text block, then
 // optionally a positional `cachePoint` marker.
-const lowerSystem = (
-  breakpoints: BedrockCache.Breakpoints,
-  system: ReadonlyArray<LLMRequest["system"][number]>,
-): BedrockSystemBlock[] => system.flatMap((part) => textWithCache(breakpoints, part.text, part.cache))
+const lowerSystem = (system: ReadonlyArray<LLMRequest["system"][number]>): BedrockSystemBlock[] =>
+  system.flatMap((part) => textWithCache(part.text, part.cache))

 const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request: LLMRequest) {
  const toolChoice = request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined
  const generation = request.generation
-  // Bedrock-Claude shares Anthropic's 4-breakpoint cap. Spend the budget in
-  // tools → system → messages order to favour the highest-impact prefixes.
-  const breakpoints = BedrockCache.breakpoints()
-  const toolConfig =
-    request.tools.length > 0 && request.toolChoice?.type !== "none"
-      ? { tools: lowerTools(breakpoints, request.tools), toolChoice }
-      : undefined
-  const system = request.system.length === 0 ? undefined : lowerSystem(breakpoints, request.system)
-  const messages = yield* lowerMessages(request, breakpoints)
-  if (breakpoints.dropped > 0) {
-    yield* Effect.logWarning(
-      `Bedrock Converse: dropped ${breakpoints.dropped} cache breakpoint(s); the API allows at most ${BedrockCache.BEDROCK_BREAKPOINT_CAP} per request.`,
-    )
-  }
  return {
    modelId: request.model.id,
-    messages,
-    system,
+    messages: yield* lowerMessages(request),
+    system: request.system.length === 0 ? undefined : lowerSystem(request.system),
    inferenceConfig:
      generation?.maxTokens === undefined &&
      generation?.temperature === undefined &&
@@ -380,7 +345,10 @@ const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request:
            topP: generation?.topP,
            stopSequences: generation?.stop,
          },
-    toolConfig,
+    toolConfig:
+      request.tools.length > 0 && request.toolChoice?.type !== "none"
+        ? { tools: request.tools.map(lowerTool), toolChoice }
+        : undefined,
  }
 })

@@ -395,15 +363,22 @@ const mapFinishReason = (reason: string): FinishReason => {
  return "unknown"
 }

+// AWS Bedrock Converse reports `inputTokens` (inclusive total) with
+// `cacheReadInputTokens` and `cacheWriteInputTokens` as subsets. Pass
+// the total through and derive the non-cached breakdown. Bedrock does
+// not break reasoning out of `outputTokens` for any current model.
 const mapUsage = (usage: BedrockUsageSchema | undefined): Usage | undefined => {
  if (!usage) return undefined
+  const cacheTotal = (usage.cacheReadInputTokens ?? 0) + (usage.cacheWriteInputTokens ?? 0)
+  const nonCached = ProviderShared.subtractTokens(usage.inputTokens, cacheTotal)
  return new Usage({
    inputTokens: usage.inputTokens,
    outputTokens: usage.outputTokens,
-    totalTokens: ProviderShared.totalTokens(usage.inputTokens, usage.outputTokens, usage.totalTokens),
+    nonCachedInputTokens: nonCached,
    cacheReadInputTokens: usage.cacheReadInputTokens,
    cacheWriteInputTokens: usage.cacheWriteInputTokens,
-    native: usage,
+    totalTokens: ProviderShared.totalTokens(usage.inputTokens, usage.outputTokens, usage.totalTokens),
+    providerMetadata: { bedrock: usage },
  })
 }

@@ -281,15 +281,28 @@ const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMReque
 // =============================================================================
 // Stream Parsing
 // =============================================================================
+// Gemini reports `promptTokenCount` (inclusive total) with a
+// `cachedContentTokenCount` subset. `candidatesTokenCount` is *exclusive*
+// of `thoughtsTokenCount` — visible-only, not a total — so we sum the two
+// to produce the inclusive `outputTokens` the rest of the contract expects.
+// Output is left undefined when the visible component is missing, so we
+// don't fabricate an inclusive number from a partial breakdown.
 const mapUsage = (usage: GeminiUsage | undefined) => {
  if (!usage) return undefined
+  const cached = usage.cachedContentTokenCount
+  const nonCached = ProviderShared.subtractTokens(usage.promptTokenCount, cached)
+  const outputTokens =
+    usage.candidatesTokenCount !== undefined
+      ? usage.candidatesTokenCount + (usage.thoughtsTokenCount ?? 0)
+      : undefined
  return new Usage({
    inputTokens: usage.promptTokenCount,
-    outputTokens: usage.candidatesTokenCount,
+    outputTokens,
+    nonCachedInputTokens: nonCached,
+    cacheReadInputTokens: cached,
    reasoningTokens: usage.thoughtsTokenCount,
-    cacheReadInputTokens: usage.cachedContentTokenCount,
-    totalTokens: ProviderShared.totalTokens(usage.promptTokenCount, usage.candidatesTokenCount, usage.totalTokenCount),
-    native: usage,
+    totalTokens: ProviderShared.totalTokens(usage.promptTokenCount, outputTokens, usage.totalTokenCount),
+    providerMetadata: { google: usage },
  })
 }

@@ -290,15 +290,24 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
  return "unknown"
 }

+// OpenAI Chat reports `prompt_tokens` (inclusive total) with a
+// `cached_tokens` subset, and `completion_tokens` (inclusive total) with
+// a `reasoning_tokens` subset. We pass the inclusive totals through and
+// derive the non-cached breakdown so the `LLM.Usage` contract is
+// satisfied on both sides.
 const mapUsage = (usage: OpenAIChatEvent["usage"]): Usage | undefined => {
  if (!usage) return undefined
+  const cached = usage.prompt_tokens_details?.cached_tokens
+  const reasoning = usage.completion_tokens_details?.reasoning_tokens
+  const nonCached = ProviderShared.subtractTokens(usage.prompt_tokens, cached)
  return new Usage({
    inputTokens: usage.prompt_tokens,
    outputTokens: usage.completion_tokens,
-    reasoningTokens: usage.completion_tokens_details?.reasoning_tokens,
-    cacheReadInputTokens: usage.prompt_tokens_details?.cached_tokens,
+    nonCachedInputTokens: nonCached,
+    cacheReadInputTokens: cached,
+    reasoningTokens: reasoning,
    totalTokens: ProviderShared.totalTokens(usage.prompt_tokens, usage.completion_tokens, usage.total_tokens),
-    native: usage,
+    providerMetadata: { openai: usage },
  })
 }

@@ -276,15 +276,23 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
 // =============================================================================
 // Stream Parsing
 // =============================================================================
+// OpenAI Responses reports `input_tokens` (inclusive total) with a
+// `cached_tokens` subset, and `output_tokens` (inclusive total) with a
+// `reasoning_tokens` subset. Pass the totals through and derive the
+// non-cached breakdown.
 const mapUsage = (usage: OpenAIResponsesUsage | null | undefined) => {
  if (!usage) return undefined
+  const cached = usage.input_tokens_details?.cached_tokens
+  const reasoning = usage.output_tokens_details?.reasoning_tokens
+  const nonCached = ProviderShared.subtractTokens(usage.input_tokens, cached)
  return new Usage({
    inputTokens: usage.input_tokens,
    outputTokens: usage.output_tokens,
-    reasoningTokens: usage.output_tokens_details?.reasoning_tokens,
-    cacheReadInputTokens: usage.input_tokens_details?.cached_tokens,
+    nonCachedInputTokens: nonCached,
+    cacheReadInputTokens: cached,
+    reasoningTokens: reasoning,
    totalTokens: ProviderShared.totalTokens(usage.input_tokens, usage.output_tokens, usage.total_tokens),
-    native: usage,
+    providerMetadata: { openai: usage },
  })
 }

@@ -42,6 +42,11 @@ export interface ToolAccumulator {
 * supplied total; otherwise falls back to `inputTokens + outputTokens` only
 * when at least one is defined. Returns `undefined` when neither input nor
 * output is known so routes don't publish a misleading `0`.
+ *
+ * Under the `LLM.Usage` contract, `inputTokens` and `outputTokens` are
+ * inclusive totals, so the computed fallback already covers cache reads /
+ * writes and reasoning — used mainly for Anthropic-style providers that
+ * don't surface a top-level total.
 */
 export const totalTokens = (
  inputTokens: number | undefined,
@@ -53,6 +58,39 @@ export const totalTokens = (
  return (inputTokens ?? 0) + (outputTokens ?? 0)
 }

+/**
+ * Subtract `subtrahend` from `total`, clamping to zero if the provider
+ * reports a non-sensical breakdown (e.g. `cached_tokens > prompt_tokens`).
+ * Used by protocol mappers when deriving a non-overlapping breakdown field
+ * from a provider's inclusive total — `nonCachedInputTokens` from
+ * `inputTokens - cacheReadInputTokens - cacheWriteInputTokens`.
+ *
+ * If `total` is `undefined`, returns `undefined` (we don't fabricate
+ * counts). If `subtrahend` is `undefined`, returns `total` unchanged. The
+ * provider-native breakdown stays available on `Usage.providerMetadata`
+ * for debugging.
+ */
+export const subtractTokens = (
+  total: number | undefined,
+  subtrahend: number | undefined,
+): number | undefined => {
+  if (total === undefined) return undefined
+  if (subtrahend === undefined) return total
+  return Math.max(0, total - subtrahend)
+}
+
+/**
+ * Sum a list of optional token counts, returning `undefined` only when
+ * every value is `undefined` (so we don't fabricate a `0`). Used by
+ * protocol mappers to derive the inclusive `inputTokens` total from a
+ * provider that natively reports a non-overlapping breakdown
+ * (e.g. Anthropic, whose `input_tokens` is already non-cached only).
+ */
+export const sumTokens = (...values: ReadonlyArray<number | undefined>): number | undefined => {
+  if (values.every((value) => value === undefined)) return undefined
+  return values.reduce<number>((acc, value) => acc + (value ?? 0), 0)
+}
+
 export const eventError = (route: string, message: string, raw?: string) =>
  new LLMError({
    module: "ProviderShared",
@@ -1,37 +1,20 @@
 import { Schema } from "effect"
 import type { CacheHint } from "../../schema"
-import { newBreakpoints, ttlBucket, type Breakpoints } from "./cache"

 // Bedrock cache markers are positional: emit a `cachePoint` block immediately
-// after the content the caller wants treated as a cacheable prefix. Bedrock
-// accepts optional `ttl: "5m" | "1h"` on cachePoint, mirroring Anthropic.
+// after the content the caller wants treated as a cacheable prefix.
 export const CachePointBlock = Schema.Struct({
-  cachePoint: Schema.Struct({
-    type: Schema.tag("default"),
-    ttl: Schema.optional(Schema.Literals(["5m", "1h"])),
-  }),
+  cachePoint: Schema.Struct({ type: Schema.tag("default") }),
 })
 export type CachePointBlock = Schema.Schema.Type<typeof CachePointBlock>

-// Bedrock-Claude enforces the same 4-breakpoint cap as the Anthropic Messages
-// API. Callers pass a shared counter through every `block()` call site so the
-// budget is respected across `system`, `messages`, and `tools`.
-export const BEDROCK_BREAKPOINT_CAP = 4
+// Bedrock recently added optional `ttl: "5m" | "1h"` on cachePoint. Map
+// `CacheHint.ttlSeconds` here once a recorded cassette validates the wire shape.
+const DEFAULT: CachePointBlock = { cachePoint: { type: "default" } }

-export type { Breakpoints } from "./cache"
-export const breakpoints = () => newBreakpoints(BEDROCK_BREAKPOINT_CAP)
-
-const DEFAULT_5M: CachePointBlock = { cachePoint: { type: "default" } }
-const DEFAULT_1H: CachePointBlock = { cachePoint: { type: "default", ttl: "1h" } }
-
-export const block = (breakpoints: Breakpoints, cache: CacheHint | undefined): CachePointBlock | undefined => {
+export const block = (cache: CacheHint | undefined): CachePointBlock | undefined => {
  if (cache?.type !== "ephemeral" && cache?.type !== "persistent") return undefined
-  if (breakpoints.remaining <= 0) {
-    breakpoints.dropped += 1
-    return undefined
-  }
-  breakpoints.remaining -= 1
-  return ttlBucket(cache.ttlSeconds) === "1h" ? DEFAULT_1H : DEFAULT_5M
+  return DEFAULT
 }

 export * as BedrockCache from "./bedrock-cache"
@@ -1,16 +0,0 @@
-// Shared helpers for provider cache-marker lowering. Anthropic and Bedrock
-// both enforce a 4-breakpoint cap per request and accept the same `5m`/`1h`
-// TTL buckets, so the counter and TTL mapping live here.
-
-export interface Breakpoints {
-  remaining: number
-  dropped: number
-}
-
-export const newBreakpoints = (cap: number): Breakpoints => ({ remaining: cap, dropped: 0 })
-
-// Returns `"1h"` for any `ttlSeconds >= 3600`, otherwise `undefined` (the
-// provider default 5m). Anthropic & Bedrock both treat anything shorter than
-// an hour as 5m.
-export const ttlBucket = (ttlSeconds: number | undefined): "1h" | undefined =>
-  ttlSeconds !== undefined && ttlSeconds >= 3600 ? "1h" : undefined
@@ -3,15 +3,70 @@ import { ContentBlockID, FinishReason, ProtocolID, ProviderMetadata, ResponseID,
 import { ModelRef } from "./options"
 import { ToolResultValue } from "./messages"

+/**
+ * Token usage reported by an LLM provider.
+ *
+ * **Inclusive totals** (match AI SDK / OpenAI / LangChain convention — a
+ * reader from any of those ecosystems sees the number they expect):
+ *
+ * - `inputTokens` — total prompt tokens, *including* cached reads/writes.
+ * - `outputTokens` — total output tokens, *including* reasoning.
+ * - `totalTokens` — provider-supplied total, or `inputTokens + outputTokens`.
+ *
+ * **Non-overlapping breakdown** (every field is independently meaningful;
+ * consumers never have to subtract):
+ *
+ * - `nonCachedInputTokens` — the "fresh" portion of the prompt.
+ * - `cacheReadInputTokens` — input tokens served from cache.
+ * - `cacheWriteInputTokens` — input tokens written to cache.
+ * - `reasoningTokens` — subset of `outputTokens` spent on hidden reasoning.
+ *
+ * **Invariant**: `nonCachedInputTokens + cacheReadInputTokens +
+ * cacheWriteInputTokens = inputTokens`, and `reasoningTokens ≤ outputTokens`.
+ * Each protocol mapper computes whichever side it doesn't get natively,
+ * with `Math.max(0, …)` clamping for defense against provider bugs. Because
+ * every breakdown field is stored independently, downstream consumers can
+ * read whatever they need (cost-by-category, context-pressure, AI-SDK-style
+ * inclusive total) without ever subtracting — eliminating the underflow
+ * class of bug where a clamped difference would silently store the wrong
+ * value.
+ *
+ * **Semantics by provider**:
+ *
+ * - OpenAI Chat / Responses / Gemini / Bedrock: provider reports inclusive
+ *   `inputTokens` and an inclusive `outputTokens`; mapper subtracts to
+ *   derive the breakdown.
+ * - Anthropic: provider reports the breakdown natively (`input_tokens` is
+ *   non-cached only); mapper sums to derive the inclusive `inputTokens`.
+ *   Anthropic does *not* break extended-thinking out of `output_tokens`, so
+ *   `reasoningTokens` is `undefined` and `outputTokens` carries the
+ *   combined total — a documented limitation of the Anthropic API.
+ *
+ * `providerMetadata` always carries the provider's raw usage payload —
+ * keyed by provider name (`{ openai: ... }`, `{ anthropic: ... }`, etc.)
+ * — for fields we don't normalize and for billing-level audit trails.
+ * Matches the same escape-hatch field on `LLMEvent`.
+ */
 export class Usage extends Schema.Class<Usage>("LLM.Usage")({
  inputTokens: Schema.optional(Schema.Number),
  outputTokens: Schema.optional(Schema.Number),
-  reasoningTokens: Schema.optional(Schema.Number),
+  nonCachedInputTokens: Schema.optional(Schema.Number),
  cacheReadInputTokens: Schema.optional(Schema.Number),
  cacheWriteInputTokens: Schema.optional(Schema.Number),
+  reasoningTokens: Schema.optional(Schema.Number),
  totalTokens: Schema.optional(Schema.Number),
-  native: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
-}) {}
+  providerMetadata: Schema.optional(ProviderMetadata),
+}) {
+  /**
+   * Visible output tokens — `outputTokens` minus `reasoningTokens`, clamped
+   * to zero. The one place subtraction happens in this contract; the clamp
+   * means a provider reporting `reasoningTokens > outputTokens` produces a
+   * harmless zero rather than a negative that crashes downstream schemas.
+   */
+  get visibleOutputTokens() {
+    return Math.max(0, (this.outputTokens ?? 0) - (this.reasoningTokens ?? 0))
+  }
+}

 export const RequestStart = Schema.Struct({
  type: Schema.tag("request-start"),
@@ -79,7 +79,6 @@ export const ToolResultPart = Object.assign(
    name: Schema.String,
    result: ToolResultValue,
    providerExecuted: Schema.optional(Schema.Boolean),
-    cache: Schema.optional(CacheHint),
    metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
    providerMetadata: Schema.optional(ProviderMetadata),
  }).annotate({ identifier: "LLM.Content.ToolResult" }),
@@ -95,7 +94,6 @@ export const ToolResultPart = Object.assign(
      name: input.name,
      result: ToolResultValue.make(input.result, input.resultType),
      providerExecuted: input.providerExecuted,
-      cache: input.cache,
      metadata: input.metadata,
      providerMetadata: input.providerMetadata,
    }),
@@ -153,7 +151,6 @@ export class ToolDefinition extends Schema.Class<ToolDefinition>("LLM.ToolDefini
  name: Schema.String,
  description: Schema.String,
  inputSchema: JsonSchema,
-  cache: Schema.optional(CacheHint),
  metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
  native: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
 }) {}
@@ -0,0 +1,38 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "anthropic-messages/anthropic-haiku-4-5-text",
+    "recordedAt": "2026-05-11T02:02:03.804Z",
+    "provider": "anthropic",
+    "route": "anthropic-messages",
+    "transport": "http",
+    "model": "claude-haiku-4-5-20251001",
+    "tags": [
+      "prefix:anthropic-messages",
+      "provider:anthropic",
+      "text",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.anthropic.com/v1/messages",
+        "headers": {
+          "anthropic-version": "2023-06-01",
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"claude-haiku-4-5-20251001\",\"system\":[{\"type\":\"text\",\"text\":\"You are concise.\"}],\"messages\":[{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"Reply exactly with: Hello!\"}]}],\"stream\":true,\"max_tokens\":40,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "event: message_start\ndata: {\"type\":\"message_start\",\"message\":{\"model\":\"claude-haiku-4-5-20251001\",\"id\":\"msg_01SvRWwb75gDuhBpVMHjnFaf\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[],\"stop_reason\":null,\"stop_sequence\":null,\"stop_details\":null,\"usage\":{\"input_tokens\":18,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"output_tokens\":2,\"service_tier\":\"standard\",\"inference_geo\":\"not_available\"}}   }\n\nevent: content_block_start\ndata: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"text\",\"text\":\"\"}            }\n\nevent: ping\ndata: {\"type\": \"ping\"}\n\nevent: content_block_delta\ndata: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"text_delta\",\"text\":\"Hello!\"}  }\n\nevent: content_block_stop\ndata: {\"type\":\"content_block_stop\",\"index\":0          }\n\nevent: message_delta\ndata: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"end_turn\",\"stop_sequence\":null,\"stop_details\":null},\"usage\":{\"input_tokens\":18,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"output_tokens\":5}  }\n\nevent: message_stop\ndata: {\"type\":\"message_stop\"          }\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,39 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "anthropic-messages/anthropic-haiku-4-5-tool-call",
+    "recordedAt": "2026-05-11T02:02:04.363Z",
+    "provider": "anthropic",
+    "route": "anthropic-messages",
+    "transport": "http",
+    "model": "claude-haiku-4-5-20251001",
+    "tags": [
+      "prefix:anthropic-messages",
+      "provider:anthropic",
+      "tool",
+      "tool-call",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.anthropic.com/v1/messages",
+        "headers": {
+          "anthropic-version": "2023-06-01",
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"claude-haiku-4-5-20251001\",\"system\":[{\"type\":\"text\",\"text\":\"Call tools exactly as requested.\"}],\"messages\":[{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"Call get_weather with city exactly Paris.\"}]}],\"tools\":[{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"input_schema\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}],\"tool_choice\":{\"type\":\"tool\",\"name\":\"get_weather\"},\"stream\":true,\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "event: message_start\ndata: {\"type\":\"message_start\",\"message\":{\"model\":\"claude-haiku-4-5-20251001\",\"id\":\"msg_01Lu38yDM3WD8QBQTcg3dBaF\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[],\"stop_reason\":null,\"stop_sequence\":null,\"stop_details\":null,\"usage\":{\"input_tokens\":677,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"output_tokens\":16,\"service_tier\":\"standard\",\"inference_geo\":\"not_available\"}}}\n\nevent: content_block_start\ndata: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"tool_use\",\"id\":\"toolu_017Dqk9SAAsHHfiLKsUyitaQ\",\"name\":\"get_weather\",\"input\":{},\"caller\":{\"type\":\"direct\"}}   }\n\nevent: ping\ndata: {\"type\": \"ping\"}\n\nevent: content_block_delta\ndata: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"input_json_delta\",\"partial_json\":\"\"}    }\n\nevent: content_block_delta\ndata: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"input_json_delta\",\"partial_json\":\"{\\\"cit\"}          }\n\nevent: content_block_delta\ndata: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"input_json_delta\",\"partial_json\":\"y\\\": \\\"Paris\\\"}\"}           }\n\nevent: content_block_stop\ndata: {\"type\":\"content_block_stop\",\"index\":0}\n\nevent: message_delta\ndata: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"tool_use\",\"stop_sequence\":null,\"stop_details\":null},\"usage\":{\"input_tokens\":677,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"output_tokens\":33}      }\n\nevent: message_stop\ndata: {\"type\":\"message_stop\"             }\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,59 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "anthropic-messages/anthropic-opus-4-7-tool-loop",
+    "recordedAt": "2026-05-11T02:02:07.788Z",
+    "provider": "anthropic",
+    "route": "anthropic-messages",
+    "transport": "http",
+    "model": "claude-opus-4-7",
+    "tags": [
+      "prefix:anthropic-messages",
+      "provider:anthropic",
+      "flagship",
+      "tool",
+      "tool-loop",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.anthropic.com/v1/messages",
+        "headers": {
+          "anthropic-version": "2023-06-01",
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"claude-opus-4-7\",\"system\":[{\"type\":\"text\",\"text\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"}],\"messages\":[{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"What is the weather in Paris?\"}]}],\"tools\":[{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"input_schema\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}],\"stream\":true,\"max_tokens\":80}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "event: message_start\ndata: {\"type\":\"message_start\",\"message\":{\"model\":\"claude-opus-4-7\",\"id\":\"msg_01GK5kgi8AuVfRCnQFcXEfV8\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[],\"stop_reason\":null,\"stop_sequence\":null,\"stop_details\":null,\"usage\":{\"input_tokens\":812,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"output_tokens\":0,\"service_tier\":\"standard\",\"inference_geo\":\"global\"}} }\n\nevent: content_block_start\ndata: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"tool_use\",\"id\":\"toolu_01BnVDAp13NU8ZdJ9JeJ7byF\",\"name\":\"get_weather\",\"input\":{},\"caller\":{\"type\":\"direct\"}}}\n\nevent: ping\ndata: {\"type\": \"ping\"}\n\nevent: content_block_delta\ndata: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"input_json_delta\",\"partial_json\":\"\"}        }\n\nevent: content_block_delta\ndata: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"input_json_delta\",\"partial_json\":\"{\\\"c\"}      }\n\nevent: content_block_delta\ndata: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"input_json_delta\",\"partial_json\":\"ity\\\": \\\"\"}        }\n\nevent: content_block_delta\ndata: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"input_json_delta\",\"partial_json\":\"Paris\"} }\n\nevent: content_block_delta\ndata: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"input_json_delta\",\"partial_json\":\"\\\"}\"}         }\n\nevent: content_block_stop\ndata: {\"type\":\"content_block_stop\",\"index\":0       }\n\nevent: message_delta\ndata: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"tool_use\",\"stop_sequence\":null,\"stop_details\":null},\"usage\":{\"input_tokens\":812,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"output_tokens\":66}   }\n\nevent: message_stop\ndata: {\"type\":\"message_stop\"          }\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.anthropic.com/v1/messages",
+        "headers": {
+          "anthropic-version": "2023-06-01",
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"claude-opus-4-7\",\"system\":[{\"type\":\"text\",\"text\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"}],\"messages\":[{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"What is the weather in Paris?\"}]},{\"role\":\"assistant\",\"content\":[{\"type\":\"tool_use\",\"id\":\"toolu_01BnVDAp13NU8ZdJ9JeJ7byF\",\"name\":\"get_weather\",\"input\":{\"city\":\"Paris\"}}]},{\"role\":\"user\",\"content\":[{\"type\":\"tool_result\",\"tool_use_id\":\"toolu_01BnVDAp13NU8ZdJ9JeJ7byF\",\"content\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}]}],\"tools\":[{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"input_schema\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}],\"stream\":true,\"max_tokens\":80}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "event: message_start\ndata: {\"type\":\"message_start\",\"message\":{\"model\":\"claude-opus-4-7\",\"id\":\"msg_01237VTnjPeYSRh31UjXWaEa\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[],\"stop_reason\":null,\"stop_sequence\":null,\"stop_details\":null,\"usage\":{\"input_tokens\":909,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"output_tokens\":8,\"service_tier\":\"standard\",\"inference_geo\":\"global\"}}       }\n\nevent: content_block_start\ndata: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"text\",\"text\":\"\"}               }\n\nevent: ping\ndata: {\"type\": \"ping\"}\n\nevent: content_block_delta\ndata: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"text_delta\",\"text\":\"Paris is sunny.\"}     }\n\nevent: content_block_stop\ndata: {\"type\":\"content_block_stop\",\"index\":0      }\n\nevent: message_delta\ndata: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"end_turn\",\"stop_sequence\":null,\"stop_details\":null},\"usage\":{\"input_tokens\":909,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"output_tokens\":12}      }\n\nevent: message_stop\ndata: {\"type\":\"message_stop\"       }\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,35 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "anthropic-messages/rejects-malformed-assistant-tool-order",
+    "recordedAt": "2026-05-11T02:01:44.544Z",
+    "tags": [
+      "prefix:anthropic-messages",
+      "provider:anthropic",
+      "protocol:anthropic-messages",
+      "tool",
+      "sad-path"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.anthropic.com/v1/messages",
+        "headers": {
+          "anthropic-version": "2023-06-01",
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"claude-haiku-4-5-20251001\",\"messages\":[{\"role\":\"assistant\",\"content\":[{\"type\":\"tool_use\",\"id\":\"call_1\",\"name\":\"get_weather\",\"input\":{\"city\":\"Paris\"}},{\"type\":\"text\",\"text\":\"I will check the weather.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"tool_result\",\"tool_use_id\":\"call_1\",\"content\":\"{\\\"temperature\\\":\\\"72F\\\"}\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"Use that result to answer briefly.\"}]}],\"tools\":[{\"name\":\"get_weather\",\"description\":\"Get weather\",\"input_schema\":{\"type\":\"object\",\"properties\":{}}}],\"stream\":true,\"max_tokens\":4096}"
+      },
+      "response": {
+        "status": 400,
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"messages.1: `tool_use` ids were found without `tool_result` blocks immediately after: call_1. Each `tool_use` block must have a corresponding `tool_result` block in the next message.\"},\"request_id\":\"req_011CauxVdQf3N2PPFJ5aH8Bh\"}"
+      }
+    }
+  ]
+}
@@ -0,0 +1,37 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "gemini/gemini-2-5-flash-text",
+    "recordedAt": "2026-05-11T02:02:08.410Z",
+    "provider": "google",
+    "route": "gemini",
+    "transport": "http",
+    "model": "gemini-2.5-flash",
+    "tags": [
+      "prefix:gemini",
+      "provider:google",
+      "text",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:streamGenerateContent?alt=sse",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"contents\":[{\"role\":\"user\",\"parts\":[{\"text\":\"Reply exactly with: Hello!\"}]}],\"systemInstruction\":{\"parts\":[{\"text\":\"You are concise.\"}]},\"generationConfig\":{\"maxOutputTokens\":80,\"temperature\":0}}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"candidates\": [{\"content\": {\"parts\": [{\"text\": \"Hello!\"}],\"role\": \"model\"},\"finishReason\": \"STOP\",\"index\": 0}],\"usageMetadata\": {\"promptTokenCount\": 11,\"candidatesTokenCount\": 2,\"totalTokenCount\": 13,\"promptTokensDetails\": [{\"modality\": \"TEXT\",\"tokenCount\": 11}],\"serviceTier\": \"standard\"},\"modelVersion\": \"gemini-2.5-flash\",\"responseId\": \"nzgBatP3OZmW-8YP567bqQs\"}\r\n\r\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,38 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "gemini/gemini-2-5-flash-tool-call",
+    "recordedAt": "2026-05-11T02:02:09.308Z",
+    "provider": "google",
+    "route": "gemini",
+    "transport": "http",
+    "model": "gemini-2.5-flash",
+    "tags": [
+      "prefix:gemini",
+      "provider:google",
+      "tool",
+      "tool-call",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:streamGenerateContent?alt=sse",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"contents\":[{\"role\":\"user\",\"parts\":[{\"text\":\"Call get_weather with city exactly Paris.\"}]}],\"systemInstruction\":{\"parts\":[{\"text\":\"Call tools exactly as requested.\"}]},\"tools\":[{\"functionDeclarations\":[{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"required\":[\"city\"],\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}}}}]}],\"toolConfig\":{\"functionCallingConfig\":{\"mode\":\"ANY\",\"allowedFunctionNames\":[\"get_weather\"]}},\"generationConfig\":{\"maxOutputTokens\":80,\"temperature\":0}}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"candidates\": [{\"content\": {\"parts\": [{\"functionCall\": {\"name\": \"get_weather\",\"args\": {\"city\": \"Paris\"}},\"thoughtSignature\": \"CiQBDDnWx/X6sWeX2joSugyWO3L/lt0AgIPCvhpqf3845fj+H70KXwEMOdbHB/cnaqYCro0pU+yLWoA55jhuwoLmTcnYm4Qzcm5DuW/v0NUyz8RDx6DFh61juENveUztly6yc6/XiWJHtsgncd9YgcZhuQKqtp5KZTkGYpT3g6v3yP9GK4AoCoUBAQw51sePh3WuWovHnwIotKLVZiU9pwh34k4FY7ugPOxyDAG9j69cy7BYYzSchI10LEjLoLlCMZuNIPootBgI02QWY/4h2PIv33BAADrFPM2T3aE4cAuMoa3GCu2nztJ/95junDhIhuXZSQ/Mh9EVxpx7ml99Z7Hxb7OtDsSZZLeCBuGmSw==\"}],\"role\": \"model\"},\"finishReason\": \"STOP\",\"index\": 0,\"finishMessage\": \"Model generated function call(s).\"}],\"usageMetadata\": {\"promptTokenCount\": 55,\"candidatesTokenCount\": 15,\"totalTokenCount\": 115,\"promptTokensDetails\": [{\"modality\": \"TEXT\",\"tokenCount\": 55}],\"thoughtsTokenCount\": 45,\"serviceTier\": \"standard\"},\"modelVersion\": \"gemini-2.5-flash\",\"responseId\": \"oDgBauj6HZ3B-8YPpaOc6QU\"}\r\n\r\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,37 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "openai-chat/openai-chat-gpt-4o-mini-text",
+    "recordedAt": "2026-05-11T02:01:46.536Z",
+    "provider": "openai",
+    "route": "openai-chat",
+    "transport": "http",
+    "model": "gpt-4o-mini",
+    "tags": [
+      "prefix:openai-chat",
+      "provider:openai",
+      "text",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.openai.com/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gpt-4o-mini\",\"messages\":[{\"role\":\"system\",\"content\":\"You are concise.\"},{\"role\":\"user\",\"content\":\"Reply exactly with: Hello!\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":40,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"chatcmpl-DeAFmC9UiBnVu6dAjBvWVBVypm2gc\",\"object\":\"chat.completion.chunk\",\"created\":1778464906,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_3ef558a83f\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"refusal\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"wHajUz1js\"}\n\ndata: {\"id\":\"chatcmpl-DeAFmC9UiBnVu6dAjBvWVBVypm2gc\",\"object\":\"chat.completion.chunk\",\"created\":1778464906,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_3ef558a83f\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"Hello\"},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"lVHzGq\"}\n\ndata: {\"id\":\"chatcmpl-DeAFmC9UiBnVu6dAjBvWVBVypm2gc\",\"object\":\"chat.completion.chunk\",\"created\":1778464906,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_3ef558a83f\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"!\"},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"3gaQRWEjE7\"}\n\ndata: {\"id\":\"chatcmpl-DeAFmC9UiBnVu6dAjBvWVBVypm2gc\",\"object\":\"chat.completion.chunk\",\"created\":1778464906,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_3ef558a83f\",\"choices\":[{\"index\":0,\"delta\":{},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":null,\"obfuscation\":\"vuKPZ\"}\n\ndata: {\"id\":\"chatcmpl-DeAFmC9UiBnVu6dAjBvWVBVypm2gc\",\"object\":\"chat.completion.chunk\",\"created\":1778464906,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_3ef558a83f\",\"choices\":[],\"usage\":{\"prompt_tokens\":21,\"completion_tokens\":2,\"total_tokens\":23,\"prompt_tokens_details\":{\"cached_tokens\":0,\"audio_tokens\":0},\"completion_tokens_details\":{\"reasoning_tokens\":0,\"audio_tokens\":0,\"accepted_prediction_tokens\":0,\"rejected_prediction_tokens\":0}},\"obfuscation\":\"kfFsssdujmM\"}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,38 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "openai-chat/openai-chat-gpt-4o-mini-tool-call",
+    "recordedAt": "2026-05-11T02:01:47.484Z",
+    "provider": "openai",
+    "route": "openai-chat",
+    "transport": "http",
+    "model": "gpt-4o-mini",
+    "tags": [
+      "prefix:openai-chat",
+      "provider:openai",
+      "tool",
+      "tool-call",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.openai.com/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gpt-4o-mini\",\"messages\":[{\"role\":\"system\",\"content\":\"Call tools exactly as requested.\"},{\"role\":\"user\",\"content\":\"Call get_weather with city exactly Paris.\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}}],\"tool_choice\":{\"type\":\"function\",\"function\":{\"name\":\"get_weather\"}},\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"chatcmpl-DeAFnVMdELyuymMRlTqN5dBdM7eGd\",\"object\":\"chat.completion.chunk\",\"created\":1778464907,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_d88f6e55bb\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":null,\"tool_calls\":[{\"index\":0,\"id\":\"call_sH8T7MPdJXS5KJginKrzexL5\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"\"}}],\"refusal\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"d0rJ\"}\n\ndata: {\"id\":\"chatcmpl-DeAFnVMdELyuymMRlTqN5dBdM7eGd\",\"object\":\"chat.completion.chunk\",\"created\":1778464907,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_d88f6e55bb\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"{\\\"\"}}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"hL82erd6bBoy91\"}\n\ndata: {\"id\":\"chatcmpl-DeAFnVMdELyuymMRlTqN5dBdM7eGd\",\"object\":\"chat.completion.chunk\",\"created\":1778464907,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_d88f6e55bb\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"city\"}}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"e3aoKH4tvJlXW\"}\n\ndata: {\"id\":\"chatcmpl-DeAFnVMdELyuymMRlTqN5dBdM7eGd\",\"object\":\"chat.completion.chunk\",\"created\":1778464907,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_d88f6e55bb\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"\\\":\\\"\"}}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"T8EzFVNaUCLE\"}\n\ndata: {\"id\":\"chatcmpl-DeAFnVMdELyuymMRlTqN5dBdM7eGd\",\"object\":\"chat.completion.chunk\",\"created\":1778464907,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_d88f6e55bb\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"Paris\"}}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"lelRBxF08Zes\"}\n\ndata: {\"id\":\"chatcmpl-DeAFnVMdELyuymMRlTqN5dBdM7eGd\",\"object\":\"chat.completion.chunk\",\"created\":1778464907,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_d88f6e55bb\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"\\\"}\"}}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"F1xO8sVMyZt4BU\"}\n\ndata: {\"id\":\"chatcmpl-DeAFnVMdELyuymMRlTqN5dBdM7eGd\",\"object\":\"chat.completion.chunk\",\"created\":1778464907,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_d88f6e55bb\",\"choices\":[{\"index\":0,\"delta\":{},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":null,\"obfuscation\":\"YL3pl\"}\n\ndata: {\"id\":\"chatcmpl-DeAFnVMdELyuymMRlTqN5dBdM7eGd\",\"object\":\"chat.completion.chunk\",\"created\":1778464907,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_d88f6e55bb\",\"choices\":[],\"usage\":{\"prompt_tokens\":67,\"completion_tokens\":5,\"total_tokens\":72,\"prompt_tokens_details\":{\"cached_tokens\":0,\"audio_tokens\":0},\"completion_tokens_details\":{\"reasoning_tokens\":0,\"audio_tokens\":0,\"accepted_prediction_tokens\":0,\"rejected_prediction_tokens\":0}},\"obfuscation\":\"7cov9qkofwo\"}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,56 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "openai-chat/openai-chat-gpt-4o-mini-tool-loop",
+    "recordedAt": "2026-05-11T02:01:50.433Z",
+    "provider": "openai",
+    "route": "openai-chat",
+    "transport": "http",
+    "model": "gpt-4o-mini",
+    "tags": [
+      "prefix:openai-chat",
+      "provider:openai",
+      "tool",
+      "tool-loop",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.openai.com/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gpt-4o-mini\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"chatcmpl-DeAFo2rbsuuMn0ftzxwfDpqnYANrT\",\"object\":\"chat.completion.chunk\",\"created\":1778464908,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":null,\"tool_calls\":[{\"index\":0,\"id\":\"call_HrDZhrMUauvVddKWzVQFJ69Q\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"\"}}],\"refusal\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"aAqC\"}\n\ndata: {\"id\":\"chatcmpl-DeAFo2rbsuuMn0ftzxwfDpqnYANrT\",\"object\":\"chat.completion.chunk\",\"created\":1778464908,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"{\\\"\"}}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"haZ3pKk14oDay0\"}\n\ndata: {\"id\":\"chatcmpl-DeAFo2rbsuuMn0ftzxwfDpqnYANrT\",\"object\":\"chat.completion.chunk\",\"created\":1778464908,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"city\"}}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"ZUJHQytFyDezp\"}\n\ndata: {\"id\":\"chatcmpl-DeAFo2rbsuuMn0ftzxwfDpqnYANrT\",\"object\":\"chat.completion.chunk\",\"created\":1778464908,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"\\\":\\\"\"}}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"3buqnStPteyG\"}\n\ndata: {\"id\":\"chatcmpl-DeAFo2rbsuuMn0ftzxwfDpqnYANrT\",\"object\":\"chat.completion.chunk\",\"created\":1778464908,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"Paris\"}}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"7epPqHEU3OeA\"}\n\ndata: {\"id\":\"chatcmpl-DeAFo2rbsuuMn0ftzxwfDpqnYANrT\",\"object\":\"chat.completion.chunk\",\"created\":1778464908,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"\\\"}\"}}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"gSU5gyl9K7sUCv\"}\n\ndata: {\"id\":\"chatcmpl-DeAFo2rbsuuMn0ftzxwfDpqnYANrT\",\"object\":\"chat.completion.chunk\",\"created\":1778464908,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{},\"logprobs\":null,\"finish_reason\":\"tool_calls\"}],\"usage\":null,\"obfuscation\":\"P3oSvByfy60JCsz\"}\n\ndata: {\"id\":\"chatcmpl-DeAFo2rbsuuMn0ftzxwfDpqnYANrT\",\"object\":\"chat.completion.chunk\",\"created\":1778464908,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[],\"usage\":{\"prompt_tokens\":71,\"completion_tokens\":14,\"total_tokens\":85,\"prompt_tokens_details\":{\"cached_tokens\":0,\"audio_tokens\":0},\"completion_tokens_details\":{\"reasoning_tokens\":0,\"audio_tokens\":0,\"accepted_prediction_tokens\":0,\"rejected_prediction_tokens\":0}},\"obfuscation\":\"5AhQ5ToYra\"}\n\ndata: [DONE]\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.openai.com/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gpt-4o-mini\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"},{\"role\":\"assistant\",\"content\":null,\"tool_calls\":[{\"id\":\"call_HrDZhrMUauvVddKWzVQFJ69Q\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}]},{\"role\":\"tool\",\"tool_call_id\":\"call_HrDZhrMUauvVddKWzVQFJ69Q\",\"content\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"chatcmpl-DeAFpaPsUwxt5hjwGgC4BBmlALxwR\",\"object\":\"chat.completion.chunk\",\"created\":1778464909,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"refusal\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"2QWjdWPSe\"}\n\ndata: {\"id\":\"chatcmpl-DeAFpaPsUwxt5hjwGgC4BBmlALxwR\",\"object\":\"chat.completion.chunk\",\"created\":1778464909,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"Paris\"},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"sm5IpQ\"}\n\ndata: {\"id\":\"chatcmpl-DeAFpaPsUwxt5hjwGgC4BBmlALxwR\",\"object\":\"chat.completion.chunk\",\"created\":1778464909,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\" is\"},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"cWsGxqHw\"}\n\ndata: {\"id\":\"chatcmpl-DeAFpaPsUwxt5hjwGgC4BBmlALxwR\",\"object\":\"chat.completion.chunk\",\"created\":1778464909,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\" sunny\"},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"8hNCi\"}\n\ndata: {\"id\":\"chatcmpl-DeAFpaPsUwxt5hjwGgC4BBmlALxwR\",\"object\":\"chat.completion.chunk\",\"created\":1778464909,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\".\"},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null,\"obfuscation\":\"4yKCNDRcwq\"}\n\ndata: {\"id\":\"chatcmpl-DeAFpaPsUwxt5hjwGgC4BBmlALxwR\",\"object\":\"chat.completion.chunk\",\"created\":1778464909,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[{\"index\":0,\"delta\":{},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":null,\"obfuscation\":\"NvRgg\"}\n\ndata: {\"id\":\"chatcmpl-DeAFpaPsUwxt5hjwGgC4BBmlALxwR\",\"object\":\"chat.completion.chunk\",\"created\":1778464909,\"model\":\"gpt-4o-mini-2024-07-18\",\"service_tier\":\"default\",\"system_fingerprint\":\"fp_c83829e916\",\"choices\":[],\"usage\":{\"prompt_tokens\":103,\"completion_tokens\":5,\"total_tokens\":108,\"prompt_tokens_details\":{\"cached_tokens\":0,\"audio_tokens\":0},\"completion_tokens_details\":{\"reasoning_tokens\":0,\"audio_tokens\":0,\"accepted_prediction_tokens\":0,\"rejected_prediction_tokens\":0}},\"obfuscation\":\"4s7mLtNpZ\"}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,37 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "openai-compatible-chat/deepseek-chat-text",
+    "recordedAt": "2026-05-11T02:02:10.220Z",
+    "provider": "deepseek",
+    "route": "openai-compatible-chat",
+    "transport": "http",
+    "model": "deepseek-chat",
+    "tags": [
+      "prefix:openai-compatible-chat",
+      "provider:deepseek",
+      "text",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepseek.com/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"deepseek-chat\",\"messages\":[{\"role\":\"system\",\"content\":\"You are concise.\"},{\"role\":\"user\",\"content\":\"Reply exactly with: Hello!\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":40,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"159019a5-f981-4103-8b3c-2be37cefc505\",\"object\":\"chat.completion.chunk\",\"created\":1778464929,\"model\":\"deepseek-v4-flash\",\"system_fingerprint\":\"fp_058df29938_prod0820_fp8_kvcache_20260402\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\"},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"id\":\"159019a5-f981-4103-8b3c-2be37cefc505\",\"object\":\"chat.completion.chunk\",\"created\":1778464929,\"model\":\"deepseek-v4-flash\",\"system_fingerprint\":\"fp_058df29938_prod0820_fp8_kvcache_20260402\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"Hello\"},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"id\":\"159019a5-f981-4103-8b3c-2be37cefc505\",\"object\":\"chat.completion.chunk\",\"created\":1778464929,\"model\":\"deepseek-v4-flash\",\"system_fingerprint\":\"fp_058df29938_prod0820_fp8_kvcache_20260402\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"!\"},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"id\":\"159019a5-f981-4103-8b3c-2be37cefc505\",\"object\":\"chat.completion.chunk\",\"created\":1778464929,\"model\":\"deepseek-v4-flash\",\"system_fingerprint\":\"fp_058df29938_prod0820_fp8_kvcache_20260402\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"\"},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":14,\"completion_tokens\":2,\"total_tokens\":16,\"prompt_tokens_details\":{\"cached_tokens\":0},\"prompt_cache_hit_tokens\":0,\"prompt_cache_miss_tokens\":14}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,37 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "openai-compatible-chat/togetherai-llama-3-3-70b-text",
+    "recordedAt": "2026-05-11T02:02:13.341Z",
+    "provider": "togetherai",
+    "route": "openai-compatible-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": [
+      "prefix:openai-compatible-chat",
+      "provider:togetherai",
+      "text",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.together.xyz/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"You are concise.\"},{\"role\":\"user\",\"content\":\"Reply exactly with: Hello!\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":40,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream;charset=utf-8"
+        },
+        "body": "data: {\"id\":\"oibreET-3pDw3Z-9f9d9996ce37066b\",\"object\":\"chat.completion.chunk\",\"created\":1778464931,\"choices\":[{\"index\":0,\"text\":\"Hello\",\"logprobs\":null,\"finish_reason\":null,\"seed\":null,\"delta\":{\"token_id\":9906,\"role\":\"assistant\",\"content\":\"Hello\"}}],\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"usage\":null}\n\ndata: {\"id\":\"oibreET-3pDw3Z-9f9d9996ce37066b\",\"object\":\"chat.completion.chunk\",\"created\":1778464931,\"choices\":[{\"index\":0,\"text\":\"!\",\"logprobs\":null,\"finish_reason\":null,\"seed\":null,\"delta\":{\"token_id\":null,\"role\":\"assistant\",\"content\":\"!\"}}],\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"usage\":null}\n\ndata: {\"id\":\"oibreET-3pDw3Z-9f9d9996ce37066b\",\"object\":\"chat.completion.chunk\",\"created\":1778464931,\"choices\":[{\"index\":0,\"text\":\"\",\"logprobs\":null,\"finish_reason\":\"stop\",\"seed\":12144769634208630000,\"delta\":{\"token_id\":128009,\"role\":\"assistant\",\"content\":\"\"}}],\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"usage\":{\"prompt_tokens\":45,\"completion_tokens\":3,\"total_tokens\":48,\"cached_tokens\":0}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,38 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "openai-compatible-chat/togetherai-llama-3-3-70b-tool-call",
+    "recordedAt": "2026-05-11T02:02:14.453Z",
+    "provider": "togetherai",
+    "route": "openai-compatible-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": [
+      "prefix:openai-compatible-chat",
+      "provider:togetherai",
+      "tool",
+      "tool-call",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.together.xyz/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"messages\":[{\"role\":\"system\",\"content\":\"Call tools exactly as requested.\"},{\"role\":\"user\",\"content\":\"Call get_weather with city exactly Paris.\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}}],\"tool_choice\":{\"type\":\"function\",\"function\":{\"name\":\"get_weather\"}},\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":80,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream;charset=utf-8"
+        },
+        "body": "data: {\"id\":\"oibreu3-6Ng1vN-9f9d99a9ba73066b\",\"object\":\"chat.completion.chunk\",\"created\":1778464933,\"choices\":[{\"index\":0,\"role\":\"assistant\",\"text\":\"\",\"logprobs\":null,\"finish_reason\":null,\"delta\":{\"token_id\":null,\"role\":\"assistant\",\"content\":\"\"}}],\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\"}\n\ndata: {\"id\":\"oibreu3-6Ng1vN-9f9d99a9ba73066b\",\"object\":\"chat.completion.chunk\",\"created\":1778464933,\"choices\":[{\"index\":0,\"text\":\"\",\"logprobs\":null,\"finish_reason\":null,\"delta\":{\"token_id\":null,\"role\":\"assistant\",\"content\":\"\",\"tool_calls\":[{\"index\":0,\"id\":\"call_jue52dtu6iozr0ny9rq357u5\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"\"}}]}}],\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\"}\n\ndata: {\"id\":\"oibreu3-6Ng1vN-9f9d99a9ba73066b\",\"object\":\"chat.completion.chunk\",\"created\":1778464933,\"choices\":[{\"index\":0,\"text\":\"\",\"logprobs\":null,\"finish_reason\":\"tool_calls\",\"delta\":{\"token_id\":null,\"role\":\"assistant\",\"content\":\"\",\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}]}}],\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\"}\n\ndata: {\"id\":\"oibreu3-6Ng1vN-9f9d99a9ba73066b\",\"object\":\"chat.completion.chunk\",\"created\":1778464933,\"choices\":[{\"index\":0,\"text\":\"\",\"logprobs\":null,\"finish_reason\":\"tool_calls\",\"seed\":17440360718047570000,\"delta\":{\"token_id\":128009,\"role\":\"assistant\",\"content\":\"\"}}],\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"usage\":{\"prompt_tokens\":194,\"completion_tokens\":19,\"total_tokens\":213,\"cached_tokens\":0}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,143 @@
+{
+  "version": 1,
+  "metadata": {
+    "name": "openai-responses-websocket/openai-responses-websocket-gpt-4-1-mini-tool-loop",
+    "recordedAt": "2026-05-11T02:02:03.284Z",
+    "provider": "openai",
+    "route": "openai-responses-websocket",
+    "transport": "websocket",
+    "model": "gpt-4.1-mini",
+    "tags": [
+      "prefix:openai-responses-websocket",
+      "provider:openai",
+      "transport:websocket",
+      "tool",
+      "tool-loop",
+      "golden"
+    ]
+  },
+  "interactions": [
+    {
+      "transport": "websocket",
+      "open": {
+        "url": "wss://api.openai.com/v1/responses",
+        "headers": {}
+      },
+      "client": [
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-4.1-mini\",\"input\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}],\"store\":false,\"max_output_tokens\":80,\"temperature\":0}"
+        }
+      ],
+      "server": [
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_0f8d81b5d3287513016a013897415481a28585571d6710366f\",\"object\":\"response\",\"created_at\":1778464919,\"status\":\"in_progress\",\"background\":false,\"completed_at\":null,\"error\":null,\"frequency_penalty\":0.0,\"incomplete_details\":null,\"instructions\":null,\"max_output_tokens\":80,\"max_tool_calls\":null,\"model\":\"gpt-4.1-mini-2025-04-14\",\"moderation\":null,\"output\":[],\"parallel_tool_calls\":true,\"presence_penalty\":0.0,\"previous_response_id\":null,\"prompt_cache_key\":null,\"prompt_cache_retention\":\"in_memory\",\"reasoning\":{\"effort\":null,\"summary\":null},\"safety_identifier\":null,\"service_tier\":\"auto\",\"store\":false,\"temperature\":0.0,\"text\":{\"format\":{\"type\":\"text\"},\"verbosity\":\"medium\"},\"tool_choice\":\"auto\",\"tools\":[{\"type\":\"function\",\"description\":\"Get current weather for a city.\",\"name\":\"get_weather\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":true}],\"top_logprobs\":0,\"top_p\":1.0,\"truncation\":\"disabled\",\"usage\":null,\"user\":null,\"metadata\":{}},\"sequence_number\":0}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.in_progress\",\"response\":{\"id\":\"resp_0f8d81b5d3287513016a013897415481a28585571d6710366f\",\"object\":\"response\",\"created_at\":1778464919,\"status\":\"in_progress\",\"background\":false,\"completed_at\":null,\"error\":null,\"frequency_penalty\":0.0,\"incomplete_details\":null,\"instructions\":null,\"max_output_tokens\":80,\"max_tool_calls\":null,\"model\":\"gpt-4.1-mini-2025-04-14\",\"moderation\":null,\"output\":[],\"parallel_tool_calls\":true,\"presence_penalty\":0.0,\"previous_response_id\":null,\"prompt_cache_key\":null,\"prompt_cache_retention\":\"in_memory\",\"reasoning\":{\"effort\":null,\"summary\":null},\"safety_identifier\":null,\"service_tier\":\"auto\",\"store\":false,\"temperature\":0.0,\"text\":{\"format\":{\"type\":\"text\"},\"verbosity\":\"medium\"},\"tool_choice\":\"auto\",\"tools\":[{\"type\":\"function\",\"description\":\"Get current weather for a city.\",\"name\":\"get_weather\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":true}],\"top_logprobs\":0,\"top_p\":1.0,\"truncation\":\"disabled\",\"usage\":null,\"user\":null,\"metadata\":{}},\"sequence_number\":1}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"id\":\"fc_0f8d81b5d3287513016a01389815f081a2a68583ed9b19c405\",\"type\":\"function_call\",\"status\":\"in_progress\",\"arguments\":\"\",\"call_id\":\"call_wRJoUdcb6w5SIpieBFOKxF6z\",\"name\":\"get_weather\"},\"output_index\":0,\"sequence_number\":2}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.function_call_arguments.delta\",\"delta\":\"{\\\"\",\"item_id\":\"fc_0f8d81b5d3287513016a01389815f081a2a68583ed9b19c405\",\"obfuscation\":\"EAdvUfYaysnNd3\",\"output_index\":0,\"sequence_number\":3}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.function_call_arguments.delta\",\"delta\":\"city\",\"item_id\":\"fc_0f8d81b5d3287513016a01389815f081a2a68583ed9b19c405\",\"obfuscation\":\"b0QWJJlscLyl\",\"output_index\":0,\"sequence_number\":4}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.function_call_arguments.delta\",\"delta\":\"\\\":\\\"\",\"item_id\":\"fc_0f8d81b5d3287513016a01389815f081a2a68583ed9b19c405\",\"obfuscation\":\"m5EV1wzceCkjb\",\"output_index\":0,\"sequence_number\":5}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.function_call_arguments.delta\",\"delta\":\"Paris\",\"item_id\":\"fc_0f8d81b5d3287513016a01389815f081a2a68583ed9b19c405\",\"obfuscation\":\"QM3FjPoN9Gi\",\"output_index\":0,\"sequence_number\":6}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.function_call_arguments.delta\",\"delta\":\"\\\"}\",\"item_id\":\"fc_0f8d81b5d3287513016a01389815f081a2a68583ed9b19c405\",\"obfuscation\":\"lug74orAYXtg0e\",\"output_index\":0,\"sequence_number\":7}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.function_call_arguments.done\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\",\"item_id\":\"fc_0f8d81b5d3287513016a01389815f081a2a68583ed9b19c405\",\"output_index\":0,\"sequence_number\":8}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"id\":\"fc_0f8d81b5d3287513016a01389815f081a2a68583ed9b19c405\",\"type\":\"function_call\",\"status\":\"completed\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\",\"call_id\":\"call_wRJoUdcb6w5SIpieBFOKxF6z\",\"name\":\"get_weather\"},\"output_index\":0,\"sequence_number\":9}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_0f8d81b5d3287513016a013897415481a28585571d6710366f\",\"object\":\"response\",\"created_at\":1778464919,\"status\":\"completed\",\"background\":false,\"completed_at\":1778464920,\"error\":null,\"frequency_penalty\":0.0,\"incomplete_details\":null,\"instructions\":null,\"max_output_tokens\":80,\"max_tool_calls\":null,\"model\":\"gpt-4.1-mini-2025-04-14\",\"moderation\":null,\"output\":[{\"id\":\"fc_0f8d81b5d3287513016a01389815f081a2a68583ed9b19c405\",\"type\":\"function_call\",\"status\":\"completed\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\",\"call_id\":\"call_wRJoUdcb6w5SIpieBFOKxF6z\",\"name\":\"get_weather\"}],\"parallel_tool_calls\":true,\"presence_penalty\":0.0,\"previous_response_id\":null,\"prompt_cache_key\":null,\"prompt_cache_retention\":\"in_memory\",\"reasoning\":{\"effort\":null,\"summary\":null},\"safety_identifier\":null,\"service_tier\":\"default\",\"store\":false,\"temperature\":0.0,\"text\":{\"format\":{\"type\":\"text\"},\"verbosity\":\"medium\"},\"tool_choice\":\"auto\",\"tools\":[{\"type\":\"function\",\"description\":\"Get current weather for a city.\",\"name\":\"get_weather\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":true}],\"top_logprobs\":0,\"top_p\":1.0,\"truncation\":\"disabled\",\"usage\":{\"input_tokens\":69,\"input_tokens_details\":{\"cached_tokens\":0},\"output_tokens\":15,\"output_tokens_details\":{\"reasoning_tokens\":0},\"total_tokens\":84},\"user\":null,\"metadata\":{}},\"sequence_number\":10}"
+        }
+      ]
+    },
+    {
+      "transport": "websocket",
+      "open": {
+        "url": "wss://api.openai.com/v1/responses",
+        "headers": {}
+      },
+      "client": [
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-4.1-mini\",\"input\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]},{\"type\":\"function_call\",\"call_id\":\"call_wRJoUdcb6w5SIpieBFOKxF6z\",\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_wRJoUdcb6w5SIpieBFOKxF6z\",\"output\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}],\"store\":false,\"max_output_tokens\":80,\"temperature\":0}"
+        }
+      ],
+      "server": [
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_0767cfe3f5d98b2a016a0138994258819485d082e5c78849a4\",\"object\":\"response\",\"created_at\":1778464921,\"status\":\"in_progress\",\"background\":false,\"completed_at\":null,\"error\":null,\"frequency_penalty\":0.0,\"incomplete_details\":null,\"instructions\":null,\"max_output_tokens\":80,\"max_tool_calls\":null,\"model\":\"gpt-4.1-mini-2025-04-14\",\"moderation\":null,\"output\":[],\"parallel_tool_calls\":true,\"presence_penalty\":0.0,\"previous_response_id\":null,\"prompt_cache_key\":null,\"prompt_cache_retention\":\"in_memory\",\"reasoning\":{\"effort\":null,\"summary\":null},\"safety_identifier\":null,\"service_tier\":\"auto\",\"store\":false,\"temperature\":0.0,\"text\":{\"format\":{\"type\":\"text\"},\"verbosity\":\"medium\"},\"tool_choice\":\"auto\",\"tools\":[{\"type\":\"function\",\"description\":\"Get current weather for a city.\",\"name\":\"get_weather\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":true}],\"top_logprobs\":0,\"top_p\":1.0,\"truncation\":\"disabled\",\"usage\":null,\"user\":null,\"metadata\":{}},\"sequence_number\":0}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.in_progress\",\"response\":{\"id\":\"resp_0767cfe3f5d98b2a016a0138994258819485d082e5c78849a4\",\"object\":\"response\",\"created_at\":1778464921,\"status\":\"in_progress\",\"background\":false,\"completed_at\":null,\"error\":null,\"frequency_penalty\":0.0,\"incomplete_details\":null,\"instructions\":null,\"max_output_tokens\":80,\"max_tool_calls\":null,\"model\":\"gpt-4.1-mini-2025-04-14\",\"moderation\":null,\"output\":[],\"parallel_tool_calls\":true,\"presence_penalty\":0.0,\"previous_response_id\":null,\"prompt_cache_key\":null,\"prompt_cache_retention\":\"in_memory\",\"reasoning\":{\"effort\":null,\"summary\":null},\"safety_identifier\":null,\"service_tier\":\"auto\",\"store\":false,\"temperature\":0.0,\"text\":{\"format\":{\"type\":\"text\"},\"verbosity\":\"medium\"},\"tool_choice\":\"auto\",\"tools\":[{\"type\":\"function\",\"description\":\"Get current weather for a city.\",\"name\":\"get_weather\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":true}],\"top_logprobs\":0,\"top_p\":1.0,\"truncation\":\"disabled\",\"usage\":null,\"user\":null,\"metadata\":{}},\"sequence_number\":1}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"id\":\"msg_0767cfe3f5d98b2a016a01389b1434819493f3d33e0ec6973c\",\"type\":\"message\",\"status\":\"in_progress\",\"content\":[],\"role\":\"assistant\"},\"output_index\":0,\"sequence_number\":2}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.content_part.added\",\"content_index\":0,\"item_id\":\"msg_0767cfe3f5d98b2a016a01389b1434819493f3d33e0ec6973c\",\"output_index\":0,\"part\":{\"type\":\"output_text\",\"annotations\":[],\"logprobs\":[],\"text\":\"\"},\"sequence_number\":3}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"content_index\":0,\"delta\":\"Paris\",\"item_id\":\"msg_0767cfe3f5d98b2a016a01389b1434819493f3d33e0ec6973c\",\"logprobs\":[],\"obfuscation\":\"fWbkBKTZ5oG\",\"output_index\":0,\"sequence_number\":4}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"content_index\":0,\"delta\":\" is\",\"item_id\":\"msg_0767cfe3f5d98b2a016a01389b1434819493f3d33e0ec6973c\",\"logprobs\":[],\"obfuscation\":\"BGuPlDrPchXwG\",\"output_index\":0,\"sequence_number\":5}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"content_index\":0,\"delta\":\" sunny\",\"item_id\":\"msg_0767cfe3f5d98b2a016a01389b1434819493f3d33e0ec6973c\",\"logprobs\":[],\"obfuscation\":\"7THzde2pni\",\"output_index\":0,\"sequence_number\":6}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"content_index\":0,\"delta\":\".\",\"item_id\":\"msg_0767cfe3f5d98b2a016a01389b1434819493f3d33e0ec6973c\",\"logprobs\":[],\"obfuscation\":\"Eo38bSElfyNmljj\",\"output_index\":0,\"sequence_number\":7}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.done\",\"content_index\":0,\"item_id\":\"msg_0767cfe3f5d98b2a016a01389b1434819493f3d33e0ec6973c\",\"logprobs\":[],\"output_index\":0,\"sequence_number\":8,\"text\":\"Paris is sunny.\"}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.content_part.done\",\"content_index\":0,\"item_id\":\"msg_0767cfe3f5d98b2a016a01389b1434819493f3d33e0ec6973c\",\"output_index\":0,\"part\":{\"type\":\"output_text\",\"annotations\":[],\"logprobs\":[],\"text\":\"Paris is sunny.\"},\"sequence_number\":9}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"id\":\"msg_0767cfe3f5d98b2a016a01389b1434819493f3d33e0ec6973c\",\"type\":\"message\",\"status\":\"completed\",\"content\":[{\"type\":\"output_text\",\"annotations\":[],\"logprobs\":[],\"text\":\"Paris is sunny.\"}],\"role\":\"assistant\"},\"output_index\":0,\"sequence_number\":10}"
+        },
+        {
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_0767cfe3f5d98b2a016a0138994258819485d082e5c78849a4\",\"object\":\"response\",\"created_at\":1778464921,\"status\":\"completed\",\"background\":false,\"completed_at\":1778464923,\"error\":null,\"frequency_penalty\":0.0,\"incomplete_details\":null,\"instructions\":null,\"max_output_tokens\":80,\"max_tool_calls\":null,\"model\":\"gpt-4.1-mini-2025-04-14\",\"moderation\":null,\"output\":[{\"id\":\"msg_0767cfe3f5d98b2a016a01389b1434819493f3d33e0ec6973c\",\"type\":\"message\",\"status\":\"completed\",\"content\":[{\"type\":\"output_text\",\"annotations\":[],\"logprobs\":[],\"text\":\"Paris is sunny.\"}],\"role\":\"assistant\"}],\"parallel_tool_calls\":true,\"presence_penalty\":0.0,\"previous_response_id\":null,\"prompt_cache_key\":null,\"prompt_cache_retention\":\"in_memory\",\"reasoning\":{\"effort\":null,\"summary\":null},\"safety_identifier\":null,\"service_tier\":\"default\",\"store\":false,\"temperature\":0.0,\"text\":{\"format\":{\"type\":\"text\"},\"verbosity\":\"medium\"},\"tool_choice\":\"auto\",\"tools\":[{\"type\":\"function\",\"description\":\"Get current weather for a city.\",\"name\":\"get_weather\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":true}],\"top_logprobs\":0,\"top_p\":1.0,\"truncation\":\"disabled\",\"usage\":{\"input_tokens\":99,\"input_tokens_details\":{\"cached_tokens\":0},\"output_tokens\":6,\"output_tokens_details\":{\"reasoning_tokens\":0},\"total_tokens\":105},\"user\":null,\"metadata\":{}},\"sequence_number\":11}"
+        }
+      ]
+    }
+  ]
+}
@@ -1,48 +0,0 @@
-import { Redactor } from "@opencode-ai/http-recorder"
-import { describe, expect } from "bun:test"
-import { Effect } from "effect"
-import { CacheHint, LLM } from "../../src"
-import { LLMClient } from "../../src/route"
-import * as AnthropicMessages from "../../src/protocols/anthropic-messages"
-import { LARGE_CACHEABLE_SYSTEM } from "../recorded-scenarios"
-import { recordedTests } from "../recorded-test"
-
-const model = AnthropicMessages.model({
-  id: "claude-haiku-4-5-20251001",
-  apiKey: process.env.ANTHROPIC_API_KEY ?? "fixture",
-})
-
-// Two identical generations in a row. The first call writes the prefix into
-// Anthropic's cache; the second should report a cache read against the same
-// prefix. Cassette captures both interactions in order.
-const cacheRequest = LLM.request({
-  id: "recorded_anthropic_cache",
-  model,
-  system: [{ type: "text", text: LARGE_CACHEABLE_SYSTEM, cache: new CacheHint({ type: "ephemeral" }) }],
-  prompt: "Say hi.",
-  generation: { maxTokens: 16, temperature: 0 },
-})
-
-const recorded = recordedTests({
-  prefix: "anthropic-messages-cache",
-  provider: "anthropic",
-  protocol: "anthropic-messages",
-  requires: ["ANTHROPIC_API_KEY"],
-  options: { redactor: Redactor.defaults({ requestHeaders: { allow: ["content-type", "anthropic-version"] } }) },
-})
-
-describe("Anthropic Messages cache recorded", () => {
-  recorded.effect.with("writes then reads cache_control on identical second call", { tags: ["cache"] }, () =>
-    Effect.gen(function* () {
-      const first = yield* LLMClient.generate(cacheRequest)
-      // The first call may write the cache (cacheWriteInputTokens > 0) or it
-      // may be a fresh miss (both fields 0) depending on whether the prefix is
-      // already warm on Anthropic's side. The assertion that matters is that
-      // the SECOND call reports a non-zero cache read.
-      expect(first.usage?.cacheReadInputTokens ?? 0).toBeGreaterThanOrEqual(0)
-
-      const second = yield* LLMClient.generate(cacheRequest)
-      expect(second.usage?.cacheReadInputTokens ?? 0).toBeGreaterThan(0)
-    }),
-  )
-})
@@ -1,6 +1,6 @@
 import { describe, expect } from "bun:test"
 import { Effect } from "effect"
-import { CacheHint, LLM, LLMError } from "../../src"
+import { CacheHint, LLM, LLMError, Usage } from "../../src"
 import { LLMClient } from "../../src/route"
 import * as AnthropicMessages from "../../src/protocols/anthropic-messages"
 import { it } from "../lib/effect"
@@ -110,10 +110,11 @@ describe("Anthropic Messages route", () => {
      expect(response.text).toBe("Hello!")
      expect(response.reasoning).toBe("thinking")
      expect(response.usage).toMatchObject({
-        inputTokens: 5,
+        inputTokens: 6,
        outputTokens: 2,
+        nonCachedInputTokens: 5,
        cacheReadInputTokens: 1,
-        totalTokens: 7,
+        totalTokens: 8,
      })
      expect(response.events.find((event) => event.type === "reasoning-end")).toMatchObject({
        providerMetadata: { anthropic: { signature: "sig_1" } },
@@ -152,7 +153,13 @@ describe("Anthropic Messages route", () => {
        {
          type: "request-finish",
          reason: "tool-calls",
-          usage: { inputTokens: 5, outputTokens: 1, totalTokens: 6, native: { input_tokens: 5, output_tokens: 1 } },
+          usage: new Usage({
+            inputTokens: 5,
+            outputTokens: 1,
+            nonCachedInputTokens: 5,
+            totalTokens: 6,
+            providerMetadata: { anthropic: { input_tokens: 5, output_tokens: 1 } },
+          }),
        },
      ])
    }),
@@ -374,134 +381,4 @@ describe("Anthropic Messages route", () => {
      expect(error.message).toContain("Anthropic Messages user messages only support text content for now")
    }),
  )
-
-  it.effect("maps ttlSeconds >= 3600 to cache_control ttl: '1h'", () =>
-    Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
-        LLM.request({
-          model,
-          system: { type: "text", text: "system", cache: new CacheHint({ type: "ephemeral", ttlSeconds: 3600 }) },
-          prompt: "hi",
-        }),
-      )
-
-      expect(prepared.body).toMatchObject({
-        system: [{ type: "text", text: "system", cache_control: { type: "ephemeral", ttl: "1h" } }],
-      })
-    }),
-  )
-
-  it.effect("emits cache_control on tool definitions and tool-result blocks", () =>
-    Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
-        LLM.request({
-          model,
-          tools: [
-            {
-              name: "lookup",
-              description: "lookup tool",
-              inputSchema: { type: "object", properties: {} },
-              cache: new CacheHint({ type: "ephemeral" }),
-            },
-          ],
-          messages: [
-            LLM.user("What's the weather?"),
-            LLM.assistant([LLM.toolCall({ id: "call_1", name: "lookup", input: {} })]),
-            LLM.toolMessage({
-              id: "call_1",
-              name: "lookup",
-              result: { temp: 72 },
-              cache: new CacheHint({ type: "ephemeral" }),
-            }),
-          ],
-        }),
-      )
-
-      expect(prepared.body).toMatchObject({
-        tools: [{ name: "lookup", cache_control: { type: "ephemeral" } }],
-        messages: [
-          { role: "user", content: [{ type: "text", text: "What's the weather?" }] },
-          { role: "assistant", content: [{ type: "tool_use", id: "call_1", name: "lookup" }] },
-          {
-            role: "user",
-            content: [{ type: "tool_result", tool_use_id: "call_1", cache_control: { type: "ephemeral" } }],
-          },
-        ],
-      })
-    }),
-  )
-
-  it.effect("drops cache_control breakpoints past the 4-per-request cap", () =>
-    Effect.gen(function* () {
-      const hint = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
-        LLM.request({
-          model,
-          system: [
-            { type: "text", text: "a", cache: hint },
-            { type: "text", text: "b", cache: hint },
-            { type: "text", text: "c", cache: hint },
-            { type: "text", text: "d", cache: hint },
-            { type: "text", text: "e", cache: hint },
-            { type: "text", text: "f", cache: hint },
-          ],
-          prompt: "hi",
-        }),
-      )
-
-      const system = (prepared.body as { system: Array<{ cache_control?: unknown }> }).system
-      const marked = system.filter((part) => part.cache_control !== undefined)
-      expect(marked).toHaveLength(4)
-      expect(system[4]?.cache_control).toBeUndefined()
-      expect(system[5]?.cache_control).toBeUndefined()
-    }),
-  )
-
-  it.effect("spends breakpoint budget on tools before system before messages", () =>
-    Effect.gen(function* () {
-      const hint = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
-        LLM.request({
-          model,
-          tools: [
-            {
-              name: "t1",
-              description: "t1",
-              inputSchema: { type: "object", properties: {} },
-              cache: hint,
-            },
-            {
-              name: "t2",
-              description: "t2",
-              inputSchema: { type: "object", properties: {} },
-              cache: hint,
-            },
-            {
-              name: "t3",
-              description: "t3",
-              inputSchema: { type: "object", properties: {} },
-              cache: hint,
-            },
-            {
-              name: "t4",
-              description: "t4",
-              inputSchema: { type: "object", properties: {} },
-              cache: hint,
-            },
-          ],
-          system: [{ type: "text", text: "system-tail", cache: hint }],
-          messages: [LLM.user([{ type: "text", text: "message-tail", cache: hint }])],
-        }),
-      )
-
-      const body = prepared.body as {
-        tools: Array<{ cache_control?: unknown }>
-        system: Array<{ cache_control?: unknown }>
-        messages: Array<{ content: Array<{ cache_control?: unknown }> }>
-      }
-      expect(body.tools.every((t) => t.cache_control !== undefined)).toBe(true)
-      expect(body.system[0]?.cache_control).toBeUndefined()
-      expect(body.messages[0]?.content[0]?.cache_control).toBeUndefined()
-    }),
-  )
 })
@@ -1,50 +0,0 @@
-import { describe, expect } from "bun:test"
-import { Effect } from "effect"
-import { CacheHint, LLM } from "../../src"
-import { LLMClient } from "../../src/route"
-import * as BedrockConverse from "../../src/protocols/bedrock-converse"
-import { LARGE_CACHEABLE_SYSTEM } from "../recorded-scenarios"
-import { recordedTests } from "../recorded-test"
-
-const RECORDING_REGION = process.env.BEDROCK_RECORDING_REGION ?? "us-east-1"
-
-// Use a Claude model on Bedrock — Nova has automatic prefix caching that
-// doesn't reliably surface `cacheRead`/`cacheWrite` in usage, so the second
-// call wouldn't deterministically prove cache mapping works. Override with
-// BEDROCK_CACHE_MODEL_ID if your account has access elsewhere.
-const model = BedrockConverse.model({
-  id: process.env.BEDROCK_CACHE_MODEL_ID ?? "us.anthropic.claude-haiku-4-5-20251001-v1:0",
-  credentials: {
-    region: RECORDING_REGION,
-    accessKeyId: process.env.AWS_ACCESS_KEY_ID ?? "fixture",
-    secretAccessKey: process.env.AWS_SECRET_ACCESS_KEY ?? "fixture",
-    sessionToken: process.env.AWS_SESSION_TOKEN,
-  },
-})
-
-const cacheRequest = LLM.request({
-  id: "recorded_bedrock_cache",
-  model,
-  system: [{ type: "text", text: LARGE_CACHEABLE_SYSTEM, cache: new CacheHint({ type: "ephemeral" }) }],
-  prompt: "Say hi.",
-  generation: { maxTokens: 16, temperature: 0 },
-})
-
-const recorded = recordedTests({
-  prefix: "bedrock-converse-cache",
-  provider: "amazon-bedrock",
-  protocol: "bedrock-converse",
-  requires: ["AWS_ACCESS_KEY_ID", "AWS_SECRET_ACCESS_KEY"],
-})
-
-describe("Bedrock Converse cache recorded", () => {
-  recorded.effect.with("writes then reads cachePoint on identical second call", { tags: ["cache"] }, () =>
-    Effect.gen(function* () {
-      const first = yield* LLMClient.generate(cacheRequest)
-      expect(first.usage?.cacheReadInputTokens ?? 0).toBeGreaterThanOrEqual(0)
-
-      const second = yield* LLMClient.generate(cacheRequest)
-      expect(second.usage?.cacheReadInputTokens ?? 0).toBeGreaterThan(0)
-    }),
-  )
-})
@@ -440,77 +440,6 @@ describe("Bedrock Converse route", () => {
      expect(error.message).toContain("Bedrock Converse does not support media type application/x-tar")
    }),
  )
-
-  it.effect("maps ttlSeconds >= 3600 to cachePoint ttl: '1h'", () =>
-    Effect.gen(function* () {
-      const cache = new CacheHint({ type: "ephemeral", ttlSeconds: 3600 })
-      const prepared = yield* LLMClient.prepare(
-        LLM.request({
-          model,
-          system: [{ type: "text", text: "system", cache }],
-          prompt: "hi",
-        }),
-      )
-
-      expect(prepared.body).toMatchObject({
-        system: [{ text: "system" }, { cachePoint: { type: "default", ttl: "1h" } }],
-      })
-    }),
-  )
-
-  it.effect("appends cachePoint after marked tool definitions and tool-result blocks", () =>
-    Effect.gen(function* () {
-      const cache = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
-        LLM.request({
-          model,
-          tools: [{ name: "lookup", description: "lookup", inputSchema: { type: "object", properties: {} }, cache }],
-          messages: [
-            LLM.user("What's the weather?"),
-            LLM.assistant([LLM.toolCall({ id: "call_1", name: "lookup", input: {} })]),
-            LLM.toolMessage({ id: "call_1", name: "lookup", result: { temp: 72 }, cache }),
-          ],
-        }),
-      )
-
-      expect(prepared.body).toMatchObject({
-        toolConfig: {
-          tools: [{ toolSpec: { name: "lookup" } }, { cachePoint: { type: "default" } }],
-        },
-        messages: [
-          { role: "user", content: [{ text: "What's the weather?" }] },
-          { role: "assistant", content: [{ toolUse: { toolUseId: "call_1" } }] },
-          {
-            role: "user",
-            content: [{ toolResult: { toolUseId: "call_1" } }, { cachePoint: { type: "default" } }],
-          },
-        ],
-      })
-    }),
-  )
-
-  it.effect("drops cachePoint markers past the 4-per-request cap", () =>
-    Effect.gen(function* () {
-      const cache = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
-        LLM.request({
-          model,
-          system: [
-            { type: "text", text: "a", cache },
-            { type: "text", text: "b", cache },
-            { type: "text", text: "c", cache },
-            { type: "text", text: "d", cache },
-            { type: "text", text: "e", cache },
-            { type: "text", text: "f", cache },
-          ],
-          prompt: "hi",
-        }),
-      )
-
-      const system = (prepared.body as { system: Array<{ cachePoint?: unknown }> }).system
-      expect(system.filter((part) => "cachePoint" in part)).toHaveLength(4)
-    }),
-  )
 })

 // Live recorded integration tests. Run with `RECORD=true AWS_ACCESS_KEY_ID=...
@@ -1,47 +0,0 @@
-import { describe, expect } from "bun:test"
-import { Effect } from "effect"
-import { LLM } from "../../src"
-import { LLMClient } from "../../src/route"
-import * as Gemini from "../../src/protocols/gemini"
-import { LARGE_CACHEABLE_SYSTEM } from "../recorded-scenarios"
-import { recordedTests } from "../recorded-test"
-
-const model = Gemini.model({
-  id: "gemini-2.5-flash",
-  apiKey: process.env.GEMINI_API_KEY ?? "fixture",
-})
-
-// Gemini does implicit prefix caching on 2.5+ models above ~1024 tokens. The
-// `CacheHint` is currently a no-op for Gemini (the explicit `CachedContent`
-// API is out-of-band and intentionally not wired up). This test exists to
-// pin the usage-parsing path: `cachedContentTokenCount` should surface as
-// `cacheReadInputTokens` on the second identical call.
-const cacheRequest = LLM.request({
-  id: "recorded_gemini_cache",
-  model,
-  system: LARGE_CACHEABLE_SYSTEM,
-  prompt: "Say hi.",
-  generation: { maxTokens: 16, temperature: 0 },
-})
-
-const recorded = recordedTests({
-  prefix: "gemini-cache",
-  provider: "google",
-  protocol: "gemini",
-  requires: ["GEMINI_API_KEY"],
-})
-
-describe("Gemini cache recorded", () => {
-  recorded.effect.with("reports cachedContentTokenCount on identical second call", { tags: ["cache"] }, () =>
-    Effect.gen(function* () {
-      const first = yield* LLMClient.generate(cacheRequest)
-      expect(first.usage?.cacheReadInputTokens ?? 0).toBeGreaterThanOrEqual(0)
-
-      const second = yield* LLMClient.generate(cacheRequest)
-      // Implicit caching is best-effort on Gemini's side; we assert the field
-      // is at least populated and non-negative. When re-recording, verify the
-      // cassette shows > 0 in the second response's usage.
-      expect(second.usage?.cacheReadInputTokens ?? 0).toBeGreaterThanOrEqual(0)
-    }),
-  )
-})
@@ -1,6 +1,6 @@
 import { describe, expect } from "bun:test"
 import { Effect } from "effect"
-import { LLM, LLMError } from "../../src"
+import { LLM, LLMError, Usage } from "../../src"
 import { LLMClient } from "../../src/route"
 import * as Gemini from "../../src/protocols/gemini"
 import { it } from "../lib/effect"
@@ -198,9 +198,10 @@ describe("Gemini route", () => {
      expect(response.reasoning).toBe("thinking")
      expect(response.usage).toMatchObject({
        inputTokens: 5,
-        outputTokens: 2,
-        reasoningTokens: 1,
+        outputTokens: 3,
+        nonCachedInputTokens: 4,
        cacheReadInputTokens: 1,
+        reasoningTokens: 1,
        totalTokens: 7,
      })
      expect(response.events).toEqual([
@@ -210,20 +211,23 @@ describe("Gemini route", () => {
        {
          type: "request-finish",
          reason: "stop",
-          usage: {
+          usage: new Usage({
            inputTokens: 5,
-            outputTokens: 2,
-            reasoningTokens: 1,
+            outputTokens: 3,
+            nonCachedInputTokens: 4,
            cacheReadInputTokens: 1,
+            reasoningTokens: 1,
            totalTokens: 7,
-            native: {
-              promptTokenCount: 5,
-              candidatesTokenCount: 2,
-              totalTokenCount: 7,
-              thoughtsTokenCount: 1,
-              cachedContentTokenCount: 1,
+            providerMetadata: {
+              google: {
+                promptTokenCount: 5,
+                candidatesTokenCount: 2,
+                totalTokenCount: 7,
+                thoughtsTokenCount: 1,
+                cachedContentTokenCount: 1,
+              },
            },
-          },
+          }),
        },
      ])
    }),
@@ -257,12 +261,13 @@ describe("Gemini route", () => {
        {
          type: "request-finish",
          reason: "tool-calls",
-          usage: {
+          usage: new Usage({
            inputTokens: 5,
            outputTokens: 1,
+            nonCachedInputTokens: 5,
            totalTokens: 6,
-            native: { promptTokenCount: 5, candidatesTokenCount: 1 },
-          },
+            providerMetadata: { google: { promptTokenCount: 5, candidatesTokenCount: 1 } },
+          }),
        },
      ])
    }),
@@ -1,7 +1,7 @@
 import { describe, expect } from "bun:test"
 import { Effect, Schema, Stream } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
-import { LLM, LLMError } from "../../src"
+import { LLM, LLMError, Usage } from "../../src"
 import * as Azure from "../../src/providers/azure"
 import * as OpenAI from "../../src/providers/openai"
 import * as OpenAIChat from "../../src/protocols/openai-chat"
@@ -230,20 +230,23 @@ describe("OpenAI Chat route", () => {
        {
          type: "request-finish",
          reason: "stop",
-          usage: {
+          usage: new Usage({
            inputTokens: 5,
            outputTokens: 2,
-            reasoningTokens: 0,
+            nonCachedInputTokens: 4,
            cacheReadInputTokens: 1,
+            reasoningTokens: 0,
            totalTokens: 7,
-            native: {
-              prompt_tokens: 5,
-              completion_tokens: 2,
-              total_tokens: 7,
-              prompt_tokens_details: { cached_tokens: 1 },
-              completion_tokens_details: { reasoning_tokens: 0 },
+            providerMetadata: {
+              openai: {
+                prompt_tokens: 5,
+                completion_tokens: 2,
+                total_tokens: 7,
+                prompt_tokens_details: { cached_tokens: 1 },
+                completion_tokens_details: { reasoning_tokens: 0 },
+              },
            },
-          },
+          }),
        },
      ])
    }),
@@ -1,44 +0,0 @@
-import { describe, expect } from "bun:test"
-import { Effect } from "effect"
-import { LLM } from "../../src"
-import { LLMClient } from "../../src/route"
-import * as OpenAIResponses from "../../src/protocols/openai-responses"
-import { LARGE_CACHEABLE_SYSTEM } from "../recorded-scenarios"
-import { recordedTests } from "../recorded-test"
-
-const model = OpenAIResponses.model({
-  id: "gpt-4.1-mini",
-  apiKey: process.env.OPENAI_API_KEY ?? "fixture",
-})
-
-// OpenAI caches prefixes automatically once they cross the 1024-token threshold;
-// `CacheHint` is a no-op for the wire body. The stable signal is the
-// `prompt_cache_key` routing hint, which keeps repeated calls on the same shard
-// so cache hits are observable.
-const cacheRequest = LLM.request({
-  id: "recorded_openai_responses_cache",
-  model,
-  system: LARGE_CACHEABLE_SYSTEM,
-  prompt: "Say hi.",
-  generation: { maxTokens: 16, temperature: 0 },
-  providerOptions: { openai: { promptCacheKey: "recorded-cache-test" } },
-})
-
-const recorded = recordedTests({
-  prefix: "openai-responses-cache",
-  provider: "openai",
-  protocol: "openai-responses",
-  requires: ["OPENAI_API_KEY"],
-})
-
-describe("OpenAI Responses cache recorded", () => {
-  recorded.effect.with("reports cached_tokens on identical second call", { tags: ["cache"] }, () =>
-    Effect.gen(function* () {
-      const first = yield* LLMClient.generate(cacheRequest)
-      expect(first.usage?.cacheReadInputTokens ?? 0).toBeGreaterThanOrEqual(0)
-
-      const second = yield* LLMClient.generate(cacheRequest)
-      expect(second.usage?.cacheReadInputTokens ?? 0).toBeGreaterThan(0)
-    }),
-  )
-})
@@ -1,7 +1,7 @@
 import { describe, expect } from "bun:test"
 import { ConfigProvider, Effect, Layer, Stream } from "effect"
 import { Headers, HttpClientRequest } from "effect/unstable/http"
-import { LLM, LLMError } from "../../src"
+import { LLM, LLMError, Usage } from "../../src"
 import { Auth, LLMClient, RequestExecutor, WebSocketExecutor } from "../../src/route"
 import * as Azure from "../../src/providers/azure"
 import * as OpenAI from "../../src/providers/openai"
@@ -342,20 +342,23 @@ describe("OpenAI Responses route", () => {
          type: "request-finish",
          reason: "stop",
          providerMetadata: { openai: { responseId: "resp_1", serviceTier: "default" } },
-          usage: {
+          usage: new Usage({
            inputTokens: 5,
            outputTokens: 2,
-            reasoningTokens: 0,
+            nonCachedInputTokens: 4,
            cacheReadInputTokens: 1,
+            reasoningTokens: 0,
            totalTokens: 7,
-            native: {
-              input_tokens: 5,
-              output_tokens: 2,
-              total_tokens: 7,
-              input_tokens_details: { cached_tokens: 1 },
-              output_tokens_details: { reasoning_tokens: 0 },
+            providerMetadata: {
+              openai: {
+                input_tokens: 5,
+                output_tokens: 2,
+                total_tokens: 7,
+                input_tokens_details: { cached_tokens: 1 },
+                output_tokens_details: { reasoning_tokens: 0 },
+              },
            },
-          },
+          }),
        },
      ])
    }),
@@ -411,7 +414,13 @@ describe("OpenAI Responses route", () => {
        {
          type: "request-finish",
          reason: "tool-calls",
-          usage: { inputTokens: 5, outputTokens: 1, totalTokens: 6, native: { input_tokens: 5, output_tokens: 1 } },
+          usage: new Usage({
+            inputTokens: 5,
+            outputTokens: 1,
+            nonCachedInputTokens: 5,
+            totalTokens: 6,
+            providerMetadata: { openai: { input_tokens: 5, output_tokens: 1 } },
+          }),
        },
      ])
    }),
@@ -6,18 +6,6 @@ import { tool } from "../src/tool"

 export const weatherToolName = "get_weather"

-// A deterministic system prompt long enough to clear every supported provider's
-// minimum cacheable-prefix threshold (Anthropic Haiku 3.5: 2048 tokens; Anthropic
-// Opus/Haiku 4.5: 4096 tokens; OpenAI/Gemini/Bedrock: lower). Built by repeating
-// a fixed sentence — the cassette replays bit-for-bit, so the exact text matters
-// only when re-recording with `RECORD=true`.
-export const LARGE_CACHEABLE_SYSTEM = (() => {
-  const sentence = "You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. "
-  // ~100 chars per sentence × 250 repeats ≈ 25,000 chars ≈ 5k+ tokens, safely
-  // above every provider's threshold.
-  return sentence.repeat(250)
-})()
-
 export const weatherTool = LLM.toolDefinition({
  name: weatherToolName,
  description: "Get current weather for a city.",
@@ -1,6 +1,7 @@
 import { describe, expect, test } from "bun:test"
 import { Schema } from "effect"
-import { ContentPart, LLMEvent, LLMRequest, ModelID, ModelLimits, ModelRef, ProviderID } from "../src/schema"
+import { ContentPart, LLMEvent, LLMRequest, ModelID, ModelLimits, ModelRef, ProviderID, Usage } from "../src/schema"
+import { ProviderShared } from "../src/protocols/shared"

 const model = new ModelRef({
  id: ModelID.make("fake-model"),
@@ -48,3 +49,30 @@ describe("llm schema", () => {
    expect(ContentPart.guards.media({ type: "text", text: "hi" })).toBe(false)
  })
 })
+
+describe("LLM.Usage", () => {
+  test("subtractTokens clamps non-sensical breakdowns to zero", () => {
+    // Defense against a provider reporting cached_tokens > prompt_tokens or
+    // reasoning_tokens > completion_tokens — the negative would otherwise
+    // round-trip through the pipeline and crash strict downstream schemas.
+    expect(ProviderShared.subtractTokens(5, 3)).toBe(2)
+    expect(ProviderShared.subtractTokens(5, 10)).toBe(0)
+    expect(ProviderShared.subtractTokens(5, undefined)).toBe(5)
+    expect(ProviderShared.subtractTokens(undefined, 3)).toBeUndefined()
+    expect(ProviderShared.subtractTokens(undefined, undefined)).toBeUndefined()
+  })
+
+  test("sumTokens returns undefined only when every input is undefined", () => {
+    expect(ProviderShared.sumTokens(1, 2, 3)).toBe(6)
+    expect(ProviderShared.sumTokens(1, undefined, 3)).toBe(4)
+    expect(ProviderShared.sumTokens(undefined, undefined, undefined)).toBeUndefined()
+    expect(ProviderShared.sumTokens()).toBeUndefined()
+  })
+
+  test("visibleOutputTokens clamps reasoning > output to zero", () => {
+    expect(new Usage({ outputTokens: 10, reasoningTokens: 4 }).visibleOutputTokens).toBe(6)
+    expect(new Usage({ outputTokens: 10 }).visibleOutputTokens).toBe(10)
+    expect(new Usage({ outputTokens: 4, reasoningTokens: 10 }).visibleOutputTokens).toBe(0)
+    expect(new Usage({}).visibleOutputTokens).toBe(0)
+  })
+})
@@ -9,13 +9,6 @@
 - **Output**: creates `migration/<timestamp>_<slug>/migration.sql` and `snapshot.json`.
 - **Tests**: migration tests should read the per-folder layout (no `_journal.json`).

-## Development server
-
- Running `bun dev` from `packages/opencode` starts the live interactive TUI. Do not run it as a blocking foreground command when you need to inspect the result.
- Start it in `tmux` instead: `tmux new-session -d -s opencode-dev 'bun dev'`.
- Capture the current TUI output with: `tmux capture-pane -pt opencode-dev`.
- Stop the session explicitly when done: `tmux kill-session -t opencode-dev`.
-
 # Module shape

 Do not use `export namespace Foo { ... }` for module organization. It is not
@@ -1,4 +0,0 @@
-CREATE TABLE `data_migration` (
-	`name` text PRIMARY KEY,
-	`time_completed` integer NOT NULL
-);
@@ -1,6 +1,6 @@
 {
  "$schema": "https://json.schemastore.org/package.json",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "name": "opencode",
  "type": "module",
  "license": "MIT",
@@ -93,6 +93,7 @@ const appBindingCommands = [
  "theme.mode.lock",
  "help.show",
  "docs.open",
+  "app.exit",
  "app.debug",
  "app.console",
  "app.heap_snapshot",
@@ -647,6 +648,11 @@ function App(props: { onSnapshot?: () => Promise<string[]> }) {
        title: "Exit the app",
        slashName: "exit",
        slashAliases: ["quit", "q"],
+        enabled: () => {
+          const current = promptRef.current
+          if (!current?.focused) return true
+          return current.current.input === ""
+        },
        run: () => exit(),
        category: "System",
      },
@@ -779,17 +785,6 @@ function App(props: { onSnapshot?: () => Promise<string[]> }) {
    bindings: tuiConfig.keybinds.gather("app", appBindingCommands),
  }))

-  useBindings(() => ({
-    enabled: () => {
-      const ok = command.matcher.get()
-      if (!ok) return false
-      const current = promptRef.current
-      if (!current?.focused) return true
-      return current.current.input === ""
-    },
-    bindings: tuiConfig.keybinds.gather("app_exit", ["app.exit"]),
-  }))
-
  event.on(TuiEvent.CommandExecute.type, (evt) => {
    command.run(evt.properties.command)
  })
@@ -712,6 +712,7 @@ export function Prompt(props: PromptProps) {
      ...input.traits,
      ...computePromptTraits({
        mode: store.mode,
+        disabled: !!props.disabled,
        autocompleteVisible: !!auto()?.visible,
      }),
    }
@@ -4,6 +4,7 @@ export type PromptMode = "normal" | "shell"

 export interface PromptTraitsInput {
  mode: PromptMode
+  disabled: boolean
  autocompleteVisible: boolean
 }

@@ -15,9 +16,10 @@ export type PromptTraits = EditorTraits & {
 /**
 * Compute the textarea editor traits for the prompt.
 *
- * The OpenTUI managed textarea keymap owns `traits.suspend`. Prompt traits
- * only expose capture/status metadata so focus changes cannot unsuspend the
- * keymap-managed editor mappings.
+ * `traits.suspend` gates the textarea's keybinding actions (backspace,
+ * delete-word, arrow movement, undo/redo, etc.). Shell mode is an active
+ * editing mode — only `disabled` should suspend the textarea, otherwise
+ * users can type in shell mode but cannot delete or move the cursor.
 */
 export function computePromptTraits(input: PromptTraitsInput): PromptTraits {
  const capture =
@@ -28,6 +30,7 @@ export function computePromptTraits(input: PromptTraitsInput): PromptTraits {
      : undefined
  return {
    capture,
+    suspend: input.disabled,
    status: input.mode === "shell" ? "SHELL" : undefined,
    owner: "opencode",
    role: "prompt",
@@ -8,9 +8,9 @@ import {
 import {
  KeymapProvider,
  reactiveMatcherFromSignal,
+  useBindings,
  useKeymap,
  useKeymapSelector,
-  useBindings,
 } from "@opentui/keymap/solid"
 import type { Accessor } from "solid-js"
 import type { TuiConfig } from "./config/tui"
@@ -26,28 +26,6 @@ export { reactiveMatcherFromSignal, useBindings, useKeymapSelector }

 export type OpenTuiKeymap = ReturnType<typeof useKeymap>

-const KEY_ALIASES = {
-  enter: "return",
-  esc: "escape",
-} as const
-
-function expandKeyAliases(input: string) {
-  const result = Object.entries(KEY_ALIASES).reduce(
-    (acc, [alias, key]) => acc.replace(new RegExp(`(^|[+,\\s>])${alias}(?=$|[+,\\s<])`, "gi"), `$1${key}`),
-    input,
-  )
-  if (result === input) return
-  return result
-}
-
-function registerKeyAliases(keymap: OpenTuiKeymap) {
-  return keymap.appendBindingExpander((ctx) => {
-    const key = expandKeyAliases(ctx.input)
-    if (!key) return
-    return [{ key, displays: ctx.displays }]
-  })
-}
-
 const inputCommands = [
  "input.move.left",
  "input.move.right",
@@ -120,13 +98,8 @@ export function formatKeyBindings(
  return formatCommandBindingsExtra(bindings, formatOptions(config))
 }

-export function registerOpencodeKeymap(
-  keymap: OpenTuiKeymap,
-  renderer: CliRenderer,
-  config: Pick<TuiConfig.Resolved, "keybinds" | "leader_timeout">,
-) {
+export function registerOpencodeKeymap(keymap: OpenTuiKeymap, renderer: CliRenderer, config: TuiConfig.Resolved) {
  const offCommaBindings = addons.registerCommaBindings(keymap)
-  const offAliasExpander = registerKeyAliases(keymap)
  const offBaseLayout = addons.registerBaseLayoutFallback(keymap)
  const offLeader = addons.registerTimedLeader(keymap, {
    trigger: config.keybinds.get(LEADER_TOKEN),
@@ -135,17 +108,20 @@ export function registerOpencodeKeymap(
  })
  const offEscape = addons.registerEscapeClearsPendingSequence(keymap)
  const offBackspace = addons.registerBackspacePopsPendingSequence(keymap)
-  const offInputBindings = addons.registerManagedTextareaLayer(keymap, renderer, {
+  const offInputCommands = addons.registerEditBufferCommands(keymap, renderer)
+  const offInputSuspension = addons.registerTextareaMappingSuspension(keymap, renderer)
+  const offInputBindings = keymap.registerLayer({
    enabled: () => renderer.currentFocusedEditor !== null,
    bindings: config.keybinds.gather("input", inputCommands),
  })

  return () => {
    offInputBindings()
+    offInputSuspension()
+    offInputCommands()
    offBackspace()
    offEscape()
    offLeader()
-    offAliasExpander()
    offBaseLayout()
    offCommaBindings()
  }
@@ -1,8 +1,5 @@
 import { createOpencodeClient } from "@opencode-ai/sdk/v2"
 import { SessionID } from "@/session/schema"
-import { Schema } from "effect"
-
-const decodeSessionID = Schema.decodeUnknownSync(SessionID)

 export async function validateSession(input: {
  url: string
@@ -13,11 +10,9 @@ export async function validateSession(input: {
 }) {
  if (!input.sessionID) return

-  let sessionID: SessionID
-  try {
-    sessionID = decodeSessionID(input.sessionID)
-  } catch (error) {
-    throw new Error(`Invalid session ID: ${error instanceof Error ? error.message : "unknown error"}`, { cause: error })
+  const result = SessionID.zod.safeParse(input.sessionID)
+  if (!result.success) {
+    throw new Error(`Invalid session ID: ${result.error.issues.at(0)?.message ?? "unknown error"}`)
  }

  await createOpencodeClient({
@@ -25,5 +20,5 @@ export async function validateSession(input: {
    directory: input.directory,
    fetch: input.fetch,
    headers: input.headers,
-  }).session.get({ sessionID }, { throwOnError: true })
+  }).session.get({ sessionID: result.data }, { throwOnError: true })
 }
@@ -1,6 +1,7 @@
 import { Schema } from "effect"

 import { Identifier } from "@/id/id"
+import { zod } from "@opencode-ai/core/effect-zod"
 import { withStatics } from "@opencode-ai/core/schema"

 const workspaceIdSchema = Schema.String.check(Schema.isStartsWith("wrk")).pipe(Schema.brand("WorkspaceID"))
@@ -10,5 +11,6 @@ export type WorkspaceID = typeof workspaceIdSchema.Type
 export const WorkspaceID = workspaceIdSchema.pipe(
  withStatics((schema: typeof workspaceIdSchema) => ({
    ascending: (id?: string) => schema.make(Identifier.ascending("workspace", id)),
+    zod: zod(schema),
  })),
 )
@@ -676,12 +676,14 @@ export const layer = Layer.effect(
        }

        if (input.workspaceID === null) {
-          yield* sync.run(Session.Event.Updated, {
-            sessionID: input.sessionID,
-            info: {
-              workspaceID: null,
-            },
-          })
+          yield* Effect.sync(() =>
+            SyncEvent.run(Session.Event.Updated, {
+              sessionID: input.sessionID,
+              info: {
+                workspaceID: null,
+              },
+            }),
+          )

          log.info("session warp complete", {
            workspaceID: input.workspaceID,
@@ -1,6 +0,0 @@
-import { integer, sqliteTable, text } from "drizzle-orm/sqlite-core"
-
-export const DataMigrationTable = sqliteTable("data_migration", {
-  name: text().primaryKey(),
-  time_completed: integer().notNull(),
-})
@@ -1,59 +0,0 @@
-import { Context, Effect, Layer } from "effect"
-import { Database } from "./storage/db"
-import { DataMigrationTable } from "./data-migration.sql"
-import * as Log from "@opencode-ai/core/util/log"
-import { eq } from "drizzle-orm"
-
-export type Migration<R = never> = {
-  name: string
-  run: Effect.Effect<void, unknown, R>
-}
-
-const log = Log.create({ service: "data-migration" })
-
-export interface Interface {}
-
-export class Service extends Context.Service<Service, Interface>()("@opencode/DataMigration") {}
-
-export const layer = Layer.effect(
-  Service,
-  Effect.gen(function* () {
-    const migrations: Migration[] = []
-
-    yield* Effect.gen(function* () {
-      if (migrations.length === 0) return
-
-      // Migrations run in a background fiber, so they must be resumable until
-      // their completion row is written.
-      for (const migration of migrations) {
-        const completed = Database.use((db) =>
-          db
-            .select({ name: DataMigrationTable.name })
-            .from(DataMigrationTable)
-            .where(eq(DataMigrationTable.name, migration.name))
-            .get(),
-        )
-        if (completed) continue
-
-        log.info("running data migration", { name: migration.name })
-        yield* migration.run
-        Database.use((db) =>
-          db
-            .insert(DataMigrationTable)
-            .values({ name: migration.name, time_completed: Date.now() })
-            .onConflictDoNothing()
-            .run(),
-        )
-      }
-    }).pipe(
-      Effect.tapCause((cause) => Effect.logError("failed to run data migrations", { cause })),
-      Effect.ignore,
-      Effect.forkScoped,
-    )
-    return Service.of({})
-  }),
-)
-
-export const defaultLayer = layer
-
-export * as DataMigration from "./data-migration"
@@ -54,7 +54,6 @@ import { SessionShare } from "@/share/session"
 import { SyncEvent } from "@/sync"
 import { Npm } from "@opencode-ai/core/npm"
 import { memoMap } from "@opencode-ai/core/effect/memo-map"
-import { DataMigration } from "@/data-migration"

 export const AppLayer = Layer.mergeAll(
  Npm.defaultLayer,
@@ -107,7 +106,6 @@ export const AppLayer = Layer.mergeAll(
  ShareNext.defaultLayer,
  SessionShare.defaultLayer,
  SyncEvent.defaultLayer,
-  DataMigration.defaultLayer,
 ).pipe(Layer.provideMerge(InstanceLayer.layer), Layer.provideMerge(Observability.layer))

 const rt = ManagedRuntime.make(AppLayer, { memoMap })
@@ -4,6 +4,7 @@ import * as Session from "./session"
 import { SessionID, MessageID, PartID } from "./schema"
 import { Provider } from "@/provider/provider"
 import { MessageV2 } from "./message-v2"
+import z from "zod"
 import { Token } from "@/util/token"
 import * as Log from "@opencode-ai/core/util/log"
 import { SessionProcessor } from "./processor"
@@ -18,9 +19,9 @@ import { InstanceState } from "@/effect/instance-state"
 import { isOverflow as overflow, usable } from "./overflow"
 import { makeRuntime } from "@/effect/run-service"
 import { serviceUse } from "@/effect/service-use"
-import { SyncEvent } from "@/sync"
+import { fn } from "@/util/fn"
+import { EventV2 } from "@/v2/event"
 import { SessionEvent } from "@/v2/session-event"
-import { Flag } from "@opencode-ai/core/flag/flag"

 const log = Log.create({ service: "session.compaction" })

@@ -220,7 +221,6 @@ export const layer: Layer.Layer<
  | Plugin.Service
  | SessionProcessor.Service
  | Provider.Service
-  | SyncEvent.Service
 > = Layer.effect(
  Service,
  Effect.gen(function* () {
@@ -231,7 +231,6 @@ export const layer: Layer.Layer<
    const plugin = yield* Plugin.Service
    const processors = yield* SessionProcessor.Service
    const provider = yield* Provider.Service
-    const sync = yield* SyncEvent.Service

    const isOverflow = Effect.fn("SessionCompaction.isOverflow")(function* (input: {
      tokens: MessageV2.Assistant["tokens"]
@@ -570,14 +569,12 @@ export const layer: Layer.Layer<
            parts: [],
          },
        )
-        if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-          yield* sync.run(SessionEvent.Compaction.Ended.Sync, {
-            sessionID: input.sessionID,
-            timestamp: DateTime.makeUnsafe(Date.now()),
-            text: summary ?? "",
-            include: selected.tail_start_id,
-          })
-        }
+        EventV2.run(SessionEvent.Compaction.Ended.Sync, {
+          sessionID: input.sessionID,
+          timestamp: DateTime.makeUnsafe(Date.now()),
+          text: summary ?? "",
+          include: selected.tail_start_id,
+        })
        yield* bus.publish(Event.Compacted, { sessionID: input.sessionID })
      }
      return result
@@ -606,13 +603,11 @@ export const layer: Layer.Layer<
        auto: input.auto,
        overflow: input.overflow,
      })
-      if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-        yield* sync.run(SessionEvent.Compaction.Started.Sync, {
-          sessionID: input.sessionID,
-          timestamp: DateTime.makeUnsafe(Date.now()),
-          reason: input.auto ? "auto" : "manual",
-        })
-      }
+      EventV2.run(SessionEvent.Compaction.Started.Sync, {
+        sessionID: input.sessionID,
+        timestamp: DateTime.makeUnsafe(Date.now()),
+        reason: input.auto ? "auto" : "manual",
+      })
    })

    return Service.of({
@@ -633,7 +628,6 @@ export const defaultLayer = Layer.suspend(() =>
    Layer.provide(Plugin.defaultLayer),
    Layer.provide(Bus.layer),
    Layer.provide(Config.defaultLayer),
-    Layer.provide(SyncEvent.defaultLayer),
  ),
 )

@@ -647,4 +641,15 @@ export async function prune(input: { sessionID: SessionID }) {
  return runPromise((svc) => svc.prune(input))
 }

+export const create = fn(
+  z.object({
+    sessionID: SessionID.zod,
+    agent: z.string(),
+    model: z.object({ providerID: ProviderID.zod, modelID: ModelID.zod }),
+    auto: z.boolean(),
+    overflow: z.boolean().optional(),
+  }),
+  (input) => runPromise((svc) => svc.create(input)),
+)
+
 export * as SessionCompaction from "./compaction"
@@ -21,11 +21,10 @@ import { Question } from "@/question"
 import { errorMessage } from "@/util/error"
 import * as Log from "@opencode-ai/core/util/log"
 import { isRecord } from "@/util/record"
-import { SyncEvent } from "@/sync"
+import { EventV2 } from "@/v2/event"
 import { SessionEvent } from "@/v2/session-event"
 import { Modelv2 } from "@/v2/model"
 import * as DateTime from "effect/DateTime"
-import { Flag } from "@opencode-ai/core/flag/flag"

 const DOOM_LOOP_THRESHOLD = 3
 const log = Log.create({ service: "session.processor" })
@@ -97,7 +96,6 @@ export const layer: Layer.Layer<
  | Image.Service
  | SessionSummary.Service
  | SessionStatus.Service
-  | SyncEvent.Service
 > = Layer.effect(
  Service,
  Effect.gen(function* () {
@@ -113,7 +111,6 @@ export const layer: Layer.Layer<
    const scope = yield* Scope.Scope
    const status = yield* SessionStatus.Service
    const image = yield* Image.Service
-    const sync = yield* SyncEvent.Service

    const create = Effect.fn("SessionProcessor.create")(function* (input: Input) {
      // Pre-capture snapshot before the LLM stream starts. The AI SDK
@@ -232,13 +229,11 @@ export const layer: Layer.Layer<
          case "reasoning-start":
            if (value.id in ctx.reasoningMap) return
            // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-            if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-              yield* sync.run(SessionEvent.Reasoning.Started.Sync, {
-                sessionID: ctx.sessionID,
-                reasoningID: value.id,
-                timestamp: DateTime.makeUnsafe(Date.now()),
-              })
-            }
+            EventV2.run(SessionEvent.Reasoning.Started.Sync, {
+              sessionID: ctx.sessionID,
+              reasoningID: value.id,
+              timestamp: DateTime.makeUnsafe(Date.now()),
+            })
            ctx.reasoningMap[value.id] = {
              id: PartID.ascending(),
              messageID: ctx.assistantMessage.id,
@@ -267,14 +262,12 @@ export const layer: Layer.Layer<
          case "reasoning-end":
            if (!(value.id in ctx.reasoningMap)) return
            // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-            if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-              yield* sync.run(SessionEvent.Reasoning.Ended.Sync, {
-                sessionID: ctx.sessionID,
-                reasoningID: value.id,
-                text: ctx.reasoningMap[value.id].text,
-                timestamp: DateTime.makeUnsafe(Date.now()),
-              })
-            }
+            EventV2.run(SessionEvent.Reasoning.Ended.Sync, {
+              sessionID: ctx.sessionID,
+              reasoningID: value.id,
+              text: ctx.reasoningMap[value.id].text,
+              timestamp: DateTime.makeUnsafe(Date.now()),
+            })
            // oxlint-disable-next-line no-self-assign -- reactivity trigger
            ctx.reasoningMap[value.id].text = ctx.reasoningMap[value.id].text
            ctx.reasoningMap[value.id].time = { ...ctx.reasoningMap[value.id].time, end: Date.now() }
@@ -288,14 +281,12 @@ export const layer: Layer.Layer<
              throw new Error(`Tool call not allowed while generating summary: ${value.toolName}`)
            }
            // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-            if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-              yield* sync.run(SessionEvent.Tool.Input.Started.Sync, {
-                sessionID: ctx.sessionID,
-                callID: value.id,
-                name: value.toolName,
-                timestamp: DateTime.makeUnsafe(Date.now()),
-              })
-            }
+            EventV2.run(SessionEvent.Tool.Input.Started.Sync, {
+              sessionID: ctx.sessionID,
+              callID: value.id,
+              name: value.toolName,
+              timestamp: DateTime.makeUnsafe(Date.now()),
+            })
            const part = yield* session.updatePart({
              id: ctx.toolcalls[value.id]?.partID ?? PartID.ascending(),
              messageID: ctx.assistantMessage.id,
@@ -319,14 +310,12 @@ export const layer: Layer.Layer<

          case "tool-input-end": {
            // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-            if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-              yield* sync.run(SessionEvent.Tool.Input.Ended.Sync, {
-                sessionID: ctx.sessionID,
-                callID: value.id,
-                text: "",
-                timestamp: DateTime.makeUnsafe(Date.now()),
-              })
-            }
+            EventV2.run(SessionEvent.Tool.Input.Ended.Sync, {
+              sessionID: ctx.sessionID,
+              callID: value.id,
+              text: "",
+              timestamp: DateTime.makeUnsafe(Date.now()),
+            })
            return
          }

@@ -336,19 +325,17 @@ export const layer: Layer.Layer<
            }
            const toolCall = yield* readToolCall(value.toolCallId)
            // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-            if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-              yield* sync.run(SessionEvent.Tool.Called.Sync, {
-                sessionID: ctx.sessionID,
-                callID: value.toolCallId,
-                tool: value.toolName,
-                input: value.input,
-                provider: {
-                  executed: toolCall?.part.metadata?.providerExecuted === true,
-                  ...(value.providerMetadata ? { metadata: value.providerMetadata } : {}),
-                },
-                timestamp: DateTime.makeUnsafe(Date.now()),
-              })
-            }
+            EventV2.run(SessionEvent.Tool.Called.Sync, {
+              sessionID: ctx.sessionID,
+              callID: value.toolCallId,
+              tool: value.toolName,
+              input: value.input,
+              provider: {
+                executed: toolCall?.part.metadata?.providerExecuted === true,
+                ...(value.providerMetadata ? { metadata: value.providerMetadata } : {}),
+              },
+              timestamp: DateTime.makeUnsafe(Date.now()),
+            })
            yield* updateToolCall(value.toolCallId, (match) => ({
              ...match,
              tool: value.toolName,
@@ -418,29 +405,27 @@ export const layer: Layer.Layer<
              attachments: attachments?.length ? attachments : undefined,
            }
            // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-            if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-              yield* sync.run(SessionEvent.Tool.Success.Sync, {
-                sessionID: ctx.sessionID,
-                callID: value.toolCallId,
-                structured: output.metadata,
-                content: [
-                  {
-                    type: "text",
-                    text: output.output,
-                  },
-                  ...(output.attachments?.map((item: MessageV2.FilePart) => ({
-                    type: "file",
-                    uri: item.url,
-                    mime: item.mime,
-                    name: item.filename,
-                  })) ?? []),
-                ],
-                provider: {
-                  executed: toolCall?.part.metadata?.providerExecuted === true,
+            EventV2.run(SessionEvent.Tool.Success.Sync, {
+              sessionID: ctx.sessionID,
+              callID: value.toolCallId,
+              structured: output.metadata,
+              content: [
+                {
+                  type: "text",
+                  text: output.output,
                },
-                timestamp: DateTime.makeUnsafe(Date.now()),
-              })
-            }
+                ...(output.attachments?.map((item: MessageV2.FilePart) => ({
+                  type: "file",
+                  uri: item.url,
+                  mime: item.mime,
+                  name: item.filename,
+                })) ?? []),
+              ],
+              provider: {
+                executed: toolCall?.part.metadata?.providerExecuted === true,
+              },
+              timestamp: DateTime.makeUnsafe(Date.now()),
+            })
            yield* completeToolCall(value.toolCallId, output)
            return
          }
@@ -448,20 +433,18 @@ export const layer: Layer.Layer<
          case "tool-error": {
            const toolCall = yield* readToolCall(value.toolCallId)
            // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-            if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-              yield* sync.run(SessionEvent.Tool.Failed.Sync, {
-                sessionID: ctx.sessionID,
-                callID: value.toolCallId,
-                error: {
-                  type: "unknown",
-                  message: errorMessage(value.error),
-                },
-                provider: {
-                  executed: toolCall?.part.metadata?.providerExecuted === true,
-                },
-                timestamp: DateTime.makeUnsafe(Date.now()),
-              })
-            }
+            EventV2.run(SessionEvent.Tool.Failed.Sync, {
+              sessionID: ctx.sessionID,
+              callID: value.toolCallId,
+              error: {
+                type: "unknown",
+                message: errorMessage(value.error),
+              },
+              provider: {
+                executed: toolCall?.part.metadata?.providerExecuted === true,
+              },
+              timestamp: DateTime.makeUnsafe(Date.now()),
+            })
            yield* failToolCall(value.toolCallId, value.error)
            return
          }
@@ -473,19 +456,17 @@ export const layer: Layer.Layer<
            if (!ctx.snapshot) ctx.snapshot = yield* snapshot.track()
            if (!ctx.assistantMessage.summary) {
              // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-              if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-                yield* sync.run(SessionEvent.Step.Started.Sync, {
-                  sessionID: ctx.sessionID,
-                  agent: input.assistantMessage.agent,
-                  model: {
-                    id: Modelv2.ID.make(ctx.model.id),
-                    providerID: Modelv2.ProviderID.make(ctx.model.providerID),
-                    variant: Modelv2.VariantID.make(input.assistantMessage.variant ?? "default"),
-                  },
-                  snapshot: ctx.snapshot,
-                  timestamp: DateTime.makeUnsafe(Date.now()),
-                })
-              }
+              EventV2.run(SessionEvent.Step.Started.Sync, {
+                sessionID: ctx.sessionID,
+                agent: input.assistantMessage.agent,
+                model: {
+                  id: Modelv2.ID.make(ctx.model.id),
+                  providerID: Modelv2.ProviderID.make(ctx.model.providerID),
+                  variant: Modelv2.VariantID.make(input.assistantMessage.variant ?? "default"),
+                },
+                snapshot: ctx.snapshot,
+                timestamp: DateTime.makeUnsafe(Date.now()),
+              })
            }
            yield* session.updatePart({
              id: PartID.ascending(),
@@ -505,16 +486,14 @@ export const layer: Layer.Layer<
            })
            if (!ctx.assistantMessage.summary) {
              // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-              if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-                yield* sync.run(SessionEvent.Step.Ended.Sync, {
-                  sessionID: ctx.sessionID,
-                  finish: value.finishReason,
-                  cost: usage.cost,
-                  tokens: usage.tokens,
-                  snapshot: completedSnapshot,
-                  timestamp: DateTime.makeUnsafe(Date.now()),
-                })
-              }
+              EventV2.run(SessionEvent.Step.Ended.Sync, {
+                sessionID: ctx.sessionID,
+                finish: value.finishReason,
+                cost: usage.cost,
+                tokens: usage.tokens,
+                snapshot: completedSnapshot,
+                timestamp: DateTime.makeUnsafe(Date.now()),
+              })
            }
            ctx.assistantMessage.finish = value.finishReason
            ctx.assistantMessage.cost += usage.cost
@@ -562,12 +541,10 @@ export const layer: Layer.Layer<
          case "text-start":
            if (!ctx.assistantMessage.summary) {
              // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-              if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-                yield* sync.run(SessionEvent.Text.Started.Sync, {
-                  sessionID: ctx.sessionID,
-                  timestamp: DateTime.makeUnsafe(Date.now()),
-                })
-              }
+              EventV2.run(SessionEvent.Text.Started.Sync, {
+                sessionID: ctx.sessionID,
+                timestamp: DateTime.makeUnsafe(Date.now()),
+              })
            }
            ctx.currentText = {
              id: PartID.ascending(),
@@ -609,13 +586,11 @@ export const layer: Layer.Layer<
            )).text
            if (!ctx.assistantMessage.summary) {
              // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-              if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-                yield* sync.run(SessionEvent.Text.Ended.Sync, {
-                  sessionID: ctx.sessionID,
-                  text: ctx.currentText.text,
-                  timestamp: DateTime.makeUnsafe(Date.now()),
-                })
-              }
+              EventV2.run(SessionEvent.Text.Ended.Sync, {
+                sessionID: ctx.sessionID,
+                text: ctx.currentText.text,
+                timestamp: DateTime.makeUnsafe(Date.now()),
+              })
            }
            {
              const end = Date.now()
@@ -705,16 +680,14 @@ export const layer: Layer.Layer<
        }
        if (!ctx.assistantMessage.summary) {
          // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-          if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-            yield* sync.run(SessionEvent.Step.Failed.Sync, {
-              sessionID: ctx.sessionID,
-              error: {
-                type: "unknown",
-                message: errorMessage(e),
-              },
-              timestamp: DateTime.makeUnsafe(Date.now()),
-            })
-          }
+          EventV2.run(SessionEvent.Step.Failed.Sync, {
+            sessionID: ctx.sessionID,
+            error: {
+              type: "unknown",
+              message: errorMessage(e),
+            },
+            timestamp: DateTime.makeUnsafe(Date.now()),
+          })
        }
        ctx.assistantMessage.error = error
        yield* bus.publish(Session.Event.Error, {
@@ -759,28 +732,22 @@ export const layer: Layer.Layer<
                parse,
                set: (info) => {
                  // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-                  const event = Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM
-                    ? sync.run(SessionEvent.Retried.Sync, {
-                        sessionID: ctx.sessionID,
-                        attempt: info.attempt,
-                        error: {
-                          message: info.message,
-                          isRetryable: true,
-                        },
-                        timestamp: DateTime.makeUnsafe(Date.now()),
-                      })
-                    : Effect.void
-                  return event.pipe(
-                    Effect.andThen(
-                      status.set(ctx.sessionID, {
-                        type: "retry",
-                        attempt: info.attempt,
-                        message: info.message,
-                        action: info.action,
-                        next: info.next,
-                      }),
-                    ),
-                  )
+                  EventV2.run(SessionEvent.Retried.Sync, {
+                    sessionID: ctx.sessionID,
+                    attempt: info.attempt,
+                    error: {
+                      message: info.message,
+                      isRetryable: true,
+                    },
+                    timestamp: DateTime.makeUnsafe(Date.now()),
+                  })
+                  return status.set(ctx.sessionID, {
+                    type: "retry",
+                    attempt: info.attempt,
+                    message: info.message,
+                    action: info.action,
+                    next: info.next,
+                  })
                },
              }),
            ),
@@ -821,7 +788,6 @@ export const defaultLayer = Layer.suspend(() =>
    Layer.provide(Image.defaultLayer),
    Layer.provide(Bus.layer),
    Layer.provide(Config.defaultLayer),
-    Layer.provide(SyncEvent.defaultLayer),
  ),
 )

@@ -43,6 +43,7 @@ import { Shell } from "@/shell/shell"
 import { ShellID } from "@/tool/shell/id"
 import { AppFileSystem } from "@opencode-ai/core/filesystem"
 import { Truncate } from "@/tool/truncate"
+import { Image } from "@/image/image"
 import { decodeDataUrl } from "@/util/data-url"
 import { Process } from "@/util/process"
 import { Cause, Effect, Exit, Latch, Layer, Option, Scope, Context, Schema, Types } from "effect"
@@ -53,7 +54,7 @@ import { InstanceState } from "@/effect/instance-state"
 import { TaskTool, type TaskPromptOps } from "@/tool/task"
 import { SessionRunState } from "./run-state"
 import { EffectBridge } from "@/effect/bridge"
-import { SyncEvent } from "@/sync"
+import { EventV2 } from "@/v2/event"
 import { SessionEvent } from "@/v2/session-event"
 import { Modelv2 } from "@/v2/model"
 import { AgentAttachment, FileAttachment, Source } from "@/v2/session-prompt"
@@ -80,10 +81,10 @@ const elog = EffectLogger.create({ service: "session.prompt" })

 export interface Interface {
  readonly cancel: (sessionID: SessionID) => Effect.Effect<void>
-  readonly prompt: (input: PromptInput) => Effect.Effect<MessageV2.WithParts>
+  readonly prompt: (input: PromptInput) => Effect.Effect<MessageV2.WithParts, Image.Error>
  readonly loop: (input: LoopInput) => Effect.Effect<MessageV2.WithParts>
  readonly shell: (input: ShellInput) => Effect.Effect<MessageV2.WithParts>
-  readonly command: (input: CommandInput) => Effect.Effect<MessageV2.WithParts>
+  readonly command: (input: CommandInput) => Effect.Effect<MessageV2.WithParts, Image.Error>
  readonly resolvePromptParts: (template: string) => Effect.Effect<PromptInput["parts"]>
 }

@@ -108,6 +109,7 @@ export const layer = Layer.effect(
    const lsp = yield* LSP.Service
    const registry = yield* ToolRegistry.Service
    const truncate = yield* Truncate.Service
+    const image = yield* Image.Service
    const spawner = yield* ChildProcessSpawner.ChildProcessSpawner
    const scope = yield* Scope.Scope
    const instruction = yield* Instruction.Service
@@ -116,7 +118,6 @@ export const layer = Layer.effect(
    const summary = yield* SessionSummary.Service
    const sys = yield* SystemPrompt.Service
    const llm = yield* LLM.Service
-    const sync = yield* SyncEvent.Service
    const runner = Effect.fn("SessionPrompt.runner")(function* () {
      return yield* EffectBridge.make()
    })
@@ -124,7 +125,7 @@ export const layer = Layer.effect(
      return {
        cancel: (sessionID: SessionID) => cancel(sessionID),
        resolvePromptParts: (template: string) => resolvePromptParts(template),
-        prompt: (input: PromptInput) => prompt(input),
+        prompt: (input: PromptInput) => prompt(input).pipe(Effect.catch(Effect.die)),
      } satisfies TaskPromptOps
    })

@@ -757,7 +758,7 @@ NOTE: At any point in time through this workflow you should feel free to ask the
              yield* bus.publish(Session.Event.Error, { sessionID: input.sessionID, error: error.toObject() })
              throw error
            }
-            const model = input.model ?? agent.model ?? (yield* currentModel(input.sessionID))
+            const model = input.model ?? agent.model ?? (yield* lastModel(input.sessionID))
            const userMsg: MessageV2.User = {
              id: input.messageID ?? MessageID.ascending(),
              sessionID: input.sessionID,
@@ -808,14 +809,12 @@ NOTE: At any point in time through this workflow you should feel free to ask the
              },
            }
            yield* sessions.updatePart(part)
-            if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-              yield* sync.run(SessionEvent.Shell.Started.Sync, {
-                sessionID: input.sessionID,
-                timestamp: DateTime.makeUnsafe(started),
-                callID,
-                command: input.command,
-              })
-            }
+            EventV2.run(SessionEvent.Shell.Started.Sync, {
+              sessionID: input.sessionID,
+              timestamp: DateTime.makeUnsafe(started),
+              callID,
+              command: input.command,
+            })
            return { msg, part, cwd: ctx.directory }
          }).pipe(Effect.ensuring(markReady))

@@ -831,14 +830,12 @@ NOTE: At any point in time through this workflow you should feel free to ask the
                output += "\n\n" + ["<metadata>", "User aborted the command", "</metadata>"].join("\n")
              }
              const completed = Date.now()
-              if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-                yield* sync.run(SessionEvent.Shell.Ended.Sync, {
-                  sessionID: input.sessionID,
-                  timestamp: DateTime.makeUnsafe(completed),
-                  callID: part.callID,
-                  output,
-                })
-              }
+              EventV2.run(SessionEvent.Shell.Ended.Sync, {
+                sessionID: input.sessionID,
+                timestamp: DateTime.makeUnsafe(completed),
+                callID: part.callID,
+                output,
+              })
              if (!msg.time.completed) {
                msg.time.completed = completed
                yield* sessions.updateMessage(msg)
@@ -919,17 +916,7 @@ NOTE: At any point in time through this workflow you should feel free to ask the
      return yield* Effect.failCause(exit.cause)
    })

-    const currentModel = Effect.fnUntraced(function* (sessionID: SessionID) {
-      const current = Database.use((db) =>
-        db.select({ model: SessionTable.model }).from(SessionTable).where(eq(SessionTable.id, sessionID)).get(),
-      )
-      if (current?.model) {
-        return {
-          providerID: ProviderID.make(current.model.providerID),
-          modelID: ModelID.make(current.model.id),
-          ...(current.model.variant && current.model.variant !== "default" ? { variant: current.model.variant } : {}),
-        }
-      }
+    const lastModel = Effect.fnUntraced(function* (sessionID: SessionID) {
      const match = yield* sessions.findMessage(sessionID, (m) => m.info.role === "user" && !!m.info.model)
      if (Option.isSome(match) && match.value.info.role === "user") return match.value.info.model
      return yield* provider.defaultModel()
@@ -946,14 +933,7 @@ NOTE: At any point in time through this workflow you should feel free to ask the
        throw error
      }

-      const current = Database.use((db) =>
-        db
-          .select({ agent: SessionTable.agent, model: SessionTable.model })
-          .from(SessionTable)
-          .where(eq(SessionTable.id, input.sessionID))
-          .get(),
-      )
-      const model = input.model ?? ag.model ?? (yield* currentModel(input.sessionID))
+      const model = input.model ?? ag.model ?? (yield* lastModel(input.sessionID))
      const same = ag.model && model.providerID === ag.model.providerID && model.modelID === ag.model.modelID
      const full =
        !input.variant && ag.variant && same
@@ -977,8 +957,15 @@ NOTE: At any point in time through this workflow you should feel free to ask the
        format: input.format,
      }

+      const current = Database.use((db) =>
+        db
+          .select({ agent: SessionTable.agent, model: SessionTable.model })
+          .from(SessionTable)
+          .where(eq(SessionTable.id, input.sessionID))
+          .get(),
+      )
      if (current?.agent !== info.agent) {
-        yield* sync.run(SessionEvent.AgentSwitched.Sync, {
+        EventV2.run(SessionEvent.AgentSwitched.Sync, {
          sessionID: input.sessionID,
          timestamp: DateTime.makeUnsafe(info.time.created),
          agent: info.agent,
@@ -987,9 +974,9 @@ NOTE: At any point in time through this workflow you should feel free to ask the
      if (
        current?.model?.providerID !== info.model.providerID ||
        current.model.id !== info.model.modelID ||
-        (current.model.variant === "default" ? undefined : current.model.variant) !== info.model.variant
+        current.model.variant !== info.model.variant
      ) {
-        yield* sync.run(SessionEvent.ModelSwitched.Sync, {
+        EventV2.run(SessionEvent.ModelSwitched.Sync, {
          sessionID: input.sessionID,
          timestamp: DateTime.makeUnsafe(info.time.created),
          model: {
@@ -1290,7 +1277,9 @@ NOTE: At any point in time through this workflow you should feel free to ask the
        { message: info, parts: resolvedParts },
      )

-      const parts = resolvedParts
+      const parts = yield* Effect.forEach(resolvedParts, (part) =>
+        part.type === "file" && part.mime.startsWith("image/") ? image.normalize(part) : Effect.succeed(part),
+      )

      const parsed = MessageV2.Info.zod.safeParse(info)
      if (!parsed.success) {
@@ -1364,51 +1353,47 @@ NOTE: At any point in time through this workflow you should feel free to ask the
        },
      )
      // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-      if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-        yield* sync.run(SessionEvent.Prompted.Sync, {
-          sessionID: input.sessionID,
-          timestamp: DateTime.makeUnsafe(info.time.created),
-          prompt: {
-            text: nextPrompt.text.join("\n"),
-            files: nextPrompt.files,
-            agents: nextPrompt.agents,
-          },
-        })
-      }
+      EventV2.run(SessionEvent.Prompted.Sync, {
+        sessionID: input.sessionID,
+        timestamp: DateTime.makeUnsafe(info.time.created),
+        prompt: {
+          text: nextPrompt.text.join("\n"),
+          files: nextPrompt.files,
+          agents: nextPrompt.agents,
+        },
+      })
      for (const text of nextPrompt.synthetic) {
        // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-        if (Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) {
-          yield* sync.run(SessionEvent.Synthetic.Sync, {
-            sessionID: input.sessionID,
-            timestamp: DateTime.makeUnsafe(info.time.created),
-            text,
-          })
-        }
+        EventV2.run(SessionEvent.Synthetic.Sync, {
+          sessionID: input.sessionID,
+          timestamp: DateTime.makeUnsafe(info.time.created),
+          text,
+        })
      }

      return { info, parts }
    }, Effect.scoped)

-    const prompt: (input: PromptInput) => Effect.Effect<MessageV2.WithParts> = Effect.fn("SessionPrompt.prompt")(
-      function* (input: PromptInput) {
-        const session = yield* sessions.get(input.sessionID).pipe(Effect.orDie)
-        yield* revert.cleanup(session)
-        const message = yield* createUserMessage(input)
-        yield* sessions.touch(input.sessionID)
+    const prompt: (input: PromptInput) => Effect.Effect<MessageV2.WithParts, Image.Error> = Effect.fn(
+      "SessionPrompt.prompt",
+    )(function* (input: PromptInput) {
+      const session = yield* sessions.get(input.sessionID).pipe(Effect.orDie)
+      yield* revert.cleanup(session)
+      const message = yield* createUserMessage(input)
+      yield* sessions.touch(input.sessionID)

-        const permissions: Permission.Ruleset = []
-        for (const [t, enabled] of Object.entries(input.tools ?? {})) {
-          permissions.push({ permission: t, action: enabled ? "allow" : "deny", pattern: "*" })
-        }
-        if (permissions.length > 0) {
-          session.permission = permissions
-          yield* sessions.setPermission({ sessionID: session.id, permission: permissions })
-        }
+      const permissions: Permission.Ruleset = []
+      for (const [t, enabled] of Object.entries(input.tools ?? {})) {
+        permissions.push({ permission: t, action: enabled ? "allow" : "deny", pattern: "*" })
+      }
+      if (permissions.length > 0) {
+        session.permission = permissions
+        yield* sessions.setPermission({ sessionID: session.id, permission: permissions })
+      }

-        if (input.noReply === true) return message
-        return yield* loop({ sessionID: input.sessionID })
-      },
-    )
+      if (input.noReply === true) return message
+      return yield* loop({ sessionID: input.sessionID })
+    })

    const lastAssistant = Effect.fnUntraced(function* (sessionID: SessionID) {
      const match = yield* sessions.findMessage(sessionID, (m) => m.info.role !== "user")
@@ -1719,7 +1704,7 @@ NOTE: At any point in time through this workflow you should feel free to ask the
          if (cmdAgent?.model) return cmdAgent.model
        }
        if (input.model) return Provider.parseModel(input.model)
-        return yield* currentModel(input.sessionID)
+        return yield* lastModel(input.sessionID)
      })

      yield* getModel(taskModel.providerID, taskModel.modelID, input.sessionID)
@@ -1752,7 +1737,7 @@ NOTE: At any point in time through this workflow you should feel free to ask the
      const userModel = isSubtask
        ? input.model
          ? Provider.parseModel(input.model)
-          : yield* currentModel(input.sessionID)
+          : yield* lastModel(input.sessionID)
        : taskModel

      yield* plugin.trigger(
@@ -1809,6 +1794,7 @@ export const defaultLayer = Layer.suspend(() =>
    Layer.provide(Session.defaultLayer),
    Layer.provide(SessionRevert.defaultLayer),
    Layer.provide(SessionSummary.defaultLayer),
+    Layer.provide(Image.defaultLayer),
    Layer.provide(
      Layer.mergeAll(
        Agent.defaultLayer,
@@ -1816,7 +1802,6 @@ export const defaultLayer = Layer.suspend(() =>
        LLM.defaultLayer,
        Bus.layer,
        CrossSpawnSpawner.defaultLayer,
-        SyncEvent.defaultLayer,
      ),
    ),
  ),
@@ -1,12 +1,14 @@
 import { Schema } from "effect"

 import { Identifier } from "@/id/id"
+import { zod } from "@opencode-ai/core/effect-zod"
 import { withStatics } from "@opencode-ai/core/schema"

 export const SessionID = Schema.String.check(Schema.isStartsWith("ses")).pipe(
  Schema.brand("SessionID"),
  withStatics((s) => ({
    descending: (id?: string) => s.make(Identifier.descending("session", id)),
+    zod: zod(s),
  })),
 )

@@ -16,6 +18,7 @@ export const MessageID = Schema.String.check(Schema.isStartsWith("msg")).pipe(
  Schema.brand("MessageID"),
  withStatics((s) => ({
    ascending: (id?: string) => s.make(Identifier.ascending("message", id)),
+    zod: zod(s),
  })),
 )

@@ -25,6 +28,7 @@ export const PartID = Schema.String.check(Schema.isStartsWith("prt")).pipe(
  Schema.brand("PartID"),
  withStatics((s) => ({
    ascending: (id?: string) => s.make(Identifier.ascending("part", id)),
+    zod: zod(s),
  })),
 )

@@ -1,6 +1,7 @@
 import { Identifier } from "@/id/id"
 import { SyncEvent } from "@/sync"
 import { withStatics } from "@opencode-ai/core/schema"
+import { Flag } from "@opencode-ai/core/flag/flag"
 import * as Schema from "effect/Schema"

 export const ID = Schema.String.pipe(
@@ -40,4 +41,13 @@ export function define<const Type extends string, Fields extends Schema.Struct.F
  })
 }

+export function run<Def extends SyncEvent.Definition>(
+  def: Def,
+  data: SyncEvent.Event<Def>["data"],
+  options?: { publish?: boolean },
+) {
+  if (!Flag.OPENCODE_EXPERIMENTAL_EVENT_SYSTEM) return
+  SyncEvent.run(def, data, options)
+}
+
 export * as EventV2 from "./event"
@@ -12,7 +12,6 @@ import { SessionEvent } from "./session-event"
 import { V2Schema } from "./schema"
 import { optionalOmitUndefined } from "@opencode-ai/core/schema"
 import { Modelv2 } from "./model"
-import { SyncEvent } from "@/sync"

 export const Delivery = Schema.Literals(["immediate", "deferred"]).annotate({
  identifier: "Session.Delivery",
@@ -114,7 +113,6 @@ export class Service extends Context.Service<Service, Interface>()("@opencode/v2
 export const layer = Layer.effect(
  Service,
  Effect.gen(function* () {
-    const sync = yield* SyncEvent.Service
    const decodeMessage = Schema.decodeUnknownSync(SessionMessage.Message)

    const decode = (row: typeof SessionMessageTable.$inferSelect) =>
@@ -271,14 +269,14 @@ export const layer = Layer.effect(
      shell: Effect.fn("V2Session.shell")(function* (_input) {}),
      skill: Effect.fn("V2Session.skill")(function* (_input) {}),
      switchAgent: Effect.fn("V2Session.switchAgent")(function* (input) {
-        yield* sync.run(SessionEvent.AgentSwitched.Sync, {
+        EventV2.run(SessionEvent.AgentSwitched.Sync, {
          sessionID: input.sessionID,
          timestamp: DateTime.makeUnsafe(Date.now()),
          agent: input.agent,
        })
      }),
      switchModel: Effect.fn("V2Session.switchModel")(function* (input) {
-        yield* sync.run(SessionEvent.ModelSwitched.Sync, {
+        EventV2.run(SessionEvent.ModelSwitched.Sync, {
          sessionID: input.sessionID,
          timestamp: DateTime.makeUnsafe(Date.now()),
          model: input.model,
@@ -313,6 +311,6 @@ export const layer = Layer.effect(
  }),
 )

-export const defaultLayer = layer.pipe(Layer.provide(SyncEvent.defaultLayer))
+export const defaultLayer = layer

 export * as SessionV2 from "./session"
@@ -3,27 +3,36 @@ import { computePromptTraits } from "../../../../src/cli/cmd/tui/component/promp

 describe("computePromptTraits", () => {
  test("normal mode without autocomplete only captures tab", () => {
-    const traits = computePromptTraits({ mode: "normal", autocompleteVisible: false })
+    const traits = computePromptTraits({ mode: "normal", disabled: false, autocompleteVisible: false })
    expect(traits.capture).toEqual(["tab"])
-    expect(traits.suspend).toBeUndefined()
+    expect(traits.suspend).toBe(false)
    expect(traits.status).toBeUndefined()
  })

  test("normal mode with autocomplete captures navigation keys", () => {
-    const traits = computePromptTraits({ mode: "normal", autocompleteVisible: true })
+    const traits = computePromptTraits({ mode: "normal", disabled: false, autocompleteVisible: true })
    expect(traits.capture).toEqual(["escape", "navigate", "submit", "tab"])
-    expect(traits.suspend).toBeUndefined()
+    expect(traits.suspend).toBe(false)
    expect(traits.status).toBeUndefined()
  })

-  test("shell mode does not write the keymap-owned suspend trait", () => {
-    const traits = computePromptTraits({ mode: "shell", autocompleteVisible: false })
-    expect(traits.suspend).toBeUndefined()
+  test("shell mode does not suspend the textarea", () => {
+    // Suspending the textarea would gate every keybinding action
+    // (backspace, delete-word-backward, arrow movement, etc.) — see
+    // @opentui/core 0.2.x TextareaRenderable.handleKeyPress. Shell mode is
+    // an active editing mode, so suspend must stay off.
+    const traits = computePromptTraits({ mode: "shell", disabled: false, autocompleteVisible: false })
+    expect(traits.suspend).toBe(false)
  })

  test("shell mode disables capture and labels the prompt", () => {
-    const traits = computePromptTraits({ mode: "shell", autocompleteVisible: false })
+    const traits = computePromptTraits({ mode: "shell", disabled: false, autocompleteVisible: false })
    expect(traits.capture).toBeUndefined()
    expect(traits.status).toBe("SHELL")
  })
+
+  test("disabled suspends regardless of mode", () => {
+    expect(computePromptTraits({ mode: "normal", disabled: true, autocompleteVisible: false }).suspend).toBe(true)
+    expect(computePromptTraits({ mode: "shell", disabled: true, autocompleteVisible: false }).suspend).toBe(true)
+  })
 })
@@ -24,7 +24,6 @@ import { CrossSpawnSpawner } from "@opencode-ai/core/cross-spawn-spawner"
 import { provideTmpdirServer } from "../fixture/fixture"
 import { testEffect } from "../lib/effect"
 import { raw, reply, TestLLMServer } from "../lib/llm-server"
-import { SyncEvent } from "@/sync"

 void Log.init({ print: false })

@@ -167,7 +166,6 @@ const deps = Layer.mergeAll(
  LLM.defaultLayer,
  Provider.defaultLayer,
  status,
-  SyncEvent.defaultLayer,
 ).pipe(Layer.provideMerge(infra))
 const env = Layer.mergeAll(
  TestLLMServer.layer,
@@ -50,7 +50,6 @@ import { Reference } from "../../src/reference/reference"
 import { provideTmpdirInstance, provideTmpdirServer } from "../fixture/fixture"
 import { testEffect } from "../lib/effect"
 import { reply, TestLLMServer } from "../lib/llm-server"
-import { SyncEvent } from "@/sync"

 void Log.init({ print: false })

@@ -175,7 +174,6 @@ function makeHttp() {
    mcp,
    AppFileSystem.defaultLayer,
    status,
-    SyncEvent.defaultLayer,
  ).pipe(Layer.provideMerge(infra))
  const question = Question.layer.pipe(Layer.provideMerge(deps))
  const todo = Todo.layer.pipe(Layer.provideMerge(deps))
@@ -15,20 +15,20 @@ import { WorkspaceID } from "../../src/control-plane/schema"
 // schema we assert:
 //   1. The Effect decoder (`Schema.decodeUnknownSync`) accepts valid input.
 //   2. The derived Zod (`X.zod.parse`) accepts the same input and returns the
-//      same shape for schemas that still expose Zod statics.
-//   3. Clearly-invalid input is rejected by both paths where both exist.
+//      same shape.
+//   3. Clearly-invalid input is rejected by both paths.
 //
 // The point is to lock down the Schema <-> Zod bridge so a future edit to
 // any input schema can't silently drop or widen a field on one side.

 // Representative valid IDs — the branded schemas require the right prefix
 // (see src/id/id.ts).
-const sessionID = Schema.decodeUnknownSync(SessionID)("ses_01J5Y5H0AH4Q4NXJ6P4C3P5V2K")
-const sessionIDChild = Schema.decodeUnknownSync(SessionID)("ses_01J5Y5H0AH4Q4NXJ6P4C3P5V2L")
-const messageID = Schema.decodeUnknownSync(MessageID)("msg_01J5Y5H0AH4Q4NXJ6P4C3P5V2M")
-const partID = Schema.decodeUnknownSync(PartID)("prt_01J5Y5H0AH4Q4NXJ6P4C3P5V2N")
+const sessionID = SessionID.zod.parse("ses_01J5Y5H0AH4Q4NXJ6P4C3P5V2K")
+const sessionIDChild = SessionID.zod.parse("ses_01J5Y5H0AH4Q4NXJ6P4C3P5V2L")
+const messageID = MessageID.zod.parse("msg_01J5Y5H0AH4Q4NXJ6P4C3P5V2M")
+const partID = PartID.zod.parse("prt_01J5Y5H0AH4Q4NXJ6P4C3P5V2N")
 const projectID = ProjectID.zod.parse("proj-alpha")
-const workspaceID = Schema.decodeUnknownSync(WorkspaceID)("wrk-primary")
+const workspaceID = WorkspaceID.zod.parse("wrk-primary")

 function decodeUnknown<S extends Schema.Top>(schema: S) {
  const decode = Schema.decodeUnknownSync(schema as any)
@@ -58,7 +58,6 @@ import { CrossSpawnSpawner } from "@opencode-ai/core/cross-spawn-spawner"
 import { Ripgrep } from "../../src/file/ripgrep"
 import { Format } from "../../src/format"
 import { Reference } from "../../src/reference/reference"
-import { SyncEvent } from "@/sync"

 void Log.init({ print: false })

@@ -125,7 +124,6 @@ function makeHttp() {
    mcp,
    AppFileSystem.defaultLayer,
    status,
-    SyncEvent.defaultLayer,
  ).pipe(Layer.provideMerge(infra))
  const question = Question.layer.pipe(Layer.provideMerge(deps))
  const todo = Todo.layer.pipe(Layer.provideMerge(deps))
@@ -1,7 +1,7 @@
 {
  "$schema": "https://json.schemastore.org/package.json",
  "name": "@opencode-ai/plugin",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "type": "module",
  "license": "MIT",
  "scripts": {
@@ -1,7 +1,7 @@
 {
  "$schema": "https://json.schemastore.org/package.json",
  "name": "@opencode-ai/sdk",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "type": "module",
  "license": "MIT",
  "scripts": {
@@ -1,6 +1,6 @@
 {
  "name": "@opencode-ai/slack",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "type": "module",
  "license": "MIT",
  "scripts": {
@@ -1,6 +1,6 @@
 {
  "name": "@opencode-ai/ui",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "type": "module",
  "license": "MIT",
  "exports": {
@@ -2,7 +2,7 @@
  "name": "@opencode-ai/web",
  "type": "module",
  "license": "MIT",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "scripts": {
    "dev": "astro dev",
    "dev:remote": "VITE_API_URL=https://api.opencode.ai astro dev",
@@ -2,7 +2,7 @@
  "name": "opencode",
  "displayName": "opencode",
  "description": "opencode for VS Code",
-  "version": "1.14.48",
+  "version": "1.14.46",
  "publisher": "sst-dev",
  "repository": {
    "type": "git",
Author	SHA1	Message	Date
Kit Langton	c20d070b9a	docs(llm): fix stale references in protocols/shared.ts and gemini.ts - subtractTokens JSDoc said the raw payload lives on Usage.native, but that field was renamed to providerMetadata earlier in this PR. - totalTokens JSDoc still described the abandoned "additive" first-pass contract where inputTokens/outputTokens were non-cached / visible only. We landed on inclusive totals; the fallback already covers cache and reasoning. - Removed a duplicate inline comment in Gemini's mapUsage — the function-level comment already explains the visible/reasoning sum and the undefined-when-incomplete rule.	2026-05-10 22:08:12 -04:00
Kit Langton	d048bd6f4b	test(llm): re-record golden scenarios against live providers Verifies the new Usage mapper code against live provider responses for OpenAI Chat, OpenAI Responses, Anthropic, Gemini, DeepSeek, and TogetherAI — 16 fresh recordings, all assertions pass. No existing cassettes were modified; these populate test slots that were previously skipped in replay mode. Recorded via: set -a; source .env.recorded.local; set +a RECORD=true bun test test/provider/*.recorded.test.ts Redactor stripped all auth headers; no secrets in the cassettes.	2026-05-10 22:03:02 -04:00
Kit Langton	ab9b79ef88	refactor(llm): rename Usage.native to providerMetadata Aligns the escape-hatch field name with `LLMEvent.providerMetadata` used elsewhere in this package (and with AI SDK / pydantic-ai / LangChain conventions for the same idea). Two parallel escape hatches having different names was a wart. The raw payload is now wrapped under the provider key — `{ openai: ... }`, `{ anthropic: ... }`, `{ google: ... }`, `{ bedrock: ... }` — using the existing `ProviderMetadata = Record<string, Record<string, unknown>>` schema rather than a flat record. Same shape as `LLMEvent.providerMetadata`, so consumers downstream can read both with the same code. Anthropic's `mergeUsage` merges the per-provider sub-record across `message_start` and `message_delta` instead of spreading at the top level.	2026-05-10 21:42:09 -04:00
Kit Langton	d4ff331052	refactor(llm): inclusive total + non-overlapping breakdown for Usage Final shape after considering ecosystem conventions: inputTokens — inclusive total (matches AI SDK / OpenAI / LangChain) outputTokens — inclusive total (includes reasoning) nonCachedInputTokens — breakdown: fresh prompt cacheReadInputTokens — breakdown: cache hit cacheWriteInputTokens — breakdown: cache write reasoningTokens — subset of outputTokens Invariant: nonCached + cacheRead + cacheWrite = inputTokens reasoningTokens <= outputTokens Why this shape: - `inputTokens` keeps its AI-SDK / OpenAI semantics, so a reader from any major ecosystem sees the number they expect. - The non-overlapping breakdown fields are populated alongside the inclusive totals — consumers read whichever they need without subtracting. This eliminates the underflow bug class (opencode#26620) structurally without diverging on naming. - Aligns with the AI SDK v3 spec proposal (vercel/ai#9921), which adds exactly this kind of non-overlapping breakdown to address the active ecosystem bugs around cache token double-counting and underflow (pydantic-ai#4364, langfuse#12306/#11979, vercel/ai#8349, langchain#32818, langchainjs#10249). Mappers: - OpenAI Chat / Responses / Bedrock: provider reports inclusive totals natively; mapper derives `nonCachedInputTokens` via `ProviderShared.subtractTokens`. - Gemini: `promptTokenCount` is inclusive; `candidatesTokenCount` is exclusive of `thoughtsTokenCount`, so mapper sums those to produce the inclusive `outputTokens`. Only computes the total when the visible component is reported (avoids fabricating an inclusive number from a partial breakdown). - Anthropic: `input_tokens` is non-cached natively; mapper sums it with cache reads/writes to produce the inclusive `inputTokens`. `output_tokens` is inclusive (Anthropic doesn't break thinking out, so `reasoningTokens` stays undefined). Added a `visibleOutputTokens` getter (clamped `outputTokens - reasoningTokens`) as the one safe escape hatch for consumers wanting the non-reasoning view. Added `ProviderShared.sumTokens` to derive an inclusive total from a non-overlapping breakdown, returning `undefined` when every input is undefined (so we don't fabricate a 0).	2026-05-10 20:39:22 -04:00
Kit Langton	f5d199db62	feat(llm): add Usage.totalInputTokens / totalOutputTokens getters Match the `LLMResponse.text` / `reasoning` / `toolCalls` getter pattern in the same file — `usage.totalInputTokens` reads naturally and lives where the Usage data does. Both sums are monotonic under the additive contract, so callers no longer need to remember which fields are non-overlapping. Test fixtures that previously asserted with `usage: { ... }` plain literals are now wrapped with `new Usage({...})` to match the runtime shape the mappers actually produce (an instance, not a struct).	2026-05-10 19:29:41 -04:00
Kit Langton	0d4f8d126f	refactor(llm): drop Usage.totalInput / totalOutput helpers The additive contract delivers value at the mapper boundary — every field is non-overlapping and non-negative, so any caller summing arbitrary subsets is correct by construction. Two-line helpers that just sum three or two known fields add API surface without paying for themselves, and there are no in-tree consumers today. If v2 wants them at integration time, the right place is a getter on the `Schema.Class` (matching the `LLMResponse.text` / `reasoning` / `toolCalls` pattern in the same file), not a static namespace helper.	2026-05-10 19:15:46 -04:00
Kit Langton	478f3ae50c	refactor(llm): trim Usage helpers + Bedrock subtraction Review pass: - Drop `Pick<>` type aliases on `Usage.totalInput` / `Usage.totalOutput` — the helpers can take `Usage` directly since every field is optional. - Collapse Bedrock's nested `subtractTokens(subtractTokens(...))` into a single subtraction against the summed cache subtotals. - Drop arithmetic-walkthrough comments in test fixtures (the raw fixture values are right next to the expected outputs). - Generalize the comment on `mapUsage` in `openai-chat.ts` so the rationale outlives the PR reference.	2026-05-10 13:22:49 -04:00
Kit Langton	b9451175a6	refactor(llm): make LLM.Usage a fully-additive contract Defines a single invariant for `LLM.Usage`: every field is non-negative and every meaningful aggregate is a sum, never a difference. Total billable input = inputTokens + cacheReadInputTokens + cacheWriteInputTokens. Total billable output = outputTokens + reasoningTokens. Adding two non-negatives cannot underflow, so consumers can no longer reproduce the underflow-then-clamp bug class fixed by #26620. Each protocol mapper now enforces the contract at the provider boundary via `ProviderShared.subtractTokens`, which clamps with `Math.max(0, …)` for defense against provider bugs: - OpenAI Chat / Responses: pull `cached_tokens` out of `prompt_tokens` / `input_tokens`; pull `reasoning_tokens` out of `completion_tokens` / `output_tokens`. The provider's `total_tokens` is preserved verbatim. - Gemini: pull `cachedContentTokenCount` out of `promptTokenCount`. Gemini already split visible candidates from thoughts. - Bedrock: pull `cacheReadInputTokens` and `cacheWriteInputTokens` out of `inputTokens`, matching AWS prompt-caching docs. - Anthropic: already non-overlapping per the Messages API; pass through. Adds `Usage.totalInput` / `Usage.totalOutput` helpers for callers that want the merged view, and a regression test covering the clamp behavior. The reasoning underflow fixed in #26620 was the most visible symptom of a broader semantic inconsistency in this package: providers also disagreed on whether `inputTokens` includes cache reads (Anthropic excluded; OpenAI/Gemini/Bedrock included), which would silently double-subtract the moment v2 wired LLM.Usage into Session.getUsage. Normalizing now, pre-integration, closes both holes in one move.	2026-05-10 13:07:58 -04:00