fix(ai): disable bedrock mantle responses websockets

Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>

.github/workflows/test.yml

@@ -42,7 +42,7 @@ jobs:
       - name: Find affected packages
         id: packages
         env:
-          TURBO_SCM_BASE: ${{ github.event.pull_request.base.sha || github.event.before }}
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
           TURBO_SCM_HEAD: ${{ github.sha }}
         run: |
           if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
@@ -71,6 +71,7 @@ jobs:
         uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
         with:
           token: ${{ secrets.GITHUB_TOKEN }}
+          fetch-depth: 0
 
       - name: Setup Node
         uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
@@ -110,9 +111,16 @@ jobs:
 
       - name: Run unit tests
         timeout-minutes: 20
-        run: GITHUB_ACTIONS=false bun turbo test
+        run: |
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            GITHUB_ACTIONS=false bun turbo test
+            exit 0
+          fi
+          GITHUB_ACTIONS=false bun turbo test --affected
         env:
           OPENCODE_EXPERIMENTAL_DISABLE_FILEWATCHER: ${{ runner.os == 'Windows' && 'true' || 'false' }}
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
+          TURBO_SCM_HEAD: ${{ github.sha }}
 
       - name: Verify published codemode package
         if: runner.os == 'Linux'
@@ -122,8 +130,15 @@ jobs:
       - name: Verify packed workerd SDK
         if: runner.os == 'Linux'
         timeout-minutes: 15
-        working-directory: packages/sdk
-        run: bun run verify:package
+        run: |
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            bun turbo verify:package --filter=@opencode-ai/sdk
+            exit 0
+          fi
+          bun turbo verify:package --affected --filter=@opencode-ai/sdk
+        env:
+          TURBO_SCM_BASE: ${{ github.event_name == 'pull_request' && format('{0}^1', github.sha) || github.event.before }}
+          TURBO_SCM_HEAD: ${{ github.sha }}
 
       - name: Verify compiled service lifecycle
         if: always()

bun.lock

@@ -346,9 +346,7 @@
         "@ai-sdk/amazon-bedrock": "4.0.112",
         "@ai-sdk/anthropic": "3.0.82",
         "@ai-sdk/azure": "3.0.88",
-        "@ai-sdk/cerebras": "2.0.41",
         "@ai-sdk/cohere": "3.0.27",
-        "@ai-sdk/deepinfra": "2.0.41",
         "@ai-sdk/gateway": "3.0.104",
         "@ai-sdk/google-vertex": "4.0.128",
         "@ai-sdk/groq": "3.0.31",
@@ -357,7 +355,6 @@
         "@ai-sdk/perplexity": "3.0.26",
         "@ai-sdk/provider": "3.0.8",
         "@ai-sdk/provider-utils": "4.0.23",
-        "@ai-sdk/togetherai": "2.0.41",
         "@ai-sdk/vercel": "2.0.39",
         "@aws-sdk/credential-providers": "3.1057.0",
         "@ff-labs/fff-bun": "0.10.5",
@@ -667,6 +664,7 @@
       "dependencies": {
         "@opencode-ai/client": "workspace:*",
         "@opencode-ai/core": "workspace:*",
+        "@opencode-ai/plugin": "workspace:*",
         "@opencode-ai/schema": "workspace:*",
         "@opencode-ai/server": "workspace:*",
         "@opencode-ai/util": "workspace:*",
@@ -973,6 +971,7 @@
       "dependencies": {
         "@effect/opentelemetry": "catalog:",
         "@effect/platform-node": "catalog:",
+        "@effect/platform-node-shared": "catalog:",
         "@npmcli/arborist": "catalog:",
         "@npmcli/config": "10.8.1",
         "@opentelemetry/api": "1.9.0",
@@ -1178,8 +1177,6 @@
 
     "@ai-sdk/deepgram": ["@ai-sdk/deepgram@2.0.52", "", { "dependencies": { "@ai-sdk/provider": "3.0.15", "@ai-sdk/provider-utils": "4.0.46" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-8pcrQvEQCbvrrQKnD6hclBbI0hUgSrgyADykRbabxv/g9vPurfMC6n23J7dD+KZ3EcCoW+qz3IUIfySJ58gBOg=="],
 
-    "@ai-sdk/deepinfra": ["@ai-sdk/deepinfra@2.0.41", "", { "dependencies": { "@ai-sdk/openai-compatible": "2.0.37", "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.21" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-y6RoOP7DGWmDSiSxrUSt5p18sbz+Ixe5lMVPmdE7x+Tr5rlrzvftyHhjWHfqlAtoYERZTGFbP6tPW1OfQcrb4A=="],
-
     "@ai-sdk/deepseek": ["@ai-sdk/deepseek@2.0.47", "", { "dependencies": { "@ai-sdk/provider": "3.0.14", "@ai-sdk/provider-utils": "4.0.38" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-MzcQ321JO8OY+TVLFI81A7cIIuoeLLxrLCDD+8C1E3Ro6UFyfMtRXo9bw9OhTMRSDMo6hgSDOo4Fekz8aJtQYQ=="],
 
     "@ai-sdk/elevenlabs": ["@ai-sdk/elevenlabs@2.0.52", "", { "dependencies": { "@ai-sdk/provider": "3.0.15", "@ai-sdk/provider-utils": "4.0.46" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-ZgkausouWvO9U4ZtowNJ093bSNYOvH8zqls3uLC3+oxzWvbbTZO8SOdmFk0+gGafsXFJvq2yUX9+rEeJPwOJLw=="],
@@ -1206,8 +1203,6 @@
 
     "@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.23", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.6" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-z8GlDaCmRSDlqkMF2f4/RFgWxdarvIbyuk+m6WXT1LYgsnGiXRJGTD2Z1+SDl3LqtFuRtGX1aghYvQLoHL/9pg=="],
 
-    "@ai-sdk/togetherai": ["@ai-sdk/togetherai@2.0.41", "", { "dependencies": { "@ai-sdk/openai-compatible": "2.0.37", "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.21" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-k3p9e3k0/gpDDyTtvafsK4HYR4D/aUQW/kzCwWo1+CzdBU84i4L14gWISC/mv6tgSicMXHcEUd521fPufQwNlg=="],
-
     "@ai-sdk/vercel": ["@ai-sdk/vercel@2.0.39", "", { "dependencies": { "@ai-sdk/openai-compatible": "2.0.37", "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.21" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-8eu3ljJpkCTP4ppcyYB+NcBrkcBoSOFthCSgk5VnjaxnDaOJFaxnPwfddM7wx3RwMk2CiK1O61Px/LlqNc7QkQ=="],
 
     "@ai-sdk/xai": ["@ai-sdk/xai@3.0.123", "", { "dependencies": { "@ai-sdk/openai-compatible": "2.0.69", "@ai-sdk/provider": "3.0.15", "@ai-sdk/provider-utils": "4.0.46" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-WNASvd1C516oh2qYIj9EvAVPdU+Abads8DQWU6p9lQtvFFeGh8QW+3LDOARZd1GCINUFfw5yadEK845SMQKLsA=="],
@@ -5912,10 +5907,6 @@
 
     "@ai-sdk/deepgram/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.46", "", { "dependencies": { "@ai-sdk/provider": "3.0.15", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.8", "undici": "^6.28.0" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-tEtld97plCFiYevsJuOkGkeuhQndeMWFBVrJS4AjnbD5AqrNSXRCe0p+BZ3Cju/sxDeeZ9ym3q9YUV8fASA7aQ=="],
 
-    "@ai-sdk/deepinfra/@ai-sdk/openai-compatible": ["@ai-sdk/openai-compatible@2.0.37", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.21" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-+POSFVcgiu47BK64dhsI6OpcDC0/VAE2ZSaXdXGNNhpC/ava++uSRJYks0k2bpfY0wwCTgpAWZsXn/dG2Yppiw=="],
-
-    "@ai-sdk/deepinfra/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.21", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.6" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-MtFUYI1/8mgDvRmaBDjbLJPFFrMG777AvSgyIFQtZHIMzm88R/12vYBBpnk7pfiWLFE1DSZzY4WDYzGbKAcmiw=="],
-
     "@ai-sdk/deepseek/@ai-sdk/provider": ["@ai-sdk/provider@3.0.14", "", { "dependencies": { "json-schema": "^0.4.0" } }, "sha512-5X1k57JBJ4H7H1QjX7CnJYAB1I19r/trVZTMcSms7/kLNZ8RaU4Nt2agcwZzv82Hfx6Q7/TOLU7agAKeFfc8cA=="],
 
     "@ai-sdk/deepseek/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.38", "", { "dependencies": { "@ai-sdk/provider": "3.0.14", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.8" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-/HHGmtKllqjg1OLc023v9w9kK3laW7Z6TzfZukYQWCsGBbzB9p60zTvvpXFVcs44NZBVXL3viOa1HRKUbeee8g=="],
@@ -5954,10 +5945,6 @@
 
     "@ai-sdk/perplexity/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.21", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.6" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-MtFUYI1/8mgDvRmaBDjbLJPFFrMG777AvSgyIFQtZHIMzm88R/12vYBBpnk7pfiWLFE1DSZzY4WDYzGbKAcmiw=="],
 
-    "@ai-sdk/togetherai/@ai-sdk/openai-compatible": ["@ai-sdk/openai-compatible@2.0.37", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.21" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-+POSFVcgiu47BK64dhsI6OpcDC0/VAE2ZSaXdXGNNhpC/ava++uSRJYks0k2bpfY0wwCTgpAWZsXn/dG2Yppiw=="],
-
-    "@ai-sdk/togetherai/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.21", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.6" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-MtFUYI1/8mgDvRmaBDjbLJPFFrMG777AvSgyIFQtZHIMzm88R/12vYBBpnk7pfiWLFE1DSZzY4WDYzGbKAcmiw=="],
-
     "@ai-sdk/vercel/@ai-sdk/openai-compatible": ["@ai-sdk/openai-compatible@2.0.37", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.21" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-+POSFVcgiu47BK64dhsI6OpcDC0/VAE2ZSaXdXGNNhpC/ava++uSRJYks0k2bpfY0wwCTgpAWZsXn/dG2Yppiw=="],
 
     "@ai-sdk/vercel/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.21", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.6" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-MtFUYI1/8mgDvRmaBDjbLJPFFrMG777AvSgyIFQtZHIMzm88R/12vYBBpnk7pfiWLFE1DSZzY4WDYzGbKAcmiw=="],

nix/hashes.json

@@ -1,8 +1,8 @@
 {
   "nodeModules": {
-    "x86_64-linux": "sha256-3Jx1Q7hl+Y0Log/k2vd5y6dzBpzFKWlhShPESxn1Rm4=",
-    "aarch64-linux": "sha256-EiiI6g01oBIrExCMAUgT3w82P0fvu4FAJhI32C+ze0I=",
-    "aarch64-darwin": "sha256-s+w49HRp1+ewtiTaU65tPWjUiO1NQw3kzfemMEEQZb0=",
-    "x86_64-darwin": "sha256-/Ee5V7pnL/qm3c4ZHeWEjH7FhGVXArXryOugbG5vsz8="
+    "x86_64-linux": "sha256-oQnV96kE3lIqsQaaUrH4tiEX8/5xvBWizXMGxayFSHo=",
+    "aarch64-linux": "sha256-Hkl1xdCQ7voAllwtyrm3TjQT9cfej29fvPIP3lH7zVo=",
+    "aarch64-darwin": "sha256-s0WHRB13qcD0KlgWNXO7gLpsDOnfB8xny81GnN5YeQc=",
+    "x86_64-darwin": "sha256-fnYi1AxCrnO3byW9keDfBH2ueCfGZdaweOrV6AerrGs="
   }
 }

packages/ai/script/setup-recording-env.ts

@@ -157,9 +157,9 @@ const PROVIDERS: ReadonlyArray<Provider> = [
     id: "togetherai",
     label: "TogetherAI",
     tier: "compatible",
-    note: "Existing OpenAI-compatible text/tool recorded tests",
-    vars: [{ name: "TOGETHER_AI_API_KEY" }],
-    validate: (env) => validateBearer("https://api.together.xyz/v1/models", Redacted.make(env.TOGETHER_AI_API_KEY)),
+    note: "Native Together AI text/tool recorded tests",
+    vars: [{ name: "TOGETHER_API_KEY" }],
+    validate: (env) => validateBearer("https://api.together.xyz/v1/models", Redacted.make(env.TOGETHER_API_KEY)),
   },
   {
     id: "minimax",
@@ -200,8 +200,8 @@ const PROVIDERS: ReadonlyArray<Provider> = [
   {
     id: "cerebras",
     label: "Cerebras",
-    tier: "optional",
-    note: "OpenAI-compatible bridge",
+    tier: "compatible",
+    note: "Native Cerebras text/tool/tool-loop recorded tests",
     vars: [{ name: "CEREBRAS_API_KEY" }],
     validate: (env) => validateBearer("https://api.cerebras.ai/v1/models", Redacted.make(env.CEREBRAS_API_KEY)),
   },

packages/ai/src/cache-policy.ts

@@ -36,7 +36,7 @@ const resolve = (policy: CachePolicy | undefined): CachePolicyObject => {
 // Protocols whose wire format ignores inline cache markers (OpenAI's implicit
 // prefix caching, Gemini's implicit + out-of-band CachedContent). Skip the
 // whole policy pass for these — emitting hints would be harmless but pointless.
-const RESPECTS_INLINE_HINTS = new Set(["anthropic-messages", "bedrock-converse", "openrouter"])
+const RESPECTS_INLINE_HINTS = new Set(["anthropic-messages", "google-vertex-messages", "bedrock-converse", "openrouter"])
 
 const makeHint = (ttlSeconds: number | undefined): CacheHint =>
   ttlSeconds !== undefined ? new CacheHint({ type: "ephemeral", ttlSeconds }) : new CacheHint({ type: "ephemeral" })

packages/ai/src/protocols/anthropic-messages.ts

@@ -744,9 +744,12 @@ const endsInServerToolUse = (message: LLMRequest["messages"][number]) => {
   return message.role === "assistant" && last?.type === "tool-call" && last.providerExecuted === true
 }
 
-const canUseNativeSystemUpdate = (messages: LLMRequest["messages"], index: number) => {
-  const previous = messages[index - 1]
-  const next = messages[index + 1]
+const canUseNativeSystemUpdate = (request: LLMRequest, index: number) => {
+  const previous = request.messages[index - 1]
+  const next = request.messages[index + 1]
+  // Vertex currently rejects/404s for a system message after local tool results,
+  // so fold it into the user tool-result turn across continuations and history.
+  if (request.model.route.id === "google-vertex-messages" && previous?.role === "tool") return false
   return (
     previous !== undefined &&
     previous.role !== "system" &&
@@ -793,7 +796,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
     if (message.role === "system") {
       if (splitsLocalToolResults(request.messages, index))
         return yield* invalid("Anthropic Messages system updates cannot split a local tool call from its tool result")
-      if (supportsNativeSystemUpdates(request) && canUseNativeSystemUpdate(request.messages, index)) {
+      if (supportsNativeSystemUpdates(request) && canUseNativeSystemUpdate(request, index)) {
         messages.push(yield* lowerNativeSystemUpdate(message, breakpoints))
         continue
       }

packages/ai/src/protocols/open-responses.ts

@@ -196,7 +196,7 @@ type LoweredInputItem =
 // multiple streamed summary parts into the same item before flushing.
 type OpenResponsesReasoningInput = {
   type: "reasoning"
-  id: string
+  id?: string
   summary: Array<{ type: "summary_text"; text: string }>
   encrypted_content?: string | null
 }
@@ -372,9 +372,6 @@ export const Event = Schema.StructWithRest(
 )
 export type Event = Schema.Schema.Type<typeof Event>
 
-// Which lowered input item a persisted item id is about to be attached to.
-export type ItemKind = "message" | "reasoning" | "function-call" | "hosted-tool"
-
 export interface Extension {
   readonly id: string
   readonly name: string
@@ -384,10 +381,6 @@ export interface Extension {
     readonly request: LLMRequest
   }) => MediaInput | undefined
   readonly lowerHostedToolItem?: (item: unknown) => ExtendedHostedToolItem | undefined
-  // Optional grammar check applied before a persisted item id is resent as
-  // part of replayed history. Returning false drops the id; every lowered
-  // item treats a dropped id the same as an absent one.
-  readonly acceptsItemID?: (kind: ItemKind, id: string) => boolean
 }
 
 const BASE: Extension = { id: ADAPTER, name: NAME }
@@ -447,53 +440,37 @@ export const lowerToolChoice = (protocolName: string, toolChoice: NonNullable<LL
     tool: (toolName) => ({ type: "function" as const, name: toolName }),
   })
 
-// Servers validate item ids on replayed history, and a malformed or oversized
-// id can fail an otherwise valid request. Only server-issued tokens are worth
-// resending; anything else is treated as absent so the item is resent without
-// an id (or skipped, for items that cannot be expressed without one).
-const ITEM_ID_PATTERN = /^[A-Za-z0-9_-]{1,64}$/
+// Server-issued item ids need a nonempty prefix and suffix, but the prefix is
+// provider-defined and does not necessarily identify the item's semantic type.
 const itemID = (providerMetadata: ProviderMetadata | undefined, providerMetadataKey: string) => {
   const metadata = providerMetadata?.[providerMetadataKey]
-  return ProviderShared.isRecord(metadata) &&
-    typeof metadata.itemId === "string" &&
-    ITEM_ID_PATTERN.test(metadata.itemId)
-    ? metadata.itemId
-    : undefined
+  if (!ProviderShared.isRecord(metadata) || typeof metadata.itemId !== "string") return undefined
+  const separator = metadata.itemId.indexOf("_")
+  return separator > 0 && separator < metadata.itemId.length - 1 ? metadata.itemId : undefined
 }
 
-const acceptsItemID = (extension: Extension, kind: ItemKind, id: string | undefined): id is string =>
-  id !== undefined && (extension.acceptsItemID?.(kind, id) ?? true)
-
-const lowerToolCall = (
-  part: ToolCallPart,
-  providerMetadataKey: string,
-  extension: Extension,
-): OpenResponsesInputItem => {
+const lowerToolCall = (part: ToolCallPart, providerMetadataKey: string): OpenResponsesInputItem => {
   const id = itemID(part.providerMetadata, providerMetadataKey)
   return {
     type: "function_call",
-    ...(acceptsItemID(extension, "function-call", id) ? { id } : {}),
+    ...(id === undefined ? {} : { id }),
     call_id: part.id,
     name: part.name,
     arguments: ProviderShared.encodeJson(part.input),
   }
 }
 
-const lowerReasoning = (
-  part: ReasoningPart,
-  providerMetadataKey: string,
-  extension: Extension,
-): OpenResponsesReasoningInput | undefined => {
+const lowerReasoning = (part: ReasoningPart, providerMetadataKey: string): OpenResponsesReasoningInput | undefined => {
   const metadata = part.providerMetadata?.[providerMetadataKey]
+  if (!ProviderShared.isRecord(metadata)) return undefined
   const id = itemID(part.providerMetadata, providerMetadataKey)
-  if (!ProviderShared.isRecord(metadata) || !acceptsItemID(extension, "reasoning", id)) return undefined
   const encryptedContent =
     typeof metadata.reasoningEncryptedContent === "string" || metadata.reasoningEncryptedContent === null
       ? metadata.reasoningEncryptedContent
       : undefined
   return {
     type: "reasoning",
-    id,
+    ...(id === undefined ? {} : { id }),
     summary: part.text.length > 0 ? [{ type: "summary_text", text: part.text }] : [],
     encrypted_content: encryptedContent,
   }
@@ -584,9 +561,7 @@ const lowerToolResultOutput = Effect.fnUntraced(function* (
 })
 
 const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (request: LLMRequest, extension: Extension) {
-  const system: LoweredInputItem[] =
-    request.system.length === 0 ? [] : [{ role: "system", content: ProviderShared.joinText(request.system) }]
-  const input: LoweredInputItem[] = [...system]
+  const input: LoweredInputItem[] = []
   const providerMetadataKey = request.model.route.providerMetadataKey ?? "openresponses"
 
   for (const message of request.messages) {
@@ -616,8 +591,7 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
           Array<{ id: string | undefined; phase: MessagePhase | null | undefined; parts: TextPart[] }>
         >((groups, part) => {
           const metadata = part.providerMetadata?.[providerMetadataKey]
-          const rawID = itemID(part.providerMetadata, providerMetadataKey)
-          const id = acceptsItemID(extension, "message", rawID) ? rawID : undefined
+          const id = itemID(part.providerMetadata, providerMetadataKey)
           const phase = ProviderShared.isRecord(metadata) ? messagePhase(metadata.phase) : undefined
           const group = groups.at(-1)
           if (group && group.id === id && group.phase === phase) group.parts.push(part)
@@ -642,23 +616,23 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
         }
         if (part.type === "reasoning") {
           flushText()
-          const reasoning = lowerReasoning(part, providerMetadataKey, extension)
+          const reasoning = lowerReasoning(part, providerMetadataKey)
           if (!reasoning) continue
-          const existing = reasoningItems[reasoning.id]
+          const existing = reasoning.id === undefined ? undefined : reasoningItems[reasoning.id]
           if (existing) {
             existing.summary.push(...reasoning.summary)
             if (typeof reasoning.encrypted_content === "string")
               existing.encrypted_content = reasoning.encrypted_content
             continue
           }
-          reasoningItems[reasoning.id] = reasoning
+          if (reasoning.id !== undefined) reasoningItems[reasoning.id] = reasoning
           input.push(reasoning)
           continue
         }
         if (part.type === "tool-call") {
           flushText()
           if (part.providerExecuted === true) continue
-          input.push(lowerToolCall(part, providerMetadataKey, extension))
+          input.push(lowerToolCall(part, providerMetadataKey))
           continue
         }
         if (part.type === "tool-result" && part.providerExecuted === true) {
@@ -670,7 +644,7 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
               : Schema.is(HostedToolItem)(part.result.value)
                 ? part.result.value
                 : extension.lowerHostedToolItem?.(part.result.value)
-          if (acceptsItemID(extension, "hosted-tool", id) && hosted?.id === id) {
+          if (id !== undefined && hosted?.id === id) {
             if (!hostedToolItems.has(id)) {
               input.push(hosted)
               hostedToolItems.add(id)
@@ -716,10 +690,11 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
 
 const lowerOptions = (request: LLMRequest) => {
   const options = OpenResponsesOptions.resolve(request)
+  const instructions = ProviderShared.joinText(request.system)
   const cacheKey = ProviderShared.promptCacheKey(request)
   const parallelToolCalls = resolveParallelToolCalls(request)
   return {
-    ...(options.instructions ? { instructions: options.instructions } : {}),
+    ...(instructions ? { instructions } : {}),
     ...(options.store !== undefined ? { store: options.store } : {}),
     ...(options.metadata ? { metadata: options.metadata } : {}),
     ...(options.safetyIdentifier ? { safety_identifier: options.safetyIdentifier } : {}),
@@ -945,25 +920,23 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
       events,
     ]
   }
-  if (item?.type !== "function_call" || !item.id) return [state, NO_EVENTS]
-  const metadata = providerMetadata(state, { itemId: item.id })
+  if (item?.type !== "function_call" || !item.call_id) return [state, NO_EVENTS]
+  const id = item.id ?? item.call_id
+  const metadata = item.id ? providerMetadata(state, { itemId: item.id }) : undefined
   const events: LLMEvent[] = []
   const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
   return [
     {
       ...state,
       lifecycle,
-      tools: ToolStream.start(state.tools, item.id, {
-        id: item.call_id ?? item.id,
+      tools: ToolStream.start(state.tools, id, {
+        id: item.call_id,
         name: item.name ?? "",
         input: item.arguments ?? "",
         providerMetadata: metadata,
       }),
     },
-    [
-      ...events,
-      LLMEvent.toolInputStart({ id: item.call_id ?? item.id, name: item.name ?? "", providerMetadata: metadata }),
-    ],
+    [...events, LLMEvent.toolInputStart({ id: item.call_id, name: item.name ?? "", providerMetadata: metadata })],
   ]
 }
 
@@ -1095,18 +1068,19 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
   }
 
   if (item.type === "function_call") {
-    if (!item.id || !item.call_id || !item.name) return [state, NO_EVENTS] satisfies StepResult
-    const tools = state.tools[item.id]
+    if (!item.call_id || !item.name) return [state, NO_EVENTS] satisfies StepResult
+    const id = item.id ?? item.call_id
+    const tools = state.tools[id]
       ? state.tools
-      : ToolStream.start(state.tools, item.id, {
+      : ToolStream.start(state.tools, id, {
           id: item.call_id,
           name: item.name,
-          providerMetadata: providerMetadata(state, { itemId: item.id }),
+          providerMetadata: item.id ? providerMetadata(state, { itemId: item.id }) : undefined,
         })
     const result =
       item.arguments === undefined
-        ? yield* ToolStream.finish(state.id, tools, item.id)
-        : yield* ToolStream.finishWithInput(state.id, tools, item.id, item.arguments)
+        ? yield* ToolStream.finish(state.id, tools, id)
+        : yield* ToolStream.finishWithInput(state.id, tools, id, item.arguments)
     const events: LLMEvent[] = []
     const resultEvents = result.events ?? []
     const lifecycle = resultEvents.length ? Lifecycle.stepStart(state.lifecycle, events) : state.lifecycle
@@ -1160,10 +1134,11 @@ const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (
           event.response?.output ?? [],
           () => [state, NO_EVENTS] satisfies StepResult,
           ([current, events], item) => {
+            const id = item.id ?? (item.type === "function_call" ? item.call_id : undefined)
             if (
-              !item.id ||
-              ((item.type !== "function_call" || !current.tools[item.id]) &&
-                (item.type !== "reasoning" || !current.reasoningItems[item.id]))
+              !id ||
+              ((item.type !== "function_call" || !current.tools[id]) &&
+                (item.type !== "reasoning" || !current.reasoningItems[id]))
             )
               return Effect.succeed([current, events] satisfies StepResult)
             return onOutputItemDone(current, { type: "response.output_item.done", item }).pipe(
@@ -1289,10 +1264,11 @@ export const step = (state: ParserState, input: Event) => {
   if (event.type === "response.output_item.added") {
     if (event.item?.type === "message" && !event.item.id)
       return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
+    const id = event.item?.id ?? (event.item?.type === "function_call" ? event.item.call_id : undefined)
     return Effect.succeed(
       onOutputItemAdded(
-        event.output_index !== undefined && event.item?.id
-          ? { ...state, outputItems: { ...state.outputItems, [event.output_index]: event.item.id } }
+        event.output_index !== undefined && id
+          ? { ...state, outputItems: { ...state.outputItems, [event.output_index]: id } }
           : state,
         event,
       ),

packages/ai/src/protocols/openai-chat.ts

@@ -364,8 +364,9 @@ const lowerUserMessage = Effect.fn("OpenAIChat.lowerUserMessage")(function* (
 
 const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(function* (
   message: OpenAIChatRequestMessage,
-  configuredField?: string,
-  options: LoweringOptions = {},
+  configuredField: string | undefined,
+  requireReasoning: boolean,
+  options: LoweringOptions,
 ) {
   const content: TextPart[] = []
   const reasoning: ReasoningPart[] = []
@@ -392,15 +393,17 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
   const nativeReasoning = openAICompatibleReasoningContent(message.native?.openaiCompatible)
   const fullyStructured = reasoning.every((part) => Array.isArray(part.providerMetadata?.openai?.reasoningDetails))
   const field = (() => {
-    if (configuredField !== undefined) return configuredField
-    if (reasoning.length === 0) return undefined
+    if (configuredField !== undefined && (requireReasoning || reasoning.length > 0 || nativeReasoning !== undefined))
+      return configuredField
+    if (reasoning.length === 0) return requireReasoning ? "reasoning_content" : undefined
     if (observedField !== undefined) return observedField
     if (nativeReasoning !== undefined) return "reasoning_content"
-    if (!fullyStructured) return "reasoning_content"
+    if (!fullyStructured || requireReasoning) return "reasoning_content"
   })()
   const reasoningText = (() => {
-    if (configuredField !== undefined) return reasoning.length === 0 ? (nativeReasoning ?? "") : text
-    if (reasoning.length === 0) return nativeReasoning
+    if (configuredField !== undefined)
+      return reasoning.length === 0 ? (nativeReasoning ?? (requireReasoning ? "" : undefined)) : text
+    if (reasoning.length === 0) return nativeReasoning ?? (requireReasoning ? "" : undefined)
     return text
   })()
   const cached = message.content.findLast((part) => "cache" in part && part.cache !== undefined)
@@ -454,11 +457,13 @@ const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (
 
 const lowerMessage = Effect.fn("OpenAIChat.lowerMessage")(function* (
   message: OpenAIChatRequestMessage,
-  reasoningField?: string,
-  options: LoweringOptions = {},
+  reasoningField: string | undefined,
+  requireReasoning: boolean,
+  options: LoweringOptions,
 ) {
   if (message.role === "user") return [yield* lowerUserMessage(message, options)]
-  if (message.role === "assistant") return [yield* lowerAssistantMessage(message, reasoningField, options)]
+  if (message.role === "assistant")
+    return [yield* lowerAssistantMessage(message, reasoningField, requireReasoning, options)]
   return (yield* lowerToolMessages(message, options)).messages
 })
 
@@ -480,6 +485,13 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
         : [{ role: "system", content: ProviderShared.joinText(request.system) }]
   const messages = [...system]
   const modelID = request.model.id.toLowerCase()
+  const requireReasoning =
+    request.model.compatibility?.requireReasoning ??
+    (request.model.compatibility?.reasoningField !== undefined ||
+      request.model.provider === "deepseek" ||
+      request.model.route.endpoint.baseURL?.toLowerCase().includes("deepseek.com") ||
+      modelID.includes("deepseek"))
+  const reasoningField = request.model.compatibility?.reasoningField
   const mistral = ["mistral", "devstral", "codestral", "pixtral", "mixtral"].some((family) => modelID.includes(family))
   const lowering = {
     ...options,
@@ -554,7 +566,7 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
       continue
     }
     flushImages()
-    messages.push(...(yield* lowerMessage(message, request.model.compatibility?.reasoningField, lowering)))
+    messages.push(...(yield* lowerMessage(message, reasoningField, requireReasoning, lowering)))
   }
   flushImages()
   return messages

packages/ai/src/protocols/openai-responses.ts

@@ -86,28 +86,10 @@ const OpenAIResponsesBody = Schema.Struct({
 })
 export type OpenAIResponsesBody = Schema.Schema.Type<typeof OpenAIResponsesBody>
 
-// Replayed items are paired with stored server state by id, so a foreign or
-// synthetic token can fail request validation even when `call_id` pairing is
-// intact. Only resend ids in each item kind's own grammar; hosted tool
-// items keep generic validation because every hosted tool mints its own
-// prefix. The same allowlist approach codex uses before resending history
-// (codex-rs core/src/client.rs, `prepare_response_items_for_request`).
-const ITEM_ID_PREFIXES: Record<OpenResponses.ItemKind, ReadonlyArray<string>> = {
-  message: ["msg_"],
-  reasoning: ["rs_"],
-  "function-call": ["fc_"],
-  // Every hosted tool mints its own id prefix, so items keep generic validation.
-  "hosted-tool": [],
-}
-
 const extension = {
   id: ADAPTER,
   name: NAME,
   lowerHostedToolItem: (item: unknown) => (Schema.is(OpenAIResponsesHostedToolItem)(item) ? item : undefined),
-  acceptsItemID: (kind: OpenResponses.ItemKind, id: string) => {
-    const prefixes = ITEM_ID_PREFIXES[kind]
-    return prefixes.length === 0 || prefixes.some((prefix) => id.startsWith(prefix))
-  },
 } satisfies OpenResponses.Extension
 
 const nativeImageToolInput = (tool: ToolDefinition) => {

packages/ai/src/protocols/shared.ts

@@ -210,10 +210,9 @@ export const errorText = (error: unknown) => {
  * `framing` step for Server-Sent Events. Decodes UTF-8, runs the SSE channel
  * decoder, optionally filters named events, and drops empty / `[DONE]`
  * keep-alive events so the protocol event schema sees one JSON string per
- * element. The SSE channel emits a
- * `Retry` control event on its error channel; we drop it here (we don't
- * implement client-driven retries). Decoder failures become provider output
- * errors so the public error channel stays `AIError`.
+ * element. Retry control events are ignored without interrupting the stream.
+ * Decoder failures become provider output errors so the public error channel
+ * stays `AIError`.
  */
 export const sseFraming = (
   bytes: Stream.Stream<Uint8Array, AIError>,
@@ -221,9 +220,23 @@ export const sseFraming = (
 ): Stream.Stream<string, AIError> =>
   bytes.pipe(
     Stream.decodeText(),
-    Stream.pipeThroughChannel(Sse.decode()),
-    Stream.catchTag("Retry", () => Stream.empty),
-    Stream.catchTag("SseError", (error) => Stream.fail(eventError("sse", error.message))),
+    Stream.mapAccumEffect(
+      () => {
+        const output: Sse.Event[] = []
+        return {
+          output,
+          parser: Sse.makeParser((event) => {
+            if (event._tag === "Event") output.push(event)
+          }),
+        }
+      },
+      (state, chunk) =>
+        Effect.gen(function* () {
+          const error = state.parser.feed(chunk)
+          if (error) return yield* eventError("sse", error.message)
+          return [state, state.output.splice(0)] as const
+        }),
+    ),
     Stream.filter(
       (event) =>
         (events === undefined || events.has(event.event)) &&

packages/ai/src/protocols/utils/open-responses-options.ts

@@ -56,7 +56,6 @@ export const StreamOptions = Schema.Struct({
 })
 
 export const Options = Schema.Struct({
-  instructions: Schema.optional(Schema.String),
   store: Schema.optional(Schema.Boolean),
   metadata: Schema.optional(Schema.Record(Schema.String, Schema.String)),
   safetyIdentifier: Schema.optional(Schema.String),

packages/ai/src/providers/amazon-bedrock-mantle.ts

@@ -29,6 +29,11 @@ export interface Settings extends ProviderPackage.Settings {
 const responsesRoute = OpenAIResponses.route.with({
   id: "bedrock-mantle-responses",
   provider: id,
+  transport: OpenAIResponses.channelTransport({
+    id: "bedrock-mantle-responses",
+    name: "Bedrock Mantle Responses",
+    enabled: () => false,
+  }),
 })
 
 const chatRoute = OpenAIChat.route.with({

packages/ai/src/providers/cerebras.ts

@@ -0,0 +1,58 @@
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAICompatibleChat } from "../protocols/openai-compatible-chat.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import type { RouteDefaultsInput } from "../route/client.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("cerebras")
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: OpenAIProviderOptionsInput
+}
+
+export const route = OpenAICompatibleChat.route.with({
+  id: "cerebras-chat",
+  provider: id,
+  endpoint: { baseURL: profiles.cerebras.baseURL },
+})
+
+export const routes = [route]
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: baseURL ?? profiles.cerebras.baseURL },
+    auth: AuthOptions.bearer(input, "CEREBRAS_API_KEY"),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<OpenAIProviderOptionsInput>({
+        id: modelID,
+        compatibility: { maxTokensField: "max_tokens", reasoningField: "reasoning", supportsStore: false },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)

packages/ai/src/providers/deepinfra.ts

@@ -0,0 +1,61 @@
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAICompatibleChat } from "../protocols/openai-compatible-chat.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import type { RouteDefaultsInput } from "../route/client.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("deepinfra")
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: OpenAIProviderOptionsInput
+}
+
+export const route = OpenAICompatibleChat.route.with({
+  id: "deepinfra-chat",
+  provider: id,
+  endpoint: { baseURL: profiles.deepinfra.baseURL },
+})
+
+export const routes = [route]
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const root = baseURL?.replace(/\/+$/, "")
+  const configured = route.with({
+    ...defaults,
+    endpoint: {
+      baseURL: root === undefined ? profiles.deepinfra.baseURL : root.endsWith("/openai") ? root : `${root}/openai`,
+    },
+    auth: AuthOptions.bearer(input, "DEEPINFRA_API_KEY"),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<OpenAIProviderOptionsInput>({
+        id: modelID,
+        compatibility: { maxTokensField: "max_tokens", reasoningField: "reasoning_content", supportsStore: false },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)

packages/ai/src/providers/index.ts

@@ -3,8 +3,10 @@ export * as AnthropicCompatible from "./anthropic-compatible.js"
 export * as AmazonBedrock from "./amazon-bedrock.js"
 export * as AmazonBedrockMantle from "./amazon-bedrock-mantle.js"
 export * as Azure from "./azure.js"
+export * as Cerebras from "./cerebras.js"
 export * as Cloudflare from "./cloudflare.js"
 export { CloudflareAIGateway, CloudflareWorkersAI } from "./cloudflare.js"
+export * as DeepInfra from "./deepinfra.js"
 export * as Google from "./google.js"
 export * as GoogleVertex from "./google-vertex.js"
 export * as GoogleVertexChat from "./google-vertex-chat.js"
@@ -14,5 +16,6 @@ export * as OpenAI from "./openai.js"
 export * as OpenAICompatible from "./openai-compatible.js"
 export * as OpenAICompatibleResponses from "./openai-compatible-responses.js"
 export * as OpenRouter from "./openrouter.js"
+export * as TogetherAI from "./togetherai.js"
 export * as XAI from "./xai.js"
 export * as ZAI from "./zai.js"

packages/ai/src/providers/togetherai.ts

@@ -0,0 +1,58 @@
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAICompatibleChat } from "../protocols/openai-compatible-chat.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import type { RouteDefaultsInput } from "../route/client.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("togetherai")
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: OpenAIProviderOptionsInput
+}
+
+export const route = OpenAICompatibleChat.route.with({
+  id: "togetherai-chat",
+  provider: id,
+  endpoint: { baseURL: profiles.togetherai.baseURL },
+})
+
+export const routes = [route]
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: baseURL ?? profiles.togetherai.baseURL },
+    auth: AuthOptions.bearer(input, ["TOGETHER_API_KEY", "TOGETHER_AI_API_KEY"]),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<OpenAIProviderOptionsInput>({
+        id: modelID,
+        compatibility: { maxTokensField: "max_tokens", supportsStore: false, supportsStrictMode: false },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)

packages/ai/src/schema/options.ts

@@ -153,6 +153,8 @@ export class LanguageModelCompatibility extends Schema.Class<LanguageModelCompat
 )({
   toolSchema: Schema.optional(LanguageModelToolSchemaCompatibility),
   reasoningField: Schema.optional(Schema.String),
+  /** Require every assistant message to include its reasoning field, even when empty. */
+  requireReasoning: Schema.optional(Schema.Boolean),
   maxTokensField: Schema.optional(LanguageModelMaxTokensFieldCompatibility),
   requireFinishReason: Schema.optional(Schema.Boolean),
   requireAssistantAfterTool: Schema.optional(Schema.Boolean),

packages/ai/test/cache-policy.test.ts

@@ -3,7 +3,7 @@ import { Effect } from "effect"
 import { CacheHint, LLM, Message } from "../src/index.js"
 import { Auth } from "../src/route.js"
 import { compileRequest } from "../src/route/client.js"
-import { AmazonBedrock } from "../src/providers.js"
+import { AmazonBedrock, GoogleVertexMessages } from "../src/providers.js"
 import * as AnthropicMessages from "../src/protocols/anthropic-messages.js"
 import * as Gemini from "../src/protocols/gemini.js"
 import * as OpenAIChat from "../src/protocols/openai-chat.js"
@@ -86,6 +86,27 @@ describe("applyCachePolicy", () => {
     }),
   )
 
+  it.effect("'auto' emits Anthropic cache markers on Vertex", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: GoogleVertexMessages.configure({ accessToken: "test", location: "global", project: "test" }).model(
+            "claude-opus-4-8",
+          ),
+          system: "You are concise.",
+          tools: [{ name: "lookup", description: "Look up a value", inputSchema: { type: "object", properties: {} } }],
+          prompt: "hi",
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({
+        tools: [{ name: "lookup", cache_control: { type: "ephemeral" } }],
+        system: [{ type: "text", text: "You are concise.", cache_control: { type: "ephemeral" } }],
+        messages: [{ role: "user", content: [{ type: "text", text: "hi", cache_control: { type: "ephemeral" } }] }],
+      })
+    }),
+  )
+
   it.effect("'auto' is a no-op on OpenAI (implicit caching protocol)", () =>
     Effect.gen(function* () {
       const prepared = yield* compileRequest(

packages/ai/test/fixtures/recordings/cerebras-chat/cerebras-gpt-oss-120b-text.json

@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "cerebras",
+    "route": "cerebras-chat",
+    "transport": "http",
+    "model": "gpt-oss-120b",
+    "tags": ["prefix:cerebras-chat", "provider:cerebras", "text", "golden"],
+    "name": "cerebras-chat/cerebras-gpt-oss-120b-text",
+    "recordedAt": "2026-08-25T23:55:27.619Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.cerebras.ai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gpt-oss-120b\",\"messages\":[{\"role\":\"system\",\"content\":\"You are concise.\"},{\"role\":\"user\",\"content\":\"Reply exactly with: Hello!\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":256}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"role\":\"assistant\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\"The\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\" user says: \\\"Reply exactly with: Hello!\\\" So we must output exactly\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\" \\\"Hello!\\\" with no extra characters, no formatting\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\". Ensure no extra spaces or new\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\"lines? Probably just \\\"Hello!\\\".\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"reasoning\":\" Usually we output exactly that.\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{\"content\":\"Hello!\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-cd45cd6e-886a-433e-8ba9-caca78c1f13a\",\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\",\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_e2cabf4999eb0aead3d1\",\"object\":\"chat.completion.chunk\",\"usage\":{\"total_tokens\":142,\"completion_tokens\":58,\"completion_tokens_details\":{\"accepted_prediction_tokens\":0,\"rejected_prediction_tokens\":0,\"reasoning_tokens\":46},\"prompt_tokens\":84,\"prompt_tokens_details\":{\"cached_tokens\":0}},\"time_info\":{\"created\":1787702127.645281,\"queue_time\":0.003817115,\"prompt_time\":0.001587193,\"completion_time\":0.029805929,\"total_time\":0.036823272705078125}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/cerebras-chat/cerebras-gpt-oss-120b-tool-call.json

@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "cerebras",
+    "route": "cerebras-chat",
+    "transport": "http",
+    "model": "gpt-oss-120b",
+    "tags": ["prefix:cerebras-chat", "provider:cerebras", "tool", "tool-call", "golden"],
+    "name": "cerebras-chat/cerebras-gpt-oss-120b-tool-call",
+    "recordedAt": "2026-08-25T23:55:28.454Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.cerebras.ai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"gpt-oss-120b\",\"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},\"strict\":false}}],\"tool_choice\":{\"type\":\"function\",\"function\":{\"name\":\"get_weather\"}},\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":512}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"role\":\"assistant\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"reasoning\":\"We\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"reasoning\":\" need to call the function get_weather with city \\\"\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"reasoning\":\"Paris\\\".\"},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"name\":\"get_weather\",\"arguments\":\"\"},\"type\":\"function\",\"id\":\"3d860cefe\",\"index\":0}]},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"},\"type\":\"function\",\"index\":0}]},\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"id\":\"chatcmpl-402a09cc-8668-446d-8f9c-3e4121f5fa51\",\"choices\":[{\"delta\":{},\"finish_reason\":\"tool_calls\",\"index\":0}],\"created\":1787702127,\"model\":\"gpt-oss-120b\",\"system_fingerprint\":\"fp_4cfabdd6620dc0120785\",\"object\":\"chat.completion.chunk\",\"usage\":{\"total_tokens\":174,\"completion_tokens\":37,\"completion_tokens_details\":{\"accepted_prediction_tokens\":0,\"rejected_prediction_tokens\":0,\"reasoning_tokens\":13},\"prompt_tokens\":137,\"prompt_tokens_details\":{\"cached_tokens\":0}},\"time_info\":{\"created\":1787702127.8019717,\"queue_time\":0.31196235,\"prompt_time\":0.005234764,\"completion_time\":0.020198402,\"total_time\":0.702225923538208}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/deepinfra-chat/deepinfra-llama-3-3-70b-text.json

@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "text", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-text",
+    "recordedAt": "2026-08-26T00:34:03.019Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/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: {\"service_tier\":null,\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"Hello\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"!\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":25,\"total_tokens\":28,\"completion_tokens\":3,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RbZ8MyY5pos2MShihSmoXoRe\",\"object\":\"chat.completion.chunk\",\"created\":1787704442,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":25,\"total_tokens\":28,\"completion_tokens\":3,\"estimated_cost\":3.46e-6,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/deepinfra-chat/deepinfra-llama-3-3-70b-tool-call.json

@@ -0,0 +1,32 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "tool", "tool-call", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-tool-call",
+    "recordedAt": "2026-08-26T00:34:04.173Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/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},\"strict\":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: {\"service_tier\":null,\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":[{\"index\":0,\"id\":\"call_SMfBjXa8eCmHLyjfeyARxe3a\",\"function\":{\"arguments\":\"{\\\"city\\\": \\\"Paris\\\"}\",\"name\":\"get_weather\"},\"type\":\"function\"}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":230,\"total_tokens\":244,\"completion_tokens\":14,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RFOwhlB2PhZrgMLviGrx5BQf\",\"object\":\"chat.completion.chunk\",\"created\":1787704443,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":230,\"total_tokens\":244,\"completion_tokens\":14,\"estimated_cost\":0.000027480000000000005,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/deepinfra-chat/deepinfra-llama-3-3-70b-tool-loop.json

@@ -0,0 +1,50 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "deepinfra",
+    "route": "deepinfra-chat",
+    "transport": "http",
+    "model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+    "tags": ["prefix:deepinfra-chat", "provider:deepinfra", "tool", "tool-loop", "golden"],
+    "name": "deepinfra-chat/deepinfra-llama-3-3-70b-tool-loop",
+    "recordedAt": "2026-08-26T00:34:05.656Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"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},\"strict\":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: {\"service_tier\":null,\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":[{\"index\":0,\"id\":\"call_W3stxe7FNHozlB4tDxlTVZou\",\"function\":{\"arguments\":\"{\\\"city\\\": \\\"Paris\\\"}\",\"name\":\"get_weather\"},\"type\":\"function\"}]},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":243,\"total_tokens\":257,\"completion_tokens\":14,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RxlHFSnlLbUUz6XSxqBQj7TC\",\"object\":\"chat.completion.chunk\",\"created\":1787704444,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":243,\"total_tokens\":257,\"completion_tokens\":14,\"estimated_cost\":0.000028780000000000002,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.deepinfra.com/v1/openai/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"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_W3stxe7FNHozlB4tDxlTVZou\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}],\"reasoning_content\":\"\"},{\"role\":\"tool\",\"tool_call_id\":\"call_W3stxe7FNHozlB4tDxlTVZou\",\"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},\"strict\":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: {\"service_tier\":null,\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"Paris\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\" is\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\" sunny\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\".\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":null}],\"usage\":null}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\",\"reasoning_content\":null,\"tool_calls\":null},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":276,\"total_tokens\":281,\"completion_tokens\":5,\"estimated_cost\":null,\"prompt_tokens_details\":null}}\n\ndata: {\"service_tier\":\"default\",\"id\":\"chatcmpl-RHESVChSFPybgK1eUj2cLLY0\",\"object\":\"chat.completion.chunk\",\"created\":1787704445,\"model\":\"meta-llama/Llama-3.3-70B-Instruct-Turbo\",\"choices\":[],\"usage\":{\"prompt_tokens\":276,\"total_tokens\":281,\"completion_tokens\":5,\"estimated_cost\":0.0000292,\"prompt_tokens_details\":null}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/openai-responses-websocket/continues-a-tool-call-over-one-socket.json

@@ -32,7 +32,7 @@
         {
           "direction": "client",
           "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Call get_weather once, then 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},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"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},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Call get_weather once, then reply exactly: Paris is sunny.\"}"
         },
         {
           "direction": "server",
@@ -62,7 +62,7 @@
         {
           "direction": "client",
           "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"type\":\"function_call_output\",\"call_id\":\"call_ws_weather\",\"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},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"previous_response_id\":\"resp_ws_tool_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"type\":\"function_call_output\",\"call_id\":\"call_ws_weather\",\"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},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"previous_response_id\":\"resp_ws_tool_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Call get_weather once, then reply exactly: Paris is sunny.\"}"
         },
         {
           "direction": "server",

packages/ai/test/fixtures/recordings/openai-responses-websocket/reconstructs-full-context-after-reconnect.json

@@ -32,7 +32,7 @@
         {
           "direction": "client",
           "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Follow the user's exact reply instruction.\"}"
         },
         {
           "direction": "server",
@@ -81,7 +81,7 @@
         {
           "direction": "client",
           "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_reconnect_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Alpha.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Beta.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_reconnect_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Alpha.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Beta.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Follow the user's exact reply instruction.\"}"
         },
         {
           "direction": "server",

packages/ai/test/fixtures/recordings/openai-responses-websocket/recovers-from-explicit-continuation-rejection.json

@@ -32,7 +32,7 @@
         {
           "direction": "client",
           "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Follow the user's exact reply instruction.\"}"
         },
         {
           "direction": "server",
@@ -81,7 +81,7 @@
         {
           "direction": "client",
           "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"previous_response_id\":\"resp_ws_rejection_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"previous_response_id\":\"resp_ws_rejection_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Follow the user's exact reply instruction.\"}"
         },
         {
           "direction": "server",
@@ -91,7 +91,7 @@
         {
           "direction": "client",
           "kind": "text",
-          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_rejection_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Ready.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_rejection_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Ready.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"instructions\":\"Follow the user's exact reply instruction.\"}"
         },
         {
           "direction": "server",

packages/ai/test/fixtures/recordings/openai-responses/openai-responses-gpt-5-5-reasoning-continuation.json

@@ -26,7 +26,7 @@
         "headers": {
           "content-type": "application/json"
         },
-        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Show concise reasoning when the provider supports visible reasoning summaries.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Think briefly, then reply exactly with: Hello!\"}]}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"low\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":120,\"stream\":true}"
+        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Think briefly, then reply exactly with: Hello!\"}]}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"low\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":120,\"stream\":true,\"instructions\":\"Show concise reasoning when the provider supports visible reasoning summaries.\"}"
       },
       "response": {
         "status": 200,

packages/ai/test/fixtures/recordings/openai-responses/openai-responses-gpt-5-5-reasoning.json

@@ -18,7 +18,7 @@
         "headers": {
           "content-type": "application/json"
         },
-        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Show concise reasoning when the provider supports visible reasoning summaries.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Think briefly, then reply exactly with: Hello!\"}]}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"low\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":120,\"stream\":true}"
+        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Think briefly, then reply exactly with: Hello!\"}]}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"low\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":120,\"stream\":true,\"instructions\":\"Show concise reasoning when the provider supports visible reasoning summaries.\"}"
       },
       "response": {
         "status": 200,

packages/ai/test/fixtures/recordings/openai-responses/openai-responses-gpt-5-5-tool-loop.json

@@ -25,7 +25,7 @@
         "headers": {
           "content-type": "application/json"
         },
-        "body": "{\"model\":\"gpt-5.5\",\"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},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":80,\"stream\":true}"
+        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"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},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":80,\"stream\":true,\"instructions\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"}"
       },
       "response": {
         "status": 200,
@@ -43,7 +43,7 @@
         "headers": {
           "content-type": "application/json"
         },
-        "body": "{\"model\":\"gpt-5.5\",\"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\":\"reasoning\",\"id\":\"rs_0ad67c31d9ddad95016a869efbd02487d1a51eed850e6f87f5\",\"summary\":[],\"encrypted_content\":\"gAAAAABqhp79U9UPKmTdmo9tmdil0C2KXpkFqUc4MNkYHT53Lzos9omncFPg76QzUmmSOdBcajisWBEo-xiTCvhp135uACUq8TJcdw4DluieYq6dWszijy28PFFfeO-6MmHwi7zeln1Z202zErJUEyuf1bML68VAeam5PqlMLG-a4-pmnWiH2ExWKibTUX37QoMQoArrkccJOCmxwDflV_kWDPMFxQVDfeMg9fd1gVv2u-x1Mjk0b9mJDOq0Fe5Gh-IkpWzfXgZTdptFmCM75cksvs61Rqsx6P33czal-LSixEF0WMizCvbMQmqKGs7MKGMeoa6j6vWOnB3ICIbv6FShnSaTpZWJFwejvOurkfuxa-2q6xVDZsBoQCgMWPHsqLxwAo1JKdfBk0pMvSuvpw2BRxykUZ1ULCYJ-BypST65292-EuSZFIuXPMPir-_raSCTsgsZNMscDG6ll3qksDTDS6_o5NutD7Ra-WZzaUe_HQlSLKLACTc4qv2EK1QoC4aYv4goxkTSx17WhS2D86lILgkUd-TIHjJ6iR3uxSNx7YeBNxiJgddIAEjAaSrdF-WDouSNT9k3efd5HhT3zahIOMKgb3XIQzFOYWfWgea5-SbaIdKwne9hU0QyhcBQs6yoifSg-fJZtahbPb-GCDYnOLlH-bV94vldoccb-2P1JdB3jaLj5tJUecfr2H4qiu8MgkPj0TkwYNbJynYmJo9H5Lm-XJ9gfzIXzJh0arKwsS4gwDLf4J3LOEF3WEW3mknOjjb9PrLmHRYXQQh9tTiX9ILPZpbufkyCurTUMQgWiSCitXBC6FoLXRHilSmb-6_avBnlUMziMfey-FkKvRfiPox6BaJrnOq6SGlOv11y7EKvzrn29la7HKPygYenDAkyq2mq0Zk2nLWNmJcv9sQTBrkBdFMmJYPi2J2im8XD5MmAjEL8R4FCBHoPIIZ6pENQykvH8PhpWKuzF5gJlY3Vwz4iJ0Qb9TrNI0hzBoI1U0LeB5FJ2HgjZQwCFF5x3ubh72xrUsFpuyYyYPa8GDT0Bo-LW_IlJ_mN4EwI5Nk9n-8Bt015yxsfpa5YaDeCeQFcdj8SD0UAd7QWtGACpzKcIj1-vJJU7OiwscV_v1dLvoiEe1ehI9jcvPn28TgHlo_dippe0iMN4FAm1Bf8vtWVMFDvfV1rPv1pAFFnSa9XqFszD5Exo_xzcQEoKXvQv3OnUtoiM4Db4uadClazLjoep2TQgHcJBVbTbLySTVPmok4ROFQZsU_mq4vu2M__d8HOjADfIIYz5VLVQKNpo0Hv_QkT2bn56Q==\"},{\"type\":\"function_call\",\"id\":\"fc_0ad67c31d9ddad95016a869efd126887d1b7e2f17f155cb5dc\",\"call_id\":\"call_qrzOfKDfzaq8fqbSNNVHlNsV\",\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_qrzOfKDfzaq8fqbSNNVHlNsV\",\"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},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":80,\"stream\":true}"
+        "body": "{\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]},{\"type\":\"reasoning\",\"id\":\"rs_0ad67c31d9ddad95016a869efbd02487d1a51eed850e6f87f5\",\"summary\":[],\"encrypted_content\":\"gAAAAABqhp79U9UPKmTdmo9tmdil0C2KXpkFqUc4MNkYHT53Lzos9omncFPg76QzUmmSOdBcajisWBEo-xiTCvhp135uACUq8TJcdw4DluieYq6dWszijy28PFFfeO-6MmHwi7zeln1Z202zErJUEyuf1bML68VAeam5PqlMLG-a4-pmnWiH2ExWKibTUX37QoMQoArrkccJOCmxwDflV_kWDPMFxQVDfeMg9fd1gVv2u-x1Mjk0b9mJDOq0Fe5Gh-IkpWzfXgZTdptFmCM75cksvs61Rqsx6P33czal-LSixEF0WMizCvbMQmqKGs7MKGMeoa6j6vWOnB3ICIbv6FShnSaTpZWJFwejvOurkfuxa-2q6xVDZsBoQCgMWPHsqLxwAo1JKdfBk0pMvSuvpw2BRxykUZ1ULCYJ-BypST65292-EuSZFIuXPMPir-_raSCTsgsZNMscDG6ll3qksDTDS6_o5NutD7Ra-WZzaUe_HQlSLKLACTc4qv2EK1QoC4aYv4goxkTSx17WhS2D86lILgkUd-TIHjJ6iR3uxSNx7YeBNxiJgddIAEjAaSrdF-WDouSNT9k3efd5HhT3zahIOMKgb3XIQzFOYWfWgea5-SbaIdKwne9hU0QyhcBQs6yoifSg-fJZtahbPb-GCDYnOLlH-bV94vldoccb-2P1JdB3jaLj5tJUecfr2H4qiu8MgkPj0TkwYNbJynYmJo9H5Lm-XJ9gfzIXzJh0arKwsS4gwDLf4J3LOEF3WEW3mknOjjb9PrLmHRYXQQh9tTiX9ILPZpbufkyCurTUMQgWiSCitXBC6FoLXRHilSmb-6_avBnlUMziMfey-FkKvRfiPox6BaJrnOq6SGlOv11y7EKvzrn29la7HKPygYenDAkyq2mq0Zk2nLWNmJcv9sQTBrkBdFMmJYPi2J2im8XD5MmAjEL8R4FCBHoPIIZ6pENQykvH8PhpWKuzF5gJlY3Vwz4iJ0Qb9TrNI0hzBoI1U0LeB5FJ2HgjZQwCFF5x3ubh72xrUsFpuyYyYPa8GDT0Bo-LW_IlJ_mN4EwI5Nk9n-8Bt015yxsfpa5YaDeCeQFcdj8SD0UAd7QWtGACpzKcIj1-vJJU7OiwscV_v1dLvoiEe1ehI9jcvPn28TgHlo_dippe0iMN4FAm1Bf8vtWVMFDvfV1rPv1pAFFnSa9XqFszD5Exo_xzcQEoKXvQv3OnUtoiM4Db4uadClazLjoep2TQgHcJBVbTbLySTVPmok4ROFQZsU_mq4vu2M__d8HOjADfIIYz5VLVQKNpo0Hv_QkT2bn56Q==\"},{\"type\":\"function_call\",\"id\":\"fc_0ad67c31d9ddad95016a869efd126887d1b7e2f17f155cb5dc\",\"call_id\":\"call_qrzOfKDfzaq8fqbSNNVHlNsV\",\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_qrzOfKDfzaq8fqbSNNVHlNsV\",\"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},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"},\"max_output_tokens\":80,\"stream\":true,\"instructions\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"}"
       },
       "response": {
         "status": 200,

packages/ai/test/fixtures/recordings/pdf/openai-tool-result.json

@@ -21,7 +21,7 @@
         "headers": {
           "content-type": "application/json"
         },
-        "body": "{\"model\":\"gpt-4o-mini\",\"input\":[{\"role\":\"system\",\"content\":\"Read the PDF returned by the tool and follow the user's response format exactly.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Return only the verification code from the PDF.\"}]},{\"type\":\"function_call\",\"call_id\":\"call_pdf_1\",\"name\":\"read_pdf\",\"arguments\":\"{}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_pdf_1\",\"output\":[{\"type\":\"input_text\",\"text\":\"PDF read successfully\"},{\"type\":\"input_file\",\"filename\":\"verification.pdf\",\"file_data\":\"data:application/pdf;base64,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\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"read_pdf\",\"description\":\"Read the attached PDF.\",\"parameters\":{\"type\":\"object\",\"properties\":{},\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"max_output_tokens\":40,\"temperature\":0,\"stream\":true}"
+        "body": "{\"model\":\"gpt-4o-mini\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Return only the verification code from the PDF.\"}]},{\"type\":\"function_call\",\"call_id\":\"call_pdf_1\",\"name\":\"read_pdf\",\"arguments\":\"{}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_pdf_1\",\"output\":[{\"type\":\"input_text\",\"text\":\"PDF read successfully\"},{\"type\":\"input_file\",\"filename\":\"verification.pdf\",\"file_data\":\"data:application/pdf;base64,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\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"read_pdf\",\"description\":\"Read the attached PDF.\",\"parameters\":{\"type\":\"object\",\"properties\":{},\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"max_output_tokens\":40,\"temperature\":0,\"stream\":true,\"instructions\":\"Read the PDF returned by the tool and follow the user's response format exactly.\"}"
       },
       "response": {
         "status": 200,

packages/ai/test/fixtures/recordings/pdf/xai-tool-result.json

@@ -21,7 +21,7 @@
         "headers": {
           "content-type": "application/json"
         },
-        "body": "{\"model\":\"grok-4.5\",\"input\":[{\"role\":\"system\",\"content\":\"Read the PDF returned by the tool and follow the user's response format exactly.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Return only the verification code from the PDF.\"}]},{\"type\":\"function_call\",\"call_id\":\"call_pdf_1\",\"name\":\"read_pdf\",\"arguments\":\"{}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_pdf_1\",\"output\":[{\"type\":\"input_text\",\"text\":\"PDF read successfully\"},{\"type\":\"input_file\",\"filename\":\"verification.pdf\",\"file_data\":\"data:application/pdf;base64,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\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"read_pdf\",\"description\":\"Read the attached PDF.\",\"parameters\":{\"type\":\"object\",\"properties\":{},\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"max_output_tokens\":40,\"temperature\":0,\"stream\":true}"
+        "body": "{\"model\":\"grok-4.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Return only the verification code from the PDF.\"}]},{\"type\":\"function_call\",\"call_id\":\"call_pdf_1\",\"name\":\"read_pdf\",\"arguments\":\"{}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_pdf_1\",\"output\":[{\"type\":\"input_text\",\"text\":\"PDF read successfully\"},{\"type\":\"input_file\",\"filename\":\"verification.pdf\",\"file_data\":\"data:application/pdf;base64,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\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"read_pdf\",\"description\":\"Read the attached PDF.\",\"parameters\":{\"type\":\"object\",\"properties\":{},\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"include\":[\"reasoning.encrypted_content\"],\"max_output_tokens\":40,\"temperature\":0,\"stream\":true,\"instructions\":\"Read the PDF returned by the tool and follow the user's response format exactly.\"}"
       },
       "response": {
         "status": 200,

packages/ai/test/provider-package.test.ts

@@ -26,6 +26,9 @@ describe("provider package entrypoints", () => {
       import("@opencode-ai/ai/providers/amazon-bedrock/mantle"),
       import("@opencode-ai/ai/providers/amazon-bedrock/mantle/chat"),
       import("@opencode-ai/ai/providers/amazon-bedrock/mantle/responses"),
+      import("@opencode-ai/ai/providers/togetherai"),
+      import("@opencode-ai/ai/providers/cerebras"),
+      import("@opencode-ai/ai/providers/deepinfra"),
     ])
 
     for (const module of modules) expect(module.model).toBeFunction()
@@ -35,6 +38,24 @@ describe("provider package entrypoints", () => {
     expect(modules[19].model).toBe(modules[20].model)
   })
 
+  test("maps DeepInfra package settings onto its native executable model", async () => {
+    const DeepInfra = await import("@opencode-ai/ai/providers/deepinfra")
+    const settings = {
+      apiKey: "fixture",
+      baseURL: "https://provider.example.test/v1/",
+      headers: { "x-application": "opencode" },
+      body: { service_tier: "priority" },
+      providerOptions: { reasoningEffort: "high" as const },
+    }
+    const deepinfra = DeepInfra.model("google/gemma-3-27b-it", settings)
+
+    expect(deepinfra.route.id).toBe("deepinfra-chat")
+    expect(deepinfra.route.endpoint.baseURL).toBe("https://provider.example.test/v1/openai")
+    expect(deepinfra.route.defaults.providerOptions).toEqual(settings.providerOptions)
+    expect(deepinfra.route.defaults.headers).toEqual(settings.headers)
+    expect(deepinfra.route.defaults.http?.body).toEqual(settings.body)
+  })
+
   test("maps OpenRouter and xAI package settings onto executable models", async () => {
     const OpenRouter = await import("@opencode-ai/ai/providers/openrouter")
     const XAI = await import("@opencode-ai/ai/providers/xai")

packages/ai/test/provider/anthropic-messages.test.ts

@@ -5,6 +5,7 @@ import { CacheHint, LLM, AIError, LLMRequest, Message, ToolCallPart, ToolDefinit
 import { Auth, LLMClient } from "../../src/route.js"
 import { compileRequest } from "../../src/route/client.js"
 import * as AnthropicMessages from "../../src/protocols/anthropic-messages.js"
+import { GoogleVertexMessages } from "../../src/providers.js"
 import { continuationRequest, nativeAnthropicMessagesContinuation } from "../continuation-scenarios.js"
 import { it } from "../lib/effect.js"
 import { dynamicResponse, fixedResponse } from "../lib/http.js"
@@ -27,6 +28,12 @@ const compileUnsignedReasoning = (model: LLMRequest["model"]) =>
     }),
   )
 
+const vertexOpus48 = GoogleVertexMessages.configure({
+  accessToken: "test",
+  location: "global",
+  project: "test",
+}).model("claude-opus-4-8")
+
 const request = LLM.request({
   id: "req_1",
   model,
@@ -286,6 +293,149 @@ describe("Anthropic Messages route", () => {
     }),
   )
 
+  it.effect("keeps a terminal Vertex system update in the tool-result turn", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: vertexOpus48,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })]),
+            Message.tool({ id: "call_1", name: "lookup", result: "Done." }),
+            Message.system("Operator update."),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ type: "tool_use", id: "call_1", name: "lookup", input: {} }],
+        },
+        {
+          role: "user",
+          content: [
+            {
+              type: "tool_result",
+              tool_use_id: "call_1",
+              content: '"Done."',
+              is_error: undefined,
+              cache_control: undefined,
+            },
+            {
+              type: "text",
+              text: "<system-update>\nOperator update.\n</system-update>",
+              cache_control: undefined,
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("preserves folded tool-result system updates across multi-turn Vertex history", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: vertexOpus48,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })]),
+            Message.tool({ id: "call_1", name: "lookup", result: "Done." }),
+            Message.system("Operator update."),
+            Message.assistant("Acknowledged."),
+            Message.user("Next step."),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ type: "tool_use", id: "call_1", name: "lookup", input: {} }],
+        },
+        {
+          role: "user",
+          content: [
+            {
+              type: "tool_result",
+              tool_use_id: "call_1",
+              content: '"Done."',
+              is_error: undefined,
+              cache_control: undefined,
+            },
+            {
+              type: "text",
+              text: "<system-update>\nOperator update.\n</system-update>",
+              cache_control: undefined,
+            },
+          ],
+        },
+        { role: "assistant", content: [{ type: "text", text: "Acknowledged." }] },
+        { role: "user", content: [{ type: "text", text: "Next step." }] },
+      ])
+    }),
+  )
+
+  it.effect("keeps a terminal direct Anthropic system update native", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: opus48,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })]),
+            Message.tool({ id: "call_1", name: "lookup", result: "Done." }),
+            Message.system("Operator update."),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ type: "tool_use", id: "call_1", name: "lookup", input: {} }],
+        },
+        {
+          role: "user",
+          content: [
+            {
+              type: "tool_result",
+              tool_use_id: "call_1",
+              content: '"Done."',
+              is_error: undefined,
+              cache_control: undefined,
+            },
+          ],
+        },
+        {
+          role: "system",
+          content: [{ type: "text", text: "Operator update.", cache_control: undefined }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("keeps an ordinary terminal Vertex system update native", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: vertexOpus48,
+          messages: [Message.user("Before."), Message.system("Operator update.")],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        { role: "user", content: [{ type: "text", text: "Before." }] },
+        {
+          role: "system",
+          content: [{ type: "text", text: "Operator update.", cache_control: undefined }],
+        },
+      ])
+    }),
+  )
+
   it.effect("rejects a system update between a local tool call and its result", () =>
     Effect.gen(function* () {
       const error = yield* compileRequest(

packages/ai/test/provider/bedrock-mantle.test.ts

@@ -1,11 +1,11 @@
 import { describe, expect } from "bun:test"
 import { Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
-import { LLM } from "../../src/index.js"
+import { LLM, Message } from "../../src/index.js"
 import { AmazonBedrockMantle } from "../../src/providers.js"
 import { compileRequest, LLMClient } from "../../src/route/client.js"
 import { it } from "../lib/effect.js"
-import { dynamicResponse } from "../lib/http.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
 import { sseEvents } from "../lib/sse.js"
 import { recordedTests } from "../recorded-test.js"
 
@@ -59,6 +59,27 @@ describe("Amazon Bedrock Mantle provider", () => {
     }),
   )
 
+  it.effect("keeps Responses on HTTP when a WebSocket executor is available", () =>
+    Effect.gen(function* () {
+      const model = AmazonBedrockMantle.configure({ apiKey: "test-key" }).responses("openai.gpt-oss-120b")
+
+      yield* LLMClient.generate(LLM.request({ model, prompt: "Hi" }), {
+        webSocket: { execute: () => Effect.die("unexpected Mantle WebSocket request") },
+      }).pipe(
+        Effect.provide(
+          dynamicResponse((input) => {
+            expect(input.request.url).toBe("https://bedrock-mantle.us-east-1.api.aws/v1/responses")
+            return Effect.succeed(
+              input.respond(sseEvents({ type: "response.completed", response: {} }), {
+                headers: { "content-type": "text/event-stream" },
+              }),
+            )
+          }),
+        ),
+      )
+    }),
+  )
+
   it.effect("supports bearer authentication and custom base URLs", () =>
     Effect.gen(function* () {
       const seen: Array<{ readonly url: string; readonly authorization: string | undefined }> = []
@@ -83,6 +104,39 @@ describe("Amazon Bedrock Mantle provider", () => {
       expect(seen).toEqual([{ url: "https://mantle.test/v1/chat/completions", authorization: "Bearer test-key" }])
     }),
   )
+
+  it.effect("replays reasoning with Mantle's message-prefixed item ids", () =>
+    Effect.gen(function* () {
+      const model = AmazonBedrockMantle.configure({ apiKey: "test-key" }).responses("openai.gpt-oss-120b")
+      const item = { type: "reasoning", id: "msg_95d4d0af4350432a", encrypted_content: "mantle-state" }
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Think." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", item },
+              { type: "response.reasoning_summary_text.delta", item_id: item.id, delta: "Considering." },
+              { type: "response.output_item.done", item },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      const prepared = yield* compileRequest(
+        LLM.request({ model, messages: [response.message, Message.user("Continue.")] }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        {
+          type: "reasoning",
+          id: "msg_95d4d0af4350432a",
+          summary: [{ type: "summary_text", text: "Considering." }],
+          encrypted_content: "mantle-state",
+        },
+        { role: "user", content: [{ type: "input_text", text: "Continue." }] },
+      ])
+    }),
+  )
 })
 
 const recorded = recordedTests({

packages/ai/test/provider/golden.recorded.test.ts

@@ -1,5 +1,6 @@
 import * as Anthropic from "../../src/providers/anthropic.js"
 import * as AnthropicCompatible from "../../src/providers/anthropic-compatible.js"
+import { Cerebras, DeepInfra, TogetherAI } from "../../src/providers/index.js"
 import { CloudflareAIGateway, CloudflareWorkersAI } from "../../src/providers/cloudflare.js"
 import * as Google from "../../src/providers/google.js"
 import * as OpenAI from "../../src/providers/openai.js"
@@ -47,14 +48,16 @@ const cloudflareWorkersAITools = cloudflareWorkers.model("@cf/openai/gpt-oss-20b
 const deepseek = OpenAICompatible.deepseek
   .configure({ apiKey: process.env.DEEPSEEK_API_KEY ?? "fixture" })
   .model("deepseek-chat")
-const together = OpenAICompatible.togetherai
-  .configure({
-    apiKey: process.env.TOGETHER_AI_API_KEY ?? "fixture",
-  })
-  .model("meta-llama/Llama-3.3-70B-Instruct-Turbo")
+const together = TogetherAI.configure({
+  apiKey: process.env.TOGETHER_API_KEY ?? process.env.TOGETHER_AI_API_KEY ?? "fixture",
+}).model("meta-llama/Llama-3.3-70B-Instruct-Turbo")
+const cerebras = Cerebras.configure({ apiKey: process.env.CEREBRAS_API_KEY ?? "fixture" }).model("gpt-oss-120b")
 const groq = OpenAICompatible.groq
   .configure({ apiKey: process.env.GROQ_API_KEY ?? "fixture" })
   .model("llama-3.3-70b-versatile")
+const deepInfra = DeepInfra.configure({ apiKey: process.env.DEEPINFRA_API_KEY ?? "fixture" }).model(
+  "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+)
 const openRouter = OpenRouter.configure({ apiKey: process.env.OPENROUTER_API_KEY ?? "fixture" })
 const openrouter = openRouter.model("openai/gpt-4o-mini")
 const openrouterGpt55 = openRouter.model("openai/gpt-5.5")
@@ -193,8 +196,27 @@ describeRecordedGoldenScenarios([
     name: "TogetherAI Llama 3.3 70B",
     prefix: "openai-compatible-chat",
     model: together,
-    requires: ["TOGETHER_AI_API_KEY"],
-    scenarios: ["text", "tool-call"],
+    requires: ["TOGETHER_API_KEY"],
+    scenarios: [
+      {
+        id: "text",
+        cassette: "openai-compatible-chat/togetherai-streams-text",
+        prompt: "Reply with exactly: Hello!",
+        maxTokens: 20,
+      },
+      { id: "tool-call", cassette: "openai-compatible-chat/togetherai-streams-tool-call" },
+    ],
+  },
+  {
+    name: "Cerebras GPT OSS 120B",
+    prefix: "cerebras-chat",
+    model: cerebras,
+    requires: ["CEREBRAS_API_KEY"],
+    scenarios: [
+      { id: "text", maxTokens: 256, temperature: false },
+      { id: "tool-call", maxTokens: 512, temperature: false },
+      { id: "tool-loop", maxTokens: 512, temperature: false, timeout: 30_000 },
+    ],
   },
   {
     name: "Groq Llama 3.3 70B",
@@ -203,6 +225,13 @@ describeRecordedGoldenScenarios([
     requires: ["GROQ_API_KEY"],
     scenarios: ["text", "tool-call", { id: "tool-loop", timeout: 30_000 }],
   },
+  {
+    name: "DeepInfra Llama 3.3 70B",
+    prefix: "deepinfra-chat",
+    model: deepInfra,
+    requires: ["DEEPINFRA_API_KEY"],
+    scenarios: ["text", "tool-call", { id: "tool-loop", timeout: 30_000 }],
+  },
   {
     name: "OpenRouter gpt-4o-mini",
     prefix: "openai-compatible-chat",

packages/ai/test/provider/native-providers.test.ts

@@ -0,0 +1,184 @@
+import { describe, expect } from "bun:test"
+import { ConfigProvider, Effect } from "effect"
+import { HttpClientRequest } from "effect/unstable/http"
+import { LLM, Message, ToolDefinition } from "../../src/index.js"
+import { Cerebras, DeepInfra, TogetherAI } from "../../src/providers/index.js"
+import { compileRequest } from "../../src/route/client.js"
+import { it } from "../lib/effect.js"
+import { dynamicResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
+
+describe("native OpenAI-compatible providers", () => {
+  it.effect("preserves native Together AI and Cerebras provider and route identities", () =>
+    Effect.gen(function* () {
+      const together = TogetherAI.configure({ apiKey: "fixture" }).model("meta-llama/Llama-3.3-70B")
+      const cerebras = Cerebras.configure({ apiKey: "fixture" }).model("qwen-3-235b-a22b")
+
+      expect(together).toMatchObject({
+        provider: "togetherai",
+        compatibility: { maxTokensField: "max_tokens", supportsStore: false, supportsStrictMode: false },
+        route: { id: "togetherai-chat", protocol: "openai-chat" },
+      })
+      expect(together.route.endpoint.baseURL).toBe("https://api.together.xyz/v1")
+      expect(cerebras).toMatchObject({
+        provider: "cerebras",
+        compatibility: { maxTokensField: "max_tokens", reasoningField: "reasoning", supportsStore: false },
+        route: { id: "cerebras-chat", protocol: "openai-chat" },
+      })
+      expect(cerebras.route.endpoint.baseURL).toBe("https://api.cerebras.ai/v1")
+    }),
+  )
+
+  it.effect("preserves native DeepInfra provider and route identity", () =>
+    Effect.gen(function* () {
+      const deepinfra = DeepInfra.configure({ apiKey: "fixture" }).model("google/gemma-3-27b-it")
+      expect(deepinfra).toMatchObject({
+        provider: "deepinfra",
+        compatibility: { maxTokensField: "max_tokens", reasoningField: "reasoning_content", supportsStore: false },
+        route: { id: "deepinfra-chat", protocol: "openai-chat" },
+      })
+      expect(deepinfra.route.endpoint.baseURL).toBe("https://api.deepinfra.com/v1/openai")
+    }),
+  )
+
+  it.effect("applies native provider request defaults even with a custom gateway URL", () =>
+    Effect.gen(function* () {
+      const together = yield* compileRequest(
+        LLM.request({
+          model: TogetherAI.configure({ apiKey: "fixture", baseURL: "https://gateway.example/v1" }).model("llama"),
+          prompt: "Use a tool.",
+          generation: { maxTokens: 32 },
+          tools: [
+            ToolDefinition.make({ name: "lookup", description: "Look up data", inputSchema: { type: "object" } }),
+          ],
+          providerOptions: { store: true },
+        }),
+      )
+
+      expect(together.body).toMatchObject({
+        max_tokens: 32,
+        stream_options: { include_usage: true },
+        tools: [{ function: { name: "lookup" } }],
+      })
+      expect(together.body).not.toHaveProperty("max_completion_tokens")
+      expect(together.body).not.toHaveProperty("store")
+      expect(together.body.tools?.[0]?.function).not.toHaveProperty("strict")
+
+      const cerebras = yield* compileRequest(
+        LLM.request({
+          model: Cerebras.configure({ apiKey: "fixture", baseURL: "https://gateway.example/v1" }).model("qwen"),
+          generation: { maxTokens: 48 },
+          messages: [
+            Message.user("Think first."),
+            Message.assistant([
+              { type: "reasoning", text: "A deliberate thought." },
+              { type: "text", text: "An answer." },
+            ]),
+            Message.user("Continue."),
+          ],
+          providerOptions: { store: true },
+        }),
+      )
+
+      expect(cerebras.body).toMatchObject({
+        max_tokens: 48,
+        messages: [
+          { role: "user", content: "Think first." },
+          { role: "assistant", content: "An answer.", reasoning: "A deliberate thought." },
+          { role: "user", content: "Continue." },
+        ],
+      })
+      expect(cerebras.body).not.toHaveProperty("max_completion_tokens")
+      expect(cerebras.body).not.toHaveProperty("store")
+      expect(cerebras.body.messages[1]).not.toHaveProperty("reasoning_content")
+    }),
+  )
+
+  it.effect("normalizes DeepInfra API roots without duplicating the OpenAI path", () =>
+    Effect.gen(function* () {
+      for (const baseURL of [
+        "https://gateway.example/v1",
+        "https://gateway.example/v1/",
+        "https://gateway.example/v1/openai",
+        "https://gateway.example/v1/openai/",
+      ]) {
+        expect(DeepInfra.configure({ apiKey: "fixture", baseURL }).model("gemma").route.endpoint.baseURL).toBe(
+          "https://gateway.example/v1/openai",
+        )
+      }
+    }),
+  )
+
+  it.effect("maps package settings onto native executable models", () =>
+    Effect.gen(function* () {
+      for (const native of [TogetherAI, Cerebras]) {
+        const selected = native.model("provider-model", {
+          apiKey: "fixture",
+          baseURL: "https://gateway.example/v1",
+          headers: { "x-application": "opencode" },
+          body: { service_tier: "priority" },
+          providerOptions: { reasoningEffort: "high" },
+        })
+
+        expect(selected.route.endpoint.baseURL).toBe("https://gateway.example/v1")
+        expect(selected.route.defaults.headers).toEqual({ "x-application": "opencode" })
+        expect(selected.route.defaults.http?.body).toEqual({ service_tier: "priority" })
+        expect(selected.route.defaults.providerOptions).toEqual({ reasoningEffort: "high" })
+      }
+    }),
+  )
+
+  it.effect("resolves provider environment credentials and preserves deprecated Together credentials", () =>
+    Effect.gen(function* () {
+      const scenarios = [
+        {
+          model: TogetherAI.configure().model("llama"),
+          env: { TOGETHER_API_KEY: "together-primary", TOGETHER_AI_API_KEY: "together-legacy" },
+          token: "together-primary",
+          url: "https://api.together.xyz/v1/chat/completions",
+        },
+        {
+          model: TogetherAI.configure().model("llama"),
+          env: { TOGETHER_AI_API_KEY: "together-legacy" },
+          token: "together-legacy",
+          url: "https://api.together.xyz/v1/chat/completions",
+        },
+        {
+          model: Cerebras.configure().model("qwen"),
+          env: { CEREBRAS_API_KEY: "cerebras-secret" },
+          token: "cerebras-secret",
+          url: "https://api.cerebras.ai/v1/chat/completions",
+        },
+        {
+          model: DeepInfra.configure().model("gemma"),
+          env: { DEEPINFRA_API_KEY: "deepinfra-secret" },
+          token: "deepinfra-secret",
+          url: "https://api.deepinfra.com/v1/openai/chat/completions",
+        },
+      ]
+
+      yield* Effect.forEach(scenarios, (scenario) =>
+        LLM.generate(LLM.request({ model: scenario.model, prompt: "Say hello." })).pipe(
+          Effect.provide(
+            dynamicResponse((input) =>
+              Effect.gen(function* () {
+                const request = yield* HttpClientRequest.toWeb(input.request).pipe(Effect.orDie)
+                expect(request.url).toBe(scenario.url)
+                expect(request.headers.get("authorization")).toBe(`Bearer ${scenario.token}`)
+                return input.respond(
+                  sseEvents(
+                    { id: "chatcmpl_fixture", choices: [{ delta: { content: "Hello" }, finish_reason: null }] },
+                    { id: "chatcmpl_fixture", choices: [{ delta: {}, finish_reason: "stop" }] },
+                  ),
+                  { headers: { "content-type": "text/event-stream" } },
+                )
+              }),
+            ),
+          ),
+          Effect.provide(ConfigProvider.layer(ConfigProvider.fromEnv({ env: scenario.env }))),
+          Effect.tap((response) => Effect.sync(() => expect(response.text).toBe("Hello"))),
+        ),
+      )
+    }),
+  )
+})

packages/ai/test/provider/openai-chat.test.ts

@@ -169,6 +169,56 @@ describe("OpenAI Chat route", () => {
     }),
   )
 
+  it.effect("preserves observed reasoning fields when reasoning is required", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: LanguageModel.update(model, { compatibility: { requireReasoning: true } }),
+          messages: [
+            Message.assistant([
+              {
+                type: "reasoning",
+                text: "thinking",
+                providerMetadata: { openai: { reasoningField: "reasoning_text" } },
+              },
+              { type: "text", text: "Hello" },
+            ]),
+            Message.assistant("Done"),
+          ],
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        { role: "assistant", content: "Hello", reasoning_text: "thinking" },
+        { role: "assistant", content: "Done", reasoning_content: "" },
+      ])
+    }),
+  )
+
+  it.effect("omits empty configured reasoning fields when reasoning is explicitly optional", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: LanguageModel.update(model, {
+            compatibility: { reasoningField: "reasoning_text", requireReasoning: false },
+          }),
+          messages: [
+            Message.assistant([
+              { type: "reasoning", text: "thinking" },
+              { type: "text", text: "Hello" },
+            ]),
+            Message.assistant("Done"),
+          ],
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        { role: "assistant", content: "Hello", reasoning_text: "thinking" },
+        { role: "assistant", content: "Done" },
+      ])
+    }),
+  )
+
   it.effect("rejects reasoning fields that conflict with assistant message fields", () =>
     Effect.gen(function* () {
       const error = yield* compileRequest(

packages/ai/test/provider/openai-compatible-chat.test.ts

@@ -317,6 +317,56 @@ describe("OpenAI-compatible Chat route", () => {
     }),
   )
 
+  it.effect("requires reasoning for DeepSeek models, providers, and endpoints unless explicitly overridden", () =>
+    Effect.gen(function* () {
+      const cases = [
+        { id: "DeepSeek-V3", provider: "custom", baseURL: "https://api.custom.test/v1", required: true },
+        { id: "custom-model", provider: "deepseek", baseURL: "https://api.custom.test/v1", required: true },
+        { id: "custom-model", provider: "custom", baseURL: "https://API.DeepSeek.COM/v1", required: true },
+        { id: "ordinary-model", provider: "custom", baseURL: "https://api.custom.test/v1", required: false },
+        {
+          id: "ordinary-model",
+          provider: "custom",
+          baseURL: "https://api.custom.test/v1",
+          compatibility: { requireReasoning: true, reasoningField: "reasoning" },
+          required: true,
+          field: "reasoning",
+        },
+        {
+          id: "deepseek-chat",
+          provider: "deepseek",
+          baseURL: "https://api.deepseek.com/v1",
+          compatibility: { requireReasoning: false },
+          required: false,
+        },
+      ] as const
+
+      yield* Effect.forEach(cases, (item) =>
+        Effect.gen(function* () {
+          const selected = OpenAICompatibleChat.route
+            .with({ provider: item.provider, endpoint: { baseURL: item.baseURL } })
+            .model({ id: item.id, compatibility: "compatibility" in item ? item.compatibility : undefined })
+          const prepared = yield* compileRequest(
+            LLM.request({
+              model: selected,
+              messages: [
+                Message.assistant("Hello"),
+                Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })]),
+                Message.tool({ id: "call_1", name: "lookup", result: "Sunny" }),
+              ],
+            }),
+          )
+          const field = "field" in item ? item.field : "reasoning_content"
+
+          for (const message of prepared.body.messages.filter((message) => message.role === "assistant")) {
+            if (item.required) expect(message).toHaveProperty(field, "")
+            else expect(message).not.toHaveProperty(field)
+          }
+        }),
+      )
+    }),
+  )
+
   it.effect("posts to the configured compatible endpoint and parses text usage", () =>
     Effect.gen(function* () {
       const response = yield* LLMClient.generate(request).pipe(

packages/ai/test/provider/openai-compatible-responses.test.ts

@@ -47,10 +47,8 @@ describe("Open Responses-compatible route", () => {
       })
       expect(prepared.body).toEqual({
         model: "example-model",
-        input: [
-          { role: "system", content: "You are concise." },
-          { role: "user", content: [{ type: "input_text", text: "Say hello." }] },
-        ],
+        input: [{ role: "user", content: [{ type: "input_text", text: "Say hello." }] }],
+        instructions: "You are concise.",
         stream: true,
         store: false,
         include: ["reasoning.encrypted_content"],
@@ -84,10 +82,12 @@ describe("Open Responses-compatible route", () => {
       const prepared = yield* compileRequest(
         LLM.request({
           model,
+          system: "Initial instructions.",
           messages: [Message.user("Before."), Message.system("Operator update."), Message.assistant("After.")],
         }),
       )
 
+      expect(prepared.body.instructions).toBe("Initial instructions.")
       expect(prepared.body.input).toEqual([
         { role: "user", content: [{ type: "input_text", text: "Before." }] },
         { role: "developer", content: "Operator update." },
@@ -195,15 +195,19 @@ describe("Open Responses-compatible route", () => {
           model,
           messages: [
             Message.assistant([
-              // The baseline does not enforce a provider id grammar, so a
-              // non-OpenAI but well-formed token is resent as-is.
               { type: "text", text: "Kept.", providerMetadata: { openresponses: { itemId: "history_1" } } },
-              // Shape violations are dropped even without a grammar policy.
               {
                 type: "text",
-                text: "Dropped.",
-                providerMetadata: { openresponses: { itemId: `m${"a".repeat(64)}` } },
+                text: "Long.",
+                providerMetadata: { openresponses: { itemId: `history_${"a".repeat(64)}` } },
               },
+              {
+                type: "text",
+                text: "Opaque.",
+                providerMetadata: { openresponses: { itemId: "provider_value/with+symbols" } },
+              },
+              { type: "text", text: "No suffix.", providerMetadata: { openresponses: { itemId: "msg_" } } },
+              { type: "text", text: "No prefix.", providerMetadata: { openresponses: { itemId: "_item" } } },
             ]),
           ],
         }),
@@ -218,8 +222,23 @@ describe("Open Responses-compatible route", () => {
         },
         {
           type: "message",
+          id: `history_${"a".repeat(64)}`,
           role: "assistant",
-          content: [{ type: "output_text", text: "Dropped." }],
+          content: [{ type: "output_text", text: "Long." }],
+        },
+        {
+          type: "message",
+          id: "provider_value/with+symbols",
+          role: "assistant",
+          content: [{ type: "output_text", text: "Opaque." }],
+        },
+        {
+          type: "message",
+          role: "assistant",
+          content: [
+            { type: "output_text", text: "No suffix." },
+            { type: "output_text", text: "No prefix." },
+          ],
         },
       ])
     }),
@@ -288,6 +307,43 @@ describe("Open Responses-compatible route", () => {
     }),
   )
 
+  it.effect("streams function calls without optional item ids through the shared baseline", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+        provider: "example",
+      }).model("example-model")
+      const item = { type: "function_call", call_id: "call_1", name: "lookup", arguments: "" }
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Look it up." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 1, item },
+              {
+                type: "response.function_call_arguments.delta",
+                output_index: 1,
+                item_id: "opaque_item",
+                delta: '{"query":"shared"}',
+              },
+              {
+                type: "response.output_item.done",
+                output_index: 1,
+                item: { ...item, arguments: '{"query":"complete"}' },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
+        expect.objectContaining({ id: "call_1", name: "lookup", input: { query: "complete" } }),
+      ])
+      expect(response.events.find(LLMEvent.is.toolCall)?.providerMetadata).toBeUndefined()
+    }),
+  )
+
   it.effect("finalizes pending function calls from completed response output", () =>
     Effect.gen(function* () {
       const model = configure({

packages/ai/test/provider/openai-responses-websocket.recorded.test.ts

@@ -136,6 +136,7 @@ describe("OpenAI Responses WebSocket recorded", () => {
       expect(channel.opens()).toBe(1)
       expect(channel.sent).toHaveLength(2)
       expect(channel.sent[1]).toMatchObject({
+        instructions: "Call get_weather once, then reply exactly: Paris is sunny.",
         previous_response_id: expect.any(String),
         input: [{ type: "function_call_output", call_id: call.id, output: expect.any(String) }],
       })
@@ -167,8 +168,8 @@ describe("OpenAI Responses WebSocket recorded", () => {
       expect(channel.opens()).toBe(2)
       expect(channel.sent[1]).not.toHaveProperty("previous_response_id")
       expect(channel.sent[1]).toMatchObject({
+        instructions: "Follow the user's exact reply instruction.",
         input: [
-          { role: "system", content: "Follow the user's exact reply instruction." },
           { role: "user", content: [{ type: "input_text", text: "Reply exactly: Alpha." }] },
           { role: "assistant", content: [{ type: "output_text", text: "Alpha." }] },
           { role: "user", content: [{ type: "input_text", text: "Reply exactly: Beta." }] },
@@ -204,8 +205,8 @@ describe("OpenAI Responses WebSocket recorded", () => {
       expect(channel.sent[1]).toHaveProperty("previous_response_id", expect.any(String))
       expect(channel.sent[2]).not.toHaveProperty("previous_response_id")
       expect(channel.sent[2]).toMatchObject({
+        instructions: "Follow the user's exact reply instruction.",
         input: [
-          { role: "system", content: "Follow the user's exact reply instruction." },
           { role: "user", content: [{ type: "input_text", text: "Reply exactly: Ready." }] },
           { role: "assistant", content: [{ type: "output_text", text: "Ready." }] },
           { role: "user", content: [{ type: "input_text", text: "Reply exactly: Recovered." }] },

packages/ai/test/provider/openai-responses.test.ts

@@ -112,10 +112,8 @@ describe("OpenAI Responses route", () => {
 
       expect(prepared.body).toEqual({
         model: "gpt-4.1-mini",
-        input: [
-          { role: "system", content: "You are concise." },
-          { role: "user", content: [{ type: "input_text", text: "Say hello." }] },
-        ],
+        input: [{ role: "user", content: [{ type: "input_text", text: "Say hello." }] }],
+        instructions: "You are concise.",
         store: false,
         include: ["reasoning.encrypted_content"],
         stream: true,
@@ -469,7 +467,7 @@ describe("OpenAI Responses route", () => {
     }),
   )
 
-  it.effect("continues a tool call with only the new tool output", () =>
+  it.effect("continues an item-id-less tool call with only the new tool output", () =>
     Effect.gen(function* () {
       const firstRequest = {
         type: "response.create",
@@ -485,7 +483,6 @@ describe("OpenAI Responses route", () => {
           type: "response.output_item.done",
           item: {
             type: "function_call",
-            id: "fc_1",
             status: "completed",
             call_id: "call_1",
             name: "weather",
@@ -1597,8 +1594,8 @@ describe("OpenAI Responses route", () => {
       )
 
       expect(prepared.body).toMatchObject({
+        instructions: "You are concise. Continue from the provided history.",
         input: [
-          { role: "system", content: "You are concise. Continue from the provided history." },
           {
             role: "user",
             content: [
@@ -2120,6 +2117,47 @@ describe("OpenAI Responses route", () => {
     }),
   )
 
+  it.effect("routes item-id-less function arguments by output index and prefers item completion", () =>
+    Effect.gen(function* () {
+      const item = { type: "function_call", call_id: "call_1", name: "lookup", arguments: "" }
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 2, item },
+              {
+                type: "response.function_call_arguments.delta",
+                output_index: 2,
+                item_id: "opaque_delta",
+                delta: '{"query":"streamed"}',
+              },
+              {
+                type: "response.function_call_arguments.done",
+                output_index: 2,
+                item_id: "opaque_done",
+                arguments: '{"query":"arguments-done"}',
+              },
+              {
+                type: "response.output_item.done",
+                output_index: 2,
+                item: { ...item, arguments: '{"query":"output-item-done"}' },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter((event) => event.type === "tool-input-delta")).toMatchObject([
+        { id: "call_1", text: '{"query":"streamed"}' },
+      ])
+      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
+        expect.objectContaining({ id: "call_1", name: "lookup", input: { query: "output-item-done" } }),
+      ])
+      expect(response.events.find(LLMEvent.is.toolCall)?.providerMetadata).toBeUndefined()
+    }),
+  )
+
   it.effect("routes reasoning summary events by output index", () =>
     Effect.gen(function* () {
       const response = yield* LLMClient.generate(request).pipe(
@@ -2298,13 +2336,25 @@ describe("OpenAI Responses route", () => {
   it.effect("rejects function argument events without the spec-required item id", () =>
     Effect.gen(function* () {
       const events = [
-        { type: "response.function_call_arguments.delta", delta: "{}" },
-        { type: "response.function_call_arguments.done", arguments: "{}" },
+        { type: "response.function_call_arguments.delta", output_index: 0, delta: "{}" },
+        { type: "response.function_call_arguments.done", output_index: 0, arguments: "{}" },
       ]
 
       for (const event of events) {
         const error = yield* LLMClient.generate(request).pipe(
-          Effect.provide(fixedResponse(sseEvents(event, { type: "response.completed", response: { id: "resp_1" } }))),
+          Effect.provide(
+            fixedResponse(
+              sseEvents(
+                {
+                  type: "response.output_item.added",
+                  output_index: 0,
+                  item: { type: "function_call", call_id: "call_1", name: "lookup", arguments: "" },
+                },
+                event,
+                { type: "response.completed", response: { id: "resp_1" } },
+              ),
+            ),
+          ),
           Effect.flip,
         )
 
@@ -3100,38 +3150,50 @@ describe("OpenAI Responses route", () => {
     }),
   )
 
-  it.effect("drops replayed item ids outside the server's grammar", () =>
+  it.effect("preserves provider-issued item ids and removes malformed ids without dropping items", () =>
     Effect.gen(function* () {
       const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
             Message.assistant([
-              // Fails the message id prefix.
               {
                 type: "text",
                 text: "Hello",
                 providerMetadata: { openai: { itemId: "history_1" } },
               },
-              // Oversized for the Responses item id limit.
               {
                 type: "text",
                 text: "World",
-                providerMetadata: { openai: { itemId: `m${"a".repeat(64)}` } },
+                providerMetadata: { openai: { itemId: `message_${"a".repeat(64)}` } },
               },
-              // Fails the reasoning id prefix, so the whole item is unreplayable
-              // statelessly and is skipped rather than sent malformed.
               {
                 type: "reasoning",
                 text: "Checked the diff.",
                 providerMetadata: { openai: { itemId: "thinking_1", reasoningEncryptedContent: "encrypted-state" } },
               },
+              {
+                type: "reasoning",
+                text: "Missing suffix.",
+                providerMetadata: { openai: { itemId: "rs_", reasoningEncryptedContent: "another-state" } },
+              },
+              {
+                type: "reasoning",
+                text: "No prefix separator.",
+                providerMetadata: { openai: { itemId: "550e8400-e29b-41d4-a716-446655440000" } },
+              },
               ToolCallPart.make({
                 id: "call_1",
                 name: "lookup",
                 input: { query: "weather" },
                 providerMetadata: { openai: { itemId: "toolu_01A" } },
               }),
+              ToolCallPart.make({
+                id: "call_2",
+                name: "lookup",
+                input: { query: "news" },
+                providerMetadata: { openai: { itemId: "fc_" } },
+              }),
             ]),
           ],
         }),
@@ -3140,18 +3202,44 @@ describe("OpenAI Responses route", () => {
       expect(prepared.body.input).toEqual([
         {
           type: "message",
+          id: "history_1",
           role: "assistant",
-          content: [
-            { type: "output_text", text: "Hello" },
-            { type: "output_text", text: "World" },
-          ],
+          content: [{ type: "output_text", text: "Hello" }],
+        },
+        {
+          type: "message",
+          id: `message_${"a".repeat(64)}`,
+          role: "assistant",
+          content: [{ type: "output_text", text: "World" }],
+        },
+        {
+          type: "reasoning",
+          id: "thinking_1",
+          summary: [{ type: "summary_text", text: "Checked the diff." }],
+          encrypted_content: "encrypted-state",
+        },
+        {
+          type: "reasoning",
+          summary: [{ type: "summary_text", text: "Missing suffix." }],
+          encrypted_content: "another-state",
+        },
+        {
+          type: "reasoning",
+          summary: [{ type: "summary_text", text: "No prefix separator." }],
         },
         {
           type: "function_call",
+          id: "toolu_01A",
           call_id: "call_1",
           name: "lookup",
           arguments: '{"query":"weather"}',
         },
+        {
+          type: "function_call",
+          call_id: "call_2",
+          name: "lookup",
+          arguments: '{"query":"news"}',
+        },
       ])
     }),
   )
@@ -3457,6 +3545,43 @@ describe("OpenAI Responses route", () => {
     }),
   )
 
+  it.effect("finalizes and replays a completed function call without an optional item id", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.done",
+                item: { type: "function_call", call_id: "call_1", name: "lookup", arguments: '{"query":"weather"}' },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
+        expect.objectContaining({ id: "call_1", name: "lookup", input: { query: "weather" } }),
+      ])
+      expect(response.events.find(LLMEvent.is.toolCall)?.providerMetadata).toBeUndefined()
+
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            response.message,
+            Message.tool({ id: "call_1", name: "lookup", resultType: "json", result: { forecast: "sunny" } }),
+          ],
+        }),
+      )
+      expect(prepared.body.input).toEqual([
+        { type: "function_call", call_id: "call_1", name: "lookup", arguments: '{"query":"weather"}' },
+        { type: "function_call_output", call_id: "call_1", output: '{"forecast":"sunny"}' },
+      ])
+    }),
+  )
+
   it.effect("emits only missing function arguments from the arguments done event", () =>
     Effect.gen(function* () {
       const body = sseEvents(
@@ -3681,6 +3806,37 @@ describe("OpenAI Responses route", () => {
     }),
   )
 
+  it.effect("reconciles an item-id-less pending function call from completed response output", () =>
+    Effect.gen(function* () {
+      const item = { type: "function_call", call_id: "call_1", name: "lookup", arguments: "" }
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 0, item },
+              {
+                type: "response.function_call_arguments.delta",
+                output_index: 0,
+                item_id: "opaque_delta",
+                delta: '{"query":"partial',
+              },
+              {
+                type: "response.completed",
+                response: { id: "resp_1", output: [{ ...item, arguments: '{"query":"complete"}' }] },
+              },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
+        expect.objectContaining({ id: "call_1", name: "lookup", input: { query: "complete" } }),
+      ])
+      expect(response.events.find(LLMEvent.is.toolCall)?.providerMetadata).toBeUndefined()
+      expect(response.events.filter(LLMEvent.is.toolInputEnd)).toHaveLength(1)
+    }),
+  )
+
   it.effect("lets completed response output override arguments done", () =>
     Effect.gen(function* () {
       const body = sseEvents(

packages/ai/test/recorded-golden.ts

@@ -20,6 +20,7 @@ type ScenarioInput =
       readonly name?: string
       readonly cassette?: string
       readonly tags?: ReadonlyArray<string>
+      readonly prompt?: string
       readonly maxTokens?: number
       readonly temperature?: number | false
       readonly timeout?: number
@@ -87,6 +88,7 @@ const runTarget = (target: TargetInput) => {
             yield* runGoldenScenario(input.id, {
               id: `recorded_${kebab(target.name).replaceAll("-", "_")}_${input.id.replaceAll("-", "_")}`,
               model: target.model,
+              prompt: input.prompt,
               maxTokens: input.maxTokens,
               temperature: input.temperature,
             })

packages/ai/test/recorded-scenarios.ts

@@ -164,6 +164,7 @@ export const expectGoldenWeatherToolLoop = (events: ReadonlyArray<LLMEvent>) =>
 export interface GoldenScenarioContext {
   readonly id: string
   readonly model: LanguageModel
+  readonly prompt?: string
   readonly maxTokens?: number
   readonly temperature?: number | false
 }
@@ -298,7 +299,7 @@ const runGeneratedConversation = (context: GoldenScenarioContext, steps: Readonl
 
 const runTextScenario = (context: GoldenScenarioContext) =>
   runGeneratedConversation(context, [
-    user("Reply exactly with: Hello!"),
+    user(context.prompt ?? "Reply exactly with: Hello!"),
     assistant.expectText(/^Hello!?$/, {
       system: "You are concise.",
       maxTokens: context.maxTokens ?? 40,

packages/ai/test/schema.test.ts

@@ -102,6 +102,38 @@ describe("AI.Usage", () => {
     expect(error.reason._tag).toBe("InvalidProviderOutput")
   })
 
+  test("sseFraming ignores retry directives without ending the stream", async () => {
+    const encoder = new TextEncoder()
+    const frames = await Effect.runPromise(
+      ProviderShared.sseFraming(
+        Stream.make(
+          encoder.encode("retry: 1000\n\n"),
+          encoder.encode('data: {"first":true}\n\n'),
+          encoder.encode("retry: 2000\n\n"),
+          encoder.encode('data: {"second":true}\n\n'),
+        ).pipe(Stream.rechunk(1)),
+      ).pipe(Stream.runCollect),
+    )
+
+    expect(Array.from(frames)).toEqual(['{"first":true}', '{"second":true}'])
+  })
+
+  test("sseFraming preserves event data around retry directives", async () => {
+    const encoder = new TextEncoder()
+    const frames = await Effect.runPromise(
+      ProviderShared.sseFraming(
+        Stream.make(
+          encoder.encode("event: update\ndata: first\n"),
+          encoder.encode("retry: 1000\n"),
+          encoder.encode("data: second\n\n"),
+        ).pipe(Stream.rechunk(1)),
+        new Set(["update"]),
+      ).pipe(Stream.runCollect),
+    )
+
+    expect(Array.from(frames)).toEqual(["first\nsecond"])
+  })
+
   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)

packages/app/src/session/timeline/virtualizer.tsx

@@ -92,6 +92,15 @@ export function createTimelineVirtualizer(input: Input) {
     initialOffset: () => (input.pinned() ? Number.MAX_SAFE_INTEGER : 0),
     initialMeasurementsCache: initialMeasurements,
     estimateSize: () => fallbackItemSize,
+    // Do not replace this with TanStack's default measurer: without a ResizeObserver entry,
+    // it returns the cached height instead of reading the element (TanStack/virtual#1183).
+    // Restored sessions, deferred tools, and rewrapped content can then keep stale heights;
+    // our fixed-height, overflow-clipped rows will hide their content. Keep observer entries
+    // on the cheap precomputed path, but make explicit measurements read the real height.
+    measureElement: (element, entry) => {
+      const box = entry?.borderBoxSize[0]
+      return box ? Math.round(box.blockSize) : element.offsetHeight
+    },
     scrollToFn: (offset, options, instance) => {
       if (virtualContent) virtualContent.style.height = `${instance.getTotalSize()}px`
       elementScroll(offset, options, instance)

packages/app/src/settings/workspaces/project-dialog.tsx

@@ -11,7 +11,8 @@ import { useLanguage } from "@/runtime/i18n/language"
 import { getProjectAvatarVariant, type LocalProject } from "@/shell/state/layout"
 import { ServerConnection } from "@/runtime/server/registry"
 import { LocationProvider } from "@/workspaces/location"
-import { displayName, getProjectAvatarSource } from "@/shell/layout/helpers"
+import { displayName } from "@/shell/layout/helpers"
+import { ProjectIcon } from "@/shell/layout/project-icon"
 import { createEditProjectModel } from "./project-model"
 import { ProjectSettingsExtensions } from "./project-extensions"
 import { SettingsServerDataScope } from "@/settings/server-scope"
@@ -57,11 +58,7 @@ function ProjectSettingsDialog(props: { project: LocalProject; server: ServerCon
         <Tabs.List>
           <div class="project-settings-nav">
             <Tabs.Trigger value="general">
-              <ProjectAvatar
-                fallback={projectName()}
-                variant={getProjectAvatarVariant(props.project.icon?.color)}
-                class="!size-4 shrink-0"
-              />
+              <ProjectIcon project={props.project} class="!size-4 shrink-0" />
               <span class="truncate">{projectName()}</span>
             </Tabs.Trigger>
             <Tabs.Trigger value="scripts">
@@ -114,14 +111,14 @@ function ProjectSettingsDialog(props: { project: LocalProject; server: ServerCon
                     onDragLeave={model.dragLeave}
                     onClick={model.iconClick}
                   >
-                    <ProjectAvatar
+                    <ProjectIcon
+                      project={props.project}
                       fallback={model.store.name || model.defaultName()}
-                      src={getProjectAvatarSource(props.project.id, {
+                      icon={{
                         color: model.store.color,
                         url: props.project.icon?.url,
                         override: model.store.iconOverride,
-                      })}
-                      variant={getProjectAvatarVariant(model.store.color)}
+                      }}
                       class="!size-16 [&_[data-slot=project-avatar-surface]]:!rounded-[6px] [&_[data-slot=project-avatar-surface]]:!text-[32px]"
                     />
                     <span

packages/app/src/settings/workspaces/projects.tsx

@@ -1,13 +1,12 @@
 import { Component, For, Show, createMemo, createSignal } from "solid-js"
 import { IconButton } from "@opencode-ai/ui/icon-button"
 import { Icon } from "@opencode-ai/ui/icon"
-import { ProjectAvatar } from "@opencode-ai/ui/project-avatar"
 import { useDialog } from "@opencode-ai/ui/context/dialog"
 import { useLanguage } from "@/runtime/i18n/language"
 import { useGlobal } from "@/runtime/server/runtime"
-import { getProjectAvatarVariant } from "@/shell/state/layout"
 import { ServerConnection, serverName } from "@/runtime/server/registry"
 import { displayName } from "@/shell/layout/helpers"
+import { ProjectIcon } from "@/shell/layout/project-icon"
 import { InlineServerSelect } from "@/settings/server-select"
 import { DialogEditProject } from "./project-dialog"
 import "@/settings/settings.css"
@@ -40,14 +39,13 @@ export const SettingsProjects: Component = () => {
 
   const ProjectRow: Component<{ project: ProjectItem; server: ServerConnection.Any }> = (props) => {
     const name = () => displayName(props.project)
-    const color = () => getProjectAvatarVariant(props.project.icon?.color)
     return (
       <div
         class="group flex items-center justify-between gap-5 px-4 py-2.5 rounded-lg bg-v2-background-bg-base shadow-[var(--v2-elevation-raised)] cursor-pointer transition-all hover:bg-v2-background-bg-layer-01"
         onClick={() => openProjectSettings(props.project, props.server)}
       >
         <div class="flex items-center gap-2.5 min-w-0 flex-1">
-          <ProjectAvatar fallback={name()} variant={color()} class="shrink-0" />
+          <ProjectIcon project={props.project} class="shrink-0" />
           <span class="text-13-medium text-v2-text-text-base truncate">{name()}</span>
         </div>
         <div class="flex items-center gap-2 shrink-0 opacity-0 group-hover:opacity-100 transition-opacity">

packages/app/src/shell/layout/project-icon.tsx

@@ -0,0 +1,24 @@
+import { ProjectAvatar, type ProjectAvatarProps } from "@opencode-ai/ui/project-avatar"
+import { splitProps } from "solid-js"
+import { displayName, getProjectAvatarSource } from "@/shell/layout/helpers"
+import { getProjectAvatarVariant, type LocalProject } from "@/shell/state/layout"
+
+type ProjectIconProps = Omit<ProjectAvatarProps, "fallback" | "src" | "variant"> & {
+  project: LocalProject
+  fallback?: string
+  icon?: LocalProject["icon"]
+}
+
+export function ProjectIcon(props: ProjectIconProps) {
+  const [local, rest] = splitProps(props, ["project", "fallback", "icon"])
+  const icon = () => local.icon ?? local.project.icon
+
+  return (
+    <ProjectAvatar
+      {...rest}
+      fallback={local.fallback ?? displayName(local.project)}
+      src={getProjectAvatarSource(local.project.id, icon())}
+      variant={getProjectAvatarVariant(icon()?.color)}
+    />
+  )
+}

packages/app/src/shell/routes/session-ui-provider.tsx

@@ -6,6 +6,7 @@ import { LocalProvider } from "@/providers/models/selection"
 import type { ServerConnection } from "@/runtime/server/registry"
 import { sessionHref } from "@/shell/routes/session"
 import { useData } from "@/runtime/server/current"
+import { useServerSDK } from "@/runtime/server/client"
 import { useTabs } from "@/shell/tabs/tabs"
 
 export function SessionUIProvider(
@@ -17,6 +18,7 @@ export function SessionUIProvider(
   const navigate = useNavigate()
   const params = useParams()
   const data = useData()
+  const serverSDK = useServerSDK()
   const tabs = useTabs()
   const directory = () => props.directory
   const href = (sessionID: string) => sessionHref(props.server, sessionID)
@@ -53,6 +55,7 @@ export function SessionUIProvider(
       data={sessionUIData()}
       directory={directory()}
       sessionID={params.id}
+      shellOutput={(input) => serverSDK.api.shell.output(input)}
       onNavigateToSession={navigateToSession}
       onSessionHref={href}
     >

packages/app/test-browser/solid-virtual.test.ts

@@ -113,6 +113,32 @@ test("reactive count updates preserve measured row sizes", () => {
   })
 })
 
+test("explicit measurement refreshes a cached row size with a custom measurer", () => {
+  const root = document.createElement("div")
+  const element = document.createElement("div")
+  element.dataset.index = "0"
+  Object.defineProperty(element, "offsetHeight", { value: 120 })
+
+  const virtualizer = new Virtualizer<HTMLDivElement, HTMLDivElement>({
+    count: 1,
+    estimateSize: () => 60,
+    initialRect: { width: 400, height: 200 },
+    getScrollElement: () => root,
+    scrollToFn: () => {},
+    observeElementRect: () => {},
+    observeElementOffset: () => {},
+    measureElement: (node) => node.offsetHeight,
+  })
+
+  virtualizer.getTotalSize()
+  virtualizer.resizeItem(0, 60)
+  virtualizer._willUpdate()
+  virtualizer.measureElement(element)
+
+  expect(virtualizer.itemSizeCache.get(0)).toBe(120)
+  expect(virtualizer.getTotalSize()).toBe(120)
+})
+
 test("initial rect projects rows before a scroll element connects", () => {
   createRoot((dispose) => {
     const virtualizer = createVirtualizer<HTMLDivElement, HTMLDivElement>({

packages/cli/src/commands/commands.ts

@@ -181,7 +181,7 @@ const Root = Spec.make(typeof OPENCODE_CLI_NAME === "string" ? OPENCODE_CLI_NAME
         Spec.make("add", {
           description: "Install a plugin and add it to the global configuration",
           params: {
-            package: Argument.string("package").pipe(Argument.withDescription("npm registry package specifier")),
+            package: Argument.string("package").pipe(Argument.withDescription("npm registry or Git package specifier")),
           },
         }),
         Spec.make("remove", {

packages/cli/src/commands/handlers/plugin/add.ts

@@ -13,10 +13,8 @@ import { Config } from "../../../config"
 export default Runtime.handler(
   Commands.commands.plugin.commands.add,
   Effect.fn("cli.plugin.add")(function* (input) {
-    if (!(yield* Effect.promise(() => Npm.isRegistryPackage(input.package))))
-      return yield* Effect.fail(
-        new Error("Plugin target must be an npm registry package name, version, tag, or semver range"),
-      )
+    if (!(yield* Effect.promise(() => Npm.isInstallablePackage(input.package))))
+      return yield* Effect.fail(new Error("Plugin target must be an npm registry package or Git package specifier"))
     const npm = yield* Npm.Service
     const installed = yield* npm.add(input.package, { subpaths: ["server", ""] })
     const tui = yield* npm.resolve(input.package, { subpaths: ["tui"] })

packages/client/src/promise/generated/types.ts

@@ -476,7 +476,7 @@ export type PromptFileAttachment = {
 
 export type PromptAgentAttachment = { name: string; mention?: PromptMention }
 
-export type PromptSkillAttachment = { id: string; name: string; mention?: PromptMention }
+export type PromptSkillAttachment = { id: string; name: string; text?: string; mention?: PromptMention }
 
 export type SessionMessageAssistantText = { type: "text"; text: string; state?: SessionMessageProviderState }
 
@@ -1319,6 +1319,7 @@ export type ToolContent1 = ToolTextContent | ToolFileContent1
 
 export type ModelCompatibility = {
   reasoningField?: ModelReasoningField
+  requireReasoning?: boolean
   maxTokensField?: ModelMaxTokensField
   requireFinishReason?: boolean
   requireAssistantAfterTool?: boolean
@@ -2132,6 +2133,7 @@ export type SessionMessagesResponse = {
 export type IntegrationInfo = {
   id: string
   name: string
+  metadata?: { [x: string]: any }
   methods: Array<IntegrationMethod>
   connections: Array<ConnectionInfo>
 }
@@ -2738,6 +2740,7 @@ export type SessionImportInput = {
           readonly skills?: ReadonlyArray<{
             readonly id: string
             readonly name: string
+            readonly text?: string
             readonly mention?: { readonly start: number; readonly end: number; readonly text: string }
           }>
           readonly type: "user"
@@ -3013,6 +3016,7 @@ export type SessionImportInput = {
           readonly skills?: ReadonlyArray<{
             readonly id: string
             readonly name: string
+            readonly text?: string
             readonly mention?: { readonly start: number; readonly end: number; readonly text: string }
           }>
           readonly type: "user"
@@ -3288,6 +3292,7 @@ export type SessionImportInput = {
           readonly skills?: ReadonlyArray<{
             readonly id: string
             readonly name: string
+            readonly text?: string
             readonly mention?: { readonly start: number; readonly end: number; readonly text: string }
           }>
           readonly type: "user"

packages/core/package.json

@@ -22,7 +22,7 @@
     "benchmark:location": "bun run script/benchmark-location.ts",
     "build": "bun run script/build.ts",
     "update-models-snapshot": "bun run script/update-models-snapshot.ts",
-    "test": "bun test --only-failures",
+    "test": "bun run script/test.ts",
     "typecheck": "tsgo -b tsconfig.json tsconfig.tests.json"
   },
   "exports": {
@@ -102,9 +102,7 @@
     "@ai-sdk/amazon-bedrock": "4.0.112",
     "@ai-sdk/anthropic": "3.0.82",
     "@ai-sdk/azure": "3.0.88",
-    "@ai-sdk/cerebras": "2.0.41",
     "@ai-sdk/cohere": "3.0.27",
-    "@ai-sdk/deepinfra": "2.0.41",
     "@ai-sdk/gateway": "3.0.104",
     "@ai-sdk/google-vertex": "4.0.128",
     "@ai-sdk/groq": "3.0.31",
@@ -113,7 +111,6 @@
     "@ai-sdk/perplexity": "3.0.26",
     "@ai-sdk/provider": "3.0.8",
     "@ai-sdk/provider-utils": "4.0.23",
-    "@ai-sdk/togetherai": "2.0.41",
     "@ai-sdk/vercel": "2.0.39",
     "@aws-sdk/credential-providers": "3.1057.0",
     "@lydell/node-pty": "catalog:",

packages/core/script/test.ts

@@ -0,0 +1,52 @@
+import fs from "fs/promises"
+import path from "path"
+import { tmpdir } from "../test/fixture/tmpdir"
+
+await using directory = await tmpdir("oc-")
+const home = directory.path
+const temporary = path.join(home, "tmp")
+await fs.mkdir(temporary)
+
+const environment = {
+  ...Object.fromEntries(
+    Object.entries(process.env).filter(([name]) => {
+      if (process.env.RECORD === "true" && name === "OPENAI_API_KEY") return true
+      if (
+        /^(?:AWS|AZURE|GOOGLE|GCP|GCLOUD|VERTEX|OPENAI|ANTHROPIC|GEMINI|XAI|CLOUDFLARE|CF_AIG|SNOWFLAKE|AICORE|GITLAB|NPM_CONFIG)_/i.test(
+          name,
+        )
+      ) {
+        return false
+      }
+      return !/(?:^|_)(?:API_KEY|AUTHORIZATION|TOKEN|SECRET|PASSWORD|CREDENTIALS?)$/i.test(name)
+    }),
+  ),
+  HOME: home,
+  OPENCODE_TEST_HOME: home,
+  XDG_CONFIG_HOME: path.join(home, ".config"),
+  XDG_DATA_HOME: path.join(home, ".local", "share"),
+  XDG_CACHE_HOME: path.join(home, ".cache"),
+  XDG_STATE_HOME: path.join(home, ".local", "state"),
+  OPENCODE_CONFIG_DIR: path.join(home, ".config", "opencode"),
+  OPENCODE_CONFIG: undefined,
+  OPENCODE_CONFIG_CONTENT: undefined,
+  TMPDIR: temporary,
+  ...(process.platform === "win32" ? { USERPROFILE: home, TMP: temporary, TEMP: temporary } : {}),
+}
+
+const child = Bun.spawn({
+  cmd: [process.execPath, "test", "--only-failures", ...process.argv.slice(2)],
+  env: environment,
+  stdin: "inherit",
+  stdout: "inherit",
+  stderr: "inherit",
+})
+const interrupt = () => child.kill("SIGINT")
+const terminate = () => child.kill("SIGTERM")
+process.once("SIGINT", interrupt)
+process.once("SIGTERM", terminate)
+const result = await child.exited
+process.off("SIGINT", interrupt)
+process.off("SIGTERM", terminate)
+
+process.exitCode = result

packages/core/src/aisdk-native.ts

@@ -53,6 +53,18 @@ export function map(input: MapInput): Mapping | undefined {
           ...mapOpenAIOptions(input.settings),
         },
       }
+    case "@ai-sdk/cerebras":
+    case "@ai-sdk/deepinfra":
+    case "@ai-sdk/togetherai":
+      return {
+        package: `@opencode-ai/ai/providers/${input.packageName.slice("@ai-sdk/".length)}`,
+        settings: {
+          ...baseSettings,
+          ...mapAPIKey(input.settings),
+          ...mapProviderOptions(input.settings, ["apiKey", "baseURL", "fetch", "headers", "name"]),
+        },
+        ...(isStringRecord(input.settings.headers) ? { headers: input.settings.headers } : {}),
+      }
     case "@ai-sdk/google":
       return {
         package: "@opencode-ai/ai/providers/google",

packages/core/src/database/drizzle/effect-sqlite/migrator.ts

@@ -1,14 +0,0 @@
-/* oxlint-disable */
-import type { MigrationConfig } from "drizzle-orm/migrator"
-import { readMigrationFiles } from "drizzle-orm/migrator"
-import type { AnyRelations } from "drizzle-orm/relations"
-import { migrate as coreMigrate } from "../sqlite-core/effect/session.js"
-import type { EffectSQLiteDatabase } from "./driver.js"
-
-export function migrate<TRelations extends AnyRelations>(
-  db: EffectSQLiteDatabase<TRelations>,
-  config: MigrationConfig,
-) {
-  const migrations = readMigrationFiles(config)
-  return coreMigrate(migrations, db.session, config)
-}

packages/core/src/database/drizzle/index.ts

@@ -1,6 +1,5 @@
 export { EffectLogger } from "drizzle-orm/effect-core"
 export * from "./effect-sqlite/driver.js"
 export * from "./effect-sqlite/session.js"
-export { migrate } from "./effect-sqlite/migrator.js"
 
 export * as EffectDrizzleSqlite from "./index.js"

packages/core/src/database/drizzle/sqlite-core/effect/session.ts

@@ -5,13 +5,10 @@ import type { SqlError } from "effect/unstable/sql/SqlError"
 import type { EffectCacheShape } from "drizzle-orm/cache/core/cache-effect"
 import { NoopCache, strategyFor } from "drizzle-orm/cache/core/cache"
 import type { WithCacheConfig } from "drizzle-orm/cache/core/types"
-import { MigratorInitError } from "drizzle-orm/effect-core/errors"
 import { EffectDrizzleQueryError, EffectTransactionRollbackError } from "drizzle-orm/effect-core/errors"
 import type { EffectLoggerShape } from "drizzle-orm/effect-core/logger"
 import type { QueryEffectHKTBase, QueryEffectKind } from "drizzle-orm/effect-core/query-effect"
 import { entityKind, is } from "drizzle-orm/entity"
-import type { MigrationConfig, MigrationMeta } from "drizzle-orm/migrator"
-import { getMigrationsToRun } from "drizzle-orm/migrator.utils"
 import type {
   AnyRelations,
   EmptyRelations,
@@ -20,17 +17,14 @@ import type {
 } from "drizzle-orm/relations"
 import { makeJitRqbMapper } from "drizzle-orm/relations"
 import type { PreparedQuery } from "drizzle-orm/session"
-import { fillPlaceholders, type Query, type SQL, sql } from "drizzle-orm/sql/sql"
+import { fillPlaceholders, type Query, type SQL } from "drizzle-orm/sql/sql"
 import type { SQLiteDialect } from "drizzle-orm/sqlite-core/dialect"
 import type { SelectedFieldsOrdered } from "drizzle-orm/sqlite-core/query-builders/select.types"
 import type { PreparedQueryConfig, SQLiteExecuteMethod, SQLiteTransactionConfig } from "drizzle-orm/sqlite-core/session"
-import { upgradeIfNeeded } from "../../up-migrations/effect-sqlite.js"
 import { assertUnreachable, makeJitQueryMapper, type RowsMapper } from "drizzle-orm/utils"
 import { mapResultRow, resolveNullableObjectPaths } from "../../internal/drizzle-utils.js"
 import { SQLiteEffectDatabase } from "./db.js"
 
-type MigrationConfigWithInit = MigrationConfig & { init?: boolean }
-
 type SQLiteEffectExecuteMethod = SQLiteExecuteMethod | "values"
 
 export class SQLiteEffectPreparedQuery<
@@ -429,69 +423,3 @@ export abstract class SQLiteEffectTransaction<
     return new EffectTransactionRollbackError()
   }
 }
-
-export const migrate = Effect.fn("migrate")(function* <TEffectHKT extends QueryEffectHKTBase>(
-  migrations: MigrationMeta[],
-  session: SQLiteEffectSession<TEffectHKT>,
-  config: string | MigrationConfigWithInit,
-) {
-  const migrationsTable =
-    typeof config === "string" ? "__drizzle_migrations" : (config.migrationsTable ?? "__drizzle_migrations")
-
-  const { newDb } = yield* upgradeIfNeeded(migrationsTable, session, migrations)
-
-  if (newDb) {
-    yield* session.run(sql`
-		CREATE TABLE IF NOT EXISTS ${sql.identifier(migrationsTable)} (
-			id INTEGER PRIMARY KEY,
-			hash text NOT NULL,
-			created_at numeric,
-			name text,
-			applied_at TEXT
-		)
-	`)
-  }
-
-  const dbMigrations = yield* session.all<{ id: number; hash: string; created_at: string; name: string | null }>(
-    sql`SELECT id, hash, created_at, name FROM ${sql.identifier(migrationsTable)}`,
-  )
-
-  if (typeof config === "object" && config.init) {
-    if (dbMigrations.length) {
-      return yield* new MigratorInitError({ exitCode: "databaseMigrations" })
-    }
-
-    if (migrations.length > 1) {
-      return yield* new MigratorInitError({ exitCode: "localMigrations" })
-    }
-
-    const [migration] = migrations
-    if (!migration) return
-
-    yield* session.run(
-      sql`insert into ${sql.identifier(
-        migrationsTable,
-      )} ("hash", "created_at", "name", "applied_at") values(${migration.hash}, ${migration.folderMillis}, ${migration.name}, ${new Date().toISOString()})`,
-    )
-
-    return
-  }
-
-  const migrationsToRun = getMigrationsToRun({ localMigrations: migrations, dbMigrations })
-  if (migrationsToRun.length === 0) return
-
-  yield* session.transaction((tx) =>
-    Effect.gen(function* () {
-      for (const migration of migrationsToRun) {
-        for (const stmt of migration.sql) {
-          yield* tx.run(sql.raw(stmt))
-        }
-        yield* tx.run(
-          sql`insert into ${sql.identifier(
-            migrationsTable,
-          )} ("hash", "created_at", "name", "applied_at") values(${migration.hash}, ${migration.folderMillis}, ${migration.name}, ${new Date().toISOString()})`,
-        )
-      }
-    }),
-  )
-})

packages/core/src/database/drizzle/up-migrations/effect-sqlite.ts

@@ -1,102 +0,0 @@
-/* oxlint-disable */
-import * as Effect from "effect/Effect"
-import type { SqlError } from "effect/unstable/sql/SqlError"
-import { EffectDrizzleError } from "drizzle-orm/effect-core/errors"
-import type { QueryEffectHKTBase } from "drizzle-orm/effect-core/query-effect"
-import type { MigrationMeta } from "drizzle-orm/migrator"
-import { sql } from "drizzle-orm/sql/sql"
-import type { SQLiteEffectSession } from "../sqlite-core/effect/session.js"
-import {
-  buildSQLiteMigrationBackfillStatements,
-  prepareSQLiteMigrationBackfill,
-  type SQLiteMigrationTableRow,
-} from "./sqlite.js"
-import { GET_VERSION_FOR, MIGRATIONS_TABLE_VERSIONS, type UpgradeResult } from "./utils.js"
-
-const migrationUpgradeError = (cause: unknown) =>
-  new EffectDrizzleError({
-    message:
-      typeof cause === "object" && cause !== null && "message" in cause && typeof cause.message === "string"
-        ? cause.message
-        : String(cause),
-    cause,
-  })
-
-export const upgradeIfNeeded: <TEffectHKT extends QueryEffectHKTBase>(
-  migrationsTable: string,
-  session: SQLiteEffectSession<TEffectHKT>,
-  localMigrations: MigrationMeta[],
-) => Effect.Effect<UpgradeResult, EffectDrizzleError | TEffectHKT["error"] | SqlError, TEffectHKT["context"]> =
-  Effect.fn("upgradeIfNeeded")(function* <TEffectHKT extends QueryEffectHKTBase>(
-    migrationsTable: string,
-    session: SQLiteEffectSession<TEffectHKT>,
-    localMigrations: MigrationMeta[],
-  ) {
-    const tableExists = yield* session.all(
-      sql`SELECT 1 FROM sqlite_master WHERE type = 'table' AND name = ${migrationsTable}`,
-    )
-
-    if (tableExists.length === 0) {
-      return { newDb: true }
-    }
-
-    const rows = yield* session.all<{ column_name: string }>(
-      sql`SELECT name as column_name FROM pragma_table_info(${migrationsTable})`,
-    )
-
-    const version = GET_VERSION_FOR.sqlite(rows.map((r) => r.column_name))
-
-    for (let v = version; v < MIGRATIONS_TABLE_VERSIONS.sqlite; v++) {
-      const upgradeFn = upgradeFunctions[v]
-      if (!upgradeFn) {
-        return yield* new EffectDrizzleError({
-          message: `No upgrade path from migration table version ${v} to ${v + 1}`,
-          cause: { version: v },
-        })
-      }
-      yield* upgradeFn(migrationsTable, session, localMigrations)
-    }
-
-    return { newDb: false }
-  })
-
-const upgradeFunctions: Record<
-  number,
-  <TEffectHKT extends QueryEffectHKTBase>(
-    migrationsTable: string,
-    session: SQLiteEffectSession<TEffectHKT>,
-    localMigrations: MigrationMeta[],
-  ) => Effect.Effect<void, EffectDrizzleError | TEffectHKT["error"] | SqlError, TEffectHKT["context"]>
-> = {
-  0: upgradeFromV0,
-}
-
-function upgradeFromV0<TEffectHKT extends QueryEffectHKTBase>(
-  migrationsTable: string,
-  session: SQLiteEffectSession<TEffectHKT>,
-  localMigrations: MigrationMeta[],
-): Effect.Effect<void, EffectDrizzleError | TEffectHKT["error"] | SqlError, TEffectHKT["context"]> {
-  return Effect.gen(function* () {
-    const table = sql`${sql.identifier(migrationsTable)}`
-
-    const dbRows = yield* session.all<SQLiteMigrationTableRow>(
-      sql`SELECT id, hash, created_at FROM ${table} ORDER BY id ASC`,
-    )
-    const statements = yield* Effect.try({
-      try: () =>
-        buildSQLiteMigrationBackfillStatements(
-          migrationsTable,
-          prepareSQLiteMigrationBackfill(dbRows, localMigrations),
-        ),
-      catch: migrationUpgradeError,
-    })
-
-    yield* session.transaction((tx) =>
-      Effect.gen(function* () {
-        for (const statement of statements) {
-          yield* tx.run(statement)
-        }
-      }),
-    )
-  })
-}

packages/core/src/database/drizzle/up-migrations/sqlite.ts

@@ -1,105 +0,0 @@
-/* oxlint-disable */
-import type { MigrationMeta } from "drizzle-orm/migrator"
-import { type SQL, sql } from "drizzle-orm/sql/sql"
-
-/** @internal */
-export type SQLiteMigrationTableRow = { id: number | null; hash: string; created_at: number }
-
-type SQLiteMigrationBackfillEntry = {
-  name: string
-  selector:
-    | { column: "id"; value: number }
-    | { column: "created_at"; value: number }
-    | { column: "hash"; value: string }
-}
-
-function unmatchedMigrationError(unmatched: SQLiteMigrationTableRow[]) {
-  return new Error(
-    `While upgrading your database migrations table we found ${unmatched.length} (${unmatched
-      .map((it) => `[id: ${it.id}, created_at: ${it.created_at}]`)
-      .join(
-        ", ",
-      )}) migrations in the database that do not match any local migration. This means that some migrations were applied to the database but are missing from the local environment`,
-  )
-}
-
-/** @internal */
-export function prepareSQLiteMigrationBackfill(
-  dbRows: SQLiteMigrationTableRow[],
-  localMigrations: MigrationMeta[],
-): SQLiteMigrationBackfillEntry[] {
-  const sortedLocalMigrations = [...localMigrations].sort((a, b) =>
-    a.folderMillis !== b.folderMillis ? a.folderMillis - b.folderMillis : (a.name ?? "").localeCompare(b.name ?? ""),
-  )
-  const byMillis = new Map<number, MigrationMeta[]>()
-  const byHash = new Map<string, MigrationMeta>()
-  for (const migration of sortedLocalMigrations) {
-    if (!byMillis.has(migration.folderMillis)) {
-      byMillis.set(migration.folderMillis, [])
-    }
-    byMillis.get(migration.folderMillis)!.push(migration)
-    byHash.set(migration.hash, migration)
-  }
-
-  const toApply: SQLiteMigrationBackfillEntry[] = []
-  const unmatched: SQLiteMigrationTableRow[] = []
-
-  for (const dbRow of dbRows) {
-    const stringified = String(dbRow.created_at)
-    const millis = Number(stringified.substring(0, stringified.length - 3) + "000")
-    const candidates = byMillis.get(millis)
-
-    const matchedByMillis = candidates?.length === 1 ? candidates[0] : undefined
-    const matchedByCandidateHash =
-      candidates && candidates.length > 1
-        ? candidates.find((candidate) => candidate.hash && dbRow.hash && candidate.hash === dbRow.hash)
-        : undefined
-    const matchedByHash = matchedByMillis || matchedByCandidateHash ? undefined : byHash.get(dbRow.hash)
-    const matched = matchedByMillis ?? matchedByCandidateHash ?? matchedByHash
-
-    if (matched) {
-      toApply.push({
-        name: matched.name,
-        selector:
-          dbRow.id !== null
-            ? { column: "id", value: dbRow.id }
-            : matchedByMillis
-              ? { column: "created_at", value: dbRow.created_at }
-              : { column: "hash", value: dbRow.hash },
-      })
-      continue
-    }
-
-    unmatched.push(dbRow)
-  }
-
-  if (unmatched.length > 0) {
-    throw unmatchedMigrationError(unmatched)
-  }
-
-  return toApply
-}
-
-/** @internal */
-export function buildSQLiteMigrationBackfillStatements(
-  migrationsTable: string,
-  backfillEntries: SQLiteMigrationBackfillEntry[],
-) {
-  const table = sql`${sql.identifier(migrationsTable)}`
-  const statements: SQL[] = [
-    sql`ALTER TABLE ${table} ADD COLUMN ${sql.identifier("name")} text`,
-    sql`ALTER TABLE ${table} ADD COLUMN ${sql.identifier("applied_at")} TEXT`,
-  ]
-
-  for (const backfillEntry of backfillEntries) {
-    const updateQuery = sql`UPDATE ${table} SET ${sql.identifier("name")} = ${backfillEntry.name}, ${sql.identifier(
-      "applied_at",
-    )} = NULL WHERE`
-
-    updateQuery.append(sql` ${sql.identifier(backfillEntry.selector.column)} = ${backfillEntry.selector.value}`)
-
-    statements.push(updateQuery)
-  }
-
-  return statements
-}

packages/core/src/database/drizzle/up-migrations/utils.ts

@@ -1,45 +0,0 @@
-/* oxlint-disable */
-export interface UpgradeResult {
-  newDb: boolean
-}
-
-export const MIGRATIONS_TABLE_VERSIONS = {
-  sqlite: 1,
-  pg: 1,
-  effect: 1,
-  mysql: 1,
-  mssql: 1,
-  cockroach: 1,
-  singlestore: 1,
-} as const
-
-export const GET_VERSION_FOR = {
-  mysql: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  pg: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  effect: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  mssql: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  cockroach: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  singlestore: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-  sqlite: (columns: string[]): number => {
-    if (columns.includes("name")) return 1
-    return 0
-  },
-} as const

packages/core/src/integration.ts

@@ -361,6 +361,7 @@ const layer = Layer.effect(
       Info.make({
         id: entry.ref.id,
         name: entry.ref.name,
+        ...(entry.ref.metadata === undefined ? {} : { metadata: entry.ref.metadata }),
         methods: entry.methods,
         connections,
       })

packages/core/src/mcp/index.ts

@@ -228,6 +228,7 @@ export const layer = (options?: Options) =>
           .transform((draft) => {
             draft.update(integrationID, (ref) => {
               ref.name = name
+              ref.metadata = { source: "mcp" }
             })
             draft.method.update({
               integrationID,

packages/core/src/model-resolver.ts

@@ -333,8 +333,11 @@ function usesAPIKeyAuth(packageName: string | undefined) {
   return (
     name === "@ai-sdk/openai" ||
     name === "@ai-sdk/anthropic" ||
+    name === "@ai-sdk/cerebras" ||
+    name === "@ai-sdk/deepinfra" ||
     name === "@ai-sdk/openai-compatible" ||
     name === "@ai-sdk/google" ||
+    name === "@ai-sdk/togetherai" ||
     name === "@ai-sdk/xai" ||
     name === "@openrouter/ai-sdk-provider" ||
     name === "@ai-sdk/azure" ||
@@ -342,8 +345,11 @@ function usesAPIKeyAuth(packageName: string | undefined) {
     name?.startsWith("@opencode-ai/ai/providers/openai/") === true ||
     name === "@opencode-ai/ai/providers/anthropic" ||
     name === "@opencode-ai/ai/providers/anthropic-compatible" ||
+    name === "@opencode-ai/ai/providers/cerebras" ||
+    name === "@opencode-ai/ai/providers/deepinfra" ||
     name === "@opencode-ai/ai/providers/openai-compatible" ||
     name === "@opencode-ai/ai/providers/google" ||
+    name === "@opencode-ai/ai/providers/togetherai" ||
     name === "@opencode-ai/ai/providers/xai" ||
     name === "@opencode-ai/ai/providers/openrouter" ||
     name === "@opencode-ai/ai/providers/azure" ||

packages/core/src/plugin/provider.ts

@@ -6,7 +6,6 @@ import { CerebrasPlugin } from "./provider/cerebras.js"
 import { CloudflareAIGatewayPlugin } from "./provider/cloudflare-ai-gateway.js"
 import { CloudflareWorkersAIPlugin } from "./provider/cloudflare-workers-ai.js"
 import { CoherePlugin } from "./provider/cohere.js"
-import { DeepInfraPlugin } from "./provider/deepinfra.js"
 import { DynamicProviderPlugin } from "./provider/dynamic.js"
 import { GatewayPlugin } from "./provider/gateway.js"
 import { GithubCopilotPlugin } from "./provider/github-copilot.js"
@@ -26,7 +25,6 @@ import { OpencodePlugin } from "./provider/opencode.js"
 import { OpenRouterPlugin } from "./provider/openrouter.js"
 import { PerplexityPlugin } from "./provider/perplexity.js"
 import { SapAICorePlugin } from "./provider/sap-ai-core.js"
-import { TogetherAIPlugin } from "./provider/togetherai.js"
 import { VercelPlugin } from "./provider/vercel.js"
 import { VenicePlugin } from "./provider/venice.js"
 import { VLLMPlugin } from "./provider/vllm.js"
@@ -43,7 +41,6 @@ export const ProviderPlugins: PluginInternal.InternalPlugin[] = [
   CloudflareAIGatewayPlugin,
   CloudflareWorkersAIPlugin,
   CoherePlugin,
-  DeepInfraPlugin,
   GatewayPlugin,
   GithubCopilotPlugin,
   GitLabPlugin,
@@ -62,7 +59,6 @@ export const ProviderPlugins: PluginInternal.InternalPlugin[] = [
   OpenRouterPlugin,
   PerplexityPlugin,
   SapAICorePlugin,
-  TogetherAIPlugin,
   VercelPlugin,
   VenicePlugin,
   VLLMPlugin,

packages/core/src/plugin/provider/cerebras.ts

@@ -7,20 +7,12 @@ export const CerebrasPlugin = define({
   effect: Effect.fn(function* (ctx) {
     yield* ctx.catalog.transform((evt) => {
       for (const item of evt.provider.list()) {
-        if (!Provider.isAISDK(item.provider.package)) continue
-        if (Provider.packageName(item.provider.package) !== "@ai-sdk/cerebras") continue
+        const name = Provider.packageName(item.provider.package)
+        if (name !== "@ai-sdk/cerebras" && name !== "@opencode-ai/ai/providers/cerebras") continue
         evt.provider.update(item.provider.id, (provider) => {
           provider.headers = { ...provider.headers, "X-Cerebras-3rd-Party-Integration": "opencode" }
         })
       }
     })
-    yield* ctx.aisdk.hook(
-      "sdk",
-      Effect.fn(function* (evt) {
-        if (evt.package !== "@ai-sdk/cerebras") return
-        const mod = yield* Effect.promise(() => import("@ai-sdk/cerebras"))
-        evt.sdk = mod.createCerebras(evt.options)
-      }),
-    )
   }),
 })

packages/core/src/plugin/provider/deepinfra.ts

@@ -1,10 +0,0 @@
-import { createProviderPlugin } from "./factory.js"
-
-export const DeepInfraPlugin = createProviderPlugin({
-  id: "opencode.provider.deepinfra",
-  package: "@ai-sdk/deepinfra",
-  load: async (options) => {
-    const { createDeepInfra } = await import("@ai-sdk/deepinfra")
-    return createDeepInfra(options)
-  },
-})

packages/core/src/plugin/provider/togetherai.ts

@@ -1,10 +0,0 @@
-import { createProviderPlugin } from "./factory.js"
-
-export const TogetherAIPlugin = createProviderPlugin({
-  id: "opencode.provider.togetherai",
-  package: "@ai-sdk/togetherai",
-  load: async (options) => {
-    const { createTogetherAI } = await import("@ai-sdk/togetherai")
-    return createTogetherAI(options)
-  },
-})

packages/core/src/plugin/supervisor.ts

@@ -108,7 +108,7 @@ const load = Effect.fn("PluginSupervisor.load")(function* (
   const npm = yield* Npm.Service
   const entrypoint = path.isAbsolute(operation.target)
     ? pathToFileURL(operation.target).href
-    : (yield* npm.add(operation.target, { subpaths: ["server", ""] })).entrypoint
+    : (yield* npm.add(operation.target, { subpaths: ["server", ""], refresh: true })).entrypoint
   if (!entrypoint) return yield* Effect.fail(new Error(`Plugin entrypoint not found: ${operation.target}`))
   // Bun currently ignores query parameters when caching file:// imports.
   const source =

packages/core/src/provider.ts

@@ -48,6 +48,8 @@ const builtins = new Map<string, () => Promise<unknown>>([
   ["@opencode-ai/ai/providers/azure", () => import("@opencode-ai/ai/providers/azure")],
   ["@opencode-ai/ai/providers/azure/chat", () => import("@opencode-ai/ai/providers/azure/chat")],
   ["@opencode-ai/ai/providers/azure/responses", () => import("@opencode-ai/ai/providers/azure/responses")],
+  ["@opencode-ai/ai/providers/cerebras", () => import("@opencode-ai/ai/providers/cerebras")],
+  ["@opencode-ai/ai/providers/deepinfra", () => import("@opencode-ai/ai/providers/deepinfra")],
   ["@opencode-ai/ai/providers/google", () => import("@opencode-ai/ai/providers/google")],
   ["@opencode-ai/ai/providers/google-vertex", () => import("@opencode-ai/ai/providers/google-vertex")],
   ["@opencode-ai/ai/providers/google-vertex/gemini", () => import("@opencode-ai/ai/providers/google-vertex/gemini")],
@@ -65,6 +67,7 @@ const builtins = new Map<string, () => Promise<unknown>>([
   ["@opencode-ai/ai/providers/openai/responses", () => import("@opencode-ai/ai/providers/openai/responses")],
   ["@opencode-ai/ai/providers/openai-compatible", () => import("@opencode-ai/ai/providers/openai-compatible")],
   ["@opencode-ai/ai/providers/openrouter", () => import("@opencode-ai/ai/providers/openrouter")],
+  ["@opencode-ai/ai/providers/togetherai", () => import("@opencode-ai/ai/providers/togetherai")],
   ["@opencode-ai/ai/providers/xai", () => import("@opencode-ai/ai/providers/xai")],
 ])
 

packages/core/src/session.ts

@@ -596,7 +596,11 @@ const layer = Layer.effect(
               yield* plugins.flush
               return yield* Image.Service
             }).pipe(Effect.provide(locations.get(session.location)))
-            const skills = Skill.Service.pipe(Effect.provide(locations.get(session.location)))
+            const skills = Effect.gen(function* () {
+              const plugins = yield* PluginSupervisor.Service
+              yield* plugins.flush
+              return yield* Skill.Service
+            }).pipe(Effect.provide(locations.get(session.location)))
             const prompt = yield* resolvePrompt(
               { text: input.text, files: input.files, agents: input.agents, skills: input.skills },
               image,
@@ -718,7 +722,7 @@ const layer = Layer.effect(
       skill: Effect.fn("Session.skill")(function* (input) {
         const session = yield* result.get(input.sessionID)
         const skills = yield* Skill.Service.pipe(Effect.provide(locations.get(session.location)))
-        const skill = (yield* skills.list()).find((item) => item.id === input.skill)
+        const skill = yield* skills.get(input.skill)
         if (!skill) return yield* new SkillNotFoundError({ skill: input.skill })
         yield* bus.publish(
           SessionEvent.Skill.Activated,
@@ -970,16 +974,22 @@ const resolvePrompt = Effect.fn("Session.resolvePrompt")(function* (
   const selected = yield* Effect.gen(function* () {
     if (!requested?.length) return undefined
     const skillService = yield* skills
-    const available = yield* skillService.list()
-    return yield* Effect.forEach(requested, (attachment) => {
-      const skill = available.find((item) => item.id === attachment.id)
-      if (!skill) return Effect.fail(new SkillNotFoundError({ skill: attachment.id }))
-      return Effect.succeed({
-        id: skill.id,
-        name: skill.name,
-        mention: attachment.mention,
-      })
-    })
+    const prepared = new Map<Skill.ID, Skill.Name>()
+    return yield* Effect.forEach(requested, (attachment) =>
+      Effect.gen(function* () {
+        const name = prepared.get(attachment.id)
+        if (name !== undefined) return { id: attachment.id, name, mention: attachment.mention }
+        const skill = yield* skillService.get(attachment.id)
+        if (!skill) return yield* new SkillNotFoundError({ skill: attachment.id })
+        prepared.set(skill.id, skill.name)
+        return {
+          id: skill.id,
+          name: skill.name,
+          text: (yield* Skill.prepare(fs, skill).pipe(Effect.orDie)).output,
+          mention: attachment.mention,
+        }
+      }),
+    )
   })
   return Prompt.make({ text: input.text, agents: input.agents, files, skills: selected?.length ? selected : undefined })
 })

packages/core/src/session/compaction.ts

@@ -103,6 +103,7 @@ export type Outcome =
   | Pick<SessionMessage.CompactionFailed, "status" | "error">
 
 export interface Interface extends State.Transformable<Draft> {
+  readonly enabled: () => boolean
   readonly required: (input: RequiredInput) => boolean
   readonly compact: (input: AutoInput) => Effect.Effect<Outcome>
   readonly compactManual: (input: ManualInput) => Effect.Effect<Outcome>
@@ -138,7 +139,11 @@ const serialize = (message: SessionMessage.Info) => {
         (file) =>
           `[Attached ${file.mime}: ${file.name ?? (file.source.type === "uri" ? file.source.uri : "inline attachment")}]`,
       ) ?? []
-    return [`[User]: ${message.text}`, ...files].join("\n")
+    const skills =
+      message.skills?.flatMap((skill) =>
+        skill.text === undefined ? [] : [`[Skill activated: ${skill.name}]\n${skill.text}`],
+      ) ?? []
+    return [...skills, `[User]: ${message.text}`, ...files].join("\n")
   }
   if (message.type === "location-switched")
     return `[User]: The working directory has been changed to ${message.location.directory}.`
@@ -405,6 +410,7 @@ const make = (dependencies: Dependencies) => {
   return Service.of({
     transform: state.transform,
     reload: state.reload,
+    enabled: () => state.get().auto,
     required,
     compact,
     compactManual,

packages/core/src/session/runner/llm.ts

@@ -508,6 +508,7 @@ const layer = Layer.effect(
           // restart the step instead of surfacing the provider error.
           if (
             recoverOverflow &&
+            compaction.enabled() &&
             !publisher.record().outputStarted &&
             isContextOverflowFailure(overflowFailure ?? streamFailure) &&
             (yield* restore(compaction.compact(compactionInput))).status === "completed"

packages/core/src/session/runner/to-llm-message.ts

@@ -236,6 +236,7 @@ function toLLMMessage(message: SessionMessage.Info, model: Model.Ref, providerMe
       ]
     case "user":
       const content = [
+        ...(message.skills ?? []).flatMap((skill) => (skill.text === undefined ? [] : [Message.text(skill.text)])),
         ...(message.text === "" ? [] : [Message.text(message.text)]),
         ...userAttachmentContent(message.files ?? []),
       ]

packages/core/src/session/transfer.ts

@@ -220,6 +220,7 @@ function sanitizeMessage(message: SessionMessage.Info): SessionMessage.Info {
       skills: message.skills?.map((skill, index) => ({
         ...skill,
         name: Skill.Name.make(redact("skill-name", String(index), skill.name)),
+        text: skill.text === undefined ? undefined : redact("skill", String(index), skill.text),
         mention: skill.mention
           ? { ...skill.mention, text: redact("skill-mention", String(index), skill.mention.text) }
           : undefined,

packages/core/src/skill.ts

@@ -1,6 +1,7 @@
 export * as Skill from "./skill.js"
 
 import { makeLocationNode } from "@opencode-ai/util/effect/app-node"
+import type { FSUtil } from "@opencode-ai/util/fs-util"
 import path from "path"
 import { Context, Effect, Layer, Types } from "effect"
 import { Skill } from "@opencode-ai/schema/skill"
@@ -52,6 +53,21 @@ export const toModelOutput = (skill: Info, files: ReadonlyArray<string>) => {
   ].join("\n")
 }
 
+export const prepare = Effect.fn("Skill.prepare")(function* (fs: FSUtil.Interface, skill: Info) {
+  const directory = path.dirname(skill.location)
+  const files =
+    path.basename(skill.location) === "SKILL.md"
+      ? (yield* fs.scan("**/*", { cwd: directory, absolute: true, include: "file", dot: true }))
+          .filter((file) => path.basename(file) !== "SKILL.md")
+          .toSorted()
+          .slice(0, 10)
+      : []
+  return {
+    directory,
+    output: toModelOutput(skill, files),
+  }
+})
+
 export type Data = {
   skills: Map<ID, Types.DeepMutable<Info>>
 }
@@ -64,6 +80,7 @@ export type Draft = {
 }
 
 export interface Interface extends State.Transformable<Draft> {
+  readonly get: (id: ID) => Effect.Effect<Info | undefined>
   readonly list: () => Effect.Effect<Info[]>
 }
 
@@ -98,6 +115,9 @@ const layer = Layer.effect(
     return Service.of({
       transform: state.transform,
       reload: state.reload,
+      get: Effect.fn("Skill.get")(function* (id) {
+        return state.get().skills.get(id)
+      }),
       list: Effect.fn("Skill.list")(function* () {
         return Array.from(state.get().skills.values())
       }),

packages/core/src/skill/instructions.ts

@@ -26,7 +26,6 @@ const render = (skills: ReadonlyArray<Summary>) =>
   [
     "Skills provide specialized instructions and workflows for specific tasks.",
     "Use the skill tool to load a skill when a task matches its description.",
-    "When the user references a skill with @skill-id, load that skill with the skill tool.",
     ...(skills.length === 0
       ? ["No skills are currently available."]
       : ["<available_skills>", ...entries(skills), "</available_skills>"]),

packages/core/src/tool/plugin/skill.ts

@@ -1,7 +1,6 @@
 export * as SkillTool from "./skill.js"
 
 import type { Context as PluginContext } from "@opencode-ai/plugin/effect/plugin"
-import path from "path"
 import { ToolFailure } from "@opencode-ai/ai"
 import { Effect, Schema } from "effect"
 import { FSUtil } from "@opencode-ai/util/fs-util"
@@ -9,7 +8,6 @@ import { Skill } from "../../skill.js"
 import { Permission } from "../../permission.js"
 
 export const name = "skill"
-const FILE_LIMIT = 10
 
 export const Input = Schema.Struct({
   id: Skill.ID.annotate({ description: "The ID of an available skill or a skill explicitly referenced by the user" }),
@@ -47,8 +45,7 @@ export const Plugin = {
           output: Output,
           execute: (input, context) =>
             Effect.gen(function* () {
-              const current = yield* skills.list()
-              const skill = current.find((skill) => skill.id === input.id)
+              const skill = yield* skills.get(input.id)
               if (!skill) return yield* unableToLoad(input.id)
               return yield* Effect.gen(function* () {
                 yield* permission.assert({
@@ -59,19 +56,7 @@ export const Plugin = {
                   agent: context.agent,
                   source: { type: "tool", messageID: context.messageID, id: context.id },
                 })
-                const directory = path.dirname(skill.location)
-                const files =
-                  path.basename(skill.location) === "SKILL.md"
-                    ? (yield* fs.scan("**/*", { cwd: directory, absolute: true, include: "file", dot: true }))
-                        .filter((file) => path.basename(file) !== "SKILL.md")
-                        .toSorted()
-                        .slice(0, FILE_LIMIT)
-                    : []
-                return {
-                  name: skill.name,
-                  directory,
-                  output: Skill.toModelOutput(skill, files),
-                }
+                return { name: skill.name, ...(yield* Skill.prepare(fs, skill)) }
               }).pipe(Effect.mapError((error) => unableToLoad(input.id, error)))
             }).pipe(
               Effect.map((output) => ({

packages/core/test/aisdk-native.test.ts

@@ -14,7 +14,6 @@ describe("AISDKNative", () => {
         reasoningEffort: "xhigh",
         reasoningSummary: "auto",
         include: ["reasoning.encrypted_content"],
-        instructions: "Follow the repository instructions.",
         truncation: "auto",
       }),
     ).toEqual({
@@ -26,7 +25,6 @@ describe("AISDKNative", () => {
           reasoningEffort: "xhigh",
           reasoningSummary: "auto",
           include: ["reasoning.encrypted_content"],
-          instructions: "Follow the repository instructions.",
           truncation: "auto",
         },
         organization: "org",
@@ -63,6 +61,33 @@ describe("AISDKNative", () => {
     })
   })
 
+  test("maps Cerebras, DeepInfra, and Together AI settings, headers, and reasoning options to native providers", () => {
+    for (const name of ["cerebras", "deepinfra", "togetherai"]) {
+      expect(
+        map(`@ai-sdk/${name}`, {
+          apiKey: "secret",
+          baseURL: `https://${name}.example/v1`,
+          headers: { "x-provider": name },
+          name: "custom-provider",
+          reasoningEffort: "high",
+          customOption: { enabled: true },
+        }),
+      ).toEqual({
+        package: `@opencode-ai/ai/providers/${name}`,
+        settings: {
+          apiKey: "secret",
+          baseURL: `https://${name}.example/v1`,
+          providerOptions: { reasoningEffort: "high", customOption: { enabled: true } },
+        },
+        headers: { "x-provider": name },
+      })
+      expect(map(`@ai-sdk/${name}`, {})).toEqual({
+        package: `@opencode-ai/ai/providers/${name}`,
+        settings: {},
+      })
+    }
+  })
+
   test("maps Google Vertex settings to the native provider", () => {
     expect(
       map("@ai-sdk/google-vertex", {

packages/core/test/database-drizzle.test.ts

@@ -1,4 +1,4 @@
-import { mkdir, mkdtemp, rm } from "node:fs/promises"
+import { mkdtemp, rm } from "node:fs/promises"
 import { tmpdir } from "node:os"
 import { join } from "node:path"
 import { Database } from "bun:sqlite"
@@ -27,16 +27,6 @@ const makeDb = Effect.gen(function* () {
   return db
 })
 
-const createMigrationsFolder = async () => {
-  const migrationsFolder = await mkdtemp(join(tmpdir(), "effect-drizzle-sqlite-"))
-  await mkdir(join(migrationsFolder, "20240101000000_create_migrated_users"), { recursive: true })
-  await Bun.write(
-    join(migrationsFolder, "20240101000000_create_migrated_users", "migration.sql"),
-    "create table migrated_users (id integer primary key autoincrement, name text not null);",
-  )
-  return migrationsFolder
-}
-
 test("selects rows through Effect-yieldable query builders", async () => {
   await run(
     Effect.gen(function* () {
@@ -173,25 +163,3 @@ test("supports returning and rejects empty update sets", async () => {
     }),
   )
 })
-
-test("runs migrations once and records migration metadata", async () => {
-  const migrationsFolder = await createMigrationsFolder()
-  try {
-    await run(
-      Effect.gen(function* () {
-        const db = yield* EffectDrizzleSqlite.makeWithDefaults()
-
-        yield* EffectDrizzleSqlite.migrate(db, { migrationsFolder })
-        yield* EffectDrizzleSqlite.migrate(db, { migrationsFolder })
-        yield* db.run(sql`insert into migrated_users (name) values ('Margaret')`)
-
-        expect(yield* db.all<{ name: string }>(sql`select name from migrated_users`)).toEqual([{ name: "Margaret" }])
-        expect(yield* db.all<{ name: string | null }>(sql`select name from __drizzle_migrations`)).toEqual([
-          { name: "20240101000000_create_migrated_users" },
-        ])
-      }),
-    )
-  } finally {
-    await rm(migrationsFolder, { recursive: true, force: true })
-  }
-})

packages/core/test/fixture/global.ts

@@ -12,7 +12,6 @@ export const tempGlobalLayer = Layer.unwrap(
       const data = path.join(tmp.path, "data")
       const cache = path.join(tmp.path, "cache")
       return Global.layerWith({
-        home: path.join(tmp.path, "home"),
         data,
         cache,
         config: path.join(tmp.path, "config"),

packages/core/test/integration.test.ts

@@ -52,10 +52,21 @@ describe("Integration", () => {
       const openai = Integration.ID.make("openai")
 
       yield* integrations
-        .transform((editor) => editor.update(openai, (integration) => (integration.name = "OpenAI")))
+        .transform((editor) =>
+          editor.update(openai, (integration) => {
+            integration.name = "OpenAI"
+            integration.metadata = { source: "plugin", featured: true }
+          }),
+        )
         .pipe(Scope.provide(scope))
       expect(yield* integrations.get(openai)).toEqual(
-        Integration.Info.make({ id: openai, name: "OpenAI", methods: [], connections: [] }),
+        Integration.Info.make({
+          id: openai,
+          name: "OpenAI",
+          metadata: { source: "plugin", featured: true },
+          methods: [],
+          connections: [],
+        }),
       )
 
       yield* Scope.close(scope, Exit.void)

packages/core/test/model-resolver.test.ts

@@ -1,5 +1,5 @@
 import { describe, expect } from "bun:test"
-import { LLM, LanguageModel } from "@opencode-ai/ai"
+import { LLM, LanguageModel, Message } from "@opencode-ai/ai"
 import { OpenAIChat } from "@opencode-ai/ai/protocols"
 import { compileRequest } from "@opencode-ai/ai/route/client"
 import { ConfigProvider, Effect, Layer } from "effect"
@@ -391,6 +391,7 @@ describe("ModelResolver", () => {
         model(Provider.aisdk("@ai-sdk/openai-compatible"), {
           compatibility: {
             reasoningField: "vendor_reasoning",
+            requireReasoning: true,
             maxTokensField: "max_completion_tokens",
             requireFinishReason: false,
             requireAssistantAfterTool: true,
@@ -417,6 +418,7 @@ describe("ModelResolver", () => {
       expect(headers.authorization).toBe("Bearer settings-secret")
       expect(resolved.route.id).toBe("openai-compatible-chat")
       expect(resolved.compatibility?.reasoningField).toBe("vendor_reasoning")
+      expect(resolved.compatibility?.requireReasoning).toBe(true)
       expect(resolved.compatibility?.maxTokensField).toBe("max_completion_tokens")
       expect(resolved.compatibility?.requireFinishReason).toBe(false)
       expect(resolved.compatibility?.requireAssistantAfterTool).toBe(true)
@@ -677,7 +679,15 @@ describe("ModelResolver", () => {
           metadata: { accountID: "acct_123" },
         }),
       )
-      const request = LLM.request({ model: resolved, prompt: "Hello" })
+      const request = LLM.request({
+        model: resolved,
+        system: [
+          { type: "text", text: "Base instructions." },
+          { type: "text", text: "Project instructions." },
+        ],
+        messages: [Message.user("Hello"), Message.system("Updated instructions.")],
+      })
+      const prepared = yield* compileRequest(request)
       const headers = yield* resolved.route.auth.apply({
         request,
         method: "POST",
@@ -692,6 +702,13 @@ describe("ModelResolver", () => {
       })
       expect(resolved.route.defaults.headers).toMatchObject({ "chatgpt-account-id": "acct_123" })
       expect(headers.authorization).toBe("Bearer chatgpt-token")
+      expect(prepared.body).toMatchObject({
+        instructions: "Base instructions.\nProject instructions.",
+        input: [
+          { role: "user", content: [{ type: "input_text", text: "Hello" }] },
+          { role: "developer", content: "Updated instructions." },
+        ],
+      })
     }),
   )
 
@@ -862,6 +879,18 @@ describe("ModelResolver", () => {
           { thinking: { type: "adaptive", display: "summarized" }, effort: "high" },
           { thinking: { type: "adaptive", display: "summarized" }, effort: "high" },
         ],
+        [
+          "@ai-sdk/cerebras",
+          "@opencode-ai/ai/providers/cerebras",
+          { reasoningEffort: "high" },
+          { reasoningEffort: "high" },
+        ],
+        [
+          "@ai-sdk/deepinfra",
+          "@opencode-ai/ai/providers/deepinfra",
+          { reasoningEffort: "none" },
+          { reasoningEffort: "none" },
+        ],
         [
           "@ai-sdk/openai-compatible",
           "@opencode-ai/ai/providers/openai-compatible",
@@ -886,6 +915,12 @@ describe("ModelResolver", () => {
           { reasoning: { effort: "high" } },
           { reasoning: { effort: "high" } },
         ],
+        [
+          "@ai-sdk/togetherai",
+          "@opencode-ai/ai/providers/togetherai",
+          { reasoningEffort: "high" },
+          { reasoningEffort: "high" },
+        ],
         ["@ai-sdk/xai", "@opencode-ai/ai/providers/xai", { reasoningEffort: "high" }, { reasoningEffort: "high" }],
       ] as const
 
@@ -933,12 +968,15 @@ describe("ModelResolver", () => {
           "openai.gpt-oss-120b",
         ],
         ["@ai-sdk/azure", "@opencode-ai/ai/providers/azure/responses", "api-model"],
+        ["@ai-sdk/cerebras", "@opencode-ai/ai/providers/cerebras", "api-model"],
+        ["@ai-sdk/deepinfra", "@opencode-ai/ai/providers/deepinfra", "api-model"],
         ["@ai-sdk/google", "@opencode-ai/ai/providers/google", "api-model"],
         ["@ai-sdk/google-vertex", "@opencode-ai/ai/providers/google-vertex", "api-model"],
         ["@ai-sdk/google-vertex/anthropic", "@opencode-ai/ai/providers/google-vertex/messages", "claude-sonnet-4-6"],
         ["@ai-sdk/openai", "@opencode-ai/ai/providers/openai", "api-model"],
         ["@ai-sdk/openai-compatible", "@opencode-ai/ai/providers/openai-compatible", "api-model"],
         ["@openrouter/ai-sdk-provider", "@opencode-ai/ai/providers/openrouter", "api-model"],
+        ["@ai-sdk/togetherai", "@opencode-ai/ai/providers/togetherai", "api-model"],
         ["@ai-sdk/xai", "@opencode-ai/ai/providers/xai", "api-model"],
       ] as const
 
@@ -1053,6 +1091,17 @@ describe("ModelResolver", () => {
           settings: { reasoning: { effort: "high" } },
         }),
       )
+      const cerebras = yield* ModelResolver.fromCatalogModel(
+        model(Provider.aisdk("@ai-sdk/cerebras"), { settings: { reasoningEffort: "high" } }),
+      )
+      const deepinfra = yield* ModelResolver.fromCatalogModel(
+        model(Provider.aisdk("@ai-sdk/deepinfra"), {
+          settings: { baseURL: "https://deepinfra.example/provider-root", reasoningEffort: "none" },
+        }),
+      )
+      const togetherai = yield* ModelResolver.fromCatalogModel(
+        model(Provider.aisdk("@ai-sdk/togetherai"), { settings: { reasoningEffort: "high" } }),
+      )
       const xai = yield* ModelResolver.fromCatalogModel(
         model(Provider.aisdk("@ai-sdk/xai"), { settings: { reasoningEffort: "high" } }),
       )
@@ -1073,6 +1122,16 @@ describe("ModelResolver", () => {
       expect(google.route.defaults.providerOptions).toEqual({ thinkingConfig: { thinkingBudget: 1_024 } })
       expect(openrouter.route.id).toBe("openrouter")
       expect(openrouter.route.defaults.providerOptions).toEqual({ reasoning: { effort: "high" } })
+      expect(cerebras.route.id).toBe("cerebras-chat")
+      expect(cerebras.route.defaults.providerOptions).toEqual({ reasoningEffort: "high" })
+      expect(String(cerebras.provider)).toBe("test-provider")
+      expect(deepinfra.route.id).toBe("deepinfra-chat")
+      expect(deepinfra.route.endpoint.baseURL).toBe("https://deepinfra.example/provider-root/openai")
+      expect(deepinfra.route.defaults.providerOptions).toEqual({ reasoningEffort: "none" })
+      expect(String(deepinfra.provider)).toBe("test-provider")
+      expect(togetherai.route.id).toBe("togetherai-chat")
+      expect(togetherai.route.defaults.providerOptions).toEqual({ reasoningEffort: "high" })
+      expect(String(togetherai.provider)).toBe("test-provider")
       expect(xai.route.id).toBe("openai-responses")
       expect(xai.route.defaults.providerOptions).toEqual({
         reasoningEffort: "high",

packages/core/test/npm.test.ts

@@ -1,5 +1,6 @@
 import fs from "fs/promises"
 import path from "path"
+import { pathToFileURL } from "url"
 import { describe, expect, test } from "bun:test"
 import { Effect } from "effect"
 import { AppNodeBuilder } from "@opencode-ai/core/effect/app-node-builder"
@@ -21,6 +22,39 @@ const writePackage = (dir: string, pkg: Record<string, unknown>) =>
 const npmLayer = (cache: string) =>
   AppNodeBuilder.build(Npm.node, [[Global.node, Global.layerWith({ cache, state: path.join(cache, "state") })]])
 
+async function createGitFixture(directory: string) {
+  const repository = path.join(directory, "repository")
+  await fs.mkdir(path.join(repository, "dependency"), { recursive: true })
+  await writePackage(repository, {
+    name: "fixture-git-plugin",
+    exports: "./index.js",
+    dependencies: { "fixture-dependency": "file:./dependency" },
+  })
+  await writePackage(path.join(repository, "dependency"), { name: "fixture-dependency", exports: "./index.js" })
+  await Bun.write(path.join(repository, "index.js"), "export default { root: true }\n")
+  await Bun.write(path.join(repository, "dependency", "index.js"), "export const dependency = true\n")
+
+  const subdirectory = path.join(repository, "packages", "subdirectory-plugin")
+  await fs.mkdir(path.join(subdirectory, "dependency"), { recursive: true })
+  await writePackage(subdirectory, {
+    name: "fixture-subdirectory-plugin",
+    exports: "./index.js",
+    dependencies: { "fixture-subdirectory-dependency": "file:./dependency" },
+  })
+  await writePackage(path.join(subdirectory, "dependency"), {
+    name: "fixture-subdirectory-dependency",
+    exports: "./index.js",
+  })
+  await Bun.write(path.join(subdirectory, "index.js"), "export default { subdirectory: true }\n")
+  await Bun.write(path.join(subdirectory, "dependency", "index.js"), "export const dependency = true\n")
+
+  await Bun.$`git init -q -b fixture-branch ${repository}`
+  await Bun.$`git -C ${repository} add .`
+  await Bun.$`git -C ${repository} -c user.name=fixture -c user.email=fixture@example.com commit -qm fixture`
+  const commit = await Bun.$`git -C ${repository} rev-parse HEAD`.text().then((value) => value.trim())
+  return { repository, commit }
+}
+
 describe("Npm.sanitize", () => {
   test("keeps normal scoped package specs unchanged", () => {
     expect(Npm.sanitize("@opencode/acme")).toBe("@opencode/acme")
@@ -46,6 +80,33 @@ describe("Npm.isRegistryPackage", () => {
   })
 })
 
+describe("Npm.isInstallablePackage", () => {
+  test("accepts registry and npm-compatible Git specs", async () => {
+    expect(await Npm.isInstallablePackage("plugin@^1.2.0")).toBe(true)
+    expect(await Npm.isInstallablePackage("github:acme/plugin#main")).toBe(true)
+    expect(await Npm.isInstallablePackage("git+ssh://git@github.com/acme/plugin.git#main")).toBe(true)
+    expect(await Npm.isInstallablePackage("git@github.com:acme/plugin.git")).toBe(true)
+    expect(
+      await Npm.isInstallablePackage(
+        "git+https://github.com/acme/plugins.git#0123456789abcdef0123456789abcdef01234567::path:packages/plugin",
+      ),
+    ).toBe(true)
+    expect(await Npm.isInstallablePackage("./plugin")).toBe(false)
+    expect(await Npm.isInstallablePackage("https://example.com/plugin.tgz")).toBe(false)
+    expect(await Npm.isInstallablePackage("alias@npm:plugin@1.0.0")).toBe(false)
+  })
+})
+
+describe("Npm.cacheKey", () => {
+  test("preserves registry keys and hashes Git specs", async () => {
+    expect(await Npm.cacheKey("@opencode/acme@1.0.0")).toBe(Npm.sanitize("@opencode/acme@1.0.0"))
+    const spec = "git+ssh://git@github.com/acme/plugin.git#main"
+    expect(await Npm.cacheKey(spec)).toMatch(/^git-[a-f0-9]{64}$/)
+    expect(await Npm.cacheKey(spec)).toBe(await Npm.cacheKey(spec))
+    expect(await Npm.cacheKey(`${spec}-other`)).not.toBe(await Npm.cacheKey(spec))
+  })
+})
+
 describe("Npm.add", () => {
   test("resolves cached scoped package specs without reifying", async () => {
     await using tmp = await tmpdir()
@@ -116,6 +177,92 @@ describe("Npm.add", () => {
     expect(entries.tui.entrypoint).toEndWith("/tui.js")
     expect(entries.fallback.entrypoint).toEndWith("/index.js")
   })
+
+  test("installs and resolves named and unnamed Git packages with dependencies", async () => {
+    await using tmp = await tmpdir()
+    const fixture = await createGitFixture(tmp.path)
+    const cache = path.join(tmp.path, "cache")
+    const specs = [
+      `git+file://${fixture.repository}#${fixture.commit}`,
+      `fixture-named-plugin@git+file://${fixture.repository}#fixture-branch`,
+    ]
+
+    for (const spec of specs) {
+      const entries = await Effect.gen(function* () {
+        const npm = yield* Npm.Service
+        return {
+          added: yield* npm.add(spec),
+          cached: yield* npm.add(spec),
+          resolved: yield* npm.resolve(spec),
+        }
+      }).pipe(Effect.scoped, Effect.provide(npmLayer(cache)), Effect.runPromise)
+
+      expect(entries.added.entrypoint).toEndWith("/index.js")
+      expect(entries.cached).toEqual(entries.added)
+      expect(entries.resolved).toEqual(entries.added)
+      expect(
+        await fs.stat(path.join(path.dirname(entries.added.directory), "fixture-dependency", "package.json")),
+      ).toBeTruthy()
+      expect(entries.added.directory).toContain(path.join("packages", await Npm.cacheKey(spec), "node_modules"))
+    }
+  })
+
+  test("installs a Git package from an npm ::path: subdirectory", async () => {
+    await using tmp = await tmpdir()
+    const fixture = await createGitFixture(tmp.path)
+    const spec = `git+file://${fixture.repository}#${fixture.commit}::path:packages/subdirectory-plugin`
+    const entry = await Effect.gen(function* () {
+      const npm = yield* Npm.Service
+      return yield* npm.add(spec)
+    }).pipe(Effect.scoped, Effect.provide(npmLayer(path.join(tmp.path, "cache"))), Effect.runPromise)
+
+    expect(entry.directory).toEndWith(path.join("node_modules", "fixture-subdirectory-plugin"))
+    expect(entry.entrypoint).toEndWith("/index.js")
+    expect(
+      await fs.stat(path.join(path.dirname(entry.directory), "fixture-subdirectory-dependency", "package.json")),
+    ).toBeTruthy()
+  })
+
+  test("refreshes mutable Git packages once per service lifetime and preserves pinned or cached installs", async () => {
+    await using tmp = await tmpdir()
+    const fixture = await createGitFixture(tmp.path)
+    const cache = path.join(tmp.path, "cache")
+    const repository = pathToFileURL(fixture.repository).href
+    const mutable = `git+${repository}#fixture-branch`
+    const pinned = `git+${repository}#${fixture.commit}`
+
+    const first = await Effect.gen(function* () {
+      const npm = yield* Npm.Service
+      const mutableEntry = yield* npm.add(mutable, { refresh: true })
+      const pinnedEntry = yield* npm.add(pinned, { refresh: true })
+      yield* Effect.promise(async () => {
+        await Bun.write(path.join(fixture.repository, "index.js"), 'export default { root: "second" }\n')
+        await Bun.$`git -C ${fixture.repository} add .`
+        await Bun.$`git -C ${fixture.repository} -c user.name=fixture -c user.email=fixture@example.com commit -qm second`
+      })
+      yield* npm.add(mutable, { refresh: true })
+      return { mutable: mutableEntry, pinned: pinnedEntry }
+    }).pipe(Effect.scoped, Effect.provide(npmLayer(cache)), Effect.runPromise)
+    expect(await Bun.file(path.join(first.mutable.directory, "index.js")).text()).toContain("root: true")
+    expect(await Bun.file(path.join(first.pinned.directory, "index.js")).text()).toContain("root: true")
+
+    const second = await Effect.gen(function* () {
+      const npm = yield* Npm.Service
+      return {
+        mutable: yield* npm.add(mutable, { refresh: true }),
+        pinned: yield* npm.add(pinned, { refresh: true }),
+      }
+    }).pipe(Effect.scoped, Effect.provide(npmLayer(cache)), Effect.runPromise)
+    expect(await Bun.file(path.join(second.mutable.directory, "index.js")).text()).toContain('root: "second"')
+    expect(await Bun.file(path.join(second.pinned.directory, "index.js")).text()).toContain("root: true")
+
+    await fs.rename(fixture.repository, `${fixture.repository}-offline`)
+    const offline = await Effect.gen(function* () {
+      const npm = yield* Npm.Service
+      return yield* npm.add(mutable, { refresh: true })
+    }).pipe(Effect.scoped, Effect.provide(npmLayer(cache)), Effect.runPromise)
+    expect(await Bun.file(path.join(offline.directory, "index.js")).text()).toContain('root: "second"')
+  })
 })
 
 describe("Npm.resolve", () => {

packages/core/test/plugin/provider-cerebras.test.ts

@@ -1,8 +1,6 @@
-import { AISDK } from "@opencode-ai/core/aisdk"
-import { describe, expect, mock } from "bun:test"
+import { describe, expect } from "bun:test"
 import { Effect } from "effect"
 import { Catalog } from "@opencode-ai/core/catalog"
-import { Model } from "@opencode-ai/core/model"
 import { Plugin } from "@opencode-ai/core/plugin"
 import { PluginHost } from "@opencode-ai/core/plugin/host"
 import { CerebrasPlugin } from "@opencode-ai/core/plugin/provider/cerebras"
@@ -10,26 +8,14 @@ import { Provider } from "@opencode-ai/core/provider"
 import { testEffect } from "../lib/effect"
 import { PluginTestLayer } from "./fixture"
 
-const cerebrasOptions: Record<string, unknown>[] = []
 const it = testEffect(PluginTestLayer)
 
 const addPlugin = Effect.fn(function* () {
   const plugin = yield* Plugin.Service
-  const aisdk = yield* AISDK.Service
   const host = yield* PluginHost.make(plugin)
   yield* CerebrasPlugin.effect(host)
 })
 
-void mock.module("@ai-sdk/cerebras", () => ({
-  createCerebras: (options: Record<string, unknown>) => {
-    const snapshot = { ...options }
-    cerebrasOptions.push(snapshot)
-    return {
-      languageModel: (modelID: string) => ({ modelID, provider: snapshot.name, specificationVersion: "v3" }),
-    }
-  },
-}))
-
 describe("CerebrasPlugin", () => {
   it.effect("applies the legacy integration header", () =>
     Effect.gen(function* () {
@@ -57,62 +43,21 @@ describe("CerebrasPlugin", () => {
     }),
   )
 
-  it.effect("creates a bundled Cerebras SDK with the model provider ID as the SDK name", () =>
+  it.effect("applies the integration header to custom native Cerebras providers", () =>
     Effect.gen(function* () {
-      cerebrasOptions.length = 0
-      const plugin = yield* Plugin.Service
-      const aisdk = yield* AISDK.Service
-      yield* addPlugin()
-      const result = yield* aisdk.runSDK({
-        model: Model.Info.make({
-          ...Model.Info.default(Provider.ID.make("custom-cerebras"), Model.ID.make("llama-4-scout-17b-16e-instruct")),
-          modelID: Model.ID.make("llama-4-scout-17b-16e-instruct"),
-          package: "aisdk:test-provider",
-        }),
-        package: "@ai-sdk/cerebras",
-        options: { name: "custom-cerebras", apiKey: "test" },
+      const catalog = yield* Catalog.Service
+      const providerID = Provider.ID.make("custom-cerebras")
+      yield* catalog.transform((catalog) => {
+        catalog.provider.update(providerID, (item) => {
+          item.package = "@opencode-ai/ai/providers/cerebras"
+          item.headers = { Existing: "1" }
+        })
       })
-      expect(cerebrasOptions).toEqual([{ name: "custom-cerebras", apiKey: "test" }])
-      expect(result.sdk.languageModel("llama-4-scout-17b-16e-instruct").provider).toBe("custom-cerebras")
-    }),
-  )
-
-  it.effect("preserves an explicit bundled Cerebras SDK name option", () =>
-    Effect.gen(function* () {
-      cerebrasOptions.length = 0
-      const plugin = yield* Plugin.Service
-      const aisdk = yield* AISDK.Service
       yield* addPlugin()
-      yield* aisdk.runSDK({
-        model: Model.Info.make({
-          ...Model.Info.default(Provider.ID.make("custom-cerebras"), Model.ID.make("llama-4-scout-17b-16e-instruct")),
-          modelID: Model.ID.make("llama-4-scout-17b-16e-instruct"),
-          package: "aisdk:test-provider",
-        }),
-        package: "@ai-sdk/cerebras",
-        options: { name: "configured-cerebras", apiKey: "test" },
+      expect((yield* catalog.provider.get(providerID))?.headers).toEqual({
+        Existing: "1",
+        "X-Cerebras-3rd-Party-Integration": "opencode",
       })
-      expect(cerebrasOptions).toEqual([{ name: "configured-cerebras", apiKey: "test" }])
-    }),
-  )
-
-  it.effect("ignores non-Cerebras SDK packages", () =>
-    Effect.gen(function* () {
-      cerebrasOptions.length = 0
-      const plugin = yield* Plugin.Service
-      const aisdk = yield* AISDK.Service
-      yield* addPlugin()
-      const result = yield* aisdk.runSDK({
-        model: Model.Info.make({
-          ...Model.Info.default(Provider.ID.make("custom-cerebras"), Model.ID.make("llama-4-scout-17b-16e-instruct")),
-          modelID: Model.ID.make("llama-4-scout-17b-16e-instruct"),
-          package: "aisdk:test-provider",
-        }),
-        package: "@ai-sdk/groq",
-        options: { name: "custom-cerebras", apiKey: "test" },
-      })
-      expect(cerebrasOptions).toEqual([])
-      expect(result.sdk).toBeUndefined()
     }),
   )
 })

packages/core/test/plugin/provider-factory.test.ts

@@ -6,12 +6,10 @@ import { Plugin } from "@opencode-ai/core/plugin"
 import { PluginHost } from "@opencode-ai/core/plugin/host"
 import { AlibabaPlugin } from "@opencode-ai/core/plugin/provider/alibaba"
 import { CoherePlugin } from "@opencode-ai/core/plugin/provider/cohere"
-import { DeepInfraPlugin } from "@opencode-ai/core/plugin/provider/deepinfra"
 import { GatewayPlugin } from "@opencode-ai/core/plugin/provider/gateway"
 import { GroqPlugin } from "@opencode-ai/core/plugin/provider/groq"
 import { MistralPlugin } from "@opencode-ai/core/plugin/provider/mistral"
 import { PerplexityPlugin } from "@opencode-ai/core/plugin/provider/perplexity"
-import { TogetherAIPlugin } from "@opencode-ai/core/plugin/provider/togetherai"
 import { VenicePlugin } from "@opencode-ai/core/plugin/provider/venice"
 import { Provider } from "@opencode-ai/core/provider"
 import { testEffect } from "../lib/effect"
@@ -22,12 +20,10 @@ const options = { name: "custom-provider", apiKey: "test", baseURL: "https://exa
 const providers = [
   { id: "alibaba", plugin: AlibabaPlugin, package: "@ai-sdk/alibaba", provider: "alibaba.chat" },
   { id: "cohere", plugin: CoherePlugin, package: "@ai-sdk/cohere", provider: "cohere.chat" },
-  { id: "deepinfra", plugin: DeepInfraPlugin, package: "@ai-sdk/deepinfra", provider: "deepinfra.chat" },
   { id: "gateway", plugin: GatewayPlugin, package: "@ai-sdk/gateway", provider: "gateway" },
   { id: "groq", plugin: GroqPlugin, package: "@ai-sdk/groq", provider: "groq.chat" },
   { id: "mistral", plugin: MistralPlugin, package: "@ai-sdk/mistral", provider: "mistral.chat" },
   { id: "perplexity", plugin: PerplexityPlugin, package: "@ai-sdk/perplexity", provider: "perplexity" },
-  { id: "togetherai", plugin: TogetherAIPlugin, package: "@ai-sdk/togetherai", provider: "togetherai.chat" },
   { id: "venice", plugin: VenicePlugin, package: "venice-ai-sdk-provider", provider: "custom-provider.chat" },
 ] as const
 

packages/core/test/preload.test.ts

@@ -0,0 +1,28 @@
+import { describe, expect, test } from "bun:test"
+import os from "os"
+import path from "path"
+import { Global } from "@opencode-ai/util/global"
+
+describe("Core test environment", () => {
+  test("isolates global home and XDG roots", () => {
+    const home = process.env.OPENCODE_TEST_HOME
+    expect(home).toBeDefined()
+    if (!home) return
+
+    expect(os.homedir()).toBe(home)
+    expect(Global.Path.home).toBe(home)
+    expect(Global.Path.config).toBe(path.join(home, ".config", "opencode"))
+    expect(Global.Path.data).toBe(path.join(home, ".local", "share", "opencode"))
+    expect(Global.Path.cache).toBe(path.join(home, ".cache", "opencode"))
+    expect(Global.Path.state).toBe(path.join(home, ".local", "state", "opencode"))
+    expect(os.tmpdir()).toBe(path.join(home, "tmp"))
+    expect(process.env.OPENCODE_CONFIG_DIR).toBe(Global.Path.config)
+    expect(process.env.OPENCODE_CONFIG).toBeUndefined()
+    expect(process.env.OPENCODE_CONFIG_CONTENT).toBeUndefined()
+    expect(process.env.AWS_REGION).toBeUndefined()
+    expect(process.env.GOOGLE_VERTEX_PROJECT).toBeUndefined()
+    expect(process.env.NPM_CONFIG_REGISTRY).toBeUndefined()
+    expect(process.env.UIDOTSH_AUTHORIZATION).toBeUndefined()
+    if (process.env.RECORD !== "true") expect(process.env.OPENAI_API_KEY).toBeUndefined()
+  })
+})

packages/core/test/provider.test.ts

@@ -3,13 +3,16 @@ import { Effect } from "effect"
 import { Provider } from "@opencode-ai/core/provider"
 
 describe("Provider", () => {
-  test("loads Vertex native provider entrypoints", async () => {
+  test("loads bundled native provider entrypoints", async () => {
     const packages = [
+      "@opencode-ai/ai/providers/cerebras",
+      "@opencode-ai/ai/providers/deepinfra",
       "@opencode-ai/ai/providers/google-vertex",
       "@opencode-ai/ai/providers/google-vertex/gemini",
       "@opencode-ai/ai/providers/google-vertex/chat",
       "@opencode-ai/ai/providers/google-vertex/responses",
       "@opencode-ai/ai/providers/google-vertex/messages",
+      "@opencode-ai/ai/providers/togetherai",
     ]
 
     for (const specifier of packages) {

packages/core/test/session-compaction.test.ts

@@ -23,6 +23,7 @@ import { PluginHooks } from "@opencode-ai/core/plugin/hooks"
 import { Location } from "@opencode-ai/core/location"
 import { AbsolutePath } from "@opencode-ai/core/schema"
 import { Money } from "@opencode-ai/schema/money"
+import { Skill } from "@opencode-ai/schema/skill"
 import { DateTime, Effect, Fiber, Layer, Schema, Stream } from "effect"
 import { asc, eq } from "drizzle-orm"
 import { testEffect } from "./lib/effect"
@@ -231,6 +232,13 @@ it.effect("manual compaction summarizes short context instead of no-op", () =>
       id: SessionMessage.ID.create(),
       type: "user" as const,
       text: "Manual compaction should include this short conversation.",
+      skills: [
+        {
+          id: Skill.ID.make("effect"),
+          name: Skill.Name.make("Effect"),
+          text: "Use Effect services and generators.",
+        },
+      ],
       time: { created: DateTime.makeUnsafe(0) },
     }
     const session = yield* insertSession(sessionID, { parent_id: parentID })
@@ -261,6 +269,7 @@ it.effect("manual compaction summarizes short context instead of no-op", () =>
     })
     expect(requests[0]?.generation).toBeUndefined()
     expect(JSON.stringify(requests[0]?.messages)).toContain("Manual compaction should include this short conversation.")
+    expect(JSON.stringify(requests[0]?.messages)).toContain("Use Effect services and generators.")
     expect(yield* store.context(sessionID)).toMatchObject([
       { type: "compaction", reason: "manual", summary: "manual summary", recent: "" },
     ])

packages/core/test/session-create.test.ts

@@ -5,6 +5,7 @@ import path from "path"
 import { DateTime, Effect, Layer, Stream } from "effect"
 import { Money } from "@opencode-ai/schema/money"
 import { Shell } from "@opencode-ai/schema/shell"
+import { Skill } from "@opencode-ai/schema/skill"
 import { Agent } from "@opencode-ai/core/agent"
 import { asc, eq } from "drizzle-orm"
 import { Database } from "@opencode-ai/core/database/database"
@@ -1178,6 +1179,13 @@ describe("SessionTransfer", () => {
               id: sourceMessageID,
               type: "user",
               text: "Imported message",
+              skills: [
+                {
+                  id: Skill.ID.make("effect"),
+                  name: Skill.Name.make("Effect"),
+                  text: "Private skill instructions from /private/project",
+                },
+              ],
               time: { created: DateTime.makeUnsafe(100) },
             },
             {
@@ -1207,7 +1215,11 @@ describe("SessionTransfer", () => {
       const sanitized = yield* transfer.export({ sessionID, sanitize: true })
       expect(sanitized.info.time).toMatchObject({ idle: DateTime.makeUnsafe(200), viewed: DateTime.makeUnsafe(150) })
       expect(sanitized.messages).toMatchObject([
-        { id: sourceMessageID, text: `[redacted:text:${sourceMessageID}]` },
+        {
+          id: sourceMessageID,
+          text: `[redacted:text:${sourceMessageID}]`,
+          skills: [{ id: "effect", name: "[redacted:skill-name:0]", text: "[redacted:skill:0]" }],
+        },
         { id: errorMessageID, error: { type: "test_error", message: "Original error" } },
       ])
 

packages/core/test/session-remove.test.ts

@@ -16,6 +16,7 @@ import { SessionEnvironment } from "@opencode-ai/core/session/environment"
 import { LocationServiceMap } from "@opencode-ai/core/location-services"
 import { testEffect } from "./lib/effect"
 import { globalProjectLayer } from "./lib/project"
+import { tmpdir } from "./fixture/tmpdir"
 
 const closed: Session.ID[] = []
 const transport = Layer.succeed(
@@ -44,17 +45,22 @@ const it = testEffect(
     ],
   ),
 )
-const location = Location.Ref.make({ directory: AbsolutePath.make(import.meta.dir) })
 
 describe("Session.remove", () => {
   it.effect("removes a session and its children", () =>
     Effect.gen(function* () {
+      const temporary = yield* Effect.acquireRelease(
+        Effect.promise(() => tmpdir()),
+        (directory) => Effect.promise(() => directory[Symbol.asyncDispose]()),
+      )
+      const location = Location.Ref.make({ directory: AbsolutePath.make(temporary.path) })
       const session = yield* Session.Service
       const parent = yield* session.create({ location })
       const child = yield* session.create({ parentID: parent.id })
       yield* session.environment({ sessionID: parent.id, variables: { SESSION_ENV: "parent" } })
       yield* session.environment({ sessionID: child.id, variables: { SESSION_ENV: "child" } })
-      yield* (yield* LocationServiceMap.Service).contextEffect(location)
+      const locations = yield* LocationServiceMap.Service
+      yield* Effect.acquireRelease(locations.contextEffect(location), () => locations.invalidate(location))
       closed.length = 0
 
       yield* session.remove(parent.id)

packages/core/test/session-runner-message.test.ts

@@ -205,6 +205,44 @@ Recent work
     })
   })
 
+  test("lowers each prepared skill once before the prompt", () => {
+    const effect = SkillAttachment.make({
+      id: Skill.ID.make("effect"),
+      name: Skill.Name.make("Effect"),
+      text: "<skill_content>Use Effect</skill_content>",
+    })
+    const api = SkillAttachment.make({
+      id: Skill.ID.make("api-design"),
+      name: Skill.Name.make("API design"),
+      text: "<skill_content>Design APIs</skill_content>",
+    })
+    const messages = toLLMMessages(
+      [
+        SessionMessage.User.make({
+          id: id("user-skill-content"),
+          type: "user",
+          text: "Use @effect and @api-design",
+          skills: [effect, api, SkillAttachment.make({ id: effect.id, name: effect.name })],
+          time: { created },
+        }),
+      ],
+      model,
+    )
+
+    expect(messages).toEqual([
+      Message.make({
+        id: id("user-skill-content"),
+        role: "user",
+        content: [
+          { type: "text", text: "<skill_content>Use Effect</skill_content>" },
+          { type: "text", text: "<skill_content>Design APIs</skill_content>" },
+          { type: "text", text: "Use @effect and @api-design" },
+        ],
+        metadata: {},
+      }),
+    ])
+  })
+
   test("does not inject skill content for reference-only attachments", () => {
     const messages = toLLMMessages(
       [

packages/core/test/session-runner.test.ts

@@ -33,6 +33,7 @@ import { Session } from "@opencode-ai/core/session"
 import { Snapshot } from "@opencode-ai/core/snapshot"
 import { SessionEvent } from "@opencode-ai/core/session/event"
 import { SessionContext } from "@opencode-ai/core/session/context"
+import { SessionCompaction } from "@opencode-ai/core/session/compaction"
 import { SessionInbox } from "@opencode-ai/core/session/inbox"
 import { SessionMessage } from "@opencode-ai/core/session/message"
 import { SessionModelRequest } from "@opencode-ai/core/session/model-request"
@@ -458,6 +459,7 @@ const it = testEffect(
       Config.node,
       Snapshot.node,
       SessionContext.node,
+      SessionCompaction.node,
       SessionModelRequest.node,
       SessionRunnerLLM.node,
       SessionExecution.node,
@@ -2560,6 +2562,34 @@ describe("SessionRunnerLLM", () => {
     }),
   )
 
+  it.effect("does not recover provider context overflow when automatic compaction is disabled", () =>
+    Effect.gen(function* () {
+      const session = yield* setupOverflowRecovery
+      const compaction = yield* SessionCompaction.Service
+      yield* compaction.transform((draft) => draft.configure({ auto: false }))
+      yield* TestLLM.push(
+        [LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" })],
+        TestLLM.text("Must not compact", "text-unexpected-summary"),
+        TestLLM.text("Must not retry", "text-unexpected-retry"),
+      )
+      yield* admit(session, "Continue")
+      expect((yield* session.resume(sessionID).pipe(Effect.flip)).message).toBe("prompt too long")
+
+      expect(requests).toHaveLength(1)
+      expect(yield* session.context(sessionID)).toContainEqual(
+        expect.objectContaining({
+          type: "assistant",
+          finish: "error",
+          error: expect.objectContaining({ message: "prompt too long" }),
+        }),
+      )
+      expect(yield* session.context(sessionID)).not.toContainEqual(expect.objectContaining({ type: "compaction" }))
+      expect(yield* recordedEventTypes(sessionID)).not.toContain(
+        Bus.versionedType(SessionEvent.Compaction.Started.type, 1),
+      )
+    }),
+  )
+
   it.effect("recovers from provider context overflow without a configured context limit", () =>
     Effect.gen(function* () {
       const session = yield* setupOverflowRecovery

packages/core/test/session-skill.test.ts

@@ -9,6 +9,7 @@ import { Location } from "@opencode-ai/core/location"
 import { LocationServiceMap } from "@opencode-ai/core/location-service-map"
 import type { LocationServices } from "@opencode-ai/core/location-services"
 import { Project } from "@opencode-ai/core/project"
+import { PluginSupervisor } from "@opencode-ai/core/plugin/supervisor-service"
 import { AbsolutePath } from "@opencode-ai/core/schema"
 import { Session } from "@opencode-ai/core/session"
 import { SessionExecution } from "@opencode-ai/core/session/execution"
@@ -23,18 +24,20 @@ const location = Location.Ref.make({ directory: AbsolutePath.make("/project") })
 const projects = Layer.mock(Project.Service, {
   resolve: (directory) => Effect.succeed({ id: Project.ID.global, directory, canonical: directory }),
 })
-const skills = Layer.mock(Skill.Service, {
-  list: () =>
-    Effect.succeed([
-      Skill.Info.make({
-        id: Skill.ID.make("effect"),
-        name: Skill.Name.make("Effect"),
-        description: "Effect guidance",
-        location: AbsolutePath.make(path.resolve("/skills/effect/SKILL.md")),
-        content: "Use Effect",
-      }),
-    ]),
+const info = Skill.Info.make({
+  id: Skill.ID.make("effect"),
+  name: Skill.Name.make("Effect"),
+  description: "Effect guidance",
+  location: AbsolutePath.make(path.resolve("/skills/effect.md")),
+  content: "Use Effect",
 })
+const skills = Layer.merge(
+  Layer.mock(Skill.Service, {
+    get: (id) => Effect.succeed(id === info.id ? info : undefined),
+    list: () => Effect.succeed([info]),
+  }),
+  Layer.succeed(PluginSupervisor.Service, { flush: Effect.void }),
+)
 const locations = Layer.effect(
   LocationServiceMap.Service,
   LayerMap.make(
@@ -56,7 +59,7 @@ const it = testEffect(
 )
 
 describe("Session.skill", () => {
-  it.effect("keeps skill mentions as references on a normal prompt", () =>
+  it.effect("materializes mentioned skills on their owning prompt", () =>
     Effect.gen(function* () {
       const sessions = yield* Session.Service
       const database = yield* Database.Service
@@ -67,26 +70,63 @@ describe("Session.skill", () => {
       yield* sessions.prompt({
         id,
         sessionID: session.id,
-        text: "Apply @effect",
-        skills: [{ id: Skill.ID.make("effect"), mention: { start: 6, end: 13, text: "@effect" } }],
+        text: "Apply @effect and @effect",
+        skills: [
+          { id: Skill.ID.make("effect"), mention: { start: 6, end: 13, text: "@effect" } },
+          { id: Skill.ID.make("effect"), mention: { start: 18, end: 25, text: "@effect" } },
+        ],
         resume: false,
       })
+      expect(yield* sessions.messages({ sessionID: session.id })).toEqual([])
       yield* SessionInbox.promote(database.db, bus, session.id, "steer")
 
-      expect(yield* sessions.messages({ sessionID: session.id })).toContainEqual(
+      expect(yield* sessions.messages({ sessionID: session.id })).toEqual([
         expect.objectContaining({
           id,
           type: "user",
-          text: "Apply @effect",
+          text: "Apply @effect and @effect",
           skills: [
             {
               id: "effect",
               name: "Effect",
+              text: Skill.toModelOutput(info, []),
               mention: { start: 6, end: 13, text: "@effect" },
             },
+            {
+              id: "effect",
+              name: "Effect",
+              mention: { start: 18, end: 25, text: "@effect" },
+            },
           ],
         }),
-      )
+      ])
+    }),
+  )
+
+  it.effect("excludes mentioned skills when forking before their prompt", () =>
+    Effect.gen(function* () {
+      const sessions = yield* Session.Service
+      const database = yield* Database.Service
+      const bus = yield* Bus.Service
+      const session = yield* sessions.create({ location })
+      const initial = SessionMessage.ID.make("msg_before_skill_attachment")
+      const selected = SessionMessage.ID.make("msg_fork_skill_attachment")
+
+      yield* sessions.prompt({ id: initial, sessionID: session.id, text: "Before the skill", resume: false })
+      yield* SessionInbox.promote(database.db, bus, session.id, "steer")
+      yield* sessions.prompt({
+        id: selected,
+        sessionID: session.id,
+        text: "Apply @effect",
+        skills: [{ id: info.id, mention: { start: 6, end: 13, text: "@effect" } }],
+        resume: false,
+      })
+      yield* SessionInbox.promote(database.db, bus, session.id, "steer")
+      const forked = yield* sessions.fork({ sessionID: session.id, boundary: { type: "before", messageID: selected } })
+
+      expect(yield* sessions.messages({ sessionID: forked.id })).toEqual([
+        expect.objectContaining({ type: "user", text: "Before the skill" }),
+      ])
     }),
   )
 

packages/core/test/skill.test.ts

@@ -31,6 +31,8 @@ describe("Skill", () => {
       })
 
       expect(yield* skill.list()).toEqual([info("review", "Second"), info("deploy", "Deploy")])
+      expect(yield* skill.get(Skill.ID.make("review"))).toEqual(info("review", "Second"))
+      expect(yield* skill.get(Skill.ID.make("missing"))).toBeUndefined()
     }),
   )
 

packages/core/test/skill/instructions.test.ts

@@ -59,7 +59,6 @@ describe("SkillInstructions", () => {
         [
           "Skills provide specialized instructions and workflows for specific tasks.",
           "Use the skill tool to load a skill when a task matches its description.",
-          "When the user references a skill with @skill-id, load that skill with the skill tool.",
           "<available_skills>",
           "  <skill>",
           "    <id>effect</id>",