docs(plugin): correct plugin state reference

Co-authored-by: rekram1-node <rekram1-node@users.noreply.github.com>
Co-authored-by: Aiden Cline <63023139+rekram1-node@users.noreply.github.com>

.github/workflows/test.yml

@@ -94,12 +94,6 @@ jobs:
           git config --global user.email "bot@opencode.ai"
           git config --global user.name "opencode"
 
-      - name: Install ffmpeg
-        if: runner.os == 'Linux'
-        run: |
-          sudo apt-get update
-          sudo apt-get install --yes ffmpeg
-
       - name: Cache Turbo
         uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
         with:

bun.lock

@@ -118,7 +118,7 @@
         "opencode2": "./bin/opencode2.cjs",
       },
       "dependencies": {
-        "@agentclientprotocol/sdk": "1.2.1",
+        "@agentclientprotocol/sdk": "1.6.0",
         "@clack/core": "1.0.0-alpha.1",
         "@clack/prompts": "1.0.0-alpha.1",
         "@effect/platform-node": "catalog:",
@@ -1230,7 +1230,7 @@
 
     "@adobe/css-tools": ["@adobe/css-tools@4.5.0", "", {}, "sha512-6OzddxPio9UiWTCemp4N8cYLV2ZN1ncRnV1cVGtve7dhPOtRkleRyx32GQCYSwDYgaHU3USMm84tNsvKzRCa1Q=="],
 
-    "@agentclientprotocol/sdk": ["@agentclientprotocol/sdk@1.2.1", "", { "peerDependencies": { "zod": "^3.25.0 || ^4.0.0" } }, "sha512-jwYUdOQR7tc+Zfch53VL4JJyUNK/46q03uUTYb+PjECsmnNl94XFXOfYLJ8RBpMNidXd1rpOAVgb0vqD98xImA=="],
+    "@agentclientprotocol/sdk": ["@agentclientprotocol/sdk@1.6.0", "", { "peerDependencies": { "zod": "^3.25.0 || ^4.0.0" } }, "sha512-XxXrmX7aZkDgOB0Rg9cu+ZFyiUUc5lF2n9seO3Gc4OR+MTdfZOwIqF6m3LvsmmM8K3qgPmXkDH8/IFM2u9vdcQ=="],
 
     "@ai-sdk/cohere": ["@ai-sdk/cohere@3.0.27", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.21" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-OqcCq2PiFY1dbK/0Ck45KuvE8jfdxRuuAE9Y5w46dAk6U+9vPOeg1CDcmR+ncqmrYrhRl3nmyDttyDahyjCzAw=="],
 

nix/hashes.json

@@ -1,7 +1,7 @@
 {
   "nodeModules": {
-    "x86_64-linux": "sha256-yCdtDQsXERjfL9bJg7YXNloMBOPwFkPMFVNAzXlarBM=",
-    "aarch64-linux": "sha256-iQ1bLIszETcFoD4CENpoaDZJZF/KBKZFeTp9r7ogJ9Y=",
-    "aarch64-darwin": "sha256-G7oIrTXEFQ5iEF8KSpA4xJPw9xwqRDpbIR472iIoqC4="
+    "x86_64-linux": "sha256-g3k0cAFGqzmRYlcIkg1NDvlx1WxHYhnYPL0/a8E+qTg=",
+    "aarch64-linux": "sha256-a+3ymqdxOONGe2Tpq4GUccl1b+Dwzxlb9LFXgE1gZ+0=",
+    "aarch64-darwin": "sha256-h8xIzuMmaWfJqjHCO74xUDCWNKQLFrIGoKYZ+2TauYc="
   }
 }

packages/ai/src/cache-policy.ts

@@ -49,6 +49,7 @@ const RESPECTS_INLINE_HINTS = new Set([
   "zai-coding-messages",
   "bedrock-converse",
   "openrouter",
+  "digitalocean",
 ])
 
 // OpenRouter upstreams other than Anthropic and Alibaba Qwen cache without breakpoints. Gemini uses only the last

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

@@ -807,15 +807,12 @@ const requireThinkingSignature = (request: LLMRequest) => {
 // Mid-conversation system messages became available with Opus 4.8 and version
 // 5 of the other supported Claude families. Treat later family versions as
 // compatible without assuming that every Anthropic Messages model is Claude.
+// Opus 4.8 and every Claude 5 model accept mid-conversation system messages; later versions inherit support.
 const supportsNativeSystemUpdates = (request: LLMRequest) => {
-  const match = /(?:^|[./])claude-(fable|haiku|mythos|opus|sonnet)-(\d+)(?:[.-](\d+))?/.exec(
-    String(request.model.id).toLowerCase(),
-  )
-  if (!match) return false
-  const major = Number(match[2])
-  if (match[1] !== "opus") return major >= 5
-  if (major !== 4) return major >= 5
-  return match[3] !== undefined && match[3].length <= 2 && Number(match[3]) >= 8
+  const version = claudeVersion(String(request.model.id))
+  if (version === undefined) return false
+  if (version.family === "opus" && version.major === 4) return version.minor >= 8
+  return version.major >= 5
 }
 
 const endsInServerToolUse = (message: LLMRequest["messages"][number]) => {
@@ -992,13 +989,13 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
   return messages
 })
 
+// Per-turn effort started with Claude Opus 5 and every Claude 5.1 model; later versions of any family inherit it.
 const supportsEffortUpdates = (model: LLMRequest["model"]) => {
   const override = model.compatibility?.supportsEffortUpdates
   if (override !== undefined) return override
   const version = claudeVersion(model.id)
   if (version === undefined) return false
-  if (version.family === "opus") return version.major >= 5
-  if (version.family !== "fable" && version.family !== "mythos") return false
+  if (version.family === "opus" && version.major >= 5) return true
   return version.major > 5 || (version.major === 5 && version.minor >= 1)
 }
 

packages/ai/src/protocols/bedrock-converse.ts

@@ -319,6 +319,9 @@ const lowerToolResult = Effect.fn("BedrockConverse.lowerToolResult")(function* (
   } satisfies BedrockToolResultBlock
 })
 
+// Keep Claude and Nova tool-result images inline; put other models' images beside the result.
+const keepToolImagesInline = (id: string) => id.includes("anthropic.claude-") || id.includes("amazon.nova-")
+
 const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
   request: LLMRequest,
   breakpoints: BedrockCache.Breakpoints,
@@ -328,8 +331,19 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
   // Mistral can reject replay IDs even when they satisfy Converse's broader ID syntax.
   const normalizeID = request.model.id.includes("mistral.") ? MistralToolID.normalizer(request) : (id: string) => id
   const providerMetadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)
+  const hoistImages = !keepToolImagesInline(request.model.id)
+  // Bedrock expects parallel tool results before any images hoisted beside them.
+  const pendingImages: BedrockMedia.ImageBlock[] = []
+  const flushImages = () => {
+    if (pendingImages.length === 0) return
+    const previous = messages.at(-1)
+    if (previous?.role === "user")
+      messages[messages.length - 1] = { role: "user", content: [...previous.content, ...pendingImages] }
+    pendingImages.length = 0
+  }
 
   for (const message of request.messages) {
+    if (message.role !== "tool") flushImages()
     if (message.role === "system") {
       const part = yield* ProviderShared.wrappedSystemUpdate("Bedrock Converse", message)
       const content = textWithCache(breakpoints, part.text, part.cache)
@@ -403,7 +417,22 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
     for (const part of message.content) {
       if (!ProviderShared.supportsContent(part, ["tool-result"]))
         return yield* ProviderShared.unsupportedContent("Bedrock Converse", "tool", ["tool-result"])
-      content.push(yield* lowerToolResult(part, documentNames, normalizeID))
+      const result = yield* lowerToolResult(part, documentNames, normalizeID)
+      const images: BedrockMedia.ImageBlock[] = hoistImages
+        ? result.toolResult.content.filter((item) => "image" in item)
+        : []
+      const nonImageContent = result.toolResult.content.filter((item) => !("image" in item))
+      content.push(
+        images.length === 0
+          ? result
+          : {
+              toolResult: {
+                ...result.toolResult,
+                content: nonImageContent.length > 0 ? nonImageContent : [{ text: "See attached image." }],
+              },
+            },
+      )
+      pendingImages.push(...images)
       const cachePoint = BedrockCache.block(breakpoints, part.cache)
       if (cachePoint) content.push(cachePoint)
     }
@@ -413,6 +442,7 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
     else messages.push({ role: "user", content })
   }
 
+  flushImages()
   return messages
 })
 

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

@@ -212,9 +212,11 @@ const OpenAIChatUsage = Schema.StructWithRest(
     prompt_tokens: optionalNull(Schema.Number),
     completion_tokens: optionalNull(Schema.Number),
     total_tokens: optionalNull(Schema.Number),
-    // Zai reports cache hits as top-level `cached_tokens`; DeepSeek uses `prompt_cache_hit_tokens`.
+    // Provider-specific cache accounting fields.
     cached_tokens: optionalNull(Schema.Number),
     prompt_cache_hit_tokens: optionalNull(Schema.Number),
+    cache_read_input_tokens: optionalNull(Schema.Number),
+    cache_created_input_tokens: optionalNull(Schema.Number),
     prompt_tokens_details: optionalNull(
       Schema.StructWithRest(
         Schema.Struct({
@@ -903,16 +905,19 @@ const mapFinishReason = Effect.fn("OpenAIChat.mapFinishReason")(function* (event
 // satisfied on both sides.
 // Providers differ on cache-hit location: OpenAI uses
 // `prompt_tokens_details.cached_tokens`, DeepSeek uses
-// `prompt_cache_hit_tokens`, and Zai uses top-level `cached_tokens`.
+// `prompt_cache_hit_tokens`, Zai uses top-level `cached_tokens`, and
+// DigitalOcean uses top-level `cache_read_input_tokens` / `cache_created_input_tokens`.
 const mapUsage = (usage: OpenAIChatEvent["usage"], providerMetadataKey: string): Usage | undefined => {
   if (!usage) return undefined
   const input = usage.prompt_tokens ?? undefined
   const output = usage.completion_tokens ?? undefined
-  const cached = (usage.prompt_tokens_details?.cached_tokens ??
-    (usage as { prompt_cache_hit_tokens?: number | null }).prompt_cache_hit_tokens ??
-    (usage as { cached_tokens?: number | null }).cached_tokens ??
-    undefined) as number | undefined
-  const cacheWrite = usage.prompt_tokens_details?.cache_write_tokens ?? undefined
+  const cached =
+    usage.prompt_tokens_details?.cached_tokens ??
+    usage.prompt_cache_hit_tokens ??
+    usage.cached_tokens ??
+    usage.cache_read_input_tokens ??
+    undefined
+  const cacheWrite = usage.prompt_tokens_details?.cache_write_tokens ?? usage.cache_created_input_tokens ?? undefined
   const reasoning = usage.completion_tokens_details?.reasoning_tokens ?? undefined
   const nonCached = ProviderShared.subtractTokens(input, ProviderShared.sumTokens(cached, cacheWrite))
   return new Usage({

packages/ai/src/protocols/utils/cache.ts

@@ -1,4 +1,5 @@
 // Shared counter and TTL mapping for provider cache-marker lowering.
+import type { CacheHint } from "../../schema/index.js"
 
 export interface Breakpoints {
   remaining: number
@@ -11,3 +12,12 @@ export const newBreakpoints = (cap: number): Breakpoints => ({ remaining: cap, d
 // requests omit the wire TTL and use the provider default.
 export const ttlBucket = (ttlSeconds: number | undefined): "1h" | undefined =>
   ttlSeconds !== undefined && ttlSeconds >= 3600 ? "1h" : undefined
+
+export const cacheControl = () => {
+  const breakpoints = newBreakpoints(4)
+  return (cache: CacheHint | undefined) => {
+    if (cache === undefined || breakpoints.remaining === 0) return undefined
+    breakpoints.remaining -= 1
+    return { type: "ephemeral" as const, ttl: ttlBucket(cache.ttlSeconds) }
+  }
+}

packages/ai/src/provider-error.ts

@@ -35,6 +35,8 @@ const patterns = [
   /exceeds the limit of \d+/i,
   /exceeds the available context size/i,
   /greater than the context length/i,
+  // Hugging Face Text Generation Inference, e.g. Together
+  /`inputs` tokens \+ `max_new_tokens` must be <= \d+/i,
   /context window exceeds limit/i,
   /exceeded model token limit/i,
   /context[_ ]length[_ ]exceeded/i,

packages/ai/src/providers/digitalocean.ts

@@ -0,0 +1,76 @@
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAIChat } from "../protocols/openai-chat.js"
+import { cacheControl } from "../protocols/utils/cache.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import { Route, type RouteDefaultsInput } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
+import { Protocol } from "../route/protocol.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("digitalocean")
+const baseURL = "https://inference.do-ai.run/v1"
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
+
+export type Settings = ProviderPackage.Settings & OpenAIProviderOptionsInput & { readonly apiKey?: string }
+
+export const protocol = Protocol.make({
+  id: "digitalocean-chat",
+  body: {
+    schema: OpenAIChat.protocol.body.schema,
+    from: (request) => OpenAIChat.fromRequest(request, { cacheControl: cacheControl() }),
+  },
+  stream: OpenAIChat.protocol.stream,
+})
+
+export const route = Route.make({
+  id: "digitalocean",
+  provider: id,
+  providerMetadataKey: "digitalocean",
+  protocol,
+  endpoint: Endpoint.path("/chat/completions", { baseURL }),
+  framing: OpenAIChat.framing,
+})
+
+export const routes = [route]
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL: endpoint, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: endpoint ?? baseURL },
+    auth: AuthOptions.bearer(input, ["DIGITALOCEAN_ACCESS_TOKEN", "DIGITALOCEAN_API_KEY", "DO_INFERENCE_API_KEY"]),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<OpenAIProviderOptionsInput>({
+        id: modelID,
+        compatibility: {
+          supportsPromptCacheKey: true,
+        },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (
+  modelID,
+  { apiKey, baseURL, body, headers, ...providerOptions },
+) =>
+  configure({
+    apiKey,
+    baseURL,
+    headers,
+    http: { body },
+    providerOptions,
+  }).model(modelID)
+
+export * as DigitalOcean from "./digitalocean.js"

packages/ai/src/providers/google-vertex-messages.ts

@@ -37,7 +37,7 @@ export type Settings = ProviderPackage.Settings &
 const route = Route.make({
   id: "google-vertex-messages",
   provider: id,
-  providerMetadataKey: "anthropic",
+  providerMetadataKey: "vertex",
   protocol: Protocol.make({
     id: AnthropicMessages.protocol.id,
     body: {

packages/ai/src/providers/groq.ts

@@ -71,7 +71,7 @@ export const protocol = Protocol.make({
 export const route = Route.make({
   id: "groq-chat",
   provider: id,
-  providerMetadataKey: "openai",
+  providerMetadataKey: "groq",
   protocol,
   endpoint: Endpoint.path("/chat/completions", { baseURL }),
   framing: OpenAIChat.framing,

packages/ai/src/providers/index.ts

@@ -14,6 +14,7 @@ export * as CloudflareWorkersAI from "./cloudflare-workers-ai.js"
 export * as DeepInfra from "./deepinfra.js"
 export * as Deepgram from "./deepgram.js"
 export * as DeepSeek from "./deepseek.js"
+export * as DigitalOcean from "./digitalocean.js"
 export * as ElevenLabs from "./elevenlabs.js"
 export * as Fal from "./fal.js"
 export * as Fireworks from "./fireworks.js"

packages/ai/src/providers/openrouter.ts

@@ -3,11 +3,11 @@ import { Route, type RouteDefaultsInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
 import { Protocol } from "../route/protocol.js"
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
-import { HttpOptions, ProviderID, type CacheHint, type ModelID, type OpenString } from "../schema/index.js"
+import { HttpOptions, ProviderID, type ModelID, type OpenString } from "../schema/index.js"
 import type { ProviderPackage } from "../provider-package.js"
 import { SystemOne } from "../experimental/system-one.js"
 import { OpenAIChat } from "../protocols/openai-chat.js"
-import { newBreakpoints, ttlBucket } from "../protocols/utils/cache.js"
+import { cacheControl } from "../protocols/utils/cache.js"
 import { isRecord, ProviderShared } from "../protocols/shared.js"
 
 export const id = ProviderID.make("openrouter")
@@ -131,18 +131,6 @@ export const protocol = Protocol.make({
   stream: OpenAIChat.protocol.stream,
 })
 
-const cacheControl = () => {
-  const breakpoints = newBreakpoints(4)
-  return (cache: CacheHint | undefined) => {
-    if (cache === undefined || breakpoints.remaining === 0) return undefined
-    breakpoints.remaining -= 1
-    return {
-      type: "ephemeral" as const,
-      ...(ttlBucket(cache.ttlSeconds) === "1h" ? { ttl: "1h" } : {}),
-    }
-  }
-}
-
 // OpenRouter forwards `reasoning.max_tokens` as the upstream thinking budget. Upstreams such as Anthropic and Alibaba
 // reject one that is not below the output limit; 1,024 is Anthropic's minimum budget.
 const fitReasoning = (reasoning: Record<string, unknown>, maxTokens: number | undefined) =>

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

@@ -8,6 +8,7 @@ import {
   AmazonBedrock,
   AnthropicCompatible,
   CloudflareAIGateway,
+  DigitalOcean,
   GoogleVertexMessages,
   Meta,
   MiniMax,
@@ -117,6 +118,33 @@ describe("applyCachePolicy", () => {
     }),
   )
 
+  it.effect("'auto' emits cache_control markers on DigitalOcean", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: DigitalOcean.configure({ apiKey: "test" }).model("anthropic-claude-fable-5.1"),
+          system: "You are concise.",
+          tools: [{ name: "lookup", description: "Look up a value", inputSchema: { type: "object", properties: {} } }],
+          prompt: "hi",
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({
+        tools: [{ type: "function", function: { name: "lookup" }, cache_control: { type: "ephemeral" } }],
+        messages: [
+          {
+            role: "system",
+            content: [{ text: "You are concise.", cache_control: { type: "ephemeral" } }],
+          },
+          {
+            role: "user",
+            content: [{ text: "hi", cache_control: { type: "ephemeral" } }],
+          },
+        ],
+      })
+    }),
+  )
+
   it.effect("'auto' emits Anthropic cache markers on Anthropic-compatible routes", () =>
     Effect.gen(function* () {
       const prepared = yield* compileRequest(

packages/ai/test/effort-updates.test.ts

@@ -197,9 +197,15 @@ describe("Anthropic Messages effort updates", () => {
     ["anthropic/claude-opus-5", true],
     ["claude-fable-5-1", true],
     ["claude-mythos-5-1", true],
+    ["claude-opus-5-5", true],
+    ["claude-sonnet-5-5", true],
+    ["anthropic/claude-sonnet-5-5", true],
+    ["claude-sonnet-6", true],
+    ["claude-haiku-6", true],
     ["claude-fable-5", false],
     ["claude-opus-4-8", false],
     ["claude-sonnet-5", false],
+    ["claude-sonnet-5-20260801", false],
     ["kimi-k2.5", false],
   ] as const) {
     it.effect(`${supported ? "lowers" : "strips"} markers for ${id}`, () =>

packages/ai/test/fixtures/recordings/digitalocean-chat/gpt-nano-continues-a-tool-call-with-cache-markers.json

@@ -0,0 +1,54 @@
+{
+  "version": 1,
+  "metadata": {
+    "model": "openai-gpt-5-nano",
+    "tags": [
+      "prefix:digitalocean-chat",
+      "provider:digitalocean",
+      "protocol:digitalocean-chat",
+      "cache",
+      "tool",
+      "tool-loop"
+    ],
+    "name": "digitalocean-chat/gpt-nano-continues-a-tool-call-with-cache-markers",
+    "recordedAt": "2026-10-01T07:36:41.717Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://inference.do-ai.run/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"openai-gpt-5-nano\",\"messages\":[{\"role\":\"system\",\"content\":[{\"type\":\"text\",\"text\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\",\"cache_control\":{\"type\":\"ephemeral\"}}]},{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"What is the weather in Paris?\",\"cache_control\":{\"type\":\"ephemeral\"}}]}],\"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},\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"store\":false,\"reasoning_effort\":\"minimal\",\"max_completion_tokens\":1024}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"choices\":[{\"delta\":{\"content\":\"\",\"role\":\"assistant\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840200,\"id\":\"chatcmpl-EU5dIREwFebLDdrnmX7ftaDIK4qrZ\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"\",\"name\":\"get_weather\"},\"id\":\"call_zPwN7xhG4uXQINkufgtjVJ7L\",\"index\":0,\"type\":\"function\"}]},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840200,\"id\":\"chatcmpl-EU5dIREwFebLDdrnmX7ftaDIK4qrZ\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"{\\\"\"},\"index\":0,\"type\":\"function\"}]},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840200,\"id\":\"chatcmpl-EU5dIREwFebLDdrnmX7ftaDIK4qrZ\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"city\"},\"index\":0,\"type\":\"function\"}]},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840200,\"id\":\"chatcmpl-EU5dIREwFebLDdrnmX7ftaDIK4qrZ\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"\\\":\\\"\"},\"index\":0,\"type\":\"function\"}]},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840200,\"id\":\"chatcmpl-EU5dIREwFebLDdrnmX7ftaDIK4qrZ\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"Paris\"},\"index\":0,\"type\":\"function\"}]},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840200,\"id\":\"chatcmpl-EU5dIREwFebLDdrnmX7ftaDIK4qrZ\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"\\\"}\"},\"index\":0,\"type\":\"function\"}]},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840200,\"id\":\"chatcmpl-EU5dIREwFebLDdrnmX7ftaDIK4qrZ\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{},\"finish_reason\":\"tool_calls\",\"index\":0,\"logprobs\":null}],\"created\":1790840200,\"id\":\"chatcmpl-EU5dIREwFebLDdrnmX7ftaDIK4qrZ\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[],\"created\":1790840200,\"id\":\"chatcmpl-EU5dIREwFebLDdrnmX7ftaDIK4qrZ\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\",\"usage\":{\"cache_created_input_tokens\":0,\"cache_creation\":{\"ephemeral_1h_input_tokens\":0,\"ephemeral_5m_input_tokens\":0},\"cache_read_input_tokens\":0,\"completion_tokens\":23,\"completion_tokens_details\":{\"reasoning_tokens\":0},\"prompt_tokens\":155,\"prompt_tokens_details\":{\"cached_tokens\":0},\"speed\":null,\"total_tokens\":178}}\n\ndata: [DONE]\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://inference.do-ai.run/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"openai-gpt-5-nano\",\"messages\":[{\"role\":\"system\",\"content\":[{\"type\":\"text\",\"text\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\",\"cache_control\":{\"type\":\"ephemeral\"}}]},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"},{\"role\":\"assistant\",\"content\":null,\"tool_calls\":[{\"id\":\"call_zPwN7xhG4uXQINkufgtjVJ7L\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}]},{\"role\":\"tool\",\"tool_call_id\":\"call_zPwN7xhG4uXQINkufgtjVJ7L\",\"content\":[{\"type\":\"text\",\"text\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\",\"cache_control\":{\"type\":\"ephemeral\"}}]}],\"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},\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"store\":false,\"reasoning_effort\":\"minimal\",\"max_completion_tokens\":1024}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"choices\":[{\"delta\":{\"content\":\"\",\"role\":\"assistant\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840201,\"id\":\"chatcmpl-EU5dJprTkE7OHEKnmvot7rR0LSo5L\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"content\":\"Paris\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840201,\"id\":\"chatcmpl-EU5dJprTkE7OHEKnmvot7rR0LSo5L\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"content\":\" is\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840201,\"id\":\"chatcmpl-EU5dJprTkE7OHEKnmvot7rR0LSo5L\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"content\":\" sunny\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840201,\"id\":\"chatcmpl-EU5dJprTkE7OHEKnmvot7rR0LSo5L\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"content\":\".\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840201,\"id\":\"chatcmpl-EU5dJprTkE7OHEKnmvot7rR0LSo5L\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\",\"index\":0,\"logprobs\":null}],\"created\":1790840201,\"id\":\"chatcmpl-EU5dJprTkE7OHEKnmvot7rR0LSo5L\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[],\"created\":1790840201,\"id\":\"chatcmpl-EU5dJprTkE7OHEKnmvot7rR0LSo5L\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\",\"usage\":{\"cache_created_input_tokens\":0,\"cache_creation\":{\"ephemeral_1h_input_tokens\":0,\"ephemeral_5m_input_tokens\":0},\"cache_read_input_tokens\":0,\"completion_tokens\":13,\"completion_tokens_details\":{\"reasoning_tokens\":0},\"prompt_tokens\":193,\"prompt_tokens_details\":{\"cached_tokens\":0},\"speed\":null,\"total_tokens\":206}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/digitalocean-chat/gpt-nano-reuses-a-cached-prompt.json

@@ -0,0 +1,53 @@
+{
+  "version": 1,
+  "metadata": {
+    "model": "openai-gpt-5-nano",
+    "tags": [
+      "prefix:digitalocean-chat",
+      "provider:digitalocean",
+      "protocol:digitalocean-chat",
+      "cache",
+      "usage"
+    ],
+    "name": "digitalocean-chat/gpt-nano-reuses-a-cached-prompt",
+    "recordedAt": "2026-10-01T07:36:40.074Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://inference.do-ai.run/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"openai-gpt-5-nano\",\"messages\":[{\"role\":\"system\",\"content\":[{\"type\":\"text\",\"text\":\"You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid fillLine truncated
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"choices\":[{\"delta\":{\"content\":\"\",\"role\":\"assistant\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840199,\"id\":\"chatcmpl-EU5dHjjHh6g7s4uSL7kMOOpLEXVki\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"content\":\"OK\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840199,\"id\":\"chatcmpl-EU5dHjjHh6g7s4uSL7kMOOpLEXVki\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\",\"index\":0,\"logprobs\":null}],\"created\":1790840199,\"id\":\"chatcmpl-EU5dHjjHh6g7s4uSL7kMOOpLEXVki\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[],\"created\":1790840199,\"id\":\"chatcmpl-EU5dHjjHh6g7s4uSL7kMOOpLEXVki\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\",\"usage\":{\"cache_created_input_tokens\":0,\"cache_creation\":{\"ephemeral_1h_input_tokens\":0,\"ephemeral_5m_input_tokens\":0},\"cache_read_input_tokens\":0,\"completion_tokens\":10,\"completion_tokens_details\":{\"reasoning_tokens\":0},\"prompt_tokens\":4765,\"prompt_tokens_details\":{\"cached_tokens\":0},\"speed\":null,\"total_tokens\":4775}}\n\ndata: [DONE]\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://inference.do-ai.run/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"openai-gpt-5-nano\",\"messages\":[{\"role\":\"system\",\"content\":[{\"type\":\"text\",\"text\":\"You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid fillLine truncated
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"choices\":[{\"delta\":{\"content\":\"\",\"role\":\"assistant\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840200,\"id\":\"chatcmpl-EU5dHYe8HWjEAEIIvsJ6Hk0EUxL9r\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"content\":\"OK\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840200,\"id\":\"chatcmpl-EU5dHYe8HWjEAEIIvsJ6Hk0EUxL9r\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\",\"index\":0,\"logprobs\":null}],\"created\":1790840200,\"id\":\"chatcmpl-EU5dHYe8HWjEAEIIvsJ6Hk0EUxL9r\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[],\"created\":1790840200,\"id\":\"chatcmpl-EU5dHYe8HWjEAEIIvsJ6Hk0EUxL9r\",\"model\":\"openai-gpt-5-nano\",\"object\":\"chat.completion.chunk\",\"usage\":{\"cache_created_input_tokens\":0,\"cache_creation\":{\"ephemeral_1h_input_tokens\":0,\"ephemeral_5m_input_tokens\":0},\"cache_read_input_tokens\":4608,\"completion_tokens\":10,\"completion_tokens_details\":{\"reasoning_tokens\":0},\"prompt_tokens\":4765,\"prompt_tokens_details\":{\"cached_tokens\":4608},\"speed\":null,\"total_tokens\":4775}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/digitalocean-chat/haiku-continues-a-tool-call-with-cache-markers.json

@@ -0,0 +1,54 @@
+{
+  "version": 1,
+  "metadata": {
+    "model": "anthropic-claude-haiku-4.5",
+    "tags": [
+      "prefix:digitalocean-chat",
+      "provider:digitalocean",
+      "protocol:digitalocean-chat",
+      "cache",
+      "tool",
+      "tool-loop"
+    ],
+    "name": "digitalocean-chat/haiku-continues-a-tool-call-with-cache-markers",
+    "recordedAt": "2026-10-01T07:36:38.471Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://inference.do-ai.run/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"anthropic-claude-haiku-4.5\",\"messages\":[{\"role\":\"system\",\"content\":[{\"type\":\"text\",\"text\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\",\"cache_control\":{\"type\":\"ephemeral\"}}]},{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"What is the weather in Paris?\",\"cache_control\":{\"type\":\"ephemeral\"}}]}],\"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},\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"store\":false,\"max_completion_tokens\":128}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"choices\":[{\"delta\":{\"content\":\"\",\"role\":\"assistant\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840197,\"id\":\"chatcmpl-ccddd771-e8a6-4b59-b819-614f79ac4624\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"\",\"name\":\"get_weather\"},\"id\":\"toolu_014LEa2F66CGw2SoLmWBeH5C\",\"index\":0,\"type\":\"function\"}]},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840197,\"id\":\"chatcmpl-ccddd771-e8a6-4b59-b819-614f79ac4624\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"\"},\"index\":0,\"type\":\"function\"}]},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840197,\"id\":\"chatcmpl-ccddd771-e8a6-4b59-b819-614f79ac4624\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"{\\\"city\\\"\"},\"index\":0,\"type\":\"function\"}]},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840197,\"id\":\"chatcmpl-ccddd771-e8a6-4b59-b819-614f79ac4624\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\": \\\"Pari\"},\"index\":0,\"type\":\"function\"}]},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840197,\"id\":\"chatcmpl-ccddd771-e8a6-4b59-b819-614f79ac4624\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"function\":{\"arguments\":\"s\\\"}\"},\"index\":0,\"type\":\"function\"}]},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840197,\"id\":\"chatcmpl-ccddd771-e8a6-4b59-b819-614f79ac4624\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{},\"finish_reason\":\"tool_calls\",\"index\":0,\"logprobs\":null}],\"created\":1790840197,\"id\":\"chatcmpl-ccddd771-e8a6-4b59-b819-614f79ac4624\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[],\"created\":1790840197,\"id\":\"chatcmpl-ccddd771-e8a6-4b59-b819-614f79ac4624\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\",\"usage\":{\"cache_created_input_tokens\":0,\"cache_creation\":{\"ephemeral_1h_input_tokens\":0,\"ephemeral_5m_input_tokens\":0},\"cache_read_input_tokens\":0,\"completion_tokens\":54,\"completion_tokens_details\":{\"reasoning_tokens\":0},\"prompt_tokens\":593,\"prompt_tokens_details\":{\"cached_tokens\":0},\"speed\":null,\"total_tokens\":647}}\n\ndata: [DONE]\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://inference.do-ai.run/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"anthropic-claude-haiku-4.5\",\"messages\":[{\"role\":\"system\",\"content\":[{\"type\":\"text\",\"text\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\",\"cache_control\":{\"type\":\"ephemeral\"}}]},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"},{\"role\":\"assistant\",\"content\":null,\"tool_calls\":[{\"id\":\"toolu_014LEa2F66CGw2SoLmWBeH5C\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}]},{\"role\":\"tool\",\"tool_call_id\":\"toolu_014LEa2F66CGw2SoLmWBeH5C\",\"content\":[{\"type\":\"text\",\"text\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\",\"cache_control\":{\"type\":\"ephemeral\"}}]}],\"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},\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"store\":false,\"max_completion_tokens\":128}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"choices\":[{\"delta\":{\"content\":\"Paris\",\"role\":\"assistant\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840198,\"id\":\"chatcmpl-3914ee28-f256-4a49-994f-49365922a7a6\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{\"content\":\" is sunny.\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840198,\"id\":\"chatcmpl-3914ee28-f256-4a49-994f-49365922a7a6\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\",\"index\":0,\"logprobs\":null}],\"created\":1790840198,\"id\":\"chatcmpl-3914ee28-f256-4a49-994f-49365922a7a6\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[],\"created\":1790840198,\"id\":\"chatcmpl-3914ee28-f256-4a49-994f-49365922a7a6\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\",\"usage\":{\"cache_created_input_tokens\":0,\"cache_creation\":{\"ephemeral_1h_input_tokens\":0,\"ephemeral_5m_input_tokens\":0},\"cache_read_input_tokens\":0,\"completion_tokens\":7,\"completion_tokens_details\":{\"reasoning_tokens\":0},\"prompt_tokens\":668,\"prompt_tokens_details\":{\"cached_tokens\":0},\"speed\":null,\"total_tokens\":675}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/digitalocean-chat/haiku-reuses-a-cached-prompt.json

@@ -0,0 +1,53 @@
+{
+  "version": 1,
+  "metadata": {
+    "model": "anthropic-claude-haiku-4.5",
+    "tags": [
+      "prefix:digitalocean-chat",
+      "provider:digitalocean",
+      "protocol:digitalocean-chat",
+      "cache",
+      "usage"
+    ],
+    "name": "digitalocean-chat/haiku-reuses-a-cached-prompt",
+    "recordedAt": "2026-10-01T07:36:36.689Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://inference.do-ai.run/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"anthropic-claude-haiku-4.5\",\"messages\":[{\"role\":\"system\",\"content\":[{\"type\":\"text\",\"text\":\"You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and aLine truncated
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"choices\":[{\"delta\":{\"content\":\"OK\",\"role\":\"assistant\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840195,\"id\":\"chatcmpl-94f208e5-82fa-4ec6-b014-1b4ea9fbb98c\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\",\"index\":0,\"logprobs\":null}],\"created\":1790840195,\"id\":\"chatcmpl-94f208e5-82fa-4ec6-b014-1b4ea9fbb98c\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[],\"created\":1790840195,\"id\":\"chatcmpl-94f208e5-82fa-4ec6-b014-1b4ea9fbb98c\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\",\"usage\":{\"cache_created_input_tokens\":5759,\"cache_creation\":{\"ephemeral_1h_input_tokens\":0,\"ephemeral_5m_input_tokens\":5759},\"cache_read_input_tokens\":0,\"completion_tokens\":4,\"completion_tokens_details\":{\"reasoning_tokens\":0},\"prompt_tokens\":5762,\"prompt_tokens_details\":{\"cached_tokens\":0},\"speed\":null,\"total_tokens\":5766}}\n\ndata: [DONE]\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://inference.do-ai.run/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"anthropic-claude-haiku-4.5\",\"messages\":[{\"role\":\"system\",\"content\":[{\"type\":\"text\",\"text\":\"You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and avoid filler. Cite numbers when known. You are a concise, factual assistant. Answer precisely and aLine truncated
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"choices\":[{\"delta\":{\"content\":\"OK\",\"role\":\"assistant\"},\"finish_reason\":null,\"index\":0,\"logprobs\":null}],\"created\":1790840196,\"id\":\"chatcmpl-eee2d578-6fab-4fc1-8117-043193225bed\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\",\"index\":0,\"logprobs\":null}],\"created\":1790840196,\"id\":\"chatcmpl-eee2d578-6fab-4fc1-8117-043193225bed\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\"}\n\ndata: {\"choices\":[],\"created\":1790840196,\"id\":\"chatcmpl-eee2d578-6fab-4fc1-8117-043193225bed\",\"model\":\"anthropic-claude-haiku-4.5\",\"object\":\"chat.completion.chunk\",\"usage\":{\"cache_created_input_tokens\":0,\"cache_creation\":{\"ephemeral_1h_input_tokens\":0,\"ephemeral_5m_input_tokens\":0},\"cache_read_input_tokens\":5759,\"completion_tokens\":4,\"completion_tokens_details\":{\"reasoning_tokens\":0},\"prompt_tokens\":5762,\"prompt_tokens_details\":{\"cached_tokens\":5759},\"speed\":null,\"total_tokens\":5766}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

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

@@ -15,6 +15,7 @@ describe("provider error classification", () => {
       "Prompt has 5,958,968 tokens, but the configured context size is 256,000 tokens",
       "Range of input length should be [1, 129024]",
       "Too many tokens",
+      "Input validation error: `inputs` tokens + `max_new_tokens` must be <= 131073. Given: 600035 `inputs` tokens and 16 `max_new_tokens`",
       "Token limit exceeded",
     ]
 

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

@@ -314,6 +314,9 @@ describe("Anthropic Messages route", () => {
         "claude-haiku-5-1",
         "claude-fable-6",
         "anthropic/claude-mythos-7.2",
+        "claude-sonnet-5-5",
+        "claude-opus-4-8@20260101",
+        "claude-nova-6",
       ]
 
       const prepared = yield* Effect.forEach(ids, (id) =>

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

@@ -648,6 +648,177 @@ describe("Bedrock Converse route", () => {
     }),
   )
 
+  it.effect("hoists tool-result images beside the result for Bedrock GPT models", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: AmazonBedrock.configure({ baseURL: "https://bedrock-runtime.test", apiKey: "test-bearer" }).model(
+            "global.openai.gpt-6-sol",
+          ),
+          messages: [
+            Message.user("What is in this image?"),
+            Message.assistant([ToolCallPart.make({ id: "tool_1", name: "read", input: {} })]),
+            Message.tool({
+              id: "tool_1",
+              name: "read",
+              result: {
+                type: "content",
+                value: [
+                  { type: "text", text: "Image loaded." },
+                  { type: "file", uri: "data:image/png;base64,AAAA", mime: "image/png" },
+                  { type: "file", uri: "data:application/pdf;base64,QkI=", mime: "application/pdf", name: "note.pdf" },
+                ],
+              },
+            }),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages[2]).toEqual({
+        role: "user",
+        content: [
+          {
+            toolResult: {
+              toolUseId: "tool_1",
+              content: [
+                { text: "Image loaded." },
+                { text: 'Attached file "note.pdf" has document label "note".' },
+                { document: { format: "pdf", name: "note", source: { bytes: "QkI=" } } },
+              ],
+              status: "success",
+            },
+          },
+          { image: { format: "png", source: { bytes: "AAAA" } } },
+        ],
+      })
+    }),
+  )
+
+  it.effect("hoists tool-result images for other Bedrock model families", () =>
+    Effect.gen(function* () {
+      for (const id of [
+        "qwen.qwen3-vl-235b-a22b",
+        "global.xai.grok-4.7",
+        "global.moonshotai.kimi-k3",
+        "us.meta.llama4-scout-17b-instruct-v1:0",
+      ]) {
+        const prepared = yield* compileRequest(
+          LLM.request({
+            model: AmazonBedrock.configure({ baseURL: "https://bedrock-runtime.test", apiKey: "test-bearer" }).model(
+              id,
+            ),
+            messages: [
+              Message.assistant([ToolCallPart.make({ id: "tool_1", name: "read", input: {} })]),
+              Message.tool({
+                id: "tool_1",
+                name: "read",
+                result: {
+                  type: "content",
+                  value: [{ type: "file", uri: "data:image/png;base64,AAAA", mime: "image/png" }],
+                },
+              }),
+            ],
+            cache: "none",
+          }),
+        )
+        expect(prepared.body.messages[1]).toEqual({
+          role: "user",
+          content: [
+            { toolResult: { toolUseId: "tool_1", content: [{ text: "See attached image." }], status: "success" } },
+            { image: { format: "png", source: { bytes: "AAAA" } } },
+          ],
+        })
+      }
+    }),
+  )
+  ;["global.anthropic.claude-sonnet-4-5-20250929-v1:0", "us.amazon.nova-pro-v1:0"].forEach((id) => {
+    it.effect(`keeps ${id} tool images inside the result`, () =>
+      Effect.gen(function* () {
+        const prepared = yield* compileRequest(
+          LLM.request({
+            model: AmazonBedrock.configure({ baseURL: "https://bedrock-runtime.test", apiKey: "test-bearer" }).model(
+              id,
+            ),
+            messages: [
+              Message.assistant([ToolCallPart.make({ id: "tool_1", name: "read", input: {} })]),
+              Message.tool({
+                id: "tool_1",
+                name: "read",
+                result: {
+                  type: "content",
+                  value: [{ type: "file", uri: "data:image/png;base64,AAAA", mime: "image/png" }],
+                },
+              }),
+            ],
+            cache: "none",
+          }),
+        )
+
+        expect(prepared.body.messages[1]).toEqual({
+          role: "user",
+          content: [
+            {
+              toolResult: {
+                toolUseId: "tool_1",
+                content: [{ image: { format: "png", source: { bytes: "AAAA" } } }],
+                status: "success",
+              },
+            },
+          ],
+        })
+      }),
+    )
+  })
+
+  it.effect("keeps parallel tool results before hoisted images and gives image-only results text", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: AmazonBedrock.configure({ baseURL: "https://bedrock-runtime.test", apiKey: "test-bearer" }).model(
+            "global.openai.gpt-6-sol",
+          ),
+          messages: [
+            Message.assistant([
+              ToolCallPart.make({ id: "tool_1", name: "first", input: {} }),
+              ToolCallPart.make({ id: "tool_2", name: "second", input: {} }),
+            ]),
+            Message.tool({
+              id: "tool_1",
+              name: "first",
+              result: {
+                type: "content",
+                value: [{ type: "file", uri: "data:image/png;base64,AAAA", mime: "image/png" }],
+              },
+            }),
+            Message.tool({
+              id: "tool_2",
+              name: "second",
+              result: {
+                type: "content",
+                value: [
+                  { type: "text", text: "Second image." },
+                  { type: "file", uri: "data:image/jpeg;base64,BBBB", mime: "image/jpeg" },
+                ],
+              },
+            }),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages[1]).toEqual({
+        role: "user",
+        content: [
+          { toolResult: { toolUseId: "tool_1", content: [{ text: "See attached image." }], status: "success" } },
+          { toolResult: { toolUseId: "tool_2", content: [{ text: "Second image." }], status: "success" } },
+          { image: { format: "png", source: { bytes: "AAAA" } } },
+          { image: { format: "jpeg", source: { bytes: "BBBB" } } },
+        ],
+      })
+    }),
+  )
+
   it.effect("decodes text-delta + messageStop + metadata usage from binary event stream", () =>
     Effect.gen(function* () {
       const body = eventStreamBody(

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

@@ -0,0 +1,78 @@
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, LLMRequest } from "../../src/index.js"
+import { DigitalOcean } from "../../src/providers/digitalocean.js"
+import { LLMClient } from "../../src/route.js"
+import {
+  LARGE_CACHEABLE_SYSTEM,
+  expectWeatherToolLoop,
+  goldenWeatherToolLoopRequest,
+  runWeatherToolLoop,
+} from "../recorded-scenarios.js"
+import { recordedTests } from "../recorded-test.js"
+
+const recorded = recordedTests({
+  prefix: "digitalocean-chat",
+  provider: "digitalocean",
+  protocol: "digitalocean-chat",
+  requires: ["DIGITAL_OCEAN_OFFICIAL_API_KEY"],
+})
+
+for (const item of [
+  { id: "anthropic-claude-haiku-4.5", name: "Haiku", maxTokens: 128, providerOptions: undefined },
+  { id: "openai-gpt-5-nano", name: "GPT Nano", maxTokens: 1024, providerOptions: { reasoningEffort: "minimal" } },
+] as const) {
+  const model = DigitalOcean.configure({
+    apiKey: process.env.DIGITAL_OCEAN_OFFICIAL_API_KEY ?? "fixture",
+    providerOptions: item.providerOptions,
+  }).model(item.id)
+
+  describe(`DigitalOcean ${item.name} recorded`, () => {
+    recorded.effect.with(
+      `${item.name} reuses a cached prompt`,
+      { tags: ["cache", "usage"], metadata: { model: item.id } },
+      () =>
+        Effect.gen(function* () {
+          const request = LLM.request({
+            model,
+            system: LARGE_CACHEABLE_SYSTEM,
+            prompt: "Reply exactly: OK",
+            promptCacheKey: `digitalocean-recorded-${item.id}`,
+            generation: { maxTokens: item.maxTokens },
+          })
+          const first = yield* LLMClient.generate(request)
+          const second = yield* LLMClient.generate(request)
+
+          expect(first.text.trim()).toMatch(/^OK\.?$/)
+          expect(second.text.trim()).toMatch(/^OK\.?$/)
+          expect(second.usage.cacheReadInputTokens).toBeGreaterThan(0)
+          for (const response of [first, second]) {
+            expect(response.usage.inputTokens).toBeGreaterThan(4096)
+            expect(response.usage.inputTokens).toBe(
+              (response.usage.nonCachedInputTokens ?? 0) +
+                (response.usage.cacheReadInputTokens ?? 0) +
+                (response.usage.cacheWriteInputTokens ?? 0),
+            )
+          }
+        }),
+      60_000,
+    )
+
+    recorded.effect.with(
+      `${item.name} continues a tool call with cache markers`,
+      { tags: ["cache", "tool", "tool-loop"], metadata: { model: item.id } },
+      () =>
+        Effect.gen(function* () {
+          const request = goldenWeatherToolLoopRequest({
+            id: `digitalocean-${item.id}-tool-loop`,
+            model,
+            maxTokens: item.maxTokens,
+            temperature: false,
+          })
+          const events = yield* runWeatherToolLoop(LLMRequest.update(request, { cache: "auto" }))
+          expectWeatherToolLoop(events)
+        }),
+      60_000,
+    )
+  })
+}

packages/ai/test/provider/digitalocean.test.ts

@@ -0,0 +1,123 @@
+import { describe, expect, test } from "bun:test"
+import { Effect } from "effect"
+import { CacheHint, LLM } from "../../src/index.js"
+import { compileRequest } from "../../src/route/client.js"
+import { DigitalOcean } from "../../src/providers/digitalocean.js"
+import { it } from "../lib/effect.js"
+import { fixedResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
+import { LLMClient } from "../../src/route.js"
+
+describe("DigitalOcean", () => {
+  test("preserves package entrypoint headers and body overrides", () => {
+    const model = DigitalOcean.model("anthropic-claude-fable-5.1", {
+      apiKey: "test-key",
+      headers: { "X-Test": "fixture" },
+      body: { temperature: 0 },
+    })
+    expect(model.route.defaults?.http).toMatchObject({
+      headers: { "X-Test": "fixture" },
+      body: { temperature: 0 },
+    })
+  })
+
+  it.effect("prepares DigitalOcean models with default endpoint and auth", () =>
+    Effect.gen(function* () {
+      const model = DigitalOcean.configure({ apiKey: "test-key" }).model("anthropic-claude-fable-5.1")
+
+      expect(model).toMatchObject({
+        id: "anthropic-claude-fable-5.1",
+        provider: "digitalocean",
+        route: { id: "digitalocean" },
+      })
+      expect(model.route.endpoint.baseURL).toBe("https://inference.do-ai.run/v1")
+
+      const prepared = yield* compileRequest(LLM.request({ model, prompt: "Say hello.", cache: "none" }))
+
+      expect(prepared.route).toBe("digitalocean")
+      expect(prepared.body).toMatchObject({
+        model: "anthropic-claude-fable-5.1",
+        messages: [{ role: "user", content: "Say hello." }],
+        stream: true,
+      })
+    }),
+  )
+
+  it.effect("lowers the native cache policy to DigitalOcean cache_control markers", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: DigitalOcean.configure({ apiKey: "test-key" }).model("anthropic-claude-fable-5.1"),
+          system: [
+            { type: "text", text: "Base agent", cache: new CacheHint({ type: "ephemeral", ttlSeconds: 3_600 }) },
+            { type: "text", text: "Project instructions" },
+          ],
+          tools: [{ name: "lookup", description: "Lookup", inputSchema: { type: "object", properties: {} } }],
+          prompt: "Hello",
+          cache: { tools: true, system: true, messages: { tail: 1 } },
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({
+        tools: [{ cache_control: { type: "ephemeral" } }],
+        messages: [
+          {
+            role: "system",
+            content: [
+              { text: "Base agent", cache_control: { type: "ephemeral", ttl: "1h" } },
+              { text: "Project instructions", cache_control: { type: "ephemeral" } },
+            ],
+          },
+          {
+            role: "user",
+            content: [{ text: "Hello", cache_control: { type: "ephemeral" } }],
+          },
+        ],
+      })
+    }),
+  )
+
+  it.effect("parses DigitalOcean cache usage fields into AI.Usage", () =>
+    Effect.gen(function* () {
+      const model = DigitalOcean.configure({ apiKey: "test-key" }).model("anthropic-claude-fable-5.1")
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Say OK" })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                id: "chatcmpl-1",
+                object: "chat.completion.chunk",
+                created: 1,
+                model: "anthropic-claude-fable-5.1",
+                choices: [{ index: 0, delta: { content: "OK" }, finish_reason: null }],
+              },
+              {
+                id: "chatcmpl-1",
+                object: "chat.completion.chunk",
+                created: 1,
+                model: "anthropic-claude-fable-5.1",
+                choices: [{ index: 0, delta: {}, finish_reason: "stop" }],
+                usage: {
+                  prompt_tokens: 4491,
+                  completion_tokens: 8,
+                  total_tokens: 4499,
+                  cache_read_input_tokens: 4483,
+                  cache_created_input_tokens: 0,
+                },
+              },
+              "[DONE]",
+            ),
+          ),
+        ),
+      )
+
+      expect(response.usage).toMatchObject({
+        inputTokens: 4491,
+        outputTokens: 8,
+        nonCachedInputTokens: 8,
+        cacheReadInputTokens: 4483,
+        totalTokens: 4499,
+      })
+    }),
+  )
+})

packages/ai/test/provider/google-vertex.test.ts

@@ -302,6 +302,7 @@ describe("Google Vertex providers", () => {
 
       expect(model.provider).toBe("google-vertex")
       expect(response.text).toBe("Hello.")
+      expect(response.usage?.providerMetadata).toHaveProperty("vertex")
     }),
   )
 

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

@@ -171,6 +171,7 @@ describe("Groq recorded", () => {
         expect(response.text).not.toContain("<think>")
         expect(response.reasoning.length).toBeGreaterThan(0)
         expect(response.events.some(LLMEvent.is.reasoningDelta)).toBe(true)
+        expect(response.usage?.providerMetadata).toHaveProperty("groq")
         expectUsage(response)
       }),
     60_000,

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

@@ -48,7 +48,7 @@ describe("native OpenAI-compatible providers", () => {
       [GoogleVertex.configure(vertex).model("model"), "vertex"],
       [GoogleVertexChat.configure(vertex).model("model"), "vertex"],
       [GoogleVertexResponses.configure(vertex).model("model"), "vertex"],
-      [GoogleVertexMessages.configure(vertex).model("model"), "anthropic"],
+      [GoogleVertexMessages.configure(vertex).model("model"), "vertex"],
       [Anthropic.configure({ apiKey: "test" }).model("model"), "anthropic"],
       [
         AnthropicCompatible.configure({ baseURL: "https://example.test/v1", provider: "minimax" }).model("model"),
@@ -64,6 +64,7 @@ describe("native OpenAI-compatible providers", () => {
       [DeepSeek.configure({ apiKey: "test" }).model("model"), "deepseek"],
       [Fireworks.configure({ apiKey: "test" }).model("model"), "fireworks"],
       [DeepInfra.configure({ apiKey: "test" }).model("model"), "deepinfra"],
+      [Groq.configure({ apiKey: "test" }).model("model"), "groq"],
       [TogetherAI.configure({ apiKey: "test" }).model("model"), "togetherai"],
       [CloudflareAIGateway.configure({ accountId: "account" }).model("model"), "cloudflare-ai-gateway"],
       [CloudflareWorkersAI.configure({ accountId: "account" }).model("model"), "cloudflare-workers-ai"],

packages/app/e2e/regression/review.spec.ts

@@ -532,6 +532,29 @@ test("restores review state and the side-panel tab per session", async ({ page }
   await expect(review).toHaveAttribute("aria-selected", "true")
 })
 
+test("shows and restores last turn changes from the session diff", async ({ page }) => {
+  const sessionID = "ses_reviewturn"
+  await openSession(page, { name: "ReviewTurn", vcsDiff: [fileDiff("src/alpha.ts")] })
+  await page.route(`**/api/session/${sessionID}/diff**`, (route) =>
+    route.fulfill({
+      status: 200,
+      contentType: "application/json",
+      body: JSON.stringify({ data: [fileDiff("src/delta.ts")] }),
+    }),
+  )
+  const panel = page.locator("#review-panel")
+  await page.getByRole("button", { name: "Toggle review" }).click()
+  await page.getByRole("button", { name: "Git changes" }).click()
+  await page.getByRole("option", { name: "Last turn changes" }).click()
+  await expect(page.getByRole("button", { name: "Last turn changes" })).toBeVisible()
+  await expect(panel.locator('[data-slot="session-review-v2-file-name"]')).toHaveText("delta.ts")
+
+  await page.reload()
+  await expectSessionTitle(page, "ReviewTurn")
+  await expect(page.getByRole("button", { name: "Last turn changes" })).toBeVisible()
+  await expect(panel.locator('[data-slot="session-review-v2-file-name"]')).toHaveText("delta.ts")
+})
+
 test("keeps the review state a session stored before extensions", async ({ page }) => {
   const directory = "C:/OpenCode/ReviewLegacy"
   await openSession(page, {

packages/cli/package.json

@@ -24,7 +24,7 @@
     "typecheck": "tsgo -b"
   },
   "dependencies": {
-    "@agentclientprotocol/sdk": "1.2.1",
+    "@agentclientprotocol/sdk": "1.6.0",
     "@clack/core": "1.0.0-alpha.1",
     "@clack/prompts": "1.0.0-alpha.1",
     "@effect/platform-node": "catalog:",

packages/cli/src/acp/content.ts

@@ -171,7 +171,11 @@ function resourceLinkToPart(link: ResourceLink): PromptPart {
         mime: link.mimeType ?? "text/plain",
       }
   }
-  return { type: "text", text: link.uri }
+  return linkReference(link.name, link.uri)
+}
+
+export function linkReference(name: string | undefined, uri: string): PromptPart {
+  return { type: "text", text: name ? `[${name}](${uri})` : uri }
 }
 
 function filenameFromUri(uri: string | undefined): string | undefined {

packages/cli/src/acp/directories.ts

@@ -0,0 +1,83 @@
+import { isAbsolute, join, resolve } from "node:path"
+import { isDeepStrictEqual } from "node:util"
+import type { OpenCodeClient, PermissionRule, SessionInfo, SessionMetadata } from "@opencode/client/promise"
+import { FSUtil } from "@opencode/util/fs-util"
+import { Effect, Option, Schema } from "effect"
+import { ACPError } from "./error"
+import { ACPPromise } from "./promise"
+
+const key = "opencode.acp.additionalDirectories"
+const decodeStored = Schema.decodeUnknownOption(Schema.Array(Schema.String))
+
+/**
+ * Normalizes additional workspace roots without following symlinks, keeping the spelling tools will see, and
+ * drops duplicates and roots that resolve to cwd. Glob characters are rejected because permission resources
+ * would treat them as wildcards. The ACP SDK has already dropped entries that are not strings and replaced a
+ * value that is not an array with `[]`.
+ */
+export const parse = Effect.fnUntraced(function* (cwd: string, directories: readonly string[] = []) {
+  const invalid = directories.find((directory) => !isAbsolute(directory) || /[*?]/.test(directory))
+  if (invalid !== undefined) return yield* new ACPError.InvalidAdditionalDirectoryError({ directory: invalid })
+  const root = FSUtil.resolve(cwd)
+  return [...new Set(directories.map((directory) => resolve(FSUtil.windowsPath(directory))))].filter(
+    (directory) => FSUtil.resolve(directory) !== root,
+  )
+})
+
+/** Session create fields that grant these directories. */
+export function grant(directories: readonly string[]) {
+  if (directories.length === 0) return {}
+  return { permissions: rules(directories), metadata: { [key]: [...directories] } }
+}
+
+/** The additional directories ACP last activated for the session, in request order. */
+export function list(session: Pick<SessionInfo, "metadata">) {
+  return [...Option.getOrElse(decodeStored(session.metadata?.[key]), () => [])]
+}
+
+/**
+ * Replaces the rules granted for the previously activated directories with grants for these directories.
+ * Grants persist on the server session, so they also apply when it is used from other clients, and child
+ * sessions copy them when they are created; a child created earlier keeps roots its parent later dropped.
+ * Session rules are evaluated after agent and config rules, so a grant overrides config `external_directory`
+ * rules inside the root, while read, edit, and shell rules still apply.
+ */
+export const activate = Effect.fnUntraced(function* (
+  client: OpenCodeClient,
+  session: SessionInfo,
+  directories: readonly string[],
+) {
+  const previous = list(session)
+  const owned = rules(previous)
+  const current = session.permissions ?? []
+  // ACP rules lead so the session's other rules keep precedence for the same paths.
+  const permissions = [
+    ...rules(directories),
+    ...current.filter((rule) => !owned.some((item) => isDeepStrictEqual(item, rule))),
+  ]
+  const metadata: SessionMetadata = {
+    ...Object.fromEntries(Object.entries(session.metadata ?? {}).filter(([name]) => name !== key)),
+    ...(directories.length > 0 ? { [key]: [...directories] } : {}),
+  }
+  const permissionsChanged = !isDeepStrictEqual(permissions, current)
+  const metadataChanged = !isDeepStrictEqual(previous, directories)
+  if (!permissionsChanged && !metadataChanged) return
+  yield* ACPPromise.promise(() =>
+    client.session.update({
+      sessionID: session.id,
+      ...(permissionsChanged ? { permissions } : {}),
+      ...(metadataChanged ? { metadata } : {}),
+    }),
+  )
+})
+
+// Tools compare the written path without following symlinks, so both spellings of a root are granted.
+function rules(directories: readonly string[]): PermissionRule[] {
+  return [...new Set(directories.flatMap((directory) => [directory, FSUtil.resolve(directory)]))].map((directory) => ({
+    action: "external_directory",
+    resource: join(directory, "*"),
+    effect: "allow",
+  }))
+}
+
+export * as ACPDirectories from "./directories"

packages/cli/src/acp/error.ts

@@ -28,6 +28,11 @@ export class InvalidModeError extends Schema.TaggedError<InvalidModeError>()("AC
   mode: Schema.String,
 }) {}
 
+export class InvalidAdditionalDirectoryError extends Schema.TaggedError<InvalidAdditionalDirectoryError>()(
+  "ACPInvalidAdditionalDirectoryError",
+  { directory: Schema.String },
+) {}
+
 export class AuthRequiredError extends Schema.TaggedError<AuthRequiredError>()("ACPAuthRequiredError", {}) {}
 
 export class UnknownAuthMethodError extends Schema.TaggedError<UnknownAuthMethodError>()("ACPUnknownAuthMethodError", {
@@ -57,6 +62,7 @@ const Errors = Schema.Union([
   InvalidModelError,
   InvalidEffortError,
   InvalidModeError,
+  InvalidAdditionalDirectoryError,
   AuthRequiredError,
   UnknownAuthMethodError,
   InvalidRequestError,
@@ -88,6 +94,11 @@ export function toRequestError(error: Error): RequestError {
       return RequestError.invalidParams({ effort: error.effort }, `effort not found: ${error.effort}`)
     case "ACPInvalidModeError":
       return RequestError.invalidParams({ mode: error.mode }, `mode not found: ${error.mode}`)
+    case "ACPInvalidAdditionalDirectoryError":
+      return RequestError.invalidParams(
+        { additionalDirectory: error.directory },
+        `additional directory must be an absolute path without glob characters: ${error.directory}`,
+      )
     case "ACPAuthRequiredError":
       return RequestError.authRequired({}, "provider authentication required")
     case "ACPUnknownAuthMethodError":

packages/cli/src/acp/service.ts

@@ -41,6 +41,7 @@ import { OPENCODE_VERSION } from "../version"
 import type { ACPCatalog, Catalog } from "./catalog"
 import { configOptions, currentModel, DEFAULT_VARIANT_VALUE, parseModelSelection } from "./config-option"
 import type { ACPConnection } from "./connection"
+import { ACPDirectories } from "./directories"
 import { ACPError } from "./error"
 import { ACPPromise } from "./promise"
 import type { ACPSessions, Attached } from "./sessions"
@@ -171,7 +172,7 @@ export function make(input: {
           loadSession: true,
           mcpCapabilities: { http: true, sse: false },
           promptCapabilities: { embeddedContext: true, image: true },
-          sessionCapabilities: { close: {}, delete: {}, fork: {}, list: {}, resume: {} },
+          sessionCapabilities: { additionalDirectories: {}, close: {}, delete: {}, fork: {}, list: {}, resume: {} },
           _meta: { [ACPTranslate.ChildSessionUpdatesCapability]: true },
         },
         authMethods: [authMethod],
@@ -184,17 +185,23 @@ export function make(input: {
       return {}
     }),
     newSession: Effect.fnUntraced(function* (params) {
+      const directories = yield* ACPDirectories.parse(params.cwd, params.additionalDirectories)
       // Load before creating so a catalog failure leaves no session behind. Agent and model stay unset
       // so the server resolves its defaults after plugins activate.
       yield* input.catalog.get(params.cwd)
       const created = yield* ACPPromise.promise(() =>
-        input.client.session.create({ location: { directory: params.cwd } }),
+        input.client.session.create({
+          location: { directory: params.cwd },
+          ...ACPDirectories.grant(directories),
+        }),
       )
       const attached = yield* input.sessions.attach(created, params.cwd, params.mcpServers)
       return { sessionId: attached.id, configOptions: yield* currentOptions(attached) }
     }),
     loadSession: Effect.fnUntraced(function* (params) {
+      const directories = yield* ACPDirectories.parse(params.cwd, params.additionalDirectories)
       const session = yield* getSession(params.sessionId, params.cwd)
+      yield* ACPDirectories.activate(input.client, session, directories)
       const attached = yield* input.sessions.attach(session, session.location.directory, params.mcpServers)
       return yield* replay(attached).pipe(
         Effect.andThen(currentOptions(attached)),
@@ -212,12 +219,16 @@ export function make(input: {
         }),
       )
       return {
-        sessions: page.data.map((session) => ({
-          sessionId: session.id,
-          cwd: session.location.directory,
-          title: withTimestampedFallback(session),
-          updatedAt: new Date(session.time.updated).toISOString(),
-        })),
+        sessions: page.data.map((session) => {
+          const additionalDirectories = ACPDirectories.list(session)
+          return {
+            sessionId: session.id,
+            cwd: session.location.directory,
+            ...(additionalDirectories.length > 0 ? { additionalDirectories } : {}),
+            title: withTimestampedFallback(session),
+            updatedAt: new Date(session.time.updated).toISOString(),
+          }
+        }),
         ...(page.cursor.next ? { nextCursor: page.cursor.next } : {}),
       }
     }),
@@ -233,7 +244,9 @@ export function make(input: {
       return {}
     }),
     resumeSession: Effect.fnUntraced(function* (params) {
+      const directories = yield* ACPDirectories.parse(params.cwd, params.additionalDirectories)
       const session = yield* getSession(params.sessionId, params.cwd)
+      yield* ACPDirectories.activate(input.client, session, directories)
       const attached = yield* input.sessions.attach(session, session.location.directory, params.mcpServers ?? [])
       return { configOptions: yield* currentOptions(attached) }
     }),
@@ -243,7 +256,10 @@ export function make(input: {
       return {}
     }),
     forkSession: Effect.fnUntraced(function* (params) {
+      const directories = yield* ACPDirectories.parse(params.cwd, params.additionalDirectories)
       const forked = yield* ACPPromise.promise(() => input.client.session.fork({ sessionID: params.sessionId }))
+      // Forks copy the source session's rules, so the request list replaces any inherited grants.
+      yield* ACPDirectories.activate(input.client, forked, directories)
       const attached = yield* input.sessions.attach(forked, forked.location.directory, params.mcpServers ?? [])
       return yield* currentOptions(attached).pipe(
         Effect.map((configOptions) => ({ sessionId: attached.id, configOptions })),

packages/cli/src/acp/turn.ts

@@ -22,10 +22,12 @@ import {
   Scope,
   Stream,
 } from "effect"
+import { access, constants } from "node:fs/promises"
+import { fileURLToPath } from "node:url"
 import { builtinCommands, type ACPCatalog, type Catalog } from "./catalog"
 import { currentModel } from "./config-option"
 import type { ACPConnection } from "./connection"
-import { promptContentToParts } from "./content"
+import { linkReference, promptContentToParts, type PromptPart } from "./content"
 import { ACPElicitation } from "./elicitation"
 import { ACPError } from "./error"
 import { ACPPermission } from "./permission"
@@ -372,7 +374,10 @@ export const make = Effect.fnUntraced(function* (input: {
       const attached = yield* input.sessions.require(params.sessionId)
       const catalog = yield* input.catalog.get(attached.cwd)
       const childUpdates = (yield* Ref.get(input.capabilities)).childSessionUpdates
-      const prompt = preparePrompt(catalog, params.prompt, SessionMessage.ID.create())
+      const parts = yield* Effect.forEach(promptContentToParts(params.prompt), referenceUnreadableFile, {
+        concurrency: "unbounded",
+      })
+      const prompt = preparePrompt(catalog, parts, SessionMessage.ID.create())
       // Check and register in one synchronous step.
       const turn = yield* Effect.withFiber((fiber) => {
         if (FiberMap.hasUnsafe(turns, attached.id)) {
@@ -414,8 +419,7 @@ function aborted(signal: AbortSignal) {
   })
 }
 
-function preparePrompt(catalog: Catalog, prompt: PromptRequest["prompt"], messageID: string): PreparedPrompt {
-  const parts = promptContentToParts(prompt)
+function preparePrompt(catalog: Catalog, parts: readonly PromptPart[], messageID: string): PreparedPrompt {
   const visible = parts.filter((part) => part.type !== "text" || (!part.synthetic && !part.ignored))
   const synthetic = parts.flatMap((part) => (part.type === "text" && part.synthetic ? [part.text] : []))
   const text = visible.flatMap((part) => (part.type === "text" ? [part.text] : [])).join("\n")
@@ -426,6 +430,15 @@ function preparePrompt(catalog: Catalog, prompt: PromptRequest["prompt"], messag
   return { start, text, files, synthetic, slash, command }
 }
 
+// Covers only missing or permission-denied targets; the server still rejects oversized, non-regular, or unlistable ones.
+function referenceUnreadableFile(part: PromptPart) {
+  if (part.type !== "file" || !part.url.startsWith("file://")) return Effect.succeed(part)
+  return Effect.tryPromise(() => access(fileURLToPath(part.url), constants.R_OK)).pipe(
+    Effect.as(part),
+    Effect.orElseSucceed(() => linkReference(part.filename, part.url)),
+  )
+}
+
 function turnStart(messageID: string, slash: PreparedPrompt["slash"]): ACPTranslate.TurnStart {
   if (slash && builtinCommands.get(slash.name)?.start === "compaction") return { type: "compaction", id: messageID }
   return { type: "input", id: messageID }

packages/cli/src/commands/handlers/acp.ts

@@ -1,4 +1,4 @@
-import { ndJsonStream } from "@agentclientprotocol/sdk"
+import { MessageTooLargeError, ndJsonStream } from "@agentclientprotocol/sdk"
 import { OpenCode } from "@opencode/client/promise"
 import { Service } from "@opencode/client/effect/service"
 import { CrossSpawnSpawner } from "@opencode/util/cross-spawn-spawner"
@@ -16,25 +16,29 @@ export default Runtime.handler(
     const endpoint = yield* Standalone.start()
     const client = OpenCode.make({ baseUrl: endpoint.url, headers: Service.headers(endpoint) })
     const connection = yield* ACP.connect(client, ndJsonStream(Writable.toWeb(process.stdout), Bun.stdin.stream()))
-    const code = yield* Effect.raceFirst(
-      Effect.promise(() => connection.closed).pipe(Effect.as(0)),
+    const failure = yield* Effect.raceFirst(
+      Effect.promise(() => connection.closed).pipe(
+        Effect.map(() =>
+          connection.signal.reason instanceof MessageTooLargeError
+            ? `incoming message exceeded the ${connection.signal.reason.maxMessageBytes / 1024 / 1024} MiB limit`
+            : undefined,
+        ),
+      ),
       endpoint.exited.pipe(
         Effect.match({
           onSuccess: (code) => `code ${code}`,
           onFailure: (error) =>
             error.cause instanceof CrossSpawnSpawner.KilledBySignal ? `signal ${error.cause.signal}` : error.message,
         }),
-        // stdout carries ACP, so the diagnostic goes to stderr.
-        Effect.flatMap((reason) =>
-          Effect.sync(() => {
-            process.stderr.write(`opencode acp: server exited unexpectedly (${reason})\n`)
-            return 1
-          }),
-        ),
+        Effect.map((reason) => `server exited unexpectedly (${reason})`),
       ),
     )
     // Closing the handler scope would wait for the private server's graceful shutdown; its lease pipe already
     // ends the server once this process exits.
-    yield* Effect.sync(() => process.exit(code))
+    yield* Effect.sync(() => {
+      // stdout carries ACP, so the diagnostic goes to stderr.
+      if (failure) process.stderr.write(`opencode acp: ${failure}\n`)
+      process.exit(failure ? 1 : 0)
+    })
   }),
 )

packages/cli/src/commands/handlers/default.ts

@@ -4,7 +4,7 @@ import { run } from "@opencode/tui"
 import { Commands } from "../commands"
 import { Runtime } from "../../framework/runtime"
 import { Config } from "../../config"
-import { Context, Effect, Fiber, FileSystem, Option, Queue } from "effect"
+import { Context, Effect, FileSystem, Option, Queue, Schedule, Semaphore } from "effect"
 import { ServerConnection } from "../../services/server-connection"
 import { Updater } from "../../services/updater"
 import { UpdatePreflight } from "../../services/update-preflight"
@@ -61,13 +61,29 @@ export default Runtime.handler(Commands, (input) =>
       })) !== undefined
     const updater = yield* Updater.Service
     let installing: string | undefined
-    const updateListeners = new Set<(version: string) => void>()
-    const update = yield* updater
+    let latest: Updater.RunResult | undefined
+    const installListeners = new Set<(version: string) => void>()
+    const resultListeners = new Set<(result: Updater.RunResult) => void>()
+    // Background checks, `/update` lookups, and manual installs take turns so two installs never overlap.
+    const checking = yield* Semaphore.make(1)
+    yield* updater
       .run((version) => {
         installing = version
-        updateListeners.forEach((notify) => notify(version))
+        installListeners.forEach((notify) => notify(version))
       })
-      .pipe(Effect.ensuring(Effect.sync(() => (installing = undefined))), Effect.forkScoped)
+      .pipe(
+        Effect.ensuring(Effect.sync(() => (installing = undefined))),
+        Effect.tap((result) =>
+          Effect.sync(() => {
+            if (!result || (result.type === latest?.type && result.version === latest.version)) return
+            latest = result
+            resultListeners.forEach((notify) => notify(result))
+          }),
+        ),
+        checking.withPermits(1),
+        Effect.repeat(Schedule.spaced("10 minutes")),
+        Effect.forkScoped({ startImmediately: true }),
+      )
     preflight.loading()
     const config = yield* Config.Service
     const npm = yield* Npm.Service
@@ -106,21 +122,19 @@ export default Runtime.handler(Commands, (input) =>
       },
       updater: {
         remote: requestedServer !== undefined,
-        subscribe: (notify, signal) =>
-          runPromise(
-            Fiber.join(update).pipe(
-              Effect.flatMap((result) => (result === undefined ? Effect.void : Effect.sync(() => notify(result)))),
-            ),
-            { signal },
-          ),
+        subscribe: (notify) => {
+          if (latest) notify(latest)
+          resultListeners.add(notify)
+          return () => resultListeners.delete(notify)
+        },
         check: (signal, notify) => {
           if (installing) notify(installing)
-          updateListeners.add(notify)
-          return runPromise(Fiber.join(update).pipe(Effect.flatMap(() => updater.check())), { signal }).finally(() =>
-            updateListeners.delete(notify),
+          installListeners.add(notify)
+          return runPromise(checking.withPermits(1)(updater.check()), { signal }).finally(() =>
+            installListeners.delete(notify),
           )
         },
-        apply: (version) => runPromise(updater.apply(version)),
+        apply: (version) => runPromise(checking.withPermits(1)(updater.apply(version))),
       },
       packages: {
         prepare: (spec, install = true) => runPromise(install ? npm.add(spec) : npm.resolve(spec)),

packages/cli/src/index.ts

@@ -13,6 +13,7 @@ import { Global } from "@opencode/util/global"
 import { AppProcess } from "@opencode/util/process"
 import { Config } from "./config"
 import { Npm } from "@opencode/util/npm"
+import { EffectFlock } from "@opencode/util/effect-flock"
 import { Heap } from "./heap"
 import { CpuProfile } from "./cpu-profile"
 
@@ -113,7 +114,7 @@ Effect.gen(function* () {
   Effect.provide(Config.layer),
   Effect.provide(Updater.layer),
   Effect.provide(
-    LayerNode.compile(LayerNode.group([Global.node, AppProcess.node, Npm.node]), {
+    LayerNode.compile(LayerNode.group([Global.node, AppProcess.node, Npm.node, EffectFlock.node]), {
       replacements: [
         Global.node.replace(
           Global.layerWith(process.env.OPENCODE_CONFIG_DIR ? { config: process.env.OPENCODE_CONFIG_DIR } : {}),

packages/cli/src/services/updater.ts

@@ -1,5 +1,6 @@
 import { Global } from "@opencode/util/global"
 import { AppProcess } from "@opencode/util/process"
+import { EffectFlock } from "@opencode/util/effect-flock"
 import { OPENCODE_ARTIFACT, OPENCODE_CHANNEL, OPENCODE_LOCAL, OPENCODE_VERSION } from "../version"
 import { Context, Duration, Effect, FileSystem, Layer, Option, Ref, Schema } from "effect"
 import { ChildProcess } from "effect/unstable/process"
@@ -127,6 +128,7 @@ const make = Effect.gen(function* () {
   const fs = yield* FileSystem.FileSystem
   const global = yield* Global.Service
   const appProcess = yield* AppProcess.Service
+  const flock = yield* EffectFlock.Service
   const installedVersion = yield* Ref.make(OPENCODE_VERSION)
   const channel = OPENCODE_CHANNEL.replace(/[^a-zA-Z0-9._-]/g, "-")
   const installedPackage = yield* Effect.gen(function* () {
@@ -370,6 +372,9 @@ const make = Effect.gen(function* () {
     }
     yield* Effect.scoped(
       Effect.gen(function* () {
+        // Other OpenCode processes may be installing at the same time. Wait longer than the
+        // slowest install (curl runs two 5-minute commands).
+        yield* flock.acquire("cli-upgrade", undefined, { timeoutMs: Duration.toMillis("15 minutes") })
         if (method === "bun") {
           // Bun does not prune old versions from its shared package cache.
           yield* fs.makeDirectory(global.cache, { recursive: true })

packages/cli/test/acp/additional-directories.subprocess.test.ts

@@ -0,0 +1,58 @@
+import type { NewSessionResponse, PromptResponse } from "@agentclientprotocol/sdk"
+import { describe, expect, test } from "bun:test"
+import fs from "node:fs/promises"
+import path from "node:path"
+import { createAcpFixture, expectOk, initialize } from "./subprocess"
+
+// The first completion reads the file outside cwd; the follow-up completion ends the turn.
+function readingModel(file: () => string) {
+  return (request: unknown) => {
+    if (JSON.stringify(request).includes('"role":"tool"')) return "done"
+    return new Response(toolCall(file()), { headers: { "content-type": "text/event-stream" } })
+  }
+}
+
+function toolCall(file: string) {
+  const call = { index: 0, id: "call_read", type: "function", function: { name: "read", arguments: "" } }
+  const chunks = [
+    { choices: [{ delta: { role: "assistant", tool_calls: [call] }, finish_reason: null }], usage: null },
+    {
+      choices: [{ delta: { tool_calls: [{ index: 0, function: { arguments: JSON.stringify({ path: file }) } }] } }],
+      usage: null,
+    },
+    { choices: [{ delta: {}, finish_reason: "tool_calls" }], usage: null },
+    { choices: [], usage: { prompt_tokens: 10, completion_tokens: 1, total_tokens: 11 } },
+  ]
+  return `${chunks.map((chunk) => `data: ${JSON.stringify(chunk)}\n\n`).join("")}data: [DONE]\n\n`
+}
+
+describe("acp additional directories subprocess", () => {
+  test("tools read files in an additional directory without an external directory ask", async () => {
+    const target = { file: "" }
+    await using fixture = await createAcpFixture({ respond: readingModel(() => target.file) })
+    // Keep the unresolved tmpdir spelling, which differs from the real path on macOS.
+    const shared = path.join(fixture.root, "shared")
+    target.file = path.join(shared, "notes.txt")
+    await fs.mkdir(shared)
+    await Bun.write(target.file, "shared root content\n")
+    const acp = fixture.spawn()
+    await initialize(acp)
+    const session = expectOk(
+      await acp.request<NewSessionResponse>("session/new", {
+        cwd: fixture.home,
+        additionalDirectories: [shared],
+        mcpServers: [],
+      }),
+    )
+
+    const result = expectOk(
+      await acp.request<PromptResponse>("session/prompt", {
+        sessionId: session.sessionId,
+        prompt: [{ type: "text", text: "read the shared notes" }],
+      }),
+    )
+
+    expect(result.stopReason).toBe("end_turn")
+    expect(JSON.stringify(fixture.llm.requests.at(-1))).toContain("shared root content")
+  }, 60_000)
+})

packages/cli/test/acp/additional-directories.test.ts

@@ -0,0 +1,182 @@
+import { describe, expect, test } from "bun:test"
+import fs from "node:fs/promises"
+import path from "node:path"
+import type { PermissionRule } from "@opencode/client/promise"
+import { tmpdir } from "../fixture/tmpdir"
+import { makeSession, rpcError, startWire, type Wire } from "./wire-fixture"
+
+const key = "opencode.acp.additionalDirectories"
+
+const grant = (directory: string): PermissionRule => ({
+  action: "external_directory",
+  resource: path.join(directory, "*"),
+  effect: "allow",
+})
+
+const sharedLib = path.resolve("/shared/lib")
+const productDocs = path.resolve("/product-docs")
+const old = path.resolve("/old")
+
+const userGrant: PermissionRule = { action: "external_directory", resource: "/x/**", effect: "allow" }
+
+const other: PermissionRule[] = [
+  { action: "read", resource: "*.secret", effect: "deny" },
+  { action: "external_directory", resource: "/shared/lib/private/*", effect: "deny" },
+]
+
+const updates = (acp: Wire) =>
+  acp.server.requests.filter((request) => request.method === "PATCH" && request.path.startsWith("/api/session/"))
+
+describe("acp additional directories over the wire", () => {
+  test("session/new grants normalized unique directories other than cwd and lists them", async () => {
+    await using acp = await startWire()
+    await acp.initialize()
+
+    const created = await acp.request("session/new", {
+      cwd: "/workspace",
+      additionalDirectories: ["/shared/lib/", "/workspace", "/docs/../product-docs", "/shared/lib", "/workspace/"],
+      mcpServers: [],
+    })
+
+    expect(acp.server.sessions.get(created.sessionId)).toMatchObject({
+      permissions: [grant(sharedLib), grant(productDocs)],
+      metadata: { [key]: [sharedLib, productDocs] },
+    })
+    expect((await acp.request("session/list", { cwd: "/workspace" })).sessions).toEqual([
+      expect.objectContaining({
+        sessionId: created.sessionId,
+        additionalDirectories: [sharedLib, productDocs],
+      }),
+    ])
+  })
+
+  test("grants both the written and real spelling of a symlinked root and drops links to cwd", async () => {
+    await using tmp = await tmpdir()
+    const root = await fs.realpath(tmp.path)
+    const cwd = path.join(root, "workspace")
+    const shared = path.join(root, "shared")
+    await Promise.all([fs.mkdir(cwd), fs.mkdir(shared)])
+    await Promise.all([
+      fs.symlink(shared, path.join(root, "shared-link")),
+      fs.symlink(cwd, path.join(root, "workspace-link")),
+    ])
+    await using acp = await startWire()
+    await acp.initialize()
+
+    const created = await acp.request("session/new", {
+      cwd,
+      additionalDirectories: [path.join(root, "workspace-link"), path.join(root, "shared-link")],
+      mcpServers: [],
+    })
+
+    expect(acp.server.sessions.get(created.sessionId)).toMatchObject({
+      permissions: [grant(path.join(root, "shared-link")), grant(shared)],
+      metadata: { [key]: [path.join(root, "shared-link")] },
+    })
+    expect((await acp.request("session/list", { cwd })).sessions[0]?.additionalDirectories).toEqual([
+      path.join(root, "shared-link"),
+    ])
+  })
+
+  test.each(["shared/lib", "", "/shared/*", "/shared/lib?"])(
+    "rejects %p before creating a session",
+    async (directory) => {
+      await using acp = await startWire()
+      await acp.initialize()
+
+      expect(
+        await rpcError(
+          acp.request("session/new", {
+            cwd: "/workspace",
+            additionalDirectories: ["/shared/ok", directory],
+            mcpServers: [],
+          }),
+        ),
+      ).toMatchObject({ code: -32602, data: { additionalDirectory: directory } })
+      expect(acp.server.sessions.size).toBe(0)
+    },
+  )
+
+  test("load and resume replace ACP grants and keep other session rules and metadata", async () => {
+    await using acp = await startWire()
+    acp.server.sessions.set("ses_saved", {
+      ...makeSession("ses_saved"),
+      metadata: { host: "tui", [key]: [old] },
+      permissions: [grant(old), userGrant, ...other],
+    })
+    await acp.initialize()
+
+    await acp.request("session/load", {
+      cwd: "/workspace",
+      sessionId: "ses_saved",
+      additionalDirectories: ["/shared/lib", "/product-docs"],
+      mcpServers: [],
+    })
+    expect(acp.server.sessions.get("ses_saved")).toMatchObject({
+      metadata: { host: "tui", [key]: [sharedLib, productDocs] },
+      permissions: [grant(sharedLib), grant(productDocs), userGrant, ...other],
+    })
+
+    await acp.request("session/resume", {
+      cwd: "/workspace",
+      sessionId: "ses_saved",
+      additionalDirectories: ["/shared/lib", "/product-docs"],
+    })
+    expect(updates(acp)).toHaveLength(1)
+
+    await acp.request("session/resume", { cwd: "/workspace", sessionId: "ses_saved" })
+    expect(acp.server.sessions.get("ses_saved")?.metadata).toEqual({ host: "tui" })
+    expect(acp.server.sessions.get("ses_saved")?.permissions).toEqual([userGrant, ...other])
+    expect((await acp.request("session/list", { cwd: "/workspace" })).sessions[0]).not.toHaveProperty(
+      "additionalDirectories",
+    )
+  })
+
+  test("forks replace inherited grants with the requested list", async () => {
+    await using acp = await startWire()
+    acp.server.sessions.set("ses_source", {
+      ...makeSession("ses_source"),
+      metadata: { [key]: [old] },
+      permissions: [grant(old), ...other],
+    })
+    await acp.initialize()
+
+    const plain = await acp.request("session/fork", { cwd: "/workspace", sessionId: "ses_source" })
+    const granted = await acp.request("session/fork", {
+      cwd: "/workspace",
+      sessionId: "ses_source",
+      additionalDirectories: ["/shared/lib"],
+    })
+
+    expect(acp.server.sessions.get(plain.sessionId)).toMatchObject({ metadata: {}, permissions: other })
+    expect(acp.server.sessions.get(granted.sessionId)).toMatchObject({
+      metadata: { [key]: [sharedLib] },
+      permissions: [grant(sharedLib), ...other],
+    })
+    expect(acp.server.sessions.get("ses_source")?.permissions).toEqual([grant(old), ...other])
+  })
+
+  test("leaves sessions alone without additional directories", async () => {
+    await using acp = await startWire()
+    acp.server.sessions.set("ses_saved", {
+      ...makeSession("ses_saved"),
+      metadata: { host: "tui" },
+      permissions: [userGrant, ...other],
+    })
+    await acp.initialize()
+
+    const created = await acp.request("session/new", { cwd: "/workspace", mcpServers: [] })
+    await acp.request("session/load", { cwd: "/workspace", sessionId: "ses_saved", mcpServers: [] })
+    await acp.request("session/resume", { cwd: "/workspace", sessionId: "ses_saved", additionalDirectories: [] })
+
+    const create = acp.server.requests.find((request) => request.method === "POST" && request.path === "/api/session")
+    expect(create?.body).toEqual({ location: { directory: "/workspace" } })
+    expect(acp.server.sessions.get(created.sessionId)?.permissions).toBeUndefined()
+    expect(updates(acp)).toEqual([])
+    expect(
+      (await acp.request("session/list", { cwd: "/workspace" })).sessions.map(
+        (session) => session.additionalDirectories,
+      ),
+    ).toEqual([undefined, undefined])
+  })
+})

packages/cli/test/acp/content.test.ts

@@ -80,6 +80,12 @@ describe("acp content conversion", () => {
     ])
   })
 
+  test("resource_link to another scheme becomes a markdown link", () => {
+    expect(contentBlockToParts({ type: "resource_link", uri: "https://example.com/spec", name: "spec" })).toEqual([
+      { type: "text", text: "[spec](https://example.com/spec)" },
+    ])
+  })
+
   test("resource_link zed path becomes a file URL part", () => {
     expect(
       contentBlockToParts({

packages/cli/test/acp/lifecycle.subprocess.test.ts

@@ -17,6 +17,20 @@ describe("acp lifecycle subprocess", () => {
     expect(await acp.close()).toBe(0)
   }, 60_000)
 
+  test("an incoming message over the size limit exits with an error", async () => {
+    await using fixture = await createAcpFixture()
+    const acp = fixture.spawn()
+    await initialize(acp)
+    const [code] = await Promise.all([
+      acp.exited,
+      // The agent stops reading partway through the line, so the write may fail.
+      acp.notify("opencode/oversized", { data: "a".repeat(32 * 1024 * 1024) }).catch(() => undefined),
+    ])
+    await acp[Symbol.asyncDispose]()
+    expect(code).toBe(1)
+    expect(acp.stderr()).toContain("opencode acp: incoming message exceeded the 32 MiB limit\n")
+  }, 60_000)
+
   test("close capability and close request", async () => {
     await using fixture = await createAcpFixture()
     const acp = fixture.spawn()

packages/cli/test/acp/prompt-content.subprocess.test.ts

@@ -56,7 +56,22 @@ describe("acp prompt content subprocess", () => {
       }),
     )
 
+    const missing = expectOk(
+      await acp.request<PromptResponse>("session/prompt", {
+        sessionId: session.sessionId,
+        prompt: [
+          { type: "text", text: "Use this missing file." },
+          {
+            type: "resource_link",
+            uri: pathToFileURL(path.join(fixture.home, "missing.md")).href,
+            name: "missing.md",
+          },
+        ],
+      }),
+    )
+
     expect(linked.stopReason).toBe("end_turn")
-    expect(fixture.llm.requests.length).toBeGreaterThanOrEqual(3)
+    expect(missing.stopReason).toBe("end_turn")
+    expect(fixture.llm.requests.length).toBeGreaterThanOrEqual(4)
   }, 60_000)
 })

packages/cli/test/acp/prompt.test.ts

@@ -2,6 +2,10 @@ import { describe, expect, test } from "bun:test"
 import type { StopReason } from "@agentclientprotocol/sdk"
 import type { OpenCodeEvent } from "@opencode/client/promise"
 import { Schema } from "effect"
+import { mkdir } from "node:fs/promises"
+import path from "node:path"
+import { pathToFileURL } from "node:url"
+import { tmpdir } from "../fixture/tmpdir"
 import {
   childCreated,
   delivered,
@@ -80,12 +84,15 @@ describe("acp prompt turns over the wire", () => {
   })
 
   test("submits assistant-only context as synthetic input before the visible prompt", async () => {
+    await using dir = await tmpdir()
+    const readme = pathToFileURL(path.join(dir.path, "README.md")).href
+    await Bun.write(path.join(dir.path, "README.md"), "# readme\n")
     await using acp = await startSession()
 
     await acp.prompt(acp.sessionId, [
       { type: "text", text: "visible" },
       { type: "text", text: "hidden context", annotations: { audience: ["assistant"] } },
-      { type: "resource_link", uri: "file:///workspace/README.md", name: "README.md", mimeType: "text/markdown" },
+      { type: "resource_link", uri: readme, name: "README.md", mimeType: "text/markdown" },
     ])
 
     expect(acp.server.submissions).toEqual([
@@ -100,7 +107,37 @@ describe("acp prompt turns over the wire", () => {
       expect.objectContaining({
         kind: "prompt",
         text: "visible",
-        files: [{ uri: "file:///workspace/README.md", name: "README.md" }],
+        files: [{ uri: readme, name: "README.md" }],
+      }),
+    ])
+  })
+
+  test("attaches readable file links and references unreadable ones in place", async () => {
+    await using dir = await tmpdir()
+    const file = pathToFileURL(path.join(dir.path, "notes.md")).href
+    const folder = pathToFileURL(path.join(dir.path, "src")).href
+    const missing = pathToFileURL(path.join(dir.path, "missing.md")).href
+    await Bun.write(path.join(dir.path, "notes.md"), "# notes\n")
+    await mkdir(path.join(dir.path, "src"))
+    await using acp = await startSession()
+
+    const response = await acp.prompt(acp.sessionId, [
+      { type: "text", text: "compare" },
+      { type: "resource_link", uri: file, name: "notes.md" },
+      { type: "resource_link", uri: missing, name: "missing.md" },
+      { type: "resource_link", uri: folder, name: "src" },
+      { type: "text", text: "please" },
+    ])
+
+    expect(response.stopReason).toBe("end_turn")
+    expect(acp.server.submissions).toEqual([
+      expect.objectContaining({
+        kind: "prompt",
+        text: `compare\n[missing.md](${missing})\nplease`,
+        files: [
+          { uri: file, name: "notes.md" },
+          { uri: folder, name: "src" },
+        ],
       }),
     ])
   })

packages/cli/test/acp/session.test.ts

@@ -26,7 +26,7 @@ describe("acp session lifecycle over the wire", () => {
         loadSession: true,
         mcpCapabilities: { http: true, sse: false },
         promptCapabilities: { embeddedContext: true, image: true },
-        sessionCapabilities: { close: {}, delete: {}, fork: {}, list: {}, resume: {} },
+        sessionCapabilities: { additionalDirectories: {}, close: {}, delete: {}, fork: {}, list: {}, resume: {} },
         _meta: { "opencode/child-session-updates": true },
       },
       agentInfo: { name: "OpenCode" },

packages/cli/test/acp/wire-fixture.ts

@@ -61,7 +61,16 @@ const SyntheticBody = Schema.Struct({
   delivery: Delivery,
   resume: Schema.optional(Schema.Boolean),
 })
-const CreateBody = Schema.Struct({ location: Schema.Struct({ directory: Schema.String }) })
+const Permissions = Schema.Array(
+  Schema.Struct({ action: Schema.String, resource: Schema.String, effect: Schema.Literals(["allow", "deny", "ask"]) }),
+)
+const Metadata = Schema.Record(Schema.String, Schema.MutableJson)
+const CreateBody = Schema.Struct({
+  location: Schema.Struct({ directory: Schema.String }),
+  permissions: Schema.optional(Permissions),
+  metadata: Schema.optional(Metadata),
+})
+const UpdateBody = Schema.Struct({ permissions: Schema.optional(Permissions), metadata: Schema.optional(Metadata) })
 const ModelBody = Schema.Struct({
   model: Schema.Struct({ providerID: Schema.String, id: Schema.String, variant: Schema.optional(Schema.String) }),
 })
@@ -712,7 +721,13 @@ function startServer(options: WireOptions, changed: () => void) {
           return Response.json(page(sessions, query, 100))
         }),
         POST: body(CreateBody, (_req, input) =>
-          Response.json({ data: createSession(makeSession("", { cwd: input.location.directory })) }),
+          Response.json({
+            data: createSession({
+              ...makeSession("", { cwd: input.location.directory }),
+              ...(input.permissions ? { permissions: [...input.permissions] } : {}),
+              ...(input.metadata ? { metadata: input.metadata } : {}),
+            }),
+          }),
         ),
       },
       "/api/session/:sessionID": {
@@ -727,6 +742,13 @@ function startServer(options: WireOptions, changed: () => void) {
           if (invalid) return invalid
           return fake.sessions.delete(req.params.sessionID) ? noContent() : notFound(req.params.sessionID)
         }),
+        PATCH: body(UpdateBody, (req, input) => {
+          const session = fake.sessions.get(req.params.sessionID)
+          if (!session) return notFound(req.params.sessionID)
+          if (input.permissions) session.permissions = [...input.permissions]
+          if (input.metadata) session.metadata = input.metadata
+          return noContent()
+        }),
       },
       "/api/session/:sessionID/fork": {
         POST: route((req) => {

packages/cli/test/updater-install.test.ts

@@ -1,8 +1,10 @@
 import { NodeServices } from "@effect/platform-node"
+import { EffectFlock } from "@opencode/util/effect-flock"
+import { LayerNode } from "@opencode/util/effect/layer-node"
 import { Global } from "@opencode/util/global"
 import { AppProcess } from "@opencode/util/process"
 import { expect, spyOn, test } from "bun:test"
-import { Effect, FileSystem, PlatformError, Stream } from "effect"
+import { Effect, FileSystem, Layer, PlatformError, Stream } from "effect"
 import { ChildProcess, ChildProcessSpawner } from "effect/unstable/process"
 import { existsSync, readdirSync, readFileSync } from "node:fs"
 import path from "node:path"
@@ -59,6 +61,11 @@ function fixture(
     const commands: string[][] = []
     const updater = yield* Updater.Service.pipe(
       Effect.provide(Updater.layer),
+      Effect.provide(
+        LayerNode.compile(EffectFlock.node, {
+          replacements: [Global.node.replace(Layer.succeed(Global.Service, global))],
+        }),
+      ),
       Effect.provideService(Global.Service, global),
       Effect.provideService(FileSystem.FileSystem, {
         ...fs,

packages/core/src/aisdk-native.ts

@@ -93,6 +93,7 @@ const HOSTS: Readonly<Record<string, Readonly<Record<string, string>>>> = {
   },
   "cloudflare-workers-ai": { "@ai-sdk/openai-compatible": "@opencode/ai/providers/cloudflare-workers-ai" },
   deepseek: { "@ai-sdk/openai-compatible": "@opencode/ai/providers/deepseek" },
+  digitalocean: { "@ai-sdk/openai-compatible": "@opencode/ai/providers/digitalocean" },
   "fireworks-ai": { "@ai-sdk/openai-compatible": "@opencode/ai/providers/fireworks" },
   "google-vertex": { "@ai-sdk/openai-compatible": "@opencode/ai/providers/google-vertex/chat" },
   "kimi-for-coding": protocols("moonshot"),

packages/core/src/config/plugin/instruction.ts

@@ -60,14 +60,17 @@ export const Plugin = define({
       })
 
       const globalSource = Effect.fn("ConfigInstructionPlugin.globalSource")(function* () {
-        if (!discovery.global) return []
+        if (!discovery.global || !(yield* fs.isFile(globalFile))) return []
         const file = yield* read(globalFile)
         return file ? [file] : []
       })
 
       const projectSource = Effect.fn("ConfigInstructionPlugin.projectSource")(function* () {
         if (!project) return []
-        const walked = yield* Effect.forEach(yield* fs.up({ targets: ["AGENTS.md"], start, stop }), fs.resolve)
+        const walked = yield* Effect.forEach(
+          yield* fs.up({ targets: ["AGENTS.md"], start, stop, type: "file" }),
+          fs.resolve,
+        )
         const discovered = new Set(walked.filter((file) => discovery.global || file !== globalFile))
         const files = yield* Effect.forEach(discovered, read, { concurrency: "unbounded" })
         if (files.some((file) => file === undefined)) return Instructions.unavailable

packages/core/src/models-dev.ts

@@ -24,7 +24,6 @@ type Cost = {
   readonly cache_read?: Money.USDPerMillionTokens
   readonly cache_write?: Money.USDPerMillionTokens
   readonly tiers?: readonly (Cost & { readonly tier: { readonly type: "context"; readonly size: number } })[]
-  readonly context_over_200k?: Omit<Cost, "tiers" | "context_over_200k">
 }
 
 type Modality = "text" | "audio" | "image" | "video" | "pdf"
@@ -147,19 +146,6 @@ function cost(input: SourceModel["cost"]): Model.Info["cost"] {
         write: item.cache_write ?? Money.USDPerMillionTokens.zero,
       },
     })) ?? []),
-    ...(input?.context_over_200k
-      ? [
-          {
-            tier: { type: "context" as const, size: 200_000 },
-            input: input.context_over_200k.input,
-            output: input.context_over_200k.output,
-            cache: {
-              read: input.context_over_200k.cache_read ?? Money.USDPerMillionTokens.zero,
-              write: input.context_over_200k.cache_write ?? Money.USDPerMillionTokens.zero,
-            },
-          },
-        ]
-      : []),
   ]
 }
 

packages/core/src/models-dev/snapshot.txt

@@ -1 +1 @@
-{"deepinfra":{"id":"deepinfra","env":["DEEPINFRA_API_KEY"],"npm":"@ai-sdk/deepinfra","name":"Deep Infra","doc":"https://deepinfra.com/models","models":{"tencent/Hy3":{"id":"tencent/Hy3","name":"Hy3","description":"Tencent Hy reasoning model for coding, instruction following, and agent tasks","family":"Hy","attachment":false,"reasoning":true,"reasoning_options":[],"tool_call":true,"structured_output":true,"temperature":true,"release_date":"2026-07-06","last_updated":"2026-07-06","modalities":{"input":["text"],"output":["text"]},"open_weights":true,"limit":{"context":262144,"input":192000,"output":128000},"cost":{"input":0.13,"output":0.53,"cache_read":0.033},"canonical_model_id":"tencent/hy3"},"tencent/Hy4-preview":{"id":"tencent/Hy4-preview","name":"Hy4 preview","description":"A next-generation productivity model with significantly enhanced Agent and complex task execution capabilities.","family":"Hy","attachment":false,"reasoning":true,"reasoning_options":[{"type":"effort","values":["none","high"]}],"tool_call":true,"structured_output":true,"temperature":true,"release_date":"2026-08-28","last_updated":"2026-08-28","modalities":{"input":["text"],"output":["text"]},"open_weights":true,"limit":{"context":1048576,"output":64000},"cost":{"input":0.834,"output":2.501,"cache_read":0.042},"canonical_model_id":"tencent/hy4-preview"},"meta-llama/Llama-3.3-70B-Instruct-Turbo":{"id":"meta-llama/Llama-3.3-70B-Instruct-Turbo","name":"Llama 3.3 70B Turbo","description":"Compact Llama instruction model for fast chat and local deployment","family":"llama","attachment":false,"reasoning":false,"tool_call":true,"structured_output":true,"release_date":"2024-12-06","last_updated":"2024-12-06","modalities":{"input":["text"],"output":["text"]},"open_weights":true,"limit":{"context":131072,"output":16384},"cost":{"input":0.1,"output":0.32}},"meta-llama/Llama-4-Scout-17B-16E-Instruct":{"id":"meta-llama/Llama-4-Scout-17B-16E-Instruct","name":"Llama 4 Scout 17B","description":"Open multimodal Llama model for long-context analysis and efficient agents","family":"llama","attachment":true,"reasoning":false,"tool_call":true,"structured_output":true,"release_date":"2025-04-05","last_updated":"2025-04-05","modalities":{"input":["text","image"],"output":["text"]},"open_weights":true,"limit":{"context":327680,"output":16384},"cost":{"input":0.1,"output":0.3}},"meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8":{"id":"meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8","name":"Llama 4 Maverick 17B FP8","description":"Open multimodal Llama model for strong reasoning and fast responses","family":"llama","attachment":true,"reasoning":false,"tool_call":false,"structured_output":true,"release_date":"2025-04-05","last_updated":"2025-04-05","modalities":{"input":["text","image"],"output":["text"]},"open_weights":true,"limit":{"context":1048576,"output":16384},"cost":{"input":0.2,"output":0.8}},"XiaomiMiMo/MiMo-V2.6-Pro":{"id":"XiaomiMiMo/MiMo-V2.6-Pro","name":"MiMo-V2.6-Pro","description":"Stronger MiMo Pro tier for multimodal reasoning and coding-agent execution","family":"mimo","attachment":true,"reasoning":true,"reasoning_options":[{"type":"toggle"}],"tool_call":true,"structured_output":true,"temperature":true,"release_date":"2026-09-22","last_updated":"2026-09-22","modalities":{"input":["text","image","audio","video"],"output":["text"]},"open_weights":true,"limit":{"context":1048576,"output":131072},"cost":{"input":0.43,"output":0.87,"cache_read":0.0036},"canonical_model_id":"xiaomi/mimo-v2.6-pro"},"XiaomiMiMo/MiMo-V2.5-Pro":{"id":"XiaomiMiMo/MiMo-V2.5-Pro","name":"MiMo-V2.5-Pro","description":"Stronger MiMo Pro tier for multimodal reasoning and coding-agent execution","family":"mimo","attachment":true,"reasoning":true,"reasoning_options":[{"type":"toggle"}],"tool_call":true,"interleaved":{"field":"reasoning_content"},"structured_output":true,"temperature":true,"knowledge":"2024-12","release_date":"2026-04-22","last_updated":"2026-04-22","modalities":{"input":["text","audio"],"output":["text"]},"open_weights":true,"limit":{"context":1048576,"output":16384},"status":"deprecated","cost":{"input":1,"output":3,"cache_read":0.2},"canonical_model_id":"xiaomi/mimo-v2.5-pro"},"XiaomiMiMo/MiMo-V2.6-Flash":{"id":"XiaomiMiMo/MiMo-V2.6-Flash","name":"MiMo-V2.6-Flash","description":"MiMo Flash model for multimodal coding agents and long-context automation","family":"mimo","attachment":true,"reasoning":true,"reasoning_options":[{"type":"toggle"}],"tool_call":true,"structured_output":true,"temperature":true,"release_date":"2026-09-22","last_updated":"2026-09-22","modalities":{"input":["text","image","audio","video"],"output":["text"]},"open_weights":true,"limit":{"context":1048576,"output":131072},"cost":{"input":0.14,"output":0.28,"cache_read":0.0028},"canonical_model_id":"xiaomi/mimo-v2.6-flash"},"XiaomiMiMo/MiMo-V2.5":{"id":"XiaomiMiMo/MiMo-V2.5","name":"MiMo-V2.5","description":"Open MiMo model for multimodal coding agents and long-context automation","famLine truncated
+{"deepinfra":{"id":"deepinfra","env":["DEEPINFRA_API_KEY"],"npm":"@ai-sdk/deepinfra","name":"Deep Infra","doc":"https://deepinfra.com/models","models":{"tencent/Hy3":{"id":"tencent/Hy3","name":"Hy3","description":"Tencent Hy reasoning model for coding, instruction following, and agent tasks","family":"Hy","attachment":false,"reasoning":true,"reasoning_options":[],"tool_call":true,"structured_output":true,"temperature":true,"release_date":"2026-07-06","last_updated":"2026-07-06","modalities":{"input":["text"],"output":["text"]},"open_weights":true,"limit":{"context":262144,"input":192000,"output":128000},"cost":{"input":0.13,"output":0.53,"cache_read":0.033},"canonical_model_id":"tencent/hy3"},"tencent/Hy4-preview":{"id":"tencent/Hy4-preview","name":"Hy4 preview","description":"A next-generation productivity model with significantly enhanced Agent and complex task execution capabilities.","family":"Hy","attachment":false,"reasoning":true,"reasoning_options":[{"type":"effort","values":["none","high"]}],"tool_call":true,"structured_output":true,"temperature":true,"release_date":"2026-08-28","last_updated":"2026-08-28","modalities":{"input":["text"],"output":["text"]},"open_weights":true,"limit":{"context":1048576,"output":64000},"cost":{"input":0.834,"output":2.501,"cache_read":0.042},"canonical_model_id":"tencent/hy4-preview"},"meta-llama/Llama-3.3-70B-Instruct-Turbo":{"id":"meta-llama/Llama-3.3-70B-Instruct-Turbo","name":"Llama 3.3 70B Turbo","description":"Compact Llama instruction model for fast chat and local deployment","family":"llama","attachment":false,"reasoning":false,"tool_call":true,"structured_output":true,"release_date":"2024-12-06","last_updated":"2024-12-06","modalities":{"input":["text"],"output":["text"]},"open_weights":true,"limit":{"context":131072,"output":16384},"cost":{"input":0.1,"output":0.32}},"meta-llama/Llama-4-Scout-17B-16E-Instruct":{"id":"meta-llama/Llama-4-Scout-17B-16E-Instruct","name":"Llama 4 Scout 17B","description":"Open multimodal Llama model for long-context analysis and efficient agents","family":"llama","attachment":true,"reasoning":false,"tool_call":true,"structured_output":true,"release_date":"2025-04-05","last_updated":"2025-04-05","modalities":{"input":["text","image"],"output":["text"]},"open_weights":true,"limit":{"context":327680,"output":16384},"cost":{"input":0.1,"output":0.3}},"meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8":{"id":"meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8","name":"Llama 4 Maverick 17B FP8","description":"Open multimodal Llama model for strong reasoning and fast responses","family":"llama","attachment":true,"reasoning":false,"tool_call":false,"structured_output":true,"release_date":"2025-04-05","last_updated":"2025-04-05","modalities":{"input":["text","image"],"output":["text"]},"open_weights":true,"limit":{"context":1048576,"output":16384},"cost":{"input":0.2,"output":0.8}},"XiaomiMiMo/MiMo-V2.6-Pro":{"id":"XiaomiMiMo/MiMo-V2.6-Pro","name":"MiMo-V2.6-Pro","description":"Stronger MiMo Pro tier for multimodal reasoning and coding-agent execution","family":"mimo","attachment":true,"reasoning":true,"reasoning_options":[{"type":"toggle"}],"tool_call":true,"structured_output":true,"temperature":true,"release_date":"2026-09-22","last_updated":"2026-09-22","modalities":{"input":["text","image","audio","video"],"output":["text"]},"open_weights":true,"limit":{"context":1048576,"output":131072},"cost":{"input":0.43,"output":0.87,"cache_read":0.0036},"canonical_model_id":"xiaomi/mimo-v2.6-pro"},"XiaomiMiMo/MiMo-V2.5-Pro":{"id":"XiaomiMiMo/MiMo-V2.5-Pro","name":"MiMo-V2.5-Pro","description":"Stronger MiMo Pro tier for multimodal reasoning and coding-agent execution","family":"mimo","attachment":true,"reasoning":true,"reasoning_options":[{"type":"toggle"}],"tool_call":true,"interleaved":{"field":"reasoning_content"},"structured_output":true,"temperature":true,"knowledge":"2024-12","release_date":"2026-04-22","last_updated":"2026-04-22","modalities":{"input":["text","audio"],"output":["text"]},"open_weights":true,"limit":{"context":1048576,"output":16384},"status":"deprecated","cost":{"input":1,"output":3,"cache_read":0.2},"canonical_model_id":"xiaomi/mimo-v2.5-pro"},"XiaomiMiMo/MiMo-V2.6-Flash":{"id":"XiaomiMiMo/MiMo-V2.6-Flash","name":"MiMo-V2.6-Flash","description":"MiMo Flash model for multimodal coding agents and long-context automation","family":"mimo","attachment":true,"reasoning":true,"reasoning_options":[{"type":"toggle"}],"tool_call":true,"structured_output":true,"temperature":true,"release_date":"2026-09-22","last_updated":"2026-09-22","modalities":{"input":["text","image","audio","video"],"output":["text"]},"open_weights":true,"limit":{"context":1048576,"output":131072},"cost":{"input":0.14,"output":0.28,"cache_read":0.0028},"canonical_model_id":"xiaomi/mimo-v2.6-flash"},"XiaomiMiMo/MiMo-V2.5":{"id":"XiaomiMiMo/MiMo-V2.5","name":"MiMo-V2.5","description":"Open MiMo model for multimodal coding agents and long-context automation","famLine truncated

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

@@ -1,4 +1,5 @@
-import { Clock, Effect, Schema, Semaphore, Stream } from "effect"
+import { Clock, Effect, FiberHandle, Option, Schema, Semaphore, Stream } from "effect"
+import { HttpClient, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
 import { ChildProcess } from "effect/unstable/process"
 import { define } from "@opencode/plugin/effect/plugin"
 import { Form } from "@opencode/schema/form"
@@ -7,14 +8,22 @@ import { App } from "../../app.js"
 import { Bus } from "../../bus.js"
 import { Credential } from "../../credential.js"
 import { Integration } from "../../integration.js"
+import { IntegrationConnection } from "../../integration/connection.js"
+import { Model } from "../../model.js"
 import { Provider } from "../../provider.js"
 import { iife } from "../../util/iife.js"
 import { which } from "../../util/which.js"
+import type { PluginInternal } from "../internal.js"
 import { configuredSettings } from "./configured.js"
 
 const cognitiveScope = "https://cognitiveservices.azure.com/.default"
 const foundryScope = "https://ai.azure.com/.default"
+const managementScope = "https://management.azure.com/.default"
 const methodID = Integration.MethodID.make("azure-cli")
+// A resource name becomes a hostname label and a query literal, so anything else never leaves the process. Azure
+// allows only letters, digits, and hyphens in it.
+// https://learn.microsoft.com/azure/ai-services/cognitive-services-custom-subdomains
+const resourcePattern = /^[a-zA-Z0-9][a-zA-Z0-9-]*$/
 const decodeJSON = Schema.decodeUnknownEffect(Schema.fromJsonString(Schema.Unknown))
 const decodeToken = Schema.decodeUnknownEffect(
   Schema.Struct({
@@ -23,209 +32,443 @@ const decodeToken = Schema.decodeUnknownEffect(
     expiresOn: Schema.optional(Schema.NonEmptyString),
   }),
 )
-export const AzurePlugin = define({
-  id: "opencode.provider.azure",
-  effect: Effect.fn(function* (ctx) {
-    const configured = yield* configuredSettings(Provider.ID.azure)
-    const processes = yield* AppProcess.Service
-    const bus = yield* Bus.Service
-    const tokens = new Map<string, { access: string; expires: number }>()
-    const loading = Semaphore.makeUnsafe(1)
-    const loaded: { resource?: string } = {}
+const ResourceDeployments = Schema.Struct({ data: Schema.Array(Schema.Unknown) })
+const decodeResourceDeployment = Schema.decodeUnknownOption(
+  Schema.Struct({ id: Schema.NonEmptyString, model: Schema.NonEmptyString, status: Schema.String }),
+)
+const ManagementDeployments = Schema.Struct({
+  value: Schema.Array(Schema.Unknown),
+  nextLink: Schema.optional(Schema.NonEmptyString),
+})
+const decodeManagementDeployment = Schema.decodeUnknownOption(
+  Schema.Struct({
+    name: Schema.NonEmptyString,
+    properties: Schema.Struct({
+      model: Schema.Struct({ name: Schema.NonEmptyString }),
+      provisioningState: Schema.String,
+    }),
+  }),
+)
+const ResourceQuery = Schema.Struct({ query: Schema.String })
+const Resources = Schema.Struct({ data: Schema.Array(Schema.Struct({ id: Schema.NonEmptyString })) })
 
-    const command = (args: string[]) =>
-      processes
-        .run(ChildProcess.make("az", args, { extendEnv: true, stdin: "ignore" }), { timeout: "10 seconds" })
-        .pipe(
-          Effect.flatMap(AppProcess.requireSuccess),
-          Effect.flatMap((result) => decodeJSON(result.stdout.toString("utf8"))),
+type Deployment = { readonly name: string; readonly model: string }
+
+export function make(
+  endpoints = {
+    resource: (name: string) => `https://${name}.openai.azure.com/openai`,
+    management: "https://management.azure.com",
+  },
+) {
+  return define({
+    id: "opencode.provider.azure",
+    effect: Effect.fn(function* (ctx) {
+      const configured = yield* configuredSettings(Provider.ID.azure)
+      const processes = yield* AppProcess.Service
+      const bus = yield* Bus.Service
+      const credentials = yield* Credential.Service
+      const providers = yield* Provider.Service
+      const http = HttpClient.filterStatusOk(yield* HttpClient.HttpClient)
+      const tokens = new Map<string, { access: string; expires: number }>()
+      // A resource keeps its Azure Resource Manager ID until it is deleted, so discovery looks it up once.
+      const resourceIDs = new Map<string, string>()
+      const loading = Semaphore.makeUnsafe(1)
+      const discovery = yield* FiberHandle.make<void, never>()
+      const loaded: {
+        resource?: string
+        url?: string
+        deployments?: readonly Deployment[]
+        connection?: Effect.Success<ReturnType<typeof ctx.integration.connection.active>>
+      } = {}
+
+      const command = (args: string[]) =>
+        processes
+          .run(ChildProcess.make("az", args, { extendEnv: true, stdin: "ignore" }), { timeout: "10 seconds" })
+          .pipe(
+            Effect.flatMap(AppProcess.requireSuccess),
+            Effect.flatMap((result) => decodeJSON(result.stdout.toString("utf8"))),
+          )
+
+      const token = Effect.fn("AzurePlugin.token")(function* (scope: string) {
+        const now = yield* Clock.currentTimeMillis
+        const cached = tokens.get(scope)
+        if (cached && cached.expires - now > 5 * 60_000) return cached
+        const result = yield* command(["account", "get-access-token", "--scope", scope, "--output", "json"]).pipe(
+          Effect.flatMap(decodeToken),
         )
-
-    const token = Effect.fn("AzurePlugin.token")(function* (scope: string) {
-      const now = yield* Clock.currentTimeMillis
-      const cached = tokens.get(scope)
-      if (cached && cached.expires - now > 5 * 60_000) return cached
-      const result = yield* command(["account", "get-access-token", "--scope", scope, "--output", "json"]).pipe(
-        Effect.flatMap(decodeToken),
-      )
-      const expires = result.expires_on !== undefined ? result.expires_on * 1000 : Date.parse(result.expiresOn ?? "")
-      if (!Number.isFinite(expires))
-        return yield* Effect.fail(new Error("Azure CLI returned an invalid token expiration"))
-      const refreshed = { access: result.accessToken, expires }
-      tokens.set(scope, refreshed)
-      return refreshed
-    })
-
-    const available = Boolean(which("az"))
-    const form = () =>
-      iife(() => {
-        if (resolveResourceName(configured) || typeof configured?.baseURL === "string") return
-        return Form.Fields.make([
-          {
-            type: "string",
-            key: "resourceName",
-            title: "Enter Azure Resource Name",
-            placeholder: "e.g. my-models",
-            required: true,
-          },
-        ])
+        const expires = result.expires_on !== undefined ? result.expires_on * 1000 : Date.parse(result.expiresOn ?? "")
+        if (!Number.isFinite(expires))
+          return yield* Effect.fail(new Error("Azure CLI returned an invalid token expiration"))
+        const refreshed = { access: result.accessToken, expires }
+        tokens.set(scope, refreshed)
+        return refreshed
       })
 
-    yield* ctx.integration.transform((editor) => {
-      editor.method.update({
-        integrationID: Provider.ID.azure,
-        method: { type: "key", label: "API key", form: form() },
-      })
-      if (!available) return
-      editor.method.update({
-        integrationID: Provider.ID.azure,
-        method: {
-          id: methodID,
-          type: "oauth",
-          label: "Microsoft Entra ID (Azure CLI)",
-          form: form(),
-        },
-        authorize: (answer) =>
-          Effect.succeed({
-            mode: "auto" as const,
-            url: "",
-            instructions: "Sign in with `az login` before continuing.",
-            callback: Effect.gen(function* () {
-              const resourceName =
-                typeof answer.resourceName === "string" ? answer.resourceName : resolveResourceName(configured)
-              if (!resourceName) return yield* Effect.fail(new Error("Azure resource name is required"))
-              const current = yield* token(cognitiveScope)
-              loaded.resource = resourceName
-              yield* ctx.provider.reload()
-              return Credential.OAuth.make({
-                type: "oauth",
-                methodID,
-                access: current.access,
-                refresh: "azure-cli",
-                expires: current.expires,
-                metadata: { resourceName },
-              })
-            }),
-          }),
-        refresh: (credential) =>
-          token(cognitiveScope).pipe(
-            Effect.map((current) =>
-              Credential.OAuth.make({ ...credential, access: current.access, expires: current.expires }),
+      const management = (request: HttpClientRequest.HttpClientRequest) =>
+        token(managementScope).pipe(
+          Effect.flatMap((current) =>
+            http.execute(
+              request.pipe(
+                HttpClientRequest.bearerToken(current.access),
+                HttpClientRequest.acceptJson,
+                HttpClientRequest.setHeader("User-Agent", App.useragent(ctx.app)),
+              ),
             ),
           ),
-      })
-    })
-
-    const load = Effect.fn("AzurePlugin.load")(function* () {
-      const connection = yield* ctx.integration.connection.active(Provider.ID.azure)
-      const credential = connection
-        ? yield* ctx.integration.connection.resolve(connection).pipe(Effect.orElseSucceed(() => undefined))
-        : undefined
-      if (credential?.type !== "oauth" || credential.methodID !== methodID) {
-        loaded.resource = undefined
-        return
-      }
-      const resource =
-        typeof credential.metadata?.resourceName === "string" ? credential.metadata.resourceName : undefined
-      loaded.resource = resource
-    })
-
-    yield* load()
-    yield* ctx.provider.transform((evt) => {
-      for (const item of evt.list()) {
-        if (
-          item.provider.id !== Provider.ID.azure &&
-          !item.provider.package.startsWith("@opencode/ai/providers/azure/")
         )
-          continue
-        const resourceName = resolveResourceName(item.provider.settings, loaded.resource)
-        const websocket = responsesWebSocketCapable(item.provider)
-        if (!resourceName && !websocket) continue
-        evt.update(item.provider.id, (provider) => {
-          provider.settings = {
-            ...provider.settings,
-            ...(resourceName === undefined ? {} : { resourceName }),
-            ...(websocket ? { transport: provider.settings?.transport ?? "websocket" } : {}),
-            ...(resourceName !== undefined && typeof provider.settings?.baseURL === "string"
-              ? { baseURL: expandResourceName(provider.settings.baseURL, resourceName) }
-              : {}),
-          }
+
+      const available = Boolean(which("az"))
+      const form = () =>
+        iife(() => {
+          if (resolveResourceName(configured) || typeof configured?.baseURL === "string") return
+          return Form.Fields.make([
+            {
+              type: "string",
+              key: "resourceName",
+              title: "Enter Azure Resource Name",
+              placeholder: "e.g. my-models",
+              required: true,
+            },
+          ])
         })
-      }
-    })
-    yield* ctx.model.transform((models) => {
-      for (const item of models.provider.list()) {
-        if (
-          item.provider.id !== Provider.ID.azure &&
-          !item.provider.package.startsWith("@opencode/ai/providers/azure/")
+
+      yield* ctx.integration.transform((editor) => {
+        editor.method.update({
+          integrationID: Provider.ID.azure,
+          method: { type: "key", label: "API key", form: form() },
+        })
+        if (!available) return
+        editor.method.update({
+          integrationID: Provider.ID.azure,
+          method: {
+            id: methodID,
+            type: "oauth",
+            label: "Microsoft Entra ID (Azure CLI)",
+            form: form(),
+          },
+          authorize: (answer) =>
+            Effect.succeed({
+              mode: "auto" as const,
+              url: "",
+              instructions: "Sign in with `az login` before continuing.",
+              callback: Effect.gen(function* () {
+                const resourceName =
+                  (typeof answer.resourceName === "string" ? answer.resourceName.trim() : "") ||
+                  resolveResourceName(configured)
+                if (!resourceName) return yield* Effect.fail(new Error("Azure resource name is required"))
+                const current = yield* token(cognitiveScope)
+                return Credential.OAuth.make({
+                  type: "oauth",
+                  methodID,
+                  access: current.access,
+                  refresh: "azure-cli",
+                  expires: current.expires,
+                  metadata: { resourceName },
+                })
+              }),
+            }),
+          refresh: (credential) =>
+            token(cognitiveScope).pipe(
+              Effect.map((current) =>
+                Credential.OAuth.make({ ...credential, access: current.access, expires: current.expires }),
+              ),
+            ),
+        })
+      })
+
+      const load = Effect.fn("AzurePlugin.load")(function* () {
+        const connection = yield* ctx.integration.connection.active(Provider.ID.azure)
+        // Startup awaits this, so it reads the stored credential: resolving the connection would refresh an
+        // expired token through the Azure CLI.
+        const stored =
+          connection?.type === "credential" ? yield* credentials.get(Credential.ID.make(connection.id)) : undefined
+        return { connection, resource: credentialResource(stored?.value) }
+      })
+
+      // Resource Graph searches every subscription the Azure CLI account can read, not only the selected one.
+      // https://learn.microsoft.com/rest/api/azureresourcegraph/resourcegraph/resources/resources
+      const findResource = Effect.fn("AzurePlugin.findResource")(function* (resource: string) {
+        const response = yield* HttpClientRequest.post(
+          `${endpoints.management}/providers/Microsoft.ResourceGraph/resources?api-version=2022-10-01`,
+        ).pipe(
+          HttpClientRequest.schemaBodyJson(ResourceQuery)({ query: resourceQuery(resource) }),
+          Effect.flatMap(management),
+          Effect.flatMap(HttpClientResponse.schemaBodyJson(Resources)),
+          Effect.timeout("10 seconds"),
         )
-          continue
-        const resourceName = resolveResourceName(item.provider.settings, loaded.resource)
-        for (const model of models.list(item.provider.id)) {
-          models.update(item.provider.id, model.id, (draft) => {
-            if (resourceName && typeof draft.settings?.baseURL === "string")
-              draft.settings.baseURL = expandResourceName(
-                draft.settings.baseURL,
-                resolveResourceName(draft.settings, resourceName) ?? resourceName,
-              )
+        const id = response.data[0]?.id
+        if (!id) return yield* Effect.fail(new Error(`Azure resource "${resource}" was not found`))
+        return id
+      })
+
+      const managementDeployments = Effect.fn("AzurePlugin.managementDeployments")(function* (resource: string) {
+        const key = resource.toLowerCase()
+        const id = resourceIDs.get(key) ?? (yield* findResource(resource))
+        resourceIDs.set(key, id)
+        const origin = new URL(endpoints.management).origin
+        return yield* Stream.paginate(`${endpoints.management}${id}/deployments?api-version=2024-10-01`, (url) =>
+          management(HttpClientRequest.get(url)).pipe(
+            Effect.flatMap(HttpClientResponse.schemaBodyJson(ManagementDeployments)),
+            Effect.timeout("10 seconds"),
+            Effect.flatMap((response) =>
+              // Every page carries the management token, so a page link must stay on the management endpoint.
+              // https://learn.microsoft.com/rest/api/aiservices/accountmanagement/deployments/list
+              response.nextLink !== undefined && URL.parse(response.nextLink)?.origin !== origin
+                ? Effect.fail(new Error("Azure returned a deployment page outside the management endpoint"))
+                : Effect.succeed([
+                    response.value.flatMap((raw): Deployment[] => {
+                      const item = Option.getOrUndefined(decodeManagementDeployment(raw))
+                      return item?.properties.provisioningState === "Succeeded"
+                        ? [{ name: item.name, model: item.properties.model.name }]
+                        : []
+                    }),
+                    Option.fromNullishOr(response.nextLink),
+                  ] as const),
+            ),
+          ),
+        ).pipe(
+          Stream.runCollect,
+          // A moved or recreated resource has a new ID, so the next discovery looks it up again.
+          Effect.tapError(() => Effect.sync(() => resourceIDs.delete(key))),
+        )
+      })
+
+      const resourceDeployments = Effect.fn("AzurePlugin.resourceDeployments")(function* (
+        url: string,
+        credential: Credential.Value,
+      ) {
+        return yield* http
+          .execute(
+            HttpClientRequest.get(url).pipe(
+              HttpClientRequest.acceptJson,
+              HttpClientRequest.setHeader("User-Agent", App.useragent(ctx.app)),
+              credential.type === "key"
+                ? HttpClientRequest.setHeader("api-key", credential.key)
+                : HttpClientRequest.bearerToken(credential.access),
+            ),
+          )
+          .pipe(
+            Effect.flatMap(HttpClientResponse.schemaBodyJson(ResourceDeployments)),
+            Effect.timeout("10 seconds"),
+            Effect.map((response) =>
+              response.data.flatMap((raw): Deployment[] => {
+                const item = Option.getOrUndefined(decodeResourceDeployment(raw))
+                return item?.status === "succeeded" ? [{ name: item.id, model: item.model }] : []
+              }),
+            ),
+          )
+      })
+
+      // Azure documents the management API as the deployment inventory, but only an Azure CLI session can reach it:
+      // Azure Resource Manager accepts Entra ID tokens, never resource keys.
+      // https://learn.microsoft.com/rest/api/aiservices/accountmanagement/deployments/list
+      // The resource's own inventory serves API keys and identities without Azure Resource Manager read access. Only
+      // data-plane version 2022-12-01 has it; later versions dropped `/deployments` and keep `/models`, which lists
+      // models the resource can deploy rather than its deployments.
+      // https://github.com/Azure/azure-rest-api-specs/blob/main/specification/cognitiveservices/data-plane/OpenAIAuthoring/stable/2022-12-01/azureopenai.json
+      const deployments = (url: string, resource: string, credential: Credential.Value) =>
+        credential.type === "oauth"
+          ? managementDeployments(resource).pipe(Effect.catch(() => resourceDeployments(url, credential)))
+          : resourceDeployments(url, credential)
+
+      // Local and quick, so a switch rebinds the provider before discovery for the new connection calls Azure.
+      const rebind = () =>
+        loading.withPermit(
+          Effect.gen(function* () {
+            const current = yield* load()
+            if (
+              IntegrationConnection.key(current.connection) === IntegrationConnection.key(loaded.connection) &&
+              current.resource === loaded.resource
+            )
+              return
+            Object.assign(loaded, current, { url: undefined, deployments: undefined })
+            yield* ctx.provider.reload()
+          }),
+        )
+
+      const discover = Effect.fn("AzurePlugin.discover")(function* () {
+        const connection = loaded.connection
+        const settings = (yield* providers.get(Provider.ID.azure))?.settings
+        const name = loaded.resource ?? resolveResourceName(settings)
+        // A custom endpoint may expose other deployments than the resource does, so it keeps the catalog.
+        const url =
+          connection && name !== undefined && resourcePattern.test(name) && typeof settings?.baseURL !== "string"
+            ? `${endpoints.resource(name)}/deployments?api-version=2022-12-01`
+            : undefined
+        if (loaded.connection !== connection) return
+        // Keep the last inventory through transient failures only for the same connection and resource.
+        if (loaded.url !== url) {
+          loaded.url = url
+          if (loaded.deployments) {
+            loaded.deployments = undefined
+            yield* ctx.model.reload()
+          }
+        }
+        if (!connection || !name || !url) return
+        const credential = yield* ctx.integration.connection
+          .resolve(connection)
+          .pipe(Effect.orElseSucceed(() => undefined))
+        if (!credential || (credential.type === "oauth" && credential.methodID !== methodID)) return
+        const found = yield* deployments(url, name, credential).pipe(
+          // Azure promises no order; normalize it so a reordered response does not rebuild the model list.
+          Effect.map((list) => list.toSorted((a, b) => a.name.localeCompare(b.name))),
+          Effect.catch((cause) =>
+            Effect.logWarning("failed to sync Azure deployments", { cause }).pipe(Effect.as(undefined)),
+          ),
+        )
+        if (!found) return
+        if (
+          loaded.connection !== connection ||
+          IntegrationConnection.key(connection) !==
+            IntegrationConnection.key(yield* ctx.integration.connection.active(Provider.ID.azure))
+        )
+          return
+        if (JSON.stringify(found) === JSON.stringify(loaded.deployments)) return
+        const catalog = new Map(
+          Array.from((yield* providers.snapshot()).records.get(Provider.ID.azure)?.models.keys() ?? [], (id) => [
+            id.toLowerCase(),
+            id,
+          ]),
+        )
+        const unmatched = found.filter((deployment) => !catalogModel(catalog, deployment))
+        if (unmatched.length > 0)
+          yield* Effect.logWarning("Azure deployments of models outside the catalog need explicit configuration", {
+            deployments: unmatched.map((deployment) => deployment.name),
+          })
+        loaded.deployments = found
+        yield* ctx.model.reload()
+      })
+
+      const refresh = () => rebind().pipe(Effect.andThen(FiberHandle.run(discovery, discover())))
+
+      // The connection's resource wins for Azure itself, matching the runtime merge of credentials over settings.
+      const resourceFor = (provider: Provider.Info) =>
+        provider.id === Provider.ID.azure
+          ? (loaded.resource ?? resolveResourceName(provider.settings))
+          : resolveResourceName(provider.settings, loaded.resource)
+
+      Object.assign(loaded, yield* load())
+      yield* ctx.provider.transform((evt) => {
+        for (const item of evt.list()) {
+          if (
+            item.provider.id !== Provider.ID.azure &&
+            !item.provider.package.startsWith("@opencode/ai/providers/azure/")
+          )
+            continue
+          const resourceName = resourceFor(item.provider)
+          const websocket = responsesWebSocketCapable(item.provider)
+          if (!resourceName && !websocket) continue
+          evt.update(item.provider.id, (provider) => {
+            provider.settings = {
+              ...provider.settings,
+              ...(resourceName === undefined ? {} : { resourceName }),
+              ...(websocket ? { transport: provider.settings?.transport ?? "websocket" } : {}),
+              ...(resourceName !== undefined && typeof provider.settings?.baseURL === "string"
+                ? { baseURL: expandResourceName(provider.settings.baseURL, resourceName) }
+                : {}),
+            }
           })
         }
-      }
-    })
+        const item = evt.get(Provider.ID.azure)
+        if (!item) return
+        // Bind resource settings and discovery to their account, so a switch hides them until the rebind.
+        // Keep the full templates here for explicit configuration; the model transform narrows the visible list.
+        evt.add({
+          info: item.provider,
+          models: Array.from(item.models.values()),
+          sourceConnection: loaded.connection,
+        })
+      })
+      yield* ctx.model.transform((models) => {
+        for (const item of models.provider.list()) {
+          if (
+            item.provider.id !== Provider.ID.azure &&
+            !item.provider.package.startsWith("@opencode/ai/providers/azure/")
+          )
+            continue
+          const resourceName = resourceFor(item.provider)
+          for (const model of models.list(item.provider.id)) {
+            models.update(item.provider.id, model.id, (draft) => {
+              if (resourceName && typeof draft.settings?.baseURL === "string")
+                draft.settings.baseURL = expandResourceName(
+                  draft.settings.baseURL,
+                  resolveResourceName(draft.settings, resourceName) ?? resourceName,
+                )
+            })
+          }
+        }
+        if (!loaded.deployments) return
+        // Narrowing here rather than in the provider catalog keeps every catalog model available as the base
+        // of a model the user configures explicitly; those are applied after this transform.
+        const catalog = models.list(Provider.ID.azure)
+        const deployed = deployedModels(loaded.deployments, catalog)
+        for (const model of catalog) {
+          if (!deployed.has(model.id)) models.remove(Provider.ID.azure, model.id)
+        }
+        for (const [id, model] of deployed) {
+          models.update(Provider.ID.azure, id, (draft) => Object.assign(draft, model))
+        }
+      })
 
-    const reload = () => loading.withPermit(load().pipe(Effect.andThen(ctx.provider.reload())))
-    yield* bus.subscribe(Credential.Event.Switched).pipe(
-      Stream.filter((event) => event.data.integrationID === Integration.ID.make("azure")),
-      Stream.runForEach(reload),
-      Effect.forkScoped({ startImmediately: true }),
-    )
+      // A switch interrupts discovery for the previous connection instead of waiting for its Azure calls.
+      yield* bus.subscribe(Credential.Event.Switched).pipe(
+        Stream.filter((event) => event.data.integrationID === Integration.ID.make("azure")),
+        Stream.runForEach(() => refresh()),
+        Effect.forkScoped({ startImmediately: true }),
+      )
+      // Deployments load in the background so startup never waits on Azure; the catalog serves until they arrive.
+      // Later changes load when the connection changes, so a new deployment needs a reconnect or restart.
+      yield* refresh().pipe(Effect.forkScoped)
 
-    // Entra bearer tokens are minted per request from the target URL's scope, so they are injected
-    // at the transport hooks rather than stored as a credential.
-    const bearer = Effect.fn("AzurePlugin.bearer")(function* (url: string) {
-      const connection = yield* ctx.integration.connection.active(Provider.ID.azure)
-      const credential = connection
-        ? yield* ctx.integration.connection.resolve(connection).pipe(Effect.orElseSucceed(() => undefined))
-        : undefined
-      if (credential?.type !== "oauth" || credential.methodID !== methodID) return
-      const target = new URL(url)
-      const scope =
-        target.hostname.endsWith(".services.ai.azure.com") && !target.pathname.startsWith("/models")
-          ? foundryScope
-          : cognitiveScope
-      const current = yield* token(scope).pipe(Effect.orDie)
-      return `Bearer ${current.access}`
-    })
-    yield* ctx.session.hook(
-      "http.request",
-      (evt) =>
-        Effect.gen(function* () {
-          if (evt.model.providerID !== Provider.ID.azure) return
-          const authorization = yield* bearer(evt.request.url)
-          if (!authorization) return
-          evt.request.headers.delete("api-key")
-          evt.request.headers.delete("x-api-key")
-          evt.request.headers.set("authorization", authorization)
-          evt.request.headers.set("user-agent", App.useragent(ctx.app))
-        }),
-      { providerID: Provider.ID.azure },
-    )
-    yield* ctx.session.hook(
-      "experimental.ws.handshake",
-      (evt) =>
-        Effect.gen(function* () {
-          if (evt.model.providerID !== Provider.ID.azure) return
-          const authorization = yield* bearer(evt.url)
-          if (!authorization) return
-          delete evt.headers["api-key"]
-          delete evt.headers["x-api-key"]
-          evt.headers.authorization = authorization
-          evt.headers["user-agent"] = App.useragent(ctx.app)
-        }),
-      { providerID: Provider.ID.azure },
-    )
-  }),
-})
+      // Entra bearer tokens are minted per request from the target URL's scope, so they are injected
+      // at the transport hooks rather than stored as a credential.
+      const bearer = Effect.fn("AzurePlugin.bearer")(function* (url: string) {
+        const connection = yield* ctx.integration.connection.active(Provider.ID.azure)
+        const credential = connection
+          ? yield* ctx.integration.connection.resolve(connection).pipe(Effect.orElseSucceed(() => undefined))
+          : undefined
+        if (credential?.type !== "oauth" || credential.methodID !== methodID) return
+        const target = new URL(url)
+        const scope =
+          target.hostname.endsWith(".services.ai.azure.com") && !target.pathname.startsWith("/models")
+            ? foundryScope
+            : cognitiveScope
+        const current = yield* token(scope).pipe(Effect.orDie)
+        return `Bearer ${current.access}`
+      })
+      yield* ctx.session.hook(
+        "http.request",
+        (evt) =>
+          Effect.gen(function* () {
+            if (evt.model.providerID !== Provider.ID.azure) return
+            const authorization = yield* bearer(evt.request.url)
+            if (!authorization) return
+            evt.request.headers.delete("api-key")
+            evt.request.headers.delete("x-api-key")
+            evt.request.headers.set("authorization", authorization)
+            evt.request.headers.set("user-agent", App.useragent(ctx.app))
+          }),
+        { providerID: Provider.ID.azure },
+      )
+      yield* ctx.session.hook(
+        "experimental.ws.handshake",
+        (evt) =>
+          Effect.gen(function* () {
+            if (evt.model.providerID !== Provider.ID.azure) return
+            const authorization = yield* bearer(evt.url)
+            if (!authorization) return
+            delete evt.headers["api-key"]
+            delete evt.headers["x-api-key"]
+            evt.headers.authorization = authorization
+            evt.headers["user-agent"] = App.useragent(ctx.app)
+          }),
+        { providerID: Provider.ID.azure },
+      )
+    }),
+  } satisfies PluginInternal.InternalPlugin)
+}
+
+export const AzurePlugin = make()
 
 function resolveResourceName(settings: Readonly<Record<string, unknown>> | undefined, fallback?: string) {
   const configured = settings?.resourceName
@@ -239,6 +482,60 @@ function expandResourceName(baseURL: string, resourceName: string) {
     .replaceAll("${AZURE_COGNITIVE_SERVICES_RESOURCE_NAME}", resourceName)
 }
 
+// The Azure CLI method stores the resource as credential metadata, the API key method as its form answer.
+function credentialResource(credential: Credential.Value | undefined) {
+  const resource =
+    credential?.type === "key"
+      ? credential.configuration?.resourceName
+      : credential?.methodID === methodID
+        ? credential.metadata?.resourceName
+        : undefined
+  return typeof resource === "string" && resource.trim() !== "" ? resource : undefined
+}
+
+function resourceQuery(resource: string) {
+  return [
+    "resources",
+    "| where type =~ 'microsoft.cognitiveservices/accounts' and kind in~ ('AIServices', 'OpenAI')",
+    // The custom subdomain is the resource name of every endpoint, and Entra ID authentication requires one.
+    // Subdomains are globally unique, so a name matches at most one resource.
+    // https://learn.microsoft.com/azure/ai-services/cognitive-services-custom-subdomains
+    "| extend resourceName = tostring(properties.customSubDomainName)",
+    `| where resourceName =~ '${resource}'`,
+    "| project id",
+    "| take 1",
+  ].join(" ")
+}
+
+// A deployment's ID is its name, while limits, costs, and routes come from the catalog model it deploys. Azure compares
+// names without case and may return another case, so IDs are lowercase like the catalog's.
+// https://learn.microsoft.com/azure/azure-resource-manager/management/resource-name-rules
+function deployedModels(deployments: readonly Deployment[], catalog: readonly Model.MutableInfo[]) {
+  const models = new Map(catalog.map((model) => [model.id.toLowerCase(), model]))
+  return new Map(
+    deployments.flatMap((deployment) => {
+      const model = catalogModel(models, deployment)
+      if (!model) return []
+      const id = Model.ID.make(deployment.name.toLowerCase())
+      const info: Model.MutableInfo = {
+        ...structuredClone(model),
+        id,
+        modelID: Model.ID.make(deployment.name),
+        name: id === model.id.toLowerCase() ? model.name : `${model.name} (${deployment.name})`,
+      }
+      return [[id, info] as const]
+    }),
+  )
+}
+
+// Azure spells some models unlike the catalog: a model name plus a separate version, such as `gpt-4` for GPT-4 Turbo,
+// `gpt-35-turbo`, or mixed case such as `DeepSeek-V4-Flash`. A deployment named after a catalog model then stands for
+// that model, as it did before discovery; any other is left to explicit configuration.
+// https://learn.microsoft.com/azure/foundry/openai/concepts/retired-models
+function catalogModel<T>(models: ReadonlyMap<string, T>, deployment: Deployment) {
+  return models.get(deployment.model.toLowerCase()) ?? models.get(deployment.name.toLowerCase())
+}
+
 function responsesWebSocketCapable(provider: Provider.Info) {
   if (provider.package !== "@opencode/ai/providers/azure/responses") return false
   const settings = provider.settings

packages/core/src/provider.ts

@@ -69,6 +69,7 @@ const builtins = new Map<string, () => Promise<unknown>>([
   ["@opencode/ai/providers/cloudflare-workers-ai", () => import("@opencode/ai/providers/cloudflare-workers-ai")],
   ["@opencode/ai/providers/deepinfra", () => import("@opencode/ai/providers/deepinfra")],
   ["@opencode/ai/providers/deepseek", () => import("@opencode/ai/providers/deepseek")],
+  ["@opencode/ai/providers/digitalocean", () => import("@opencode/ai/providers/digitalocean")],
   ["@opencode/ai/providers/fireworks", () => import("@opencode/ai/providers/fireworks")],
   ["@opencode/ai/providers/google", () => import("@opencode/ai/providers/google")],
   ["@opencode/ai/providers/google-vertex", () => import("@opencode/ai/providers/google-vertex")],

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

@@ -94,9 +94,10 @@ export const Plugin = {
                   targets: [FILENAME],
                   start: result.content.type === "list-page" ? resolved : dirname(resolved),
                   stop: root,
+                  type: "file",
                 })
                 const candidates = (yield* Effect.forEach(discovered, fs.resolve)).filter(
-                  (file) => !FSUtil.contains(dirname(file), root),
+                  (file) => !FSUtil.contains(dirname(file), root) && file !== resolved,
                 )
                 if (candidates.length === 0) return
                 yield* sessionInstructions.load({ sessionID: context.sessionID, paths: candidates })

packages/core/src/variant.ts

@@ -574,6 +574,7 @@ const PROTOCOLS: Readonly<Record<string, Protocol>> = {
   "@opencode/ai/providers/cloudflare-workers-ai": workersAIChat,
   "@opencode/ai/providers/deepinfra": deepinfraChat,
   "@opencode/ai/providers/deepseek": deepseekChat,
+  "@opencode/ai/providers/digitalocean": openaiChat,
   "@opencode/ai/providers/fireworks": openaiChat,
   "@opencode/ai/providers/groq": openaiChat,
   "@opencode/ai/providers/meta/chat": openaiChat,

packages/core/test/instruction-discovery.test.ts

@@ -431,7 +431,7 @@ describe("ConfigInstructionPlugin.Plugin", () => {
 
   it.effect("canonicalizes boundaries and honors project opt-out", () =>
     Effect.gen(function* () {
-      const observed: { values: { targets: string[]; start: string; stop?: string }[] } = { values: [] }
+      const observed: { values: FSUtil.UpOptions[] } = { values: [] }
       const observingFS = Layer.effect(
         FSUtil.Service,
         FSUtil.Service.pipe(
@@ -487,7 +487,7 @@ describe("ConfigInstructionPlugin.Plugin", () => {
       )
 
       const repo = path.resolve("/repo")
-      expect(observed.values).toEqual([{ targets: ["AGENTS.md"], start: repo, stop: repo }])
+      expect(observed.values).toEqual([{ targets: ["AGENTS.md"], start: repo, stop: repo, type: "file" }])
     }),
   )
 })

packages/core/test/permission.test.ts

@@ -1,4 +1,4 @@
-import { describe, expect } from "bun:test"
+import { describe, expect, test } from "bun:test"
 import { Cause, Deferred, Effect, Fiber, Layer } from "effect"
 import { Agent } from "@opencode/core/agent"
 import { Database } from "@opencode/core/database/database"
@@ -148,6 +148,14 @@ describe("Permission", () => {
     }),
   )
 
+  test("matches Windows rule resources against slash-normalized file access resources", () => {
+    const rules: Permission.Ruleset = [
+      { action: "external_directory", resource: "C:\\Users\\x\\proj\\*", effect: "allow" },
+    ]
+    expect(Permission.evaluate("external_directory", "C:/Users/x/proj/src/*", rules).effect).toBe("allow")
+    expect(Permission.evaluate("external_directory", "C:/Users/x/other/*", rules).effect).toBe("ask")
+  })
+
   it.effect("allows managed output reads without granting external directory access", () =>
     Effect.gen(function* () {
       yield* setup([

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

@@ -89,10 +89,15 @@ const identity = {
   agent: Agent.ID.make("build"),
   messageID: SessionMessage.ID.make("msg_nearby"),
 }
-const readCall = (sessionID: Session.ID, id: string, readPath: string): Parameters<Tool.Snapshot["execute"]>[0] => ({
+const readCall = (
+  sessionID: Session.ID,
+  id: string,
+  readPath: string,
+  page: ReadToolFileSystem.PageInput = {},
+): Parameters<Tool.Snapshot["execute"]>[0] => ({
   sessionID,
   ...identity,
-  call: { type: "tool-call", id, name: "read", input: { path: readPath } },
+  call: { type: "tool-call", id, name: "read", input: { path: readPath, ...page } },
 })
 
 const writeAgents = (file: string, content: string) => Effect.promise(() => fs.writeFile(file, content))
@@ -131,7 +136,7 @@ describe("SessionInstructions", () => {
       yield* writeAgents(rootPath, "root-instructions")
       yield* writeAgents(subPath, "sub-instructions")
       yield* writeAgents(deepPath, "deep-instructions")
-      yield* writeAgents(otherPath, "other-instructions")
+      yield* writeAgents(otherPath, "other-instructions\nmore rules")
       yield* Effect.promise(() => fs.writeFile(path.resolve(dir, "sub", "deep", "file.txt"), "file content"))
       yield* Effect.promise(() => fs.writeFile(path.resolve(dir, "sub", "other", "file2.txt"), "file content 2"))
 
@@ -155,13 +160,23 @@ describe("SessionInstructions", () => {
       expect(firstInjected[0]!.metadata).toEqual({ instruction: { paths: [deepPath, subPath] } })
       expect(firstInjected[0]!.text).not.toContain("root-instructions")
 
+      // Neither a full nor a partial read adds an automatic copy of the file itself.
+      const read = yield* executeTool(registry, readCall(sessionID, "call-direct", "sub/other/AGENTS.md"))
+      expect(read.content?.[0]).toMatchObject({ type: "text", text: expect.stringContaining("more rules") })
+      const partial = yield* executeTool(
+        registry,
+        readCall(sessionID, "call-partial", "sub/other/AGENTS.md", { limit: 1 }),
+      )
+      expect(partial.metadata).toEqual({ truncated: true })
+      expect(yield* synthetics(sessionID)).toHaveLength(1)
+
       // A sibling read under sub/other discovers only the new AGENTS.md; sub is already
       // injected for this session so it is not re-emitted, and the root is still excluded.
       yield* executeTool(registry, readCall(sessionID, "call-other", "sub/other/file2.txt"))
 
       const secondInjected = yield* synthetics(sessionID)
       expect(secondInjected).toHaveLength(2)
-      expect(secondInjected[1]!.text).toBe(`Instructions from: ${otherPath}\nother-instructions`)
+      expect(secondInjected[1]!.text).toBe(`Instructions from: ${otherPath}\nother-instructions\nmore rules`)
       expect(secondInjected[1]!.description).toBe(`Loaded ${path.relative(dir, otherPath)}`)
       expect(secondInjected[1]!.metadata).toEqual({ instruction: { paths: [otherPath] } })
       expect(secondInjected.some((message) => message.text.includes("root-instructions"))).toBe(false)

packages/core/test/variant.test.ts

@@ -296,6 +296,17 @@ test("spells Workers AI thinking controls through the chat template", () => {
 })
 
 test("spells Chat Completions variants for hosting providers", () => {
+  expect(
+    resolve(model("@opencode/ai/providers/digitalocean", "openai-gpt-5-nano", undefined, "digitalocean"), [
+      { type: "effort", values: ["minimal", "low", "medium", "high"] },
+    ]),
+  ).toEqual([
+    { id: "minimal", settings: { reasoningEffort: "minimal" } },
+    { id: "low", settings: { reasoningEffort: "low" } },
+    { id: "medium", settings: { reasoningEffort: "medium" } },
+    { id: "high", settings: { reasoningEffort: "high" } },
+  ])
+
   expect(
     resolve(model("@opencode/ai/providers/openai-compatible", "deepseek-ai/deepseek-v4-pro", undefined, "nvidia"), [
       { type: "effort", values: ["none", "high", "max"] },

packages/gui-extensions/src/review/model.ts

@@ -2,7 +2,7 @@ import type { FileDiffInfo } from "@opencode/client/promise"
 import type { SessionReviewLineComment } from "@opencode/session-ui/session-review"
 import { previewSelectedLines } from "@opencode/session-ui/pierre/selection-bridge"
 import { checksum } from "@opencode/util/encode"
-import { createQuery, skipToken, useQueryClient } from "@tanstack/solid-query"
+import { createQuery, useQueryClient } from "@tanstack/solid-query"
 import { debounce } from "@solid-primitives/scheduled"
 import { createComputed, createEffect, createMemo, on, onCleanup } from "solid-js"
 import { createStore } from "solid-js/store"
@@ -17,7 +17,6 @@ import {
 } from "./kinds"
 
 export type ChangeMode = "git" | "branch" | "turn"
-type VcsMode = "git" | "branch"
 type FileSelection = { startLine: number; endLine: number; startChar: number; endChar: number }
 
 export type Demand = { tree: number; files: number; panel: number; details: number }
@@ -109,12 +108,10 @@ export function createReviewModel(input: { ctx: Context; view: SessionView; dema
     ) {
       list.push("branch")
     }
+    // Turn snapshots are captured only for Git sessions.
+    if (project?.vcs === "git" && view.id) list.push("turn")
     return list
   })
-  const vcsMode = createMemo<VcsMode | undefined>(() => {
-    const value = mode()
-    return value === "git" || value === "branch" ? value : undefined
-  })
   const vcsKey = createMemo(
     () =>
       [
@@ -130,22 +127,25 @@ export function createReviewModel(input: { ctx: Context; view: SessionView; dema
     const demand = input.demand
     return demand.tree + demand.files + demand.panel > 0
   })
-  const vcsQuery = createQuery(() => {
-    const value = vcsMode()
+  const turnKey = () => [ctx.id, view.server.id, "session-turn", view.id] as const
+  const diffQuery = createQuery(() => {
+    const value = mode()
+    const turn = value === "turn"
     return {
-      queryKey: [...vcsKey(), value] as const,
+      queryKey: turn ? turnKey() : ([...vcsKey(), value] as const),
       enabled: view.server.connected && wantsReview() && !!view.project?.vcs,
       refetchOnMount: "always" as const,
-      refetchOnWindowFocus: true,
-      queryFn: value
-        ? () =>
+      // A finished turn does not change on focus or filesystem events; refresh it when the session goes idle.
+      refetchOnWindowFocus: !turn,
+      queryFn: turn
+        ? () => view.server.client.session.diff({ sessionID: view.id })
+        : () =>
             view.server.client.vcs
               .diff({
                 location: { directory: directory() },
                 mode: value === "git" ? "working" : value,
               })
-              .then((result) => result.data)
-        : skipToken,
+              .then((result) => result.data),
     }
   })
   // The summary's changes row: the session directory's working tree, loaded only while the summary shows.
@@ -180,20 +180,17 @@ export function createReviewModel(input: { ctx: Context; view: SessionView; dema
     on(
       () => !layout.narrow() && layout.side.opened(view),
       (open, previous) => {
-        if (!open || previous || vcsQuery.isFetching) return
+        if (!open || previous || diffQuery.isFetching) return
         if (input.demand.tree > 0) {
           refresh()
           return
         }
-        if (vcsMode() && view.server.connected && view.project?.vcs) void vcsQuery.refetch()
+        if (view.server.connected && view.project?.vcs) void diffQuery.refetch()
       },
       { defer: true },
     ),
   )
-  const diffs = (): FileDiffInfo[] => {
-    if (mode() === "git" || mode() === "branch") return vcsQuery.isFetched ? (vcsQuery.data ?? []) : []
-    return []
-  }
+  const diffs = (): FileDiffInfo[] => (diffQuery.isFetched ? (diffQuery.data ?? []) : [])
   const renderable = createMemo(() => diffs().filter(filterRenderableDiff))
   const kinds = createMemo(() => reviewDiffKinds(renderable()))
   const activeFile = () => {
@@ -205,17 +202,12 @@ export function createReviewModel(input: { ctx: Context; view: SessionView; dema
   const count = () => diffs().length
   const hasChanges = () => count() > 0
   const ready = () => {
-    // A project without VCS never enables vcsQuery, so its status stays "pending" forever.
+    // A project without VCS never enables diffQuery, so its status stays "pending" forever.
     const project = view.project
     if (project && !project.vcs) return true
-    if (mode() === "git" || mode() === "branch") return !vcsQuery.isPending
-    return true
+    return !diffQuery.isPending
   }
   const loadDiff = async (path: string, version?: number): Promise<FileDiffInfo | undefined> => {
-    const value = vcsMode()
-    if (!value) return undefined
-    const root = reviewRootDirectory(view.project?.worktree ?? directory())
-    const scoped = reviewDiffDirectory(root, path)
     const source = diffs().find((diff) => diff.file === path)
     const valid = (diff: FileDiffInfo | undefined): FileDiffInfo | undefined => {
       if (!diff || !source) return undefined
@@ -223,6 +215,24 @@ export function createReviewModel(input: { ctx: Context; view: SessionView; dema
       if (reviewDiffNeedsLoad(diff)) return undefined
       return diff
     }
+    const value = mode()
+    // Oversized full-file patches come back empty; bounded context usually fits.
+    if (value === "turn") {
+      return queryClient
+        .fetchQuery({
+          queryKey: [...turnKey(), "bounded", version] as const,
+          staleTime: Number.POSITIVE_INFINITY,
+          retry: 2,
+          queryFn: () => view.server.client.session.diff({ sessionID: view.id, context: 3 }),
+        })
+        .then((result) => valid(result.find((diff) => diff.file === path)))
+        .catch((error: unknown) => {
+          console.debug("[session-review] failed to load bounded turn diff", { path, error })
+          return undefined
+        })
+    }
+    const root = reviewRootDirectory(view.project?.worktree ?? directory())
+    const scoped = reviewDiffDirectory(root, path)
     const request = (scope: string, context?: number) =>
       queryClient
         .fetchQuery({
@@ -376,6 +386,7 @@ export function createReviewModel(input: { ctx: Context; view: SessionView; dema
       (next, previous) => {
         if (next !== "idle" || previous === undefined || previous === "idle") return
         refresh()
+        void queryClient.invalidateQueries({ queryKey: turnKey() })
       },
       { defer: true },
     ),
@@ -416,7 +427,7 @@ export function createReviewModel(input: { ctx: Context; view: SessionView; dema
     count,
     deferRender: () => state.deferRender,
     details: (): FileDiffInfo[] | undefined => (detailsQuery.isFetched ? (detailsQuery.data ?? []) : undefined),
-    diffVersion: () => vcsQuery.dataUpdatedAt,
+    diffVersion: () => diffQuery.dataUpdatedAt,
     diffs,
     renderable,
     kinds,

packages/gui-extensions/src/review/parts.tsx

@@ -27,7 +27,7 @@ export function ReviewTitle(props: { review: ReviewModel }) {
 
 export function ReviewEmpty(props: { review: ReviewModel; loadingClass: string }) {
   const ctx = useExtension()
-  const loading = () => (props.review.mode() === "git" || props.review.mode() === "branch") && !props.review.ready()
+  const loading = () => !props.review.ready()
   const noGit = () => props.review.noGit()
   const text = () => {
     if (props.review.mode() === "git") return ctx.t("empty.git")
@@ -60,7 +60,7 @@ export function ReviewEmpty(props: { review: ReviewModel; loadingClass: string }
 
 export function ReviewPanelEmpty(props: { review: ReviewModel }) {
   const ctx = useExtension()
-  const loading = () => (props.review.mode() === "git" || props.review.mode() === "branch") && !props.review.ready()
+  const loading = () => !props.review.ready()
   const noGit = () => props.review.noGit()
   return (
     <Switch>

packages/stats/app/src/app.tsx

@@ -2,7 +2,6 @@ import { MetaProvider, Title } from "@solidjs/meta"
 import { Router } from "@solidjs/router"
 import { FileRoutes } from "@solidjs/start/router"
 import { Suspense } from "solid-js"
-import { PatternRipple } from "./component/pattern-ripple"
 import { I18nProvider, useI18n } from "./context/i18n"
 import { LanguageProvider } from "./context/language"
 import { strip } from "./lib/language"
@@ -24,7 +23,6 @@ export default function App() {
           <I18nProvider>
             <MetaProvider>
               <AppMeta />
-              <PatternRipple />
               <Suspense>{props.children}</Suspense>
             </MetaProvider>
           </I18nProvider>

packages/stats/app/src/component/pattern-ripple.tsx

@@ -1,97 +0,0 @@
-import { onCleanup, onMount } from "solid-js"
-
-const PATTERNS = '[data-page="stats"] [data-slot$="-pattern"]'
-const RIPPLE = {
-  duration: 4400,
-  brightness: 50,
-  trail: 400,
-  trailBody: 100,
-  front: 90,
-  ease: 4,
-  fade: 3.3,
-  reach: 1.3,
-  rearm: 400,
-}
-
-type Ripple = { x: number; y: number; radius: number; start: number }
-
-// Plays one soft wave of light across a dot pattern when the cursor moves onto its visible dots.
-export function PatternRipple() {
-  onMount(() => {
-    if (matchMedia("(prefers-reduced-motion: reduce)").matches) return
-
-    const ripples = new Map<HTMLElement, Ripple>()
-    const leftAt = new WeakMap<HTMLElement, number>()
-    const state = { current: undefined as HTMLElement | undefined, primed: false, running: false }
-
-    const frame = (now: number) => {
-      ripples.forEach((ripple, el) => {
-        const progress = (now - ripple.start) / RIPPLE.duration
-        if (progress >= 1) {
-          ripples.delete(el)
-          el.style.removeProperty("--dot-ripple")
-          return
-        }
-        el.style.setProperty("--dot-ripple", rippleGradient(ripple, progress))
-      })
-      if (ripples.size > 0) return requestAnimationFrame(frame)
-      state.running = false
-    }
-
-    const start = (el: HTMLElement, clientX: number, clientY: number) => {
-      const box = el.getBoundingClientRect()
-      const visible = visibleBox(el)
-      ripples.set(el, {
-        x: clientX - box.left,
-        y: clientY - box.top,
-        radius: Math.hypot(
-          Math.max(clientX - visible.left, visible.right - clientX),
-          Math.max(clientY - visible.top, visible.bottom - clientY),
-        ),
-        start: performance.now(),
-      })
-      if (state.running) return
-      state.running = true
-      requestAnimationFrame(frame)
-    }
-
-    // The pattern must be the top element under the cursor, so text cut-outs over a pattern count as off the dots.
-    // A cursor already resting on a pattern when the page loads does not trigger a ripple.
-    const move = (event: PointerEvent) => {
-      const now = performance.now()
-      const hit = document.elementFromPoint(event.clientX, event.clientY)
-      const el = hit instanceof HTMLElement && hit.matches(PATTERNS) ? hit : undefined
-      if (el !== state.current) {
-        if (state.current) leftAt.set(state.current, now)
-        if (el && state.primed && now - (leftAt.get(el) ?? -Infinity) >= RIPPLE.rearm) start(el, event.clientX, event.clientY)
-        state.current = el
-      }
-      state.primed = true
-    }
-
-    document.addEventListener("pointermove", move, { passive: true })
-    onCleanup(() => document.removeEventListener("pointermove", move))
-  })
-
-  return null
-}
-
-function rippleGradient(ripple: Ripple, progress: number) {
-  const r = ripple.radius * RIPPLE.reach * (1 - (1 - progress) ** RIPPLE.ease)
-  const alpha = RIPPLE.brightness * (1 - progress) ** RIPPLE.fade
-  const mix = (amount: number) => `color-mix(in srgb, var(--stats-text) ${amount.toFixed(1)}%, transparent)`
-  const at = (offset: number) => `${Math.max(0, r + offset).toFixed(1)}px`
-  return `radial-gradient(circle at ${ripple.x}px ${ripple.y}px, transparent ${at(-RIPPLE.trail)}, ${mix((alpha * RIPPLE.trailBody) / 100)} ${at(-RIPPLE.trail / 2)}, ${mix(alpha)} ${at(0)}, transparent ${at(RIPPLE.front)})`
-}
-
-// Some patterns are larger than their clipping container, so the ripple only needs to reach the visible part.
-function visibleBox(el: HTMLElement) {
-  const box = el.getBoundingClientRect()
-  const clip = el.parentElement?.getBoundingClientRect() ?? box
-  return {
-    left: Math.max(box.left, clip.left),
-    top: Math.max(box.top, clip.top),
-    right: Math.min(box.right, clip.right),
-    bottom: Math.min(box.bottom, clip.bottom),
-  }
-}

packages/stats/app/src/routes/index.css

@@ -674,7 +674,7 @@ body {
 [data-page="stats"] [data-slot="footer-pattern"] {
   height: 16px;
   overflow: hidden;
-  background-color: var(--stats-hero-pattern);
+  background: var(--stats-hero-pattern);
   mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
   mask-position: center top;
   mask-repeat: repeat;
@@ -883,18 +883,13 @@ body {
   display: none;
 }
 
-/* Ripple layer drawn by PatternRipple. Pattern rules set only background-color so this layer survives. */
-[data-page="stats"] [data-slot$="-pattern"] {
-  background-image: var(--dot-ripple, none);
-}
-
 [data-page="stats"] [data-slot="hero-pattern"] {
   order: 2;
   flex: 0 0 auto;
   width: 100%;
   height: 16px;
   overflow: hidden;
-  background-color: var(--stats-hero-pattern);
+  background: var(--stats-hero-pattern);
   mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
   mask-repeat: repeat;
   mask-size: 6px 6px;
@@ -1099,6 +1094,7 @@ body {
 [data-page="stats"] [data-component="top-models-chart"] {
   --top-models-bar-gap: 12px;
   --top-models-dot-size: 6px;
+  --top-models-dot-offset: 2px;
   position: relative;
   display: grid;
   grid-template-rows: 34px minmax(0, 1fr);
@@ -1228,36 +1224,38 @@ body {
   padding-inline: calc(var(--top-models-bar-gap) / 2);
   outline: none;
   cursor: pointer;
-  /* Resolved by absolutely positioned children against this slot's width. Three 2px dot columns span the bar, so this inset makes the gap between bars equal the gap between dot columns. */
-  --top-models-bar-inset: calc((100% - 6px) / 6);
 }
 
-/* Three dot columns aligned to the bar's left edge, center, and right edge, so they read as rising out of the bar. The SVG has no viewBox, so the 2px dot keeps its size while the cell stretches. */
 [data-page="stats"] [data-slot="top-models-bar"]::before {
   position: absolute;
-  inset: 0 var(--top-models-bar-inset);
+  top: 0;
+  bottom: 0;
+  left: 50%;
+  width: calc(100% - var(--top-models-bar-gap));
+  transform: translateX(-50%);
   content: "";
   background: var(--stats-dot);
-  mask-image: url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg'%3E%3Crect width='2' height='2'/%3E%3C/svg%3E");
-  mask-position: left top;
+  mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
+  mask-position: var(--top-models-dot-offset) top;
   mask-repeat: repeat;
-  mask-size: calc((100% - 2px) / 2) var(--top-models-dot-size);
-  -webkit-mask-image: url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg'%3E%3Crect width='2' height='2'/%3E%3C/svg%3E");
-  -webkit-mask-position: left top;
+  mask-size: var(--top-models-dot-size) var(--top-models-dot-size);
+  -webkit-mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
+  -webkit-mask-position: var(--top-models-dot-offset) top;
   -webkit-mask-repeat: repeat;
-  -webkit-mask-size: calc((100% - 2px) / 2) var(--top-models-dot-size);
-  transition: background-color 120ms ease;
+  -webkit-mask-size: var(--top-models-dot-size) var(--top-models-dot-size);
 }
 
-[data-page="stats"] [data-slot="top-models-bar"]:is(:hover, [data-active="true"])::before {
-  background: color-mix(in srgb, var(--stats-dot), var(--stats-text) 25%);
+@supports (width: round(down, 100%, 1px)) {
+  [data-page="stats"] [data-slot="top-models-bar"]::before {
+    width: round(down, calc(100% - var(--top-models-bar-gap)), var(--top-models-dot-size));
+  }
 }
 
 [data-page="stats"] [data-slot="top-models-stack"] {
   position: absolute;
-  right: var(--top-models-bar-inset);
+  right: calc(var(--top-models-bar-gap) / 2);
   bottom: 0;
-  left: var(--top-models-bar-inset);
+  left: calc(var(--top-models-bar-gap) / 2);
   z-index: 1;
   display: grid;
   height: var(--top-models-bar-height);
@@ -1877,7 +1875,7 @@ body {
   height: 16px;
   margin: 16px 0;
   overflow: hidden;
-  background-color: var(--stats-hero-pattern);
+  background: var(--stats-hero-pattern);
   mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
   mask-position: center top;
   mask-repeat: repeat;
@@ -2878,7 +2876,7 @@ body {
   width: 100%;
   height: 18px;
   overflow: hidden;
-  background-color: var(--stats-hero-pattern);
+  background: var(--stats-hero-pattern);
   mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
   mask-repeat: repeat;
   mask-size: 6px 6px;
@@ -3018,7 +3016,7 @@ body {
   min-width: 64px;
   height: 64px;
   overflow: hidden;
-  background-color: color-mix(in srgb, var(--stats-text) 10%, transparent);
+  background: color-mix(in srgb, var(--stats-text) 10%, transparent);
   mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
   mask-position: center top;
   mask-repeat: repeat;
@@ -3335,7 +3333,7 @@ body {
   width: 100%;
   height: 16px;
   overflow: hidden;
-  background-color: color-mix(in srgb, var(--stats-text) 10%, transparent);
+  background: color-mix(in srgb, var(--stats-text) 10%, transparent);
   mask-position: center top;
   -webkit-mask-position: center top;
 }
@@ -3482,7 +3480,7 @@ body {
   width: 100%;
   height: 16px;
   overflow: hidden;
-  background-color: var(--stats-line);
+  background: var(--stats-line);
   mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
   mask-position: center top;
   mask-repeat: repeat;
@@ -3956,7 +3954,7 @@ body {
   height: 16px;
   margin-bottom: 40px;
   overflow: hidden;
-  background-color: color-mix(in srgb, var(--stats-text) 10%, transparent);
+  background: color-mix(in srgb, var(--stats-text) 10%, transparent);
   mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
   mask-position: center top;
   mask-repeat: repeat;
@@ -4081,7 +4079,7 @@ body {
   height: 16px;
   margin-bottom: 40px;
   overflow: hidden;
-  background-color: color-mix(in srgb, var(--stats-text) 10%, transparent);
+  background: color-mix(in srgb, var(--stats-text) 10%, transparent);
   mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
   mask-position: center top;
   mask-repeat: repeat;
@@ -4547,6 +4545,7 @@ body {
 [data-page="stats"] [data-component="model-usage-chart"] {
   --model-usage-bar-gap: 6px;
   --model-usage-dot-size: 6px;
+  --model-usage-dot-offset: 2px;
   position: relative;
   display: grid;
   grid-template-rows: 40px minmax(0, 1fr);
@@ -4604,36 +4603,38 @@ body {
   padding-inline: calc(var(--model-usage-bar-gap) / 2);
   outline: none;
   cursor: pointer;
-  /* Resolved by absolutely positioned children against this slot's width. Three 2px dot columns span the bar, so this inset makes the gap between bars equal the gap between dot columns. */
-  --model-usage-bar-inset: calc((100% - 6px) / 6);
 }
 
-/* Three dot columns aligned to the bar's left edge, center, and right edge, so they read as rising out of the bar. The SVG has no viewBox, so the 2px dot keeps its size while the cell stretches. */
 [data-page="stats"] [data-slot="model-usage-column"]::before {
   position: absolute;
-  inset: 0 var(--model-usage-bar-inset);
+  top: 0;
+  bottom: 0;
+  left: 50%;
+  width: calc(100% - var(--model-usage-bar-gap));
+  transform: translateX(-50%);
   content: "";
   background: var(--stats-dot);
-  mask-image: url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg'%3E%3Crect width='2' height='2'/%3E%3C/svg%3E");
-  mask-position: left top;
+  mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
+  mask-position: var(--model-usage-dot-offset) top;
   mask-repeat: repeat;
-  mask-size: calc((100% - 2px) / 2) var(--model-usage-dot-size);
-  -webkit-mask-image: url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg'%3E%3Crect width='2' height='2'/%3E%3C/svg%3E");
-  -webkit-mask-position: left top;
+  mask-size: var(--model-usage-dot-size) var(--model-usage-dot-size);
+  -webkit-mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
+  -webkit-mask-position: var(--model-usage-dot-offset) top;
   -webkit-mask-repeat: repeat;
-  -webkit-mask-size: calc((100% - 2px) / 2) var(--model-usage-dot-size);
-  transition: background-color 120ms ease;
+  -webkit-mask-size: var(--model-usage-dot-size) var(--model-usage-dot-size);
 }
 
-[data-page="stats"] [data-slot="model-usage-column"]:is(:hover, [data-active="true"])::before {
-  background: color-mix(in srgb, var(--stats-dot), var(--stats-text) 25%);
+@supports (width: round(down, 100%, 1px)) {
+  [data-page="stats"] [data-slot="model-usage-column"]::before {
+    width: round(down, calc(100% - var(--model-usage-bar-gap)), var(--model-usage-dot-size));
+  }
 }
 
 [data-page="stats"] [data-slot="model-usage-bar"] {
   position: absolute;
-  right: var(--model-usage-bar-inset);
+  right: calc(var(--model-usage-bar-gap) / 2);
   bottom: 0;
-  left: var(--model-usage-bar-inset);
+  left: calc(var(--model-usage-bar-gap) / 2);
   z-index: 1;
   height: var(--model-usage-fill);
   min-height: 0;
@@ -5689,7 +5690,7 @@ body {
   height: 16px;
   margin-inline: -40px;
   overflow: hidden;
-  background-color: color-mix(in srgb, var(--stats-text) 10%, transparent);
+  background: color-mix(in srgb, var(--stats-text) 10%, transparent);
   mask-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 6 6' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 0H2V2H0V0Z' fill='black'/%3E%3C/svg%3E");
   mask-position: center top;
   mask-repeat: repeat;
@@ -8195,8 +8196,7 @@ body {
   }
 
   [data-page="stats"] [data-component="model-usage-chart"] {
-    /* 18px puts three 2px dot columns 6px apart, matching the 6px bar gap. */
-    --model-usage-mobile-bar-width: 18px;
+    --model-usage-mobile-bar-width: 16px;
     --model-usage-mobile-edge-space: 48px;
     --model-usage-mobile-track-width: calc(
       var(--model-usage-count) * (var(--model-usage-mobile-bar-width) + var(--model-usage-bar-gap)) +
@@ -8237,7 +8237,7 @@ body {
   }
 
   [data-page="stats"] [data-slot="model-usage-column"]::before {
-    inset: 0;
+    width: 100%;
   }
 
   [data-page="stats"] [data-slot="model-usage-bar"] {

packages/tui/src/context/update-notification.tsx

@@ -17,7 +17,7 @@ export type UpdateState =
 
 export type UpdateSource = {
   readonly remote: boolean
-  readonly subscribe: (notify: (notice: ClientNotice) => void, signal: AbortSignal) => Promise<void>
+  readonly subscribe: (notify: (notice: ClientNotice) => void) => () => void
   readonly check: (
     signal: AbortSignal,
     onInstall: (version: string) => void,
@@ -115,15 +115,8 @@ export const { use: useUpdateNotification, provider: UpdateNotificationProvider
     }
 
     onMount(() => {
-      const updater = props.updater
-      if (!updater) return
-      const controller = new AbortController()
-      onCleanup(() => controller.abort())
-      void updater
-        .subscribe((notice) => notify({ ...notice, source: "client" }), controller.signal)
-        .catch((error) => {
-          if (!controller.signal.aborted) log.error("update check failed", { error })
-        })
+      if (!props.updater) return
+      onCleanup(props.updater.subscribe((notice) => notify({ ...notice, source: "client" })))
     })
 
     onCleanup(

packages/util/src/fs-util.ts

@@ -32,6 +32,8 @@ export namespace FSUtil {
     readonly start: string
     readonly stop?: string
     readonly mode?: "all" | "first"
+    /** Only match regular files or directories (following symlinks). By default any existing path matches. */
+    readonly type?: "file" | "directory"
   }
 
   export interface Interface extends FileSystem.FileSystem {
@@ -165,7 +167,13 @@ export namespace FSUtil {
         while (true) {
           for (const target of options.targets) {
             const search = join(current, target)
-            if (yield* fs.exists(search)) {
+            const found =
+              options.type === "file"
+                ? yield* isFile(search)
+                : options.type === "directory"
+                  ? yield* isDir(search)
+                  : yield* fs.exists(search)
+            if (found) {
               result.push(search)
               if (options.mode === "first") return result
             }

services/www/src/docs/content/build/plugins/effect.mdx

@@ -699,7 +699,7 @@ interface MCPDomain extends Pick<McpApi<unknown>, "list"> {
 
 ### Plugins
 
-List active, failed, and resolved plugins at the current location.
+List active and failed plugins at the current location.
 
 ```ts
 effect: (ctx) =>
@@ -718,7 +718,7 @@ effect: (ctx) =>
     const plugin = ctx.plugin
     const active = yield* plugin.list().pipe(
       Effect.orDie,
-      Effect.map((result) => result.data.filter((item) => item.status === "active")),
+      Effect.map((result) => result.data.filter((item) => item.state.status === "active")),
     )
     yield* Effect.logInfo("active plugin count", { count: active.length })
   }),

services/www/src/docs/content/build/plugins/index.mdx

@@ -646,7 +646,7 @@ interface MCPEditor {
 
 ### Plugins
 
-List the plugins currently active for a location.
+List active and failed plugins at the current location.
 
 ```ts
 const plugins = await ctx.plugin.list()

services/www/src/docs/content/cli/commands.mdx

@@ -3,7 +3,7 @@ title: "Commands"
 description: "Reference for the opencode command line."
 ---
 
-Every command accepts `--help` for its full flag list, for example `opencode run --help`. Commands that talk to a server also accept `--standalone` to run a private server and `--server <url>` to target a specific one.
+Every command accepts `--help` for its full flag list, for example `opencode run --help`. Some server-backed commands accept `--standalone` to run a private server or `--server <url>` to target a specific one; check the command's help for its supported flags.
 
 ## run
 
@@ -167,7 +167,7 @@ $ opencode auth login anthropic
 Log in with a specific authentication method.
 
 ```bash
-$ opencode auth login anthropic --method api-key
+$ opencode auth login anthropic --method key
 ```
 
 Log out of a saved account.
@@ -462,13 +462,13 @@ $ opencode api GET /api/session
 Call an operation ID with a query parameter.
 
 ```bash
-$ opencode api v2.session.list --param limit=10
+$ opencode api session.list --param limit=10
 ```
 
 Send a JSON body.
 
 ```bash
-$ opencode api v2.session.create --data '{"title": "New session"}'
+$ opencode api session.create --data '{"title": "New session"}'
 ```
 
 Add a request header.
@@ -540,7 +540,7 @@ $ opencode upgrade
 Upgrade to a specific version with a specific package manager.
 
 ```bash
-$ opencode upgrade 1.18.15 --method bun
+$ opencode upgrade 2.0.21 --method bun
 ```
 
 View all subcommands and flags.

services/www/src/docs/content/mcp-servers.mdx

@@ -353,7 +353,7 @@ Use permission actions to hide or deny tools without disconnecting their server.
 
 ## Context
 
-For calls made on behalf of a session, OpenCode sends the session ID in `CallToolRequest.params._meta.sessionID`. This applies to direct tools and Code Mode over stdio and Streamable HTTP:
+For calls made on behalf of a session, OpenCode sends the session ID in `CallToolRequest.params._meta["ai.opencode/sessionID"]`. This applies to direct tools and Code Mode over stdio and Streamable HTTP:
 
 ```json
 {
@@ -361,7 +361,7 @@ For calls made on behalf of a session, OpenCode sends the session ID in `CallToo
   "params": {
     "name": "lookup",
     "arguments": { "query": "example" },
-    "_meta": { "sessionID": "ses_..." }
+    "_meta": { "ai.opencode/sessionID": "ses_..." }
   }
 }
 ```

services/www/src/docs/content/providers.mdx

@@ -268,7 +268,11 @@ choose **Microsoft Entra ID (Azure CLI)** when connecting Azure.
 az login
 ```
 
-For a resource in another tenant or subscription, select both explicitly.
+Enter the resource name when connecting if it is not already supplied by configuration or the environment. A resource
+name saved with a connection takes precedence over configuration; connect again to change it.
+
+Azure CLI connections use the account selected in the Azure CLI. For a resource in another tenant or subscription,
+select both explicitly.
 
 ```bash
 az login --tenant TENANT_ID
@@ -278,8 +282,32 @@ az account set --subscription NAME_OR_ID
 Instead of configuration, `AZURE_RESOURCE_NAME` supplies the resource name to the OpenCode server. The legacy
 `AZURE_COGNITIVE_SERVICES_RESOURCE_NAME` variable also works.
 
-OpenCode does not query Azure management APIs or discover deployments. If a deployment does not match its catalog
-model name, map an OpenCode model ID to the deployment with `modelID`.
+### Deployments
+
+Azure serves a model only through a deployment, so OpenCode lists the deployments of your resource and shows those
+instead of the whole Azure catalog. The list loads in the background after startup, after connecting, and after
+switching accounts. To pick up a deployment added later, restart OpenCode or connect again.
+
+- Until the list loads for a connection, or if it fails, the whole catalog stays available.
+- A failed reload for the same connection keeps its last complete inventory.
+- Switching accounts never shows the previous account's inventory; the new account starts from the whole catalog.
+- Deployments inherit limits, costs, and capabilities from their catalog models.
+
+Deployment names are their model IDs, in lowercase because Azure ignores case in names, for example
+`azure/gpt-production`. A deployment named after its model, such as `gpt-5-mini`, keeps the catalog ID. These IDs stay
+stable when other deployments are added or removed.
+
+OpenCode matches a deployment to the catalog by the model it deploys, or by its name when Azure spells the model
+differently, as with `gpt-4` for GPT-4 Turbo. A deployment named after another model, such as `gpt-5` deploying
+`gpt-5-mini`, shows that model's name, limits, and costs.
+
+With the Azure CLI, OpenCode lists deployments through the Azure management API, which needs read access to the
+resource. If that fails, and always with an API key, it uses the resource's legacy deployment inventory. Discovery reads
+every page before publishing the list. A custom `settings.baseURL` keeps the catalog.
+
+Configure a deployment explicitly when its model is not in the catalog, such as a fine-tuned model, or to map a model ID
+to a deployment yourself. OpenCode logs a warning naming such deployments. Explicitly configured models are always
+kept.
 
 ```jsonc title="opencode.jsonc"
 {

services/www/src/docs/content/websearch.mdx

@@ -12,7 +12,7 @@ The first search asks you to allow web search and select a provider. OpenCode re
 
 ## Providers
 
-OpenCode includes four search providers:
+OpenCode includes five search providers:
 
 | Provider | ID | Environment variable |
 | --- | --- | --- |
@@ -20,6 +20,7 @@ OpenCode includes four search providers:
 | Firecrawl | `firecrawl` | `FIRECRAWL_API_KEY` |
 | Parallel | `parallel` | `PARALLEL_API_KEY` |
 | Tavily | `tavily` | `TAVILY_API_KEY` |
+| TinyFish | `tinyfish` | `TINYFISH_API_KEY` (optional) |
 
 Connect an account from the TUI with `/connect`, or set the provider's environment variable before starting OpenCode.