fix(tui): rank sessions before settings for /se

Centralize control dispatch and first-Step preparation in advanceToStep. Keep input delivery outside logical-Step retries and preserve queue ordering, Location handoff, context refresh, and durable settlement.

bun.lock

@@ -125,7 +125,7 @@
         "@effect/platform-node": "catalog:",
         "@opencode-ai/client": "workspace:*",
         "@opencode-ai/plugin": "workspace:*",
-        "@opencode-ai/pty": "0.1.8",
+        "@opencode-ai/pty": "0.1.10",
         "@opencode-ai/schema": "workspace:*",
         "@opencode-ai/server": "workspace:*",
         "@opencode-ai/tui": "workspace:*",
@@ -350,7 +350,6 @@
         "@ai-sdk/cohere": "3.0.27",
         "@ai-sdk/gateway": "3.0.104",
         "@ai-sdk/google-vertex": "4.0.128",
-        "@ai-sdk/groq": "3.0.31",
         "@ai-sdk/mistral": "3.0.51",
         "@ai-sdk/openai-compatible": "2.0.41",
         "@ai-sdk/perplexity": "3.0.26",
@@ -365,7 +364,7 @@
         "@opencode-ai/ai": "workspace:*",
         "@opencode-ai/codemode": "workspace:*",
         "@opencode-ai/plugin": "workspace:*",
-        "@opencode-ai/pty": "0.1.8",
+        "@opencode-ai/pty": "0.1.10",
         "@opencode-ai/schema": "workspace:*",
         "@opencode-ai/util": "workspace:*",
         "@parcel/watcher": "2.5.1",
@@ -732,6 +731,7 @@
       },
       "devDependencies": {
         "@happy-dom/global-registrator": "20.0.11",
+        "@playwright/test": "catalog:",
         "@tsconfig/node22": "catalog:",
         "@types/bun": "catalog:",
         "@types/luxon": "catalog:",
@@ -837,6 +837,7 @@
         "@opencode-ai/client": "workspace:*",
         "@opencode-ai/session-ui": "workspace:*",
         "@opencode-ai/ui": "workspace:*",
+        "@playwright/test": "catalog:",
         "@solidjs/meta": "catalog:",
         "@storybook/addon-a11y": "10.4.4",
         "@storybook/addon-docs": "10.4.4",
@@ -2157,19 +2158,19 @@
 
     "@opencode-ai/protocol": ["@opencode-ai/protocol@workspace:packages/protocol"],
 
-    "@opencode-ai/pty": ["@opencode-ai/pty@0.1.8", "", { "optionalDependencies": { "@opencode-ai/pty-darwin-arm64": "0.1.8", "@opencode-ai/pty-darwin-x64": "0.1.8", "@opencode-ai/pty-linux-arm64-gnu": "0.1.8", "@opencode-ai/pty-linux-arm64-musl": "0.1.8", "@opencode-ai/pty-linux-x64-gnu": "0.1.8", "@opencode-ai/pty-linux-x64-musl": "0.1.8" }, "bin": { "opencode-pty": "bin/opencode-pty.js" } }, "sha512-+OHAANB6xgtO4zVMK5Q3l50Iot2fYcHL003rNpvxxCXLkpdO90UnnRCL9jrll1SWkWoDZKaUDlR/Scw6TUTN4A=="],
+    "@opencode-ai/pty": ["@opencode-ai/pty@0.1.10", "", { "optionalDependencies": { "@opencode-ai/pty-darwin-arm64": "0.1.10", "@opencode-ai/pty-darwin-x64": "0.1.10", "@opencode-ai/pty-linux-arm64-gnu": "0.1.10", "@opencode-ai/pty-linux-arm64-musl": "0.1.10", "@opencode-ai/pty-linux-x64-gnu": "0.1.10", "@opencode-ai/pty-linux-x64-musl": "0.1.10" }, "bin": { "opencode-pty": "bin/opencode-pty.js" } }, "sha512-cEJT1ADtmnb+df2wrlUcsGny6Q7pTe9Sa7keISzCO0xN1FrL1aS6+eleBPpDimHjgM/sXqvLwJv0UiAeiAvgxQ=="],
 
-    "@opencode-ai/pty-darwin-arm64": ["@opencode-ai/pty-darwin-arm64@0.1.8", "", { "os": "darwin", "cpu": "arm64" }, "sha512-5DqVcCfa9P5zpM0QpvXD/QqZXFdXvtIoaBHVHjYnTt8oqlnmMWqG4OVowzPNQg07NudW6apwX1Jop1wuERgWGw=="],
+    "@opencode-ai/pty-darwin-arm64": ["@opencode-ai/pty-darwin-arm64@0.1.10", "", { "os": "darwin", "cpu": "arm64" }, "sha512-j7aszDFRwCIazGUT9eIy4PZwh4rltjvRmoicPRTK3kONN3v0MMflstkmAFDYYpqDPTNh3qJ6xkQmB+DugEbhAg=="],
 
-    "@opencode-ai/pty-darwin-x64": ["@opencode-ai/pty-darwin-x64@0.1.8", "", { "os": "darwin", "cpu": "x64" }, "sha512-keEDxfpy0mFJYmsTqea9IUbpLKpbe9yL4ISU8tYA0daWiKuR8ZlfLTlLGltaFNO7l7q6mURMeI4cZzwySOX/Kw=="],
+    "@opencode-ai/pty-darwin-x64": ["@opencode-ai/pty-darwin-x64@0.1.10", "", { "os": "darwin", "cpu": "x64" }, "sha512-UAMP/E4lo9RGQF7xrfIwpW2ZEemj308rCogJy14ruKYJt5MwHeGNTynGiHE/1JlDLRy+21wV50jpugADgT71ag=="],
 
-    "@opencode-ai/pty-linux-arm64-gnu": ["@opencode-ai/pty-linux-arm64-gnu@0.1.8", "", { "os": "linux", "cpu": "arm64" }, "sha512-NVJvlVtPbYRyjABwLkm1iboijpSZMXv4lfg4qda7UCzAgiDZ0cWauXeW5wtd9bbnUbHw4rdNVU6YweuwASgZXw=="],
+    "@opencode-ai/pty-linux-arm64-gnu": ["@opencode-ai/pty-linux-arm64-gnu@0.1.10", "", { "os": "linux", "cpu": "arm64" }, "sha512-lTPlZNQ66koFHZqoPmvvq0SetlepKVQYgnLryhlVfYtcryWDJM7gV4+P66V12RwqWQTjt2u8j12mtg3axSKg2w=="],
 
-    "@opencode-ai/pty-linux-arm64-musl": ["@opencode-ai/pty-linux-arm64-musl@0.1.8", "", { "os": "linux", "cpu": "arm64" }, "sha512-GAhnCs7b7B8Hqa7AEN7k5ra4IqybFYePgbCSOHn3Y4mofp/5NjDlkOG/VFM7rHAy05iEmRi5WtOPRv0Mkca8zA=="],
+    "@opencode-ai/pty-linux-arm64-musl": ["@opencode-ai/pty-linux-arm64-musl@0.1.10", "", { "os": "linux", "cpu": "arm64" }, "sha512-IDmWHRylMR/ZfMw9/AAktO/Edi4TITPC+Tq7Xx3JZHsDgSba3QdyE11uNL0zM1myTGdk6Yrt4rpdAzaItPnDjw=="],
 
-    "@opencode-ai/pty-linux-x64-gnu": ["@opencode-ai/pty-linux-x64-gnu@0.1.8", "", { "os": "linux", "cpu": "x64" }, "sha512-vSbty1L0BuOQlK7DmukuhdxNKv2vzqyxOjHmh+TULptMDbDZwE1EybASoXB7dYog9as85adFCz8ZTF7Boyfxjg=="],
+    "@opencode-ai/pty-linux-x64-gnu": ["@opencode-ai/pty-linux-x64-gnu@0.1.10", "", { "os": "linux", "cpu": "x64" }, "sha512-Q1yob0/8X2JoJZzFmNKUc32XDRAe0avKQ8PLKkpJr30qWXSrGmhltgcDmn94Q70zW9Ght9on84T7cmge9brvdQ=="],
 
-    "@opencode-ai/pty-linux-x64-musl": ["@opencode-ai/pty-linux-x64-musl@0.1.8", "", { "os": "linux", "cpu": "x64" }, "sha512-2GhmJLl+Ly/JjzMS2Eqvk7SgeuHIB0cQ4Wrmnj4WKhr2V7lS+pdUEApN6S0SG/FK0TYJlWeHeOIpNZgDqzEPPg=="],
+    "@opencode-ai/pty-linux-x64-musl": ["@opencode-ai/pty-linux-x64-musl@0.1.10", "", { "os": "linux", "cpu": "x64" }, "sha512-7RLHWQxX/wfUKJJP2ZMMtkXaPsrgoMNKzE6PL/LbnYbMBtkqfld9EDcMv1RFZ0CqjNFgI0Hg4eRk6x+ZNc/wyQ=="],
 
     "@opencode-ai/schema": ["@opencode-ai/schema@workspace:packages/schema"],
 

nix/hashes.json

@@ -1,8 +1,8 @@
 {
   "nodeModules": {
-    "x86_64-linux": "sha256-6Fc26RulNnJZtMUxeryWYqCt/7xIs1E+cOQSddAC9LA=",
-    "aarch64-linux": "sha256-p7cyVH12IGAo6HyaoSgO1fBx5fNTCZGTpvbqzIpe+A8=",
-    "aarch64-darwin": "sha256-U90pWWqe7DRiUpY5KUUp4exQpa/ddb/bQ5Ldxe46LfQ=",
-    "x86_64-darwin": "sha256-Ak6omOq2BBRLQEMxPydiQY9ajaJGQR61P6az4+xxZVM="
+    "x86_64-linux": "sha256-XHUy+Hk+RHUHREt4x0HfSzr3RlEvgBd4H/fV0rlXw2M=",
+    "aarch64-linux": "sha256-/gIaM62uV2X6KCnkSi6QjyT7t7uJ2L7u7CxXxPQYK+w=",
+    "aarch64-darwin": "sha256-PWG6ALh6kG7mnC6AzEuIAU54BEQ4IB+SyyQFGb/s+Dc=",
+    "x86_64-darwin": "sha256-TcgRDHG4CAT+XoCi0JNansONjg06ZPRmeLRsqmdJ4B4="
   }
 }

packages/ai/src/cache-policy.ts

@@ -36,7 +36,12 @@ const resolve = (policy: CachePolicy | undefined): CachePolicyObject => {
 // Protocols whose wire format ignores inline cache markers (OpenAI's implicit
 // prefix caching, Gemini's implicit + out-of-band CachedContent). Skip the
 // whole policy pass for these — emitting hints would be harmless but pointless.
-const RESPECTS_INLINE_HINTS = new Set(["anthropic-messages", "google-vertex-messages", "bedrock-converse", "openrouter"])
+const RESPECTS_INLINE_HINTS = new Set([
+  "anthropic-messages",
+  "google-vertex-messages",
+  "bedrock-converse",
+  "openrouter",
+])
 
 const makeHint = (ttlSeconds: number | undefined): CacheHint =>
   ttlSeconds !== undefined ? new CacheHint({ type: "ephemeral", ttlSeconds }) : new CacheHint({ type: "ephemeral" })

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

@@ -69,14 +69,22 @@ export interface OptionsInput {
   // SDK Metadata:2649 {user_id?: string | null}
   readonly metadata?: { readonly user_id?: string | null }
   // SDK MessageCreateParamsContainer:2596 ContainerParams|string
-  readonly container?: string | { readonly id?: string | null; readonly skills?: ReadonlyArray<Record<string, unknown>> | null }
+  readonly container?:
+    | string
+    | { readonly id?: string | null; readonly skills?: ReadonlyArray<Record<string, unknown>> | null }
   readonly inference_geo?: string | null
   readonly inferenceGeo?: string | null
   readonly cache_control?: { readonly type: "ephemeral"; readonly ttl?: "5m" | "1h" }
   readonly cacheControl?: { readonly type: "ephemeral"; readonly ttl?: "5m" | "1h" }
   // SDK OutputConfig:2684 {effort, format: JSONOutputFormat}
-  readonly output_config?: { readonly effort?: string | null; readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null }
-  readonly outputConfig?: { readonly effort?: string | null; readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null }
+  readonly output_config?: {
+    readonly effort?: string | null
+    readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null
+  }
+  readonly outputConfig?: {
+    readonly effort?: string | null
+    readonly format?: { readonly type: "json_schema"; readonly schema: Record<string, unknown> } | null
+  }
 }
 
 export type ProviderOptionsInput = OptionsInput
@@ -259,7 +267,11 @@ const AnthropicToolChoice = Schema.Union([
     type: Schema.Literals(["auto", "any", "none"]),
     disable_parallel_tool_use: Schema.optional(Schema.Boolean),
   }),
-  Schema.Struct({ type: Schema.tag("tool"), name: Schema.String, disable_parallel_tool_use: Schema.optional(Schema.Boolean) }),
+  Schema.Struct({
+    type: Schema.tag("tool"),
+    name: Schema.String,
+    disable_parallel_tool_use: Schema.optional(Schema.Boolean),
+  }),
 ])
 
 const AnthropicThinking = Schema.Union([
@@ -506,7 +518,11 @@ const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult
   // Prefer the provider-owned replay payload; fall back to the result value for
   // histories constructed directly from provider events.
   const payload = part.providerMetadata?.anthropic?.["result"] ?? part.result.value
-  return { type: wireType, tool_use_id: scrubToolCallID(part.id), content: payload } satisfies AnthropicServerToolResultBlock
+  return {
+    type: wireType,
+    tool_use_id: scrubToolCallID(part.id),
+    content: payload,
+  } satisfies AnthropicServerToolResultBlock
 })
 
 const fileIdFromMetadata = (metadata: MediaPart["metadata"]): string | undefined => {
@@ -554,9 +570,7 @@ const documentContextFromMetadata = (metadata: MediaPart["metadata"]): string |
   return undefined
 }
 
-const citationsFromMetadata = (
-  metadata: MediaPart["metadata"],
-): AnthropicDocumentBlock["citations"] | undefined => {
+const citationsFromMetadata = (metadata: MediaPart["metadata"]): AnthropicDocumentBlock["citations"] | undefined => {
   if (!ProviderShared.isRecord(metadata)) return undefined
   const raw = ProviderShared.isRecord(metadata.anthropic)
     ? (metadata.anthropic.citations ?? metadata.citations)
@@ -706,8 +720,7 @@ const lowerToolResultContent = Effect.fnUntraced(function* (part: ToolResultPart
 })
 
 const requireThinkingSignature = (request: LLMRequest) => {
-  if (request.model.compatibility?.requireSignature !== undefined)
-    return request.model.compatibility.requireSignature
+  if (request.model.compatibility?.requireSignature !== undefined) return request.model.compatibility.requireSignature
   const provider = request.model.provider.toLowerCase()
   const model = request.model.id.toLowerCase()
   const baseURL = (request.model.route.endpoint.baseURL ?? "").toLowerCase()
@@ -900,21 +913,24 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
 
 const resolveOptions = Effect.fn("AnthropicMessages.resolveOptions")(function* (request: LLMRequest) {
   const input = request.providerOptions as Record<string, unknown> | undefined
-  const rawServiceTier = (input as Record<string, unknown> | undefined)?.service_tier ?? (input as Record<string, unknown> | undefined)?.serviceTier
+  const rawServiceTier =
+    (input as Record<string, unknown> | undefined)?.service_tier ??
+    (input as Record<string, unknown> | undefined)?.serviceTier
   const service_tier =
     rawServiceTier === "auto" || rawServiceTier === "standard_only"
       ? (rawServiceTier as "auto" | "standard_only")
       : undefined
   const rawMetadata = (input as Record<string, unknown> | undefined)?.metadata
   const metadata =
-    ProviderShared.isRecord(rawMetadata) &&
-    (typeof rawMetadata.user_id === "string" || rawMetadata.user_id === null)
+    ProviderShared.isRecord(rawMetadata) && (typeof rawMetadata.user_id === "string" || rawMetadata.user_id === null)
       ? { user_id: rawMetadata.user_id as string | null }
       : undefined
   const container =
     typeof (input as Record<string, unknown> | undefined)?.container === "string" ||
     ProviderShared.isRecord((input as Record<string, unknown> | undefined)?.container)
-      ? ((input as Record<string, unknown>).container as string | { id?: string | null; skills?: ReadonlyArray<Record<string, unknown>> | null })
+      ? ((input as Record<string, unknown>).container as
+          | string
+          | { id?: string | null; skills?: ReadonlyArray<Record<string, unknown>> | null })
       : undefined
   const rawInferenceGeo =
     (input as Record<string, unknown> | undefined)?.inference_geo ??
@@ -965,8 +981,7 @@ const resolveThinking = Effect.fn("AnthropicMessages.resolveThinking")(function*
     input.display === "summarized" || input.display === "omitted"
       ? (input.display as "summarized" | "omitted")
       : undefined
-  if (input.type === "adaptive")
-    return { type: "adaptive" as const, ...(display === undefined ? {} : { display }) }
+  if (input.type === "adaptive") return { type: "adaptive" as const, ...(display === undefined ? {} : { display }) }
   if (input.type === "disabled") return { type: "disabled" as const }
   if (input.type !== "enabled") return undefined
   const budget =
@@ -1418,9 +1433,7 @@ const step = (state: ParserState, event: AnthropicEvent) => {
       if (event.index === undefined)
         return Effect.fail(ProviderShared.eventError(ADAPTER, `Anthropic ${block.type} missing index`))
       if (!block.id)
-        return Effect.fail(
-          ProviderShared.eventError(ADAPTER, `Anthropic tool_use missing id at index ${event.index}`),
-        )
+        return Effect.fail(ProviderShared.eventError(ADAPTER, `Anthropic tool_use missing id at index ${event.index}`))
     }
     return Effect.succeed(onContentBlockStart(state, { ...event, content_block: block }))
   }
@@ -1473,10 +1486,9 @@ export const route = Route.make({
   provider: "anthropic",
   providerMetadataKey: "anthropic",
   protocol,
-  endpoint: Endpoint.path(
-    (input) => (input.request.model.provider === "anthropic" ? `${PATH}?beta=true` : PATH),
-    { baseURL: DEFAULT_BASE_URL },
-  ),
+  endpoint: Endpoint.path((input) => (input.request.model.provider === "anthropic" ? `${PATH}?beta=true` : PATH), {
+    baseURL: DEFAULT_BASE_URL,
+  }),
   auth: Auth.none,
   framing,
   headers: () => ({ "anthropic-version": "2023-06-01" }),

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

@@ -652,7 +652,9 @@ const step = (state: ParserState, event: BedrockEvent) =>
         method: "stream",
         reason: classifyProviderFailure({
           message:
-            event.exception.details.message ?? event.exception.details.originalMessage ?? "Bedrock Converse stream error",
+            event.exception.details.message ??
+            event.exception.details.originalMessage ??
+            "Bedrock Converse stream error",
           code: event.exception.type,
         }),
       })

packages/ai/src/protocols/bedrock-event-stream.ts

@@ -82,7 +82,9 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
         "Failed to parse Bedrock Converse event-stream payload",
       )) as Record<string, unknown>
       delete parsed.p
-      out.push(messageType === "exception" ? { exception: { type: eventType, details: parsed } } : { [eventType]: parsed })
+      out.push(
+        messageType === "exception" ? { exception: { type: eventType, details: parsed } } : { [eventType]: parsed },
+      )
     }
     return [cursor, out] as const
   })

packages/ai/src/protocols/gemini.ts

@@ -570,7 +570,12 @@ const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {
       googleMetadata({ thoughtSignature: state.reasoningSignature }),
     )
   if (state.textSignature !== undefined)
-    lifecycle = Lifecycle.textEnd(lifecycle, events, "text-0", googleMetadata({ thoughtSignature: state.textSignature }))
+    lifecycle = Lifecycle.textEnd(
+      lifecycle,
+      events,
+      "text-0",
+      googleMetadata({ thoughtSignature: state.textSignature }),
+    )
   Lifecycle.finish(lifecycle, events, {
     reason: {
       normalized:
@@ -675,8 +680,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
           id,
           name: part.functionCall.name,
           input,
-          providerMetadata:
-            part.thoughtSignature ? googleMetadata({ thoughtSignature: part.thoughtSignature }) : undefined,
+          providerMetadata: part.thoughtSignature
+            ? googleMetadata({ thoughtSignature: part.thoughtSignature })
+            : undefined,
         }),
       )
       hasToolCalls = true

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

@@ -346,7 +346,8 @@ export const Event = Schema.StructWithRest(
     item_id: Schema.optional(Schema.String),
     output_index: Schema.optional(Schema.Number),
     summary_index: Schema.optional(Schema.Number),
-    item: Schema.optional(StreamItem),
+    // OutputItemAdded/Done permit a null item in the Open Responses OpenAPI schema.
+    item: optionalNull(StreamItem),
     response: Schema.optional(
       Schema.StructWithRest(
         Schema.Struct({
@@ -812,7 +813,7 @@ export const providerMetadata = (state: ParserState, metadata: Record<string, un
 })
 
 const isReasoningItem = (item: StreamItem): item is StreamItem & { type: "reasoning"; id: string } =>
-  item.type === "reasoning" && typeof item.id === "string" && item.id.length > 0
+  item.type === "reasoning" && typeof item.id === "string"
 
 export type StepResult = readonly [ParserState, ReadonlyArray<LLMEvent>]
 
@@ -891,7 +892,7 @@ const reasoningMetadata = (state: ParserState, item: StreamItem & { id: string }
 // best-effort, not guaranteed.
 const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
   const item = event.item
-  if (item?.type === "message" && item.id) {
+  if (item?.type === "message" && item.id !== undefined) {
     const phase = messagePhase(item.phase)
     return [
       {
@@ -922,7 +923,7 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
   }
   if (item?.type !== "function_call" || !item.call_id) return [state, NO_EVENTS]
   const id = item.id ?? item.call_id
-  const metadata = item.id ? providerMetadata(state, { itemId: item.id }) : undefined
+  const metadata = item.id !== undefined ? providerMetadata(state, { itemId: item.id }) : undefined
   const events: LLMEvent[] = []
   const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
   return [
@@ -941,7 +942,7 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
 }
 
 const onReasoningSummaryPartAdded = (state: ParserState, event: Event): StepResult => {
-  if (!event.item_id || event.summary_index === undefined) return [state, NO_EVENTS]
+  if (event.item_id === undefined || event.summary_index === undefined) return [state, NO_EVENTS]
   const item = state.reasoningItems[event.item_id]
   if (!item) return [state, NO_EVENTS]
   if (event.summary_index === 0) return [state, NO_EVENTS]
@@ -988,7 +989,7 @@ const onReasoningSummaryPartAdded = (state: ParserState, event: Event): StepResu
 }
 
 const onReasoningSummaryPartDone = (state: ParserState, event: Event): StepResult => {
-  if (!event.item_id || event.summary_index === undefined) return [state, NO_EVENTS]
+  if (event.item_id === undefined || event.summary_index === undefined) return [state, NO_EVENTS]
   const item = state.reasoningItems[event.item_id]
   if (!item) return [state, NO_EVENTS]
   return [
@@ -1013,7 +1014,7 @@ const onFunctionCallArgumentsDelta = Effect.fn("OpenResponses.onFunctionCallArgu
   state: ParserState,
   event: Event,
 ) {
-  if (!event.item_id) return [state, NO_EVENTS] satisfies StepResult
+  if (event.item_id === undefined) return [state, NO_EVENTS] satisfies StepResult
   const tool = state.tools[event.item_id]
   if (!tool) return [state, NO_EVENTS] satisfies StepResult
   const final = event.type === "response.function_call_arguments.done" ? event.arguments : undefined
@@ -1044,7 +1045,7 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
   const item = event.item
   if (!item) return [state, NO_EVENTS] satisfies StepResult
 
-  if (item.type === "message" && item.id) {
+  if (item.type === "message" && item.id !== undefined) {
     const itemPhase = messagePhase(item.phase)
     const phase = itemPhase === undefined ? state.messagePhases[item.id] : itemPhase
     const events: LLMEvent[] = []
@@ -1075,7 +1076,7 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
       : ToolStream.start(state.tools, id, {
           id: item.call_id,
           name: item.name,
-          providerMetadata: item.id ? providerMetadata(state, { itemId: item.id }) : undefined,
+          providerMetadata: item.id !== undefined ? providerMetadata(state, { itemId: item.id }) : undefined,
         })
     const result =
       item.arguments === undefined
@@ -1136,7 +1137,7 @@ const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (
           ([current, events], item) => {
             const id = item.id ?? (item.type === "function_call" ? item.call_id : undefined)
             if (
-              !id ||
+              id === undefined ||
               ((item.type !== "function_call" || !current.tools[id]) &&
                 (item.type !== "reasoning" || !current.reasoningItems[id]))
             )
@@ -1219,12 +1220,13 @@ export const providerFailure = (id: string, event: Event, fallback: string) => {
 const providerError = (state: ParserState, event: Event, fallback: string) => providerFailure(state.id, event, fallback)
 
 export const step = (state: ParserState, input: Event) => {
+  // The OpenAPI requires string IDs but imposes no minLength; empty is not missing.
   const event =
-    input.item_id && outputItemID(state, input) !== input.item_id
+    input.item_id !== undefined && outputItemID(state, input) !== input.item_id
       ? { ...input, item_id: outputItemID(state, input) }
       : input
   if (event.type === "response.output_text.delta" || event.type === "response.output_text.done") {
-    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
+    if (event.item_id === undefined) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
     return Effect.succeed(
       event.type === "response.output_text.delta"
         ? onOutputTextDelta(state, event, event.item_id)
@@ -1233,7 +1235,7 @@ export const step = (state: ParserState, input: Event) => {
   }
   if (event.type === "response.refusal.delta" || event.type === "response.refusal.done") {
     const value = event.type === "response.refusal.delta" ? event.delta : event.refusal
-    if (!event.item_id || typeof value !== "string")
+    if (event.item_id === undefined || typeof value !== "string")
       return ProviderShared.eventError(state.id, `${event.type} is malformed`)
     return Effect.succeed(
       event.type === "response.refusal.delta"
@@ -1242,7 +1244,7 @@ export const step = (state: ParserState, input: Event) => {
     )
   }
   if (event.type === "response.reasoning.delta" || event.type === "response.reasoning_summary_text.delta") {
-    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
+    if (event.item_id === undefined) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
     return Effect.succeed(onReasoningDelta(state, event, event.item_id))
   }
   if (
@@ -1250,24 +1252,24 @@ export const step = (state: ParserState, input: Event) => {
     event.type === "response.reasoning_summary_text.done" ||
     event.type === "response.reasoning_text.done"
   ) {
-    if (!event.item_id) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
+    if (event.item_id === undefined) return ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
     return Effect.succeed(onReasoningDone(state, event, event.item_id))
   }
   if (event.type === "response.reasoning_summary_part.added")
-    return event.item_id
+    return event.item_id !== undefined
       ? Effect.succeed(onReasoningSummaryPartAdded(state, event))
       : ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
   if (event.type === "response.reasoning_summary_part.done")
-    return event.item_id
+    return event.item_id !== undefined
       ? Effect.succeed(onReasoningSummaryPartDone(state, event))
       : ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
   if (event.type === "response.output_item.added") {
-    if (event.item?.type === "message" && !event.item.id)
+    if (event.item?.type === "message" && event.item.id === undefined)
       return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
     const id = event.item?.id ?? (event.item?.type === "function_call" ? event.item.call_id : undefined)
     return Effect.succeed(
       onOutputItemAdded(
-        event.output_index !== undefined && id
+        event.output_index !== undefined && id !== undefined
           ? { ...state, outputItems: { ...state.outputItems, [event.output_index]: id } }
           : state,
         event,
@@ -1275,11 +1277,11 @@ export const step = (state: ParserState, input: Event) => {
     )
   }
   if (event.type === "response.function_call_arguments.delta" || event.type === "response.function_call_arguments.done")
-    return event.item_id
+    return event.item_id !== undefined
       ? onFunctionCallArgumentsDelta(state, event)
       : ProviderShared.eventError(state.id, `${event.type} is missing item_id`)
   if (event.type === "response.output_item.done") {
-    if (event.item?.type === "message" && !event.item.id)
+    if (event.item?.type === "message" && event.item.id === undefined)
       return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
     return onOutputItemDone(state, event)
   }

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

@@ -496,7 +496,11 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
   const lowering = {
     ...options,
     toolCallID: (id: string) => {
-      if (mistral) return id.replace(/[^a-zA-Z0-9]/g, "").slice(0, 9).padEnd(9, "0")
+      if (mistral)
+        return id
+          .replace(/[^a-zA-Z0-9]/g, "")
+          .slice(0, 9)
+          .padEnd(9, "0")
       if (modelID.includes("claude")) return id.replace(/[^a-zA-Z0-9_-]/g, "_")
       if (request.model.provider === "openai" || request.model.provider === "azure" || modelID.startsWith("openai/"))
         return id.slice(0, 40)
@@ -505,7 +509,8 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
   }
   const requireAssistantAfterTool = request.model.compatibility?.requireAssistantAfterTool ?? mistral
   const bridgeTools = () => {
-    if (requireAssistantAfterTool && messages.at(-1)?.role === "tool") messages.push({ role: "assistant", content: "Done." })
+    if (requireAssistantAfterTool && messages.at(-1)?.role === "tool")
+      messages.push({ role: "assistant", content: "Done." })
   }
   const pendingImages: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
   const flushImages = () => {
@@ -557,7 +562,10 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
         )
       continue
     }
-    if (message.role === "assistant" && message.content.every((part) => part.type === "text" && part.text.trim() === ""))
+    if (
+      message.role === "assistant" &&
+      message.content.every((part) => part.type === "text" && part.text.trim() === "")
+    )
       continue
     if (message.role === "tool") {
       const lowered = yield* lowerToolMessages(message, lowering)
@@ -588,7 +596,10 @@ const hasToolHistory = (messages: ReadonlyArray<LLMRequest["messages"][number]>)
 // models.dev provider naming: DeepSeek, Moonshot AI, Together AI, ZAI
 // (Zhipu + Coding Plan variants), Nvidia, Cerebras, Chutes, etc. still
 // require `max_tokens`.
-const detectMaxTokensField = (provider: string, baseURL: string | undefined): "max_tokens" | "max_completion_tokens" => {
+const detectMaxTokensField = (
+  provider: string,
+  baseURL: string | undefined,
+): "max_tokens" | "max_completion_tokens" => {
   const p = provider.toLowerCase()
   const url = (baseURL ?? "").toLowerCase()
   if (
@@ -638,7 +649,8 @@ const detectSupportsStore = (provider: string, baseURL: string | undefined): boo
   const isChutes = p === "chutes" || url.includes("chutes.ai")
   const isCloudflareWorkersAI = p === "cloudflare-workers-ai" || url.includes("api.cloudflare.com")
   const isCloudflareAiGateway = p === "cloudflare-ai-gateway" || url.includes("gateway.ai.cloudflare.com")
-  const isVercelAiGateway = p === "vercel-ai-gateway" || url.includes("ai-gateway.vercel.sh") || url.includes("vercel.sh")
+  const isVercelAiGateway =
+    p === "vercel-ai-gateway" || url.includes("ai-gateway.vercel.sh") || url.includes("vercel.sh")
   const isAntLing = p === "ant-ling" || url.includes("api.ant-ling.com")
   const isOpencode = p === "opencode" || url.includes("opencode.ai")
   const isNonStandard =
@@ -670,11 +682,7 @@ const detectSupportsStrictMode = (provider: string, baseURL: string | undefined)
   return !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia
 }
 
-const detectZaiToolStream = (
-  provider: string,
-  baseURL: string | undefined,
-  modelID: string,
-): boolean => {
+const detectZaiToolStream = (provider: string, baseURL: string | undefined, modelID: string): boolean => {
   const p = provider.toLowerCase()
   const url = (baseURL ?? "").toLowerCase()
   const isZai =
@@ -724,10 +732,10 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
   const supportsStore = request.model.compatibility?.supportsStore ?? detectSupportsStore(provider, baseURL)
   const supportsUsageInStreaming =
     request.model.compatibility?.supportsUsageInStreaming ?? detectSupportsUsageInStreaming()
-  const supportsStrictMode = request.model.compatibility?.supportsStrictMode ?? detectSupportsStrictMode(provider, baseURL)
+  const supportsStrictMode =
+    request.model.compatibility?.supportsStrictMode ?? detectSupportsStrictMode(provider, baseURL)
   const zaiToolStream =
-    request.model.compatibility?.zaiToolStream ??
-    detectZaiToolStream(provider, baseURL, request.model.id)
+    request.model.compatibility?.zaiToolStream ?? detectZaiToolStream(provider, baseURL, request.model.id)
   const hasHistory = hasToolHistory(request.messages)
   const hasActiveTools = request.tools.length > 0
   return {
@@ -816,11 +824,10 @@ const mapUsage = (usage: OpenAIChatEvent["usage"]): 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 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 reasoning = usage.completion_tokens_details?.reasoning_tokens ?? undefined
   const nonCached = ProviderShared.subtractTokens(input, ProviderShared.sumTokens(cached, cacheWrite))
@@ -936,13 +943,12 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
     const choiceUsage = (choice as unknown as { usage?: OpenAIChatEvent["usage"] })?.usage
     const usage = mapUsage(event.usage) ?? (choiceUsage ? mapUsage(choiceUsage) : undefined) ?? state.usage
     const rawFinishReason = choice?.finish_reason
-    const finishReason =
-      rawFinishReason
-        ? {
-            normalized: yield* mapFinishReason(event, rawFinishReason),
-            raw: choice?.native_finish_reason ?? rawFinishReason,
-          }
-        : state.finishReason
+    const finishReason = rawFinishReason
+      ? {
+          normalized: yield* mapFinishReason(event, rawFinishReason),
+          raw: choice?.native_finish_reason ?? rawFinishReason,
+        }
+      : state.finishReason
     const delta = choice?.delta
     const toolDeltas = delta?.tool_calls ?? []
     let tools = state.tools

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

@@ -184,7 +184,7 @@ const HOSTED_TOOLS = {
 
 const step = (state: OpenResponses.ParserState, event: OpenResponses.Event) => {
   if (event.type === "response.reasoning_text.delta")
-    return event.item_id
+    return event.item_id !== undefined
       ? Effect.succeed(
           OpenResponses.onReasoningDelta(state, event, OpenResponses.outputItemID(state, event) ?? event.item_id),
         )

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

@@ -29,10 +29,9 @@ export type ResponseIncludable = (typeof ResponseIncludables)[number] | (string
 
 export const ServiceTiers = ["auto", "default", "flex", "priority"] as const
 export type ServiceTier = (typeof ServiceTiers)[number] | (string & {})
-export const ServiceTier = Schema.declare<ServiceTier>(
-  (value): value is ServiceTier => typeof value === "string",
-  { title: "ServiceTier" },
-)
+export const ServiceTier = Schema.declare<ServiceTier>((value): value is ServiceTier => typeof value === "string", {
+  title: "ServiceTier",
+})
 
 export const Truncations = ["auto", "disabled"] as const
 export type Truncation = (typeof Truncations)[number]

packages/ai/src/protocols/utils/responses-hosted-tools.ts

@@ -34,37 +34,35 @@ export const onDone: (
   state: OpenResponses.ParserState,
   item: Item,
   tools: Definitions,
-) => Effect.Effect<OpenResponses.StepResult, AIError> = Effect.fn("ResponsesHostedTools.onDone")(function* (
-  state,
-  item,
-  tools,
-) {
-  const tool = tools[item.type]
-  if (!tool) return [state, []] satisfies OpenResponses.StepResult
-  const providerMetadata = OpenResponses.providerMetadata(state, { itemId: item.id })
-  const events: LLMEvent[] = []
-  const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
-  events.push(
-    LLMEvent.toolCall({
-      id: item.id,
-      name: tool.name,
-      input: tool.input(item),
-      providerExecuted: true,
-      providerMetadata,
-    }),
-    LLMEvent.toolResult({
-      id: item.id,
-      name: tool.name,
-      result: tool.result
-        ? yield* tool.result(item)
-        : item.error !== undefined && item.error !== null
-          ? { type: "error", value: item.error }
-          : { type: "json", value: item },
-      providerExecuted: true,
-      providerMetadata,
-    }),
-  )
-  return [{ ...state, lifecycle }, events] satisfies OpenResponses.StepResult
-})
+) => Effect.Effect<OpenResponses.StepResult, AIError> = Effect.fn("ResponsesHostedTools.onDone")(
+  function* (state, item, tools) {
+    const tool = tools[item.type]
+    if (!tool) return [state, []] satisfies OpenResponses.StepResult
+    const providerMetadata = OpenResponses.providerMetadata(state, { itemId: item.id })
+    const events: LLMEvent[] = []
+    const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
+    events.push(
+      LLMEvent.toolCall({
+        id: item.id,
+        name: tool.name,
+        input: tool.input(item),
+        providerExecuted: true,
+        providerMetadata,
+      }),
+      LLMEvent.toolResult({
+        id: item.id,
+        name: tool.name,
+        result: tool.result
+          ? yield* tool.result(item)
+          : item.error !== undefined && item.error !== null
+            ? { type: "error", value: item.error }
+            : { type: "json", value: item },
+        providerExecuted: true,
+        providerMetadata,
+      }),
+    )
+    return [{ ...state, lifecycle }, events] satisfies OpenResponses.StepResult
+  },
+)
 
 export * as ResponsesHostedTools from "./responses-hosted-tools.js"

packages/ai/src/providers/groq.ts

@@ -0,0 +1,116 @@
+import { Effect, Schema } from "effect"
+import type { ProviderPackage } from "../provider-package.js"
+import { OpenAIChat } from "../protocols/openai-chat.js"
+import { ProviderShared } from "../protocols/shared.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 { Framing } from "../route/framing.js"
+import { Protocol } from "../route/protocol.js"
+import { ProviderID, type ModelID, type LLMRequest } from "../schema/index.js"
+import { profiles } from "./openai-compatible-profile.js"
+import type { OpenAIProviderOptionsInput } from "./openai-options.js"
+
+export const id = ProviderID.make("groq")
+
+export type ProviderOptions = Pick<OpenAIProviderOptionsInput, "reasoningEffort"> & {
+  /** Controls visible reasoning on GPT-OSS; other models always use parsed reasoning. */
+  readonly includeReasoning?: boolean
+  readonly parallelToolCalls?: boolean
+  readonly serviceTier?: "on_demand" | "flex" | "auto" | "performance" | (string & {})
+  readonly user?: string
+}
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: ProviderOptions
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: ProviderOptions
+}
+
+const Options = Schema.Struct({
+  includeReasoning: Schema.optional(Schema.Boolean),
+  parallelToolCalls: Schema.optional(Schema.Boolean),
+  serviceTier: Schema.optional(Schema.String),
+  user: Schema.optional(Schema.String),
+})
+
+export const protocol = Protocol.make({
+  id: "groq-chat",
+  body: {
+    schema: Schema.Struct({
+      ...OpenAIChat.bodyFields,
+      reasoning_format: Schema.optional(Schema.Literal("parsed")),
+      include_reasoning: Schema.optional(Schema.Boolean),
+      parallel_tool_calls: Schema.optional(Schema.Boolean),
+      service_tier: Schema.optional(Schema.String),
+      user: Schema.optional(Schema.String),
+    }),
+    from: Effect.fn("Groq.fromRequest")(function* (request: LLMRequest) {
+      const options = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Options))(
+        request.providerOptions ?? {},
+      )
+      const gptOSS = request.model.id.startsWith("openai/gpt-oss-")
+      return {
+        ...(yield* OpenAIChat.fromRequest(request)),
+        reasoning_format: gptOSS ? undefined : ("parsed" as const),
+        include_reasoning: gptOSS ? options.includeReasoning : undefined,
+        parallel_tool_calls: options.parallelToolCalls,
+        service_tier: options.serviceTier,
+        user: options.user,
+      }
+    }),
+  },
+  stream: OpenAIChat.protocol.stream,
+})
+
+export const route = Route.make({
+  id: "groq-chat",
+  provider: id,
+  providerMetadataKey: "openai",
+  protocol,
+  endpoint: Endpoint.path("/chat/completions", { baseURL: profiles.groq.baseURL }),
+  framing: Framing.sse,
+})
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: baseURL ?? profiles.groq.baseURL },
+    auth: AuthOptions.bearer(input, "GROQ_API_KEY"),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) =>
+      configured.model<ProviderOptions>({
+        id: modelID,
+        compatibility: {
+          maxTokensField: "max_completion_tokens",
+          reasoningField: "reasoning",
+          requireReasoning: false,
+          supportsStore: false,
+          supportsStrictMode: false,
+        },
+      }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, ProviderOptions>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)
+
+export * as Groq from "./groq.js"

packages/ai/src/providers/index.ts

@@ -12,6 +12,7 @@ export * as GoogleVertex from "./google-vertex.js"
 export * as GoogleVertexChat from "./google-vertex-chat.js"
 export * as GoogleVertexMessages from "./google-vertex-messages.js"
 export * as GoogleVertexResponses from "./google-vertex-responses.js"
+export * as Groq from "./groq.js"
 export * as OpenAI from "./openai.js"
 export * as OpenAICompatible from "./openai-compatible.js"
 export * as OpenAICompatibleResponses from "./openai-compatible-responses.js"

packages/ai/src/route/client.ts

@@ -339,9 +339,7 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
                 return onHalt
                   ? parsed.pipe(
                       Stream.concat(
-                        Stream.suspend(() =>
-                          Stream.unwrap(onHalt(state).pipe(Effect.map(Stream.fromIterable))),
-                        ),
+                        Stream.suspend(() => Stream.unwrap(onHalt(state).pipe(Effect.map(Stream.fromIterable)))),
                       ),
                     )
                   : parsed

packages/ai/test/fixtures/recordings/azure/chat-streams-text.json

@@ -1,10 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
     "name": "azure/chat-streams-text",
     "recordedAt": "2026-08-23T17:21:53.198Z"
   },

packages/ai/test/fixtures/recordings/azure/responses-calls-a-tool.json

@@ -1,10 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
     "name": "azure/responses-calls-a-tool",
     "recordedAt": "2026-08-23T17:21:55.170Z"
   },

packages/ai/test/fixtures/recordings/azure/responses-continues-after-a-tool-result.json

@@ -1,10 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
     "name": "azure/responses-continues-after-a-tool-result",
     "recordedAt": "2026-08-23T17:21:56.397Z"
   },

packages/ai/test/fixtures/recordings/azure/responses-streams-text.json

@@ -1,10 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:azure",
-      "provider:azure"
-    ],
+    "tags": ["prefix:azure", "provider:azure"],
     "name": "azure/responses-streams-text",
     "recordedAt": "2026-08-23T17:21:54.158Z"
   },

packages/ai/test/fixtures/recordings/bedrock-mantle/streams-text.json

@@ -2,11 +2,7 @@
   "version": 1,
   "metadata": {
     "model": "openai.gpt-oss-120b",
-    "tags": [
-      "prefix:bedrock-mantle",
-      "provider:amazon-bedrock",
-      "protocol:openai-responses"
-    ],
+    "tags": ["prefix:bedrock-mantle", "provider:amazon-bedrock", "protocol:openai-responses"],
     "name": "bedrock-mantle/streams-text",
     "recordedAt": "2026-08-25T03:29:02.968Z"
   },

packages/ai/test/fixtures/recordings/cloudflare-workers-ai/cloudflare-workers-ai-gpt-oss-20b-tools-tool-call.json

@@ -7,13 +7,7 @@
     "route": "cloudflare-workers-ai",
     "transport": "http",
     "model": "@cf/openai/gpt-oss-20b",
-    "tags": [
-      "prefix:cloudflare-workers-ai",
-      "provider:cloudflare-workers-ai",
-      "tool",
-      "tool-call",
-      "golden"
-    ]
+    "tags": ["prefix:cloudflare-workers-ai", "provider:cloudflare-workers-ai", "tool", "tool-call", "golden"]
   },
   "interactions": [
     {
@@ -35,4 +29,4 @@
       }
     }
   ]
-}
+}

packages/ai/test/fixtures/recordings/google-vertex/calls-a-tool.json

@@ -1,11 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
     "name": "google-vertex/calls-a-tool",
     "recordedAt": "2026-08-23T17:21:51.036Z"
   },

packages/ai/test/fixtures/recordings/google-vertex/continues-after-a-tool-result.json

@@ -1,11 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
     "name": "google-vertex/continues-after-a-tool-result",
     "recordedAt": "2026-08-23T17:21:51.853Z"
   },

packages/ai/test/fixtures/recordings/google-vertex/streams-text.json

@@ -1,11 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:google-vertex",
-      "provider:google-vertex",
-      "protocol:gemini"
-    ],
+    "tags": ["prefix:google-vertex", "provider:google-vertex", "protocol:gemini"],
     "name": "google-vertex/streams-text",
     "recordedAt": "2026-08-23T17:21:50.112Z"
   },

packages/ai/test/fixtures/recordings/groq-chat/streams-text-with-usage.json

@@ -0,0 +1,29 @@
+{
+  "version": 1,
+  "metadata": {
+    "model": "openai/gpt-oss-20b",
+    "tags": ["prefix:groq-chat", "provider:groq", "protocol:groq-chat", "text", "usage"],
+    "name": "groq-chat/streams-text-with-usage",
+    "recordedAt": "2026-08-26T14:40:09.833Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.groq.com/openai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"openai/gpt-oss-20b\",\"messages\":[{\"role\":\"user\",\"content\":\"Reply with exactly one word: hello\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"reasoning_effort\":\"low\",\"max_completion_tokens\":512,\"include_reasoning\":false,\"service_tier\":\"on_demand\",\"user\":\"recorded-test\"}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream"
+        },
+        "body": "data: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\"},\"logprobs\":null,\"finish_reason\":null}],\"x_groq\":{\"id\":\"req_01m0z87915eep9bpf10gg7331e\",\"seed\":94036161}}\n\ndata: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"hello\"},\"logprobs\":null,\"finish_reason\":null}]}\n\ndata: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[{\"index\":0,\"delta\":{},\"logprobs\":null,\"finish_reason\":\"stop\"}],\"x_groq\":{\"id\":\"req_01m0z87915eep9bpf10gg7331e\",\"usage\":{\"queue_time\":0.10886435,\"prompt_tokens\":78,\"prompt_time\":0.003693734,\"completion_tokens\":20,\"completion_time\":0.020459983,\"total_tokens\":98,\"total_time\":0.024153717,\"completion_tokens_details\":{\"reasoning_tokens\":10}}},\"usage\":{\"queue_time\":0.10886435,\"prompt_tokens\":78,\"prompt_time\":0.003693734,\"completion_tokens\":20,\"completion_time\":0.020459983,\"total_tokens\":98,\"total_time\":0.024153717,\"completion_tokens_details\":{\"reasoning_tokens\":10}}}\n\ndata: {\"id\":\"chatcmpl-660048b0-3c99-4215-a582-e116b77eb881\",\"object\":\"chat.completion.chunk\",\"created\":1787755209,\"model\":\"openai/gpt-oss-20b\",\"system_fingerprint\":\"fp_66891002f6\",\"choices\":[],\"usage\":{\"queue_time\":0.10886435,\"prompt_tokens\":78,\"prompt_time\":0.003693734,\"completion_tokens\":20,\"completion_time\":0.020459983,\"total_tokens\":98,\"total_time\":0.024153717,\"completion_tokens_details\":{\"reasoning_tokens\":10}},\"service_tier\":\"on_demand\"}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}

packages/ai/test/fixtures/recordings/openai-responses-cache/reports-cached-tokens-on-identical-second-call.json

@@ -1,12 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:openai-responses-cache",
-      "provider:openai",
-      "protocol:openai-responses",
-      "cache"
-    ],
+    "tags": ["prefix:openai-responses-cache", "provider:openai", "protocol:openai-responses", "cache"],
     "name": "openai-responses-cache/reports-cached-tokens-on-identical-second-call",
     "recordedAt": "2026-08-25T03:29:25.124Z"
   },

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

@@ -5,14 +5,7 @@
     "route": "openai-responses",
     "transport": "http",
     "model": "gpt-5.5",
-    "tags": [
-      "prefix:openai-responses",
-      "provider:openai",
-      "flagship",
-      "tool",
-      "tool-loop",
-      "golden"
-    ],
+    "tags": ["prefix:openai-responses", "provider:openai", "flagship", "tool", "tool-loop", "golden"],
     "name": "openai-responses/openai-responses-gpt-5-5-tool-loop",
     "recordedAt": "2026-08-20T06:30:22.262Z"
   },

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

@@ -1,14 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:openai",
-      "protocol:openai-responses",
-      "tool",
-      "tool-result"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:openai", "protocol:openai-responses", "tool", "tool-result"],
     "name": "pdf/openai-tool-result",
     "recordedAt": "2026-08-25T03:29:08.297Z"
   },

packages/ai/test/fixtures/recordings/pdf/openai-user-input.json

@@ -1,13 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:openai",
-      "protocol:openai-responses",
-      "user-input"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:openai", "protocol:openai-responses", "user-input"],
     "name": "pdf/openai-user-input",
     "recordedAt": "2026-08-25T03:29:05.645Z"
   },

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

@@ -1,14 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:xai",
-      "protocol:xai-responses",
-      "tool",
-      "tool-result"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:xai", "protocol:xai-responses", "tool", "tool-result"],
     "name": "pdf/xai-tool-result",
     "recordedAt": "2026-08-25T03:29:11.774Z"
   },

packages/ai/test/fixtures/recordings/pdf/xai-user-input.json

@@ -1,13 +1,7 @@
 {
   "version": 1,
   "metadata": {
-    "tags": [
-      "prefix:pdf",
-      "pdf",
-      "provider:xai",
-      "protocol:xai-responses",
-      "user-input"
-    ],
+    "tags": ["prefix:pdf", "pdf", "provider:xai", "protocol:xai-responses", "user-input"],
     "name": "pdf/xai-user-input",
     "recordedAt": "2026-08-25T03:29:10.612Z"
   },

packages/ai/test/fixtures/recordings/vercel-ai-gateway-reasoning-tool-loop.json

@@ -52,4 +52,4 @@
       }
     }
   ]
-}
+}

packages/ai/test/fixtures/recordings/vercel-ai-gateway-reasoning.json

@@ -2,12 +2,7 @@
   "version": 1,
   "metadata": {
     "model": "anthropic/claude-sonnet-4.6",
-    "tags": [
-      "prefix:openai-compatible-chat",
-      "provider:vercel-ai-gateway",
-      "protocol:openai-chat",
-      "reasoning"
-    ],
+    "tags": ["prefix:openai-compatible-chat", "provider:vercel-ai-gateway", "protocol:openai-chat", "reasoning"],
     "name": "vercel-ai-gateway-reasoning",
     "recordedAt": "2026-07-18T11:28:42.077Z"
   },
@@ -31,4 +26,4 @@
       }
     }
   ]
-}
+}

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

@@ -89,9 +89,7 @@ describe("provider error classification", () => {
 
   test("classifies network error text as provider internal", () => {
     expect(
-      ["network error", "network-error", "network_error"].map(
-        (message) => classifyProviderFailure({ message })._tag,
-      ),
+      ["network error", "network-error", "network_error"].map((message) => classifyProviderFailure({ message })._tag),
     ).toEqual(["ProviderInternal", "ProviderInternal", "ProviderInternal"])
   })
 

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

@@ -29,6 +29,7 @@ describe("provider package entrypoints", () => {
       import("@opencode-ai/ai/providers/togetherai"),
       import("@opencode-ai/ai/providers/cerebras"),
       import("@opencode-ai/ai/providers/deepinfra"),
+      import("@opencode-ai/ai/providers/groq"),
     ])
 
     for (const module of modules) expect(module.model).toBeFunction()

packages/ai/test/provider/gemini.test.ts

@@ -515,7 +515,10 @@ describe("Gemini route", () => {
         {
           role: "model",
           parts: [
-            { functionCall: { id: "call_image", name: "read", args: { path: "pixel.png" } }, thoughtSignature: "sig_1" },
+            {
+              functionCall: { id: "call_image", name: "read", args: { path: "pixel.png" } },
+              thoughtSignature: "sig_1",
+            },
           ],
         },
         {
@@ -606,10 +609,7 @@ describe("Gemini route", () => {
       expect(prepared.body.contents).toEqual([
         {
           role: "model",
-          parts: [
-            { functionCall: { name: "shot", args: {} } },
-            { functionCall: { name: "shot", args: {} } },
-          ],
+          parts: [{ functionCall: { name: "shot", args: {} } }, { functionCall: { name: "shot", args: {} } }],
         },
         {
           role: "user",
@@ -1071,7 +1071,9 @@ describe("Gemini route", () => {
       const prepared = yield* compileRequest(
         LLM.request({
           model,
-          messages: [Message.assistant([{ type: "text", text: "All done.", providerMetadata: delta?.providerMetadata }])],
+          messages: [
+            Message.assistant([{ type: "text", text: "All done.", providerMetadata: delta?.providerMetadata }]),
+          ],
         }),
       )
       expect(prepared.body.contents).toEqual([
@@ -1572,9 +1574,7 @@ describe("Gemini route", () => {
               { candidates: [{ content: { role: "model", parts: null } }] },
               { candidates: [{ content: null, finishReason: null }] },
               {
-                candidates: [
-                  { content: { role: "model", parts: [{ text: "Hello" }] }, finishReason: "STOP" as const },
-                ],
+                candidates: [{ content: { role: "model", parts: [{ text: "Hello" }] }, finishReason: "STOP" as const }],
               },
             ),
           ),

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

@@ -26,9 +26,7 @@ const recorded = recordedTests({
 describe("Google Vertex Gemini recorded", () => {
   recorded.effect("streams text", () =>
     Effect.gen(function* () {
-      const response = yield* LLMClient.generate(
-        LLM.request({ model, prompt: "Reply with exactly one word: hello" }),
-      )
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Reply with exactly one word: hello" }))
 
       expect(response.text.toLowerCase()).toContain("hello")
     }),

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

@@ -0,0 +1,185 @@
+import { configure } from "@opencode-ai/ai/providers/groq"
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, LLMEvent, LLMRequest, LLMResponse, Message, ToolChoice, ToolDefinition } from "../../src/index.js"
+import { LLMClient } from "../../src/route.js"
+import { compileRequest } from "../../src/route/client.js"
+import { recordedTests } from "../recorded-test.js"
+
+const apiKey = process.env.GROQ_API_KEY ?? "fixture"
+const recorded = recordedTests({
+  prefix: "groq-chat",
+  provider: "groq",
+  protocol: "groq-chat",
+  requires: ["GROQ_API_KEY"],
+})
+
+const weather = ToolDefinition.make({
+  name: "lookup_weather",
+  description: "Look up the current weather for a city",
+  inputSchema: {
+    type: "object",
+    properties: { city: { type: "string", enum: ["Paris", "London"] } },
+    required: ["city"],
+    additionalProperties: false,
+  },
+})
+
+describe("Groq recorded", () => {
+  recorded.effect.with(
+    "streams text with usage",
+    { tags: ["text", "usage"], metadata: { model: "openai/gpt-oss-20b" } },
+    () =>
+      Effect.gen(function* () {
+        const request = LLM.request({
+          model: configure({
+            apiKey,
+            providerOptions: {
+              includeReasoning: false,
+              reasoningEffort: "low",
+              serviceTier: "on_demand",
+              user: "recorded-test",
+            },
+          }).model("openai/gpt-oss-20b"),
+          prompt: "Reply with exactly one word: hello",
+          generation: { maxTokens: 512 },
+        })
+        const compiled = yield* compileRequest(request)
+        expect(compiled.body).toMatchObject({
+          max_completion_tokens: 512,
+          stream_options: { include_usage: true },
+          include_reasoning: false,
+          service_tier: "on_demand",
+          user: "recorded-test",
+        })
+        expect(compiled.body.max_tokens).toBeUndefined()
+        expect(compiled.body.store).toBeUndefined()
+        expect(compiled.body.reasoning_format).toBeUndefined()
+
+        const response = yield* LLMClient.generate(request)
+        expect(response.text.toLowerCase().trim()).toBe("hello")
+        expect(response.reasoning).toBe("")
+        expect(response.events.some(LLMEvent.is.textDelta)).toBe(true)
+        expectUsage(response)
+      }),
+    60_000,
+  )
+
+  for (const item of [
+    {
+      name: "continues Qwen parallel tool calls",
+      model: configure({ apiKey, providerOptions: { parallelToolCalls: true, reasoningEffort: "none" } }).model(
+        "qwen/qwen3.6-27b",
+      ),
+      cities: ["Paris", "London"],
+      reasoning: false,
+    },
+    {
+      name: "replays GPT OSS reasoning through a tool loop",
+      model: configure({ apiKey, providerOptions: { includeReasoning: true, reasoningEffort: "low" } }).model(
+        "openai/gpt-oss-20b",
+      ),
+      cities: ["Paris"],
+      reasoning: true,
+    },
+  ]) {
+    recorded.effect.with(
+      item.name,
+      {
+        tags: ["tool", "tool-loop", "usage", item.reasoning ? "reasoning" : "parallel"],
+        metadata: { model: item.model.id },
+      },
+      () =>
+        Effect.gen(function* () {
+          const request = LLM.request({
+            model: item.model,
+            prompt: `Look up the current weather in ${item.cities.join(" and ")}. Call lookup_weather once for each city in the same response before answering. After receiving all results, report each city's weather in one short sentence.`,
+            tools: [weather],
+            toolChoice: "required",
+            generation: { maxTokens: 1536 },
+          })
+          const compiled = yield* compileRequest(request)
+          expect(compiled.body.stream_options).toEqual({ include_usage: true })
+          expect(compiled.body.store).toBeUndefined()
+          expect(compiled.body.reasoning_format).toBe(item.reasoning ? undefined : "parsed")
+          expect(compiled.body.tools[0].function.strict).toBeUndefined()
+          if (!item.reasoning) expect(compiled.body.parallel_tool_calls).toBe(true)
+
+          const first = yield* LLMClient.generate(request)
+          expect(first.finishReason.normalized).toBe("tool-calls")
+          expect(first.toolCalls).toHaveLength(item.cities.length)
+          expect(new Set(first.toolCalls.map((call) => call.id)).size).toBe(item.cities.length)
+          expect(first.toolCalls.map((call) => call.input)).toEqual(
+            expect.arrayContaining(item.cities.map((city) => ({ city }))),
+          )
+          expect(first.toolCalls.every((call) => call.name === "lookup_weather")).toBe(true)
+          expectUsage(first)
+          if (item.reasoning) {
+            expect(first.reasoning.length).toBeGreaterThan(0)
+            expect(first.events.some(LLMEvent.is.reasoningDelta)).toBe(true)
+          }
+
+          const followUp = LLMRequest.update(request, {
+            toolChoice: ToolChoice.make("none"),
+            messages: [
+              ...request.messages,
+              first.message,
+              ...first.toolCalls.map((call) =>
+                Message.tool({ id: call.id, name: call.name, result: { condition: "sunny", temperature: "18C" } }),
+              ),
+            ],
+          })
+          const replay = yield* compileRequest(followUp)
+          if (item.reasoning) {
+            expect(replay.body.messages).toEqual(
+              expect.arrayContaining([expect.objectContaining({ role: "assistant", reasoning: first.reasoning })]),
+            )
+          }
+          expect(replay.body.reasoning_format).toBe(item.reasoning ? undefined : "parsed")
+
+          const second = yield* LLMClient.generate(followUp)
+          expect(second.finishReason.normalized).toBe("stop")
+          expect(second.toolCalls).toHaveLength(0)
+          expect(second.text.toLowerCase()).toContain("sunny")
+          item.cities.forEach((city) => expect(second.text).toContain(city))
+          expectUsage(second)
+        }),
+      60_000,
+    )
+  }
+
+  recorded.effect.with(
+    "streams Qwen parsed reasoning",
+    { tags: ["reasoning", "usage"], metadata: { model: "qwen/qwen3.6-27b" } },
+    () =>
+      Effect.gen(function* () {
+        const request = LLM.request({
+          model: configure({
+            apiKey,
+            providerOptions: { reasoningEffort: "default" },
+          }).model("qwen/qwen3.6-27b"),
+          prompt:
+            "What is 173 multiplied by 219? Think through the arithmetic, then reply with only the final integer.",
+          generation: { maxTokens: 2048 },
+        })
+        const compiled = yield* compileRequest(request)
+        expect(compiled.body).toMatchObject({ reasoning_format: "parsed", reasoning_effort: "default" })
+        expect(compiled.body.include_reasoning).toBeUndefined()
+
+        const response = yield* LLMClient.generate(request)
+        expect(response.text.replaceAll(",", "").trim()).toBe("37887")
+        expect(response.text).not.toContain("<think>")
+        expect(response.reasoning.length).toBeGreaterThan(0)
+        expect(response.events.some(LLMEvent.is.reasoningDelta)).toBe(true)
+        expectUsage(response)
+      }),
+    60_000,
+  )
+})
+
+function expectUsage(response: LLMResponse) {
+  expect(response.usage).toBeDefined()
+  expect(response.usage?.inputTokens).toBeGreaterThan(0)
+  expect(response.usage?.outputTokens).toBeGreaterThan(0)
+  expect(response.events.filter(LLMEvent.is.finish)).toHaveLength(1)
+}

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

@@ -0,0 +1,112 @@
+import { expect } from "bun:test"
+import { Effect } from "effect"
+import { LanguageModel, LLM, Message } from "../../src/index.js"
+import { OpenAIChat } from "../../src/protocols/openai-chat.js"
+import { Groq } from "../../src/providers/groq.js"
+import { compileRequest } from "../../src/route/client.js"
+import { it } from "../lib/effect.js"
+import { weatherTool } from "../recorded-scenarios.js"
+
+it.effect("Groq reuses Chat streaming and defaults to parsed reasoning", () =>
+  Effect.gen(function* () {
+    expect(Groq.protocol.stream).toBe(OpenAIChat.protocol.stream)
+    const model = Groq.configure({ apiKey: "fixture" }).model("llama-3.3-70b-versatile")
+    expect(model.route.endpoint.baseURL).toBe("https://api.groq.com/openai/v1")
+    const compiled = yield* compileRequest(
+      LLM.request({ model, prompt: "Hello", tools: [weatherTool], generation: { maxTokens: 64 } }),
+    )
+    expect(compiled.body).toMatchObject({
+      max_completion_tokens: 64,
+      stream_options: { include_usage: true },
+      reasoning_format: "parsed",
+    })
+    for (const key of ["store", "max_tokens", "include_reasoning", "parallel_tool_calls", "service_tier", "user"])
+      expect(compiled.body[key]).toBeUndefined()
+    expect(compiled.body.tools?.[0]?.function).not.toHaveProperty("strict")
+  }),
+)
+
+it.effect("Groq lowers its own options for custom catalog identities and endpoints", () =>
+  Effect.gen(function* () {
+    const model = LanguageModel.update(
+      Groq.model("qwen/qwen3.6-27b", {
+        apiKey: "fixture",
+        baseURL: "https://gateway.example/v1",
+        headers: { "x-client": "test" },
+        body: { custom: "value" },
+        providerOptions: {
+          reasoningEffort: "default",
+          parallelToolCalls: true,
+          serviceTier: "flex",
+          user: "test-user",
+        },
+      }),
+      { provider: "custom-groq" },
+    )
+    const compiled = yield* compileRequest(
+      LLM.request({ model, prompt: "Hello", providerOptions: { parallelToolCalls: false, includeReasoning: false } }),
+    )
+    expect(model.route.endpoint.baseURL).toBe("https://gateway.example/v1")
+    expect(model.route.defaults.headers).toEqual({ "x-client": "test" })
+    expect(model.route.defaults.http?.body).toEqual({ custom: "value" })
+    expect(compiled.body).toMatchObject({
+      reasoning_effort: "default",
+      reasoning_format: "parsed",
+      parallel_tool_calls: false,
+      service_tier: "flex",
+      user: "test-user",
+    })
+    expect(compiled.body.include_reasoning).toBeUndefined()
+    for (const key of ["reasoningFormat", "reasoningEffort", "parallelToolCalls", "serviceTier"])
+      expect(compiled.body).not.toHaveProperty(key)
+  }),
+)
+
+it.effect("Groq replays reasoning only when present and preserves explicit reasoning exclusion", () =>
+  Effect.gen(function* () {
+    const compiled = yield* compileRequest(
+      LLM.request({
+        model: Groq.configure({ apiKey: "fixture" }).model("openai/gpt-oss-20b"),
+        messages: [
+          Message.user("Think"),
+          Message.assistant([
+            { type: "reasoning", text: "Thinking" },
+            { type: "text", text: "Answer" },
+          ]),
+          Message.user("Again"),
+          Message.assistant("Answer only"),
+          Message.user("Continue"),
+        ],
+        providerOptions: { reasoningEffort: "low", includeReasoning: false },
+      }),
+    )
+    expect(compiled.body).toMatchObject({ reasoning_effort: "low", include_reasoning: false })
+    expect(compiled.body.reasoning_format).toBeUndefined()
+    expect(compiled.body.messages[1]).toMatchObject({ reasoning: "Thinking", content: "Answer" })
+    expect(compiled.body.messages[1]).not.toHaveProperty("reasoning_content")
+    expect(compiled.body.messages[3]).not.toHaveProperty("reasoning")
+  }),
+)
+
+it.effect("Groq omits reasoning_format for the GPT-OSS family by default", () =>
+  Effect.gen(function* () {
+    for (const id of ["openai/gpt-oss-20b", "openai/gpt-oss-120b", "openai/gpt-oss-safeguard-20b"]) {
+      const compiled = yield* compileRequest(
+        LLM.request({ model: Groq.configure({ apiKey: "fixture" }).model(id), prompt: "Hello" }),
+      )
+      expect(compiled.body.reasoning_format).toBeUndefined()
+      expect(compiled.body.include_reasoning).toBeUndefined()
+    }
+  }),
+)
+
+it.effect("Groq validates option types", () =>
+  Effect.gen(function* () {
+    for (const providerOptions of [{ includeReasoning: "false" }, { parallelToolCalls: "false" }]) {
+      const error = yield* compileRequest(
+        LLM.request({ model: Groq.configure({ apiKey: "fixture" }).model("qwen"), prompt: "Hello", providerOptions }),
+      ).pipe(Effect.flip)
+      expect(error.reason._tag).toBe("InvalidRequest")
+    }
+  }),
+)

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

@@ -2,7 +2,7 @@ import { describe, expect } from "bun:test"
 import { ConfigProvider, Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { LLM, Message, ToolDefinition } from "../../src/index.js"
-import { Cerebras, DeepInfra, TogetherAI } from "../../src/providers/index.js"
+import { Cerebras, DeepInfra, Groq, TogetherAI } from "../../src/providers/index.js"
 import { compileRequest } from "../../src/route/client.js"
 import { it } from "../lib/effect.js"
 import { dynamicResponse } from "../lib/http.js"
@@ -155,6 +155,12 @@ describe("native OpenAI-compatible providers", () => {
           token: "deepinfra-secret",
           url: "https://api.deepinfra.com/v1/openai/chat/completions",
         },
+        {
+          model: Groq.configure().model("llama"),
+          env: { GROQ_API_KEY: "groq-secret" },
+          token: "groq-secret",
+          url: "https://api.groq.com/openai/v1/chat/completions",
+        },
       ]
 
       yield* Effect.forEach(scenarios, (scenario) =>

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

@@ -307,6 +307,287 @@ describe("Open Responses-compatible route", () => {
     }),
   )
 
+  describe("stream validation", () => {
+    const request = LLM.request({
+      model: configure({ apiKey: "test-key", baseURL: "https://responses.example.test/v1" }).model("example-model"),
+      prompt: "Respond.",
+    })
+
+    const fixtures = [
+      {
+        item: { type: "message" },
+        events: [
+          { type: "response.output_text.delta", delta: "Preserved" },
+          { type: "response.output_text.done", text: "Preserved" },
+          { type: "response.refusal.delta", delta: "Preserved" },
+          { type: "response.refusal.done", refusal: "Preserved" },
+        ],
+      },
+      {
+        item: { type: "reasoning", encrypted_content: "encrypted-state" },
+        events: [
+          { type: "response.reasoning.delta", delta: "Preserved" },
+          { type: "response.reasoning.done", text: "Preserved" },
+          { type: "response.reasoning_summary_text.delta", delta: "Preserved" },
+          { type: "response.reasoning_summary_text.done", text: "Preserved" },
+          { type: "response.reasoning_text.done", text: "Preserved" },
+        ],
+      },
+      {
+        item: { type: "function_call", call_id: "call_1", name: "lookup" },
+        events: [
+          { type: "response.function_call_arguments.delta", delta: '{"query":"Preserved"}' },
+          { type: "response.function_call_arguments.done", arguments: '{"query":"Preserved"}' },
+        ],
+      },
+    ]
+
+    const routings = [
+      { name: "empty item and event IDs", id: "", item_id: "" },
+      { name: "empty event ID with registered index", id: "item_1", item_id: "", output_index: 2 },
+      { name: "empty stored ID with registered index", id: "", item_id: "wrong_item", output_index: 2 },
+      { name: "empty item and event IDs with registered index", id: "", item_id: "", output_index: 2 },
+    ]
+
+    fixtures.forEach((fixture) => {
+      fixture.events.forEach((event) => {
+        routings.forEach((routing) => {
+          it.effect(`${event.type} preserves content with ${routing.name}`, () =>
+            Effect.gen(function* () {
+              const item = { ...fixture.item, id: routing.id }
+              const response = yield* LLMClient.generate(request).pipe(
+                Effect.provide(
+                  fixedResponse(
+                    sseEvents(
+                      { type: "response.output_item.added", output_index: routing.output_index, item },
+                      { ...event, item_id: routing.item_id, output_index: routing.output_index },
+                      { type: "response.output_item.done", output_index: routing.output_index, item },
+                      { type: "response.completed", response: { id: "resp_1" } },
+                    ),
+                  ),
+                ),
+              )
+
+              const metadata = { openresponses: { itemId: routing.id } }
+              if (fixture.item.type === "function_call") {
+                expect(response.toolCalls).toEqual([
+                  expect.objectContaining({
+                    id: "call_1",
+                    name: "lookup",
+                    input: { query: "Preserved" },
+                    providerMetadata: metadata,
+                  }),
+                ])
+                return
+              }
+              if (fixture.item.type === "reasoning") {
+                expect(response.message.content).toEqual([
+                  {
+                    type: "reasoning",
+                    text: "Preserved",
+                    providerMetadata: {
+                      openresponses: { itemId: routing.id, reasoningEncryptedContent: "encrypted-state" },
+                    },
+                  },
+                ])
+                expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(1)
+                return
+              }
+              expect(response.message.content).toEqual([
+                { type: "text", text: "Preserved", providerMetadata: metadata },
+              ])
+              expect(response.events.filter(LLMEvent.is.textEnd)).toEqual([
+                expect.objectContaining({ id: routing.id, providerMetadata: metadata }),
+              ])
+            }),
+          )
+        })
+      })
+    })
+
+    routings.forEach((routing) => {
+      it.effect(`preserves reasoning summary boundaries and terminal metadata with ${routing.name}`, () =>
+        Effect.gen(function* () {
+          const address = { item_id: routing.item_id, output_index: routing.output_index }
+          const response = yield* LLMClient.generate(request).pipe(
+            Effect.provide(
+              fixedResponse(
+                sseEvents(
+                  {
+                    type: "response.output_item.added",
+                    output_index: routing.output_index,
+                    item: { type: "reasoning", id: routing.id },
+                  },
+                  { type: "response.reasoning_summary_part.added", ...address, summary_index: 0 },
+                  { type: "response.reasoning_summary_text.delta", ...address, summary_index: 0, delta: "First." },
+                  { type: "response.reasoning_summary_text.done", ...address, summary_index: 0, text: "First." },
+                  { type: "response.reasoning_summary_part.done", ...address, summary_index: 0 },
+                  { type: "response.reasoning_summary_part.added", ...address, summary_index: 1 },
+                  { type: "response.reasoning_summary_text.done", ...address, summary_index: 1, text: "Second." },
+                  { type: "response.reasoning_summary_part.done", ...address, summary_index: 1 },
+                  {
+                    type: "response.completed",
+                    response: { output: [{ type: "reasoning", id: routing.id, encrypted_content: "final-state" }] },
+                  },
+                ),
+              ),
+            ),
+          )
+
+          expect(response.message.content).toEqual([
+            {
+              type: "reasoning",
+              text: "First.",
+              providerMetadata: { openresponses: { itemId: routing.id } },
+            },
+            {
+              type: "reasoning",
+              text: "Second.",
+              providerMetadata: { openresponses: { itemId: routing.id, reasoningEncryptedContent: "final-state" } },
+            },
+          ])
+          expect(response.events.filter(LLMEvent.is.reasoningEnd)).toEqual([
+            expect.objectContaining({
+              id: `${routing.id}:0`,
+              providerMetadata: { openresponses: { itemId: routing.id } },
+            }),
+            expect.objectContaining({
+              id: `${routing.id}:1`,
+              providerMetadata: { openresponses: { itemId: routing.id, reasoningEncryptedContent: "final-state" } },
+            }),
+          ])
+        }),
+      )
+    })
+
+    it.effect("reconciles pending empty-ID function arguments from completed output", () =>
+      Effect.gen(function* () {
+        const item = { type: "function_call", id: "", call_id: "call_1", name: "lookup" }
+        const response = yield* LLMClient.generate(request).pipe(
+          Effect.provide(
+            fixedResponse(
+              sseEvents(
+                { type: "response.output_item.added", item },
+                { type: "response.function_call_arguments.delta", item_id: "", delta: '{"query":"partial' },
+                {
+                  type: "response.completed",
+                  response: { output: [{ ...item, arguments: '{"query":"complete"}' }] },
+                },
+              ),
+            ),
+          ),
+        )
+
+        expect(response.toolCalls).toEqual([
+          expect.objectContaining({
+            id: "call_1",
+            name: "lookup",
+            input: { query: "complete" },
+            providerMetadata: { openresponses: { itemId: "" } },
+          }),
+        ])
+      }),
+    )
+
+    it.effect("treats null output items as no-ops without disturbing registered items", () =>
+      Effect.gen(function* () {
+        const response = yield* LLMClient.generate(request).pipe(
+          Effect.provide(
+            fixedResponse(
+              sseEvents(
+                { type: "response.output_item.added", output_index: 0, item: null },
+                { type: "response.output_item.done", output_index: 0, item: null },
+                { type: "response.output_item.added", output_index: 0, item: { type: "message", id: "msg_1" } },
+                { type: "response.output_text.delta", output_index: 0, item_id: "wrong_item", delta: "Before " },
+                { type: "response.output_item.added", output_index: 0, item: null },
+                { type: "response.output_item.done", output_index: 0, item: null },
+                { type: "response.output_text.delta", output_index: 0, item_id: "wrong_item", delta: "after" },
+                { type: "response.output_item.done", output_index: 0, item: { type: "message", id: "msg_1" } },
+                { type: "response.completed", response: { id: "resp_1" } },
+              ),
+            ),
+          ),
+        )
+
+        expect(response.message.content).toEqual([
+          { type: "text", text: "Before after", providerMetadata: { openresponses: { itemId: "msg_1" } } },
+        ])
+        expect(response.events.map((event) => event.type)).toEqual([
+          "step-start",
+          "text-start",
+          "text-delta",
+          "text-delta",
+          "text-end",
+          "step-finish",
+          "finish",
+        ])
+      }),
+    )
+
+    it.effect("rejects missing, null, and non-string event IDs even with a registered output index", () =>
+      Effect.gen(function* () {
+        yield* Effect.forEach(
+          [
+            ...fixtures.flatMap((fixture) => fixture.events.map((event) => ({ item: fixture.item, event }))),
+            ...["response.reasoning_summary_part.added", "response.reasoning_summary_part.done"].map((type) => ({
+              item: { type: "reasoning" },
+              event: { type, summary_index: 0 },
+            })),
+          ],
+          (fixture) =>
+            Effect.forEach([undefined, null, 0, false, {}, []], (item_id) =>
+              Effect.gen(function* () {
+                const error = yield* LLMClient.generate(request).pipe(
+                  Effect.provide(
+                    fixedResponse(
+                      sseEvents(
+                        {
+                          type: "response.output_item.added",
+                          output_index: 0,
+                          item: { ...fixture.item, id: "item_1" },
+                        },
+                        { ...fixture.event, output_index: 0, item_id },
+                        { type: "response.completed", response: { id: "resp_1" } },
+                      ),
+                    ),
+                  ),
+                  Effect.flip,
+                )
+                expect(error.reason._tag).toBe("InvalidProviderOutput")
+              }),
+            ),
+        )
+      }),
+    )
+
+    it.effect("keeps malformed output item IDs invalid", () =>
+      Effect.gen(function* () {
+        yield* Effect.forEach(["response.output_item.added", "response.output_item.done"], (type) =>
+          Effect.forEach(fixtures, (fixture) =>
+            Effect.forEach(
+              fixture.item.type === "message" ? [undefined, null, 0, false, {}, []] : [null, 0, false, {}, []],
+              (id) =>
+                Effect.gen(function* () {
+                  const error = yield* LLMClient.generate(request).pipe(
+                    Effect.provide(
+                      fixedResponse(
+                        sseEvents(
+                          { type, item: { ...fixture.item, id } },
+                          { type: "response.completed", response: { id: "resp_1" } },
+                        ),
+                      ),
+                    ),
+                    Effect.flip,
+                  )
+                  expect(error.reason._tag).toBe("InvalidProviderOutput")
+                }),
+            ),
+          ),
+        )
+      }),
+    )
+  })
+
   it.effect("streams function calls without optional item ids through the shared baseline", () =>
     Effect.gen(function* () {
       const model = configure({

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

@@ -2229,6 +2229,35 @@ describe("OpenAI Responses route", () => {
     }),
   )
 
+  it.effect("accepts empty IDs for native reasoning text deltas", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", output_index: 1, item: { type: "reasoning", id: "" } },
+              { type: "response.reasoning_text.delta", output_index: 1, item_id: "", delta: "Raw" },
+              {
+                type: "response.output_item.done",
+                output_index: 1,
+                item: { type: "reasoning", id: "", encrypted_content: "state" },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([
+        {
+          type: "reasoning",
+          text: "Raw",
+          providerMetadata: { openai: { itemId: "", reasoningEncryptedContent: "state" } },
+        },
+      ])
+    }),
+  )
+
   it.effect("falls back to item ids when an output index was not registered", () =>
     Effect.gen(function* () {
       const response = yield* LLMClient.generate(request).pipe(

packages/ai/test/provider/zai-images.test.ts

@@ -80,7 +80,9 @@ describe("Z.ai Images", () => {
 
   it.effect("sanitizes unpaired surrogates in outbound image requests", () =>
     Image.generate({
-      model: ZAI.configure({ apiKey: "test", http: { body: { metadata: { source: "default\uDC00" } } } }).image("model"),
+      model: ZAI.configure({ apiKey: "test", http: { body: { metadata: { source: "default\uDC00" } } } }).image(
+        "model",
+      ),
       prompt: "A red circle \uD800 on a white background \u{1F600}",
     }).pipe(
       Effect.provide(

packages/app/.gitignore

@@ -1,3 +1,5 @@
 src/assets/theme.css
 e2e/test-results
 e2e/playwright-report
+component-tests/test-results
+component-tests/playwright-report

packages/app/component-tests/composer.spec.ts

@@ -0,0 +1,26 @@
+import { expect, story } from "../../storybook/playwright/story"
+
+// Moved from packages/app/e2e/regression/prompt-thinking-level.spec.ts
+story("shows the thinking level control while relevant", async ({ mount, page }) => {
+  const component = await mount("opencode-composer-flow--model-and-variant")
+  const composer = component.locator('[data-component="composer"]')
+  const input = composer.locator('[data-component="composer-editor"]')
+  const control = composer.getByRole("button", { name: "Choose model variant" })
+
+  await page.mouse.move(0, 0)
+  await page.evaluate(() => (document.activeElement as HTMLElement | null)?.blur())
+  await expect(control).toBeVisible()
+
+  await control.click()
+  const high = page.getByRole("menuitemradio", { name: "high" })
+  await expect(high).toBeVisible()
+  await page.mouse.move(0, 0)
+  await expect(control).toBeVisible()
+  await expect(high).toBeVisible()
+  await high.click()
+
+  await input.focus()
+  await expect(control).toBeVisible()
+  await input.blur()
+  await expect(control).toBeVisible()
+})

packages/app/e2e/regression/review-line-comment.spec.ts

@@ -12,54 +12,6 @@ test.beforeEach(async ({ page }) => {
   await openReview(page)
 })
 
-test("opens the comment editor when code is clicked", async ({ page }) => {
-  const review = page.locator('[data-component="session-review"]')
-  const line = review.getByText("export const value = 'after'", { exact: true })
-  await expectAppVisible(line)
-  await line.click()
-
-  await expect(review.getByRole("textbox")).toBeVisible()
-  await expect(review.locator('[data-slot="line-comment-editor-label"]')).toHaveText("Commenting on line 2")
-})
-
-test("opens the comment editor when a line number is clicked", async ({ page }) => {
-  const review = page.locator('[data-component="session-review"]')
-  const lineNumber = review.locator('[data-column-number="1"]').last()
-  await expectAppVisible(lineNumber)
-  await lineNumber.click()
-
-  await expect(review.getByRole("textbox")).toBeVisible()
-  await expect(review.locator('[data-slot="line-comment-editor-label"]')).toHaveText("Commenting on line 1")
-})
-
-test("opens the comment editor for a line number range", async ({ page }) => {
-  const review = page.locator('[data-component="session-review"]')
-  const start = review.locator('[data-column-number="1"]').last()
-  const end = review.locator('[data-column-number="3"]').last()
-  await expectAppVisible(start)
-  await expectAppVisible(end)
-
-  await start.dragTo(end)
-
-  await expect(review.getByRole("textbox")).toBeVisible()
-  await expect(review.locator('[data-slot="line-comment-editor-label"]')).toHaveText("Commenting on lines 1-3")
-})
-
-test("shows a comment button when a diff line is hovered", async ({ page }) => {
-  const review = page.locator('[data-component="session-review"]')
-  const line = review.getByText("export const first = 1", { exact: true })
-  await expectAppVisible(line)
-
-  const comment = review.getByRole("button", { name: "Comment", exact: true, includeHidden: true })
-  await expect(comment).toHaveCount(1)
-  await line.dispatchEvent("pointermove", { pointerType: "mouse", bubbles: true, composed: true })
-  await expect(comment).toBeVisible()
-  await expect(comment).toHaveCSS("pointer-events", "auto")
-  await comment.dispatchEvent("click")
-  await expect(review.getByRole("textbox")).toBeVisible()
-  await expect(review.locator('[data-slot="line-comment-editor-label"]')).toHaveText("Commenting on line 1")
-})
-
 test("stages a submitted line comment in the prompt context", async ({ page }) => {
   page.on("request", (request) => {
     expect.soft(request.method(), `unexpected ${request.method()} ${new URL(request.url()).pathname}`).toBe("GET")

packages/app/e2e/regression/session-queue.spec.ts

@@ -201,12 +201,15 @@ test("editing restores the existing draft and replaces only the original queue p
   await view.input.fill("my in-progress draft")
   await original.click()
   await expect(view.input).toHaveText("tighten the error copy")
+  await expect(view.input).toBeFocused()
   await view.input.press("Escape")
   await expect(view.input).toHaveText("my in-progress draft")
 
   await original.click()
   await expect(view.input).toHaveText("tighten the error copy")
+  await expect(view.input).toBeFocused()
   await view.input.fill("tighten the error copy and add a retry hint")
+  await expect(view.input).toHaveText("tighten the error copy and add a retry hint")
   await view.input.press("Enter")
 
   await expect(view.rows.locator('[data-action="session-queue-edit"]')).toHaveText([

packages/app/e2e/regression/session-timeline-accessibility.spec.ts

@@ -1,16 +1,43 @@
 import { expect, test } from "@playwright/test"
-import { assistantMessage, setupTimeline, shell, userMessage } from "../performance/timeline-stability/fixture"
+import {
+  assistantMessage,
+  setupTimeline,
+  shell,
+  textPart,
+  userMessage,
+} from "../performance/timeline-stability/fixture"
 
 test("space activates a focused timeline button instead of scrolling", async ({ page }) => {
   const shellID = "prt_space_button_shell"
   await setupTimeline(page, {
-    messages: [userMessage(), assistantMessage([shell(shellID, "completed", lines(5))])],
+    messages: [
+      userMessage(),
+      assistantMessage([
+        shell(shellID, "completed", lines(5)),
+        textPart(
+          "prt_space_following",
+          "Following content leaves room to focus the command away from the bottom. ".repeat(40),
+        ),
+      ]),
+    ],
     settings: { shellToolPartsExpanded: false },
     reducedMotion: true,
+    seedHistory: true,
   })
   const scroller = page.locator(".scroll-view__viewport", { has: page.locator("[data-timeline-row]") })
   const trigger = page.getByRole("button", { name: "Used Shell" })
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight))
+    .toBeGreaterThan(300)
+  await trigger.scrollIntoViewIfNeeded()
+  await scroller.hover()
+  await page.mouse.wheel(0, -100)
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight - element.scrollTop))
+    .toBeGreaterThan(50)
+  await expect(trigger).toBeInViewport()
   await trigger.focus()
+  await expect(trigger).toBeFocused()
   const before = await scroller.evaluate((element) => element.scrollTop)
   await trigger.press("Space")
   await expect(trigger).toHaveAttribute("aria-expanded", "true")

packages/app/e2e/regression/session-timeline-file-projection.spec.ts

@@ -8,21 +8,6 @@ import {
   userText,
 } from "../performance/timeline-stability/fixture"
 
-test("renders completed write content", async ({ page }) => {
-  const id = "prt_file_projection_write"
-  await setupTimeline(page, {
-    messages: [
-      userMessage(),
-      assistantMessage([
-        toolPart(id, "write", "completed", { path: "src/write.ts", content: "export const written = true\n" }),
-      ]),
-    ],
-    settings: { editToolPartsExpanded: true },
-  })
-
-  await expect(page.locator(`[data-timeline-part-id="${id}"] [data-component="write-content"]`)).toBeVisible()
-})
-
 test("renders a completed single-file patch", async ({ page }) => {
   const id = "prt_file_projection_single_patch"
   await setupTimeline(page, {

packages/app/e2e/regression/session-timeline-file-state.spec.ts

@@ -1,53 +0,0 @@
-import { expect, test } from "@playwright/test"
-import { assistantMessage, setupTimeline, toolPart, userMessage } from "../performance/timeline-stability/fixture"
-import { createTwoFilesPatch } from "diff"
-
-test("keeps patch file disclosures independent", async ({ page }) => {
-  const patchID = "prt_nested_patch"
-  const files = [patchFile("src/a.ts", "modified"), patchFile("src/b.ts", "added"), patchFile("src/old.ts", "deleted")]
-  await setupTimeline(page, {
-    messages: [
-      userMessage(),
-      assistantMessage([
-        toolPart(
-          patchID,
-          "patch",
-          "completed",
-          { patchText: "Update three files" },
-          { metadata: { files } },
-        ),
-      ]),
-    ],
-    settings: { editToolPartsExpanded: true },
-  })
-  const wrapper = page.locator(`[data-timeline-part-id="${patchID}"]`)
-  const modified = wrapper.locator('[data-scope="apply-patch"] [data-type="update"]')
-  const deleted = wrapper.locator('[data-scope="apply-patch"] [data-type="delete"]')
-  await expect(wrapper.locator('[data-scope="apply-patch"] [aria-expanded="false"]')).toHaveCount(3)
-  await deleted.getByRole("button").click()
-  await expect(deleted.getByRole("button")).toHaveAttribute("aria-expanded", "true")
-  await expect(modified.getByRole("button")).toHaveAttribute("aria-expanded", "false")
-  await modified.getByRole("button").click()
-  await expect(modified.getByRole("button")).toHaveAttribute("aria-expanded", "true")
-  await deleted.getByRole("button").click()
-  await expect(deleted.getByRole("button")).toHaveAttribute("aria-expanded", "false")
-  await expect(modified.getByRole("button")).toHaveAttribute("aria-expanded", "true")
-})
-
-function patchFile(file: string, status: "added" | "modified" | "deleted") {
-  const before = status === "added" ? "" : source(false)
-  const after = status === "deleted" ? "" : source(true)
-  return {
-    file,
-    status,
-    patch: createTwoFilesPatch(`a/${file}`, `b/${file}`, before, after),
-    additions: status === "deleted" ? 0 : 4,
-    deletions: status === "added" ? 0 : 3,
-  }
-}
-
-function source(changed: boolean) {
-  return Array.from({ length: 12 }, (_, index) => `export const value${index} = ${changed ? index + 1 : index}\n`).join(
-    "",
-  )
-}

packages/app/e2e/regression/session-timeline-lifecycle-state.spec.ts

@@ -122,6 +122,7 @@ test("transitions thinking and hidden reasoning through busy to idle", async ({
   await expect(page.locator('[data-timeline-row="Thinking"]')).toBeVisible()
   await expect(page.getByText("Inspecting stability", { exact: true })).toBeVisible()
   await expect(page.locator(`[data-timeline-part-id="${reasoningID}"]`)).toHaveCount(0)
+  await expect(page.locator(`[data-timeline-part-id="${renderedPartID(reasoningID)}"]`)).toHaveCount(0)
   await timeline.send(partUpdated(shell("prt_reasoning_shell", "running")), 160)
   await expect(page.locator('[data-timeline-row="Thinking"]')).toBeVisible()
   await timeline.send(partUpdated(shell("prt_reasoning_shell", "completed", "done")), 180)
@@ -129,6 +130,7 @@ test("transitions thinking and hidden reasoning through busy to idle", async ({
   await timeline.send(status("idle"), 300)
   await expect(page.locator('[data-timeline-row="Thinking"]')).toHaveCount(0)
   await expect(page.locator(`[data-timeline-part-id="${reasoningID}"]`)).toHaveCount(0)
+  await expect(page.locator(`[data-timeline-part-id="${renderedPartID(reasoningID)}"]`)).toHaveCount(0)
 })
 
 test("moves busy through retry and recovery to final idle content", async ({ page }) => {

packages/app/e2e/regression/session-timeline-notices.spec.ts

@@ -122,6 +122,7 @@ test("updates running compactions to failed and cancelled boundaries", async ({
 
   await timeline.send(compactionStarted({ sessionID, reason: "auto", recent: "" }))
   await timeline.send(compactionDelta({ sessionID, text: "Partial summary that should be discarded." }))
+  await expect(page.getByText("Partial summary that should be discarded.", { exact: true })).toBeVisible()
   await timeline.send(
     compactionFailed({
       sessionID,
@@ -140,6 +141,9 @@ test("updates running compactions to failed and cancelled boundaries", async ({
   await expect(failed).not.toContainText("Partial summary that should be discarded.")
 
   await timeline.send(compactionStarted({ sessionID, reason: "manual", recent: "" }))
+  await expect(compactions).toHaveCount(2)
+  await timeline.send(compactionDelta({ sessionID, text: "Summary before cancellation." }))
+  await expect(page.getByText("Summary before cancellation.", { exact: true })).toBeVisible()
   await timeline.send(
     compactionFailed({
       sessionID,
@@ -152,88 +156,7 @@ test("updates running compactions to failed and cancelled boundaries", async ({
   const cancelled = compactions.filter({ hasNotText: "The provider rejected the summary." })
   await expect(cancelled.getByText("Session compacted", { exact: true })).toBeVisible()
   await expect(cancelled).not.toContainText("Cancellation detail should stay hidden.")
-})
-
-test("shows a delegating row while subagent input streams", async ({ page }) => {
-  await setupTimeline(page, {
-    sessionMessages: [
-      user,
-      {
-        ...assistant(false),
-        content: [
-          {
-            type: "tool",
-            id: "call_subagent",
-            name: "subagent",
-            state: { status: "streaming", input: "" },
-            time: { created: 2 },
-          },
-        ],
-      },
-    ],
-  })
-
-  const delegating = page.locator('[data-component="task-tool-delegating"]')
-  await expect(delegating).toBeVisible()
-  const shimmer = delegating.locator('[data-component="text-shimmer"]')
-  await expect(shimmer).toHaveAttribute("aria-label", "Delegating agent...")
-  await expect(shimmer).toHaveCSS("line-height", "16px")
-  const icon = delegating.locator('[data-slot="icon-svg"]')
-  await expect(icon.locator('use[href="#opencode-v2-icon-subagent"]')).toBeVisible()
-  await expect(icon).toHaveCSS("color", "rgb(174, 174, 174)")
-  await expect(page.locator('[data-component="task-tool-card"]')).toHaveCount(0)
-  await expect(page.locator('[data-timeline-row="Thinking"]')).toHaveCount(0)
-})
-
-test("renders the moved location notice in its compact timeline style", async ({ page }) => {
-  const directory = `/Users/usrnk1/Developer/opencode/${"nested-directory/".repeat(24)}session`
-  await page.setViewportSize({ width: 480, height: 720 })
-  await setupTimeline(page, {
-    sessionMessages: [
-      user,
-      {
-        id: "msg_location",
-        type: "location-switched",
-        location: { directory },
-        time: { created: 2 },
-      },
-    ],
-  })
-
-  const notice = page.locator('[data-slot="session-timeline-notice"][data-type="location-switched"]')
-  const label = notice.locator('[data-slot="session-timeline-notice-label"]')
-  const value = notice.locator('[data-slot="session-timeline-notice-value"]')
-  const tooltipTrigger = notice.locator('[data-component="tooltip-v2-trigger"]')
-
-  await expect(label).toHaveText("Moved to")
-  await expect(value).toHaveText(directory)
-  await expect(notice).not.toContainText("·")
-  await expect(notice.locator("svg")).toHaveCount(0)
-  await expect(notice).toHaveCSS("height", "28px")
-  await expect(notice).toHaveCSS("gap", "8px")
-  await expect(notice).toHaveCSS("padding-top", "4px")
-  await expect(notice).toHaveCSS("padding-bottom", "4px")
-  await expect(label).toHaveCSS("font-size", "13px")
-  await expect(label).toHaveCSS("font-weight", "530")
-  await expect(label).toHaveCSS("line-height", "16px")
-  await expect(label).toHaveCSS("color", "rgb(128, 128, 128)")
-  await expect(value).toHaveCSS("font-size", "13px")
-  await expect(value).toHaveCSS("font-weight", "440")
-  await expect(value).toHaveCSS("line-height", "16px")
-  await expect(value).toHaveCSS("color", "rgb(128, 128, 128)")
-  await expect(value).toHaveCSS("text-overflow", "ellipsis")
-  await expect(value).toHaveCSS("white-space", "nowrap")
-  await expect(value).toHaveAttribute("dir", "ltr")
-  await expect.poll(() => value.evaluate((element) => element.scrollWidth > element.clientWidth)).toBe(true)
-
-  const tooltip = page.getByText("Session working directory changed", { exact: true })
-  await label.hover()
-  await expect(tooltip).toBeVisible()
-  await page.mouse.move(0, 0)
-  await expect(tooltip).toBeHidden()
-  await tooltipTrigger.focus()
-  await expect(tooltipTrigger).toBeFocused()
-  await expect(tooltip).toBeVisible()
+  await expect(cancelled).not.toContainText("Summary before cancellation.")
 })
 
 test("moves blocking work to the background with Ctrl+B", async ({ page }) => {
@@ -271,11 +194,6 @@ test("moves blocking work to the background with Ctrl+B", async ({ page }) => {
   await request
 })
 
-test("waits for completion before labeling requested background work", async ({ page }) => {
-  await setupTimeline(page, { sessionMessages: [user, assistant(false, true, undefined, true)] })
-  await expect(page.locator('[data-component="task-tool-card"]')).not.toContainText("(background)")
-})
-
 test("navigates from a running subagent card and hides background controls in the child", async ({ page }) => {
   const childID = "ses_running_child"
   await setupTimeline(page, {

packages/app/e2e/regression/session-timeline-projection.spec.ts

@@ -7,86 +7,9 @@ import {
   toolPart,
   userMessage,
   userText,
-  type PartSeed,
 } from "../performance/timeline-stability/fixture"
 
 test.describe("session timeline projection", () => {
-  test("renders every admitted tool family and hides timeline-only exclusions", async ({ page }) => {
-    const parts = [
-      toolPart("prt_01_read", "read", "completed", { path: "src/a.ts" }),
-      toolPart("prt_02_glob", "glob", "completed", { path: ".", pattern: "**/*.ts" }),
-      toolPart("prt_03_grep", "grep", "completed", { path: ".", pattern: "value" }),
-      toolPart("prt_04_list", "list", "completed", { path: "src" }),
-      toolPart("prt_webfetch", "webfetch", "completed", { url: "https://example.com" }),
-      toolPart(
-        "prt_websearch",
-        "websearch",
-        "completed",
-        { query: "timeline stability" },
-        { output: "https://example.com/result" },
-      ),
-      toolPart("prt_task", "subagent", "completed", {
-        description: "Inspect timeline",
-        agent: "explore",
-        prompt: "Inspect the timeline implementation.",
-      }),
-      toolPart(
-        "prt_bash",
-        "shell",
-        "completed",
-        { command: "printf stable" },
-        { output: "stable", title: "printf stable" },
-      ),
-      editPart("prt_edit"),
-      toolPart("prt_write", "write", "completed", { path: "src/new.ts", content: "export const stable = true\n" }),
-      patchPart("prt_patch"),
-      toolPart("prt_todo", "todowrite", "completed", { todos: [{ content: "Hidden", status: "pending" }] }),
-      toolPart(
-        "prt_question",
-        "question",
-        "completed",
-        { questions: [{ question: "Keep stable?", header: "Stability", options: [] }] },
-        { metadata: { answers: [["Yes"]] } },
-      ),
-      toolPart("prt_skill", "skill", "completed", { name: "stability" }),
-      toolPart("prt_custom", "custom_mcp_tool", "completed", { target: "timeline", count: 2 }),
-    ]
-    await setupTimeline(page, { messages: [userMessage(), assistantMessage(parts)] })
-
-    const first = page.locator(
-      '[data-timeline-part-ids="prt_01_read,prt_02_glob,prt_03_grep,prt_04_list,prt_webfetch,prt_websearch,prt_task,prt_bash,prt_edit,prt_write,prt_patch"]',
-    )
-    const second = page.locator('[data-timeline-part-ids="prt_skill,prt_custom"]')
-    await expect(first).toBeVisible()
-    await expect(second).toBeVisible()
-    await first.getByRole("button").click()
-    await second.getByRole("button").click()
-    for (const id of [
-      "prt_webfetch",
-      "prt_websearch",
-      "prt_task",
-      "prt_bash",
-      "prt_edit",
-      "prt_write",
-      "prt_patch",
-      "prt_question",
-      "prt_skill",
-      "prt_custom",
-    ]) {
-      await expect(page.locator(`[data-timeline-part-id="${id}"]`).first(), id).toBeVisible()
-    }
-    const patch = page.locator('[data-timeline-part-id="prt_patch"]')
-    await expect(patch.getByText("1 file", { exact: true })).toBeVisible()
-    await expect(patch.getByRole("button", { name: "Patch 1 file", exact: true })).toHaveCount(0)
-    await expect(patch.getByRole("button")).toHaveCount(1)
-    await expect(patch.locator('[data-scope="apply-patch"] button[aria-expanded="false"]')).toHaveCount(1)
-    await expect(patch.locator('[data-slot="message-part-title-filename"]')).toHaveCount(0)
-    await expect(patch.locator('[data-slot="message-part-actions"]')).toHaveCount(0)
-    const edit = page.locator('[data-timeline-part-id="prt_edit"]')
-    await expect(edit).toContainText("Edit")
-    await expect(page.locator('[data-timeline-part-id="prt_todo"]')).toHaveCount(0)
-  })
-
   test("combines adjacent patch calls and repeated files into one group", async ({ page }) => {
     const first = "prt_patch_first"
     const second = "prt_patch_second"
@@ -158,43 +81,6 @@ test.describe("session timeline projection", () => {
     await expect(page.locator(`[data-timeline-part-id="${first}"], [data-timeline-part-id="${second}"]`)).toHaveCount(0)
   })
 
-  test("combines adjacent edit calls and repeated files into one group", async ({ page }) => {
-    const first = "prt_edit_first"
-    const second = "prt_edit_second"
-    await setupTimeline(page, {
-      messages: [
-        userMessage(),
-        assistantMessage([
-          toolPart(
-            first,
-            "edit",
-            "completed",
-            { path: "src/first.ts", oldString: "one", newString: "two" },
-            {
-              metadata: { files: [patchFile("src/first.ts", "modified")] },
-            },
-          ),
-          toolPart(
-            second,
-            "edit",
-            "completed",
-            { path: "src/first.ts", oldString: "two", newString: "three" },
-            {
-              metadata: { files: [patchFile("src/first.ts", "modified")] },
-            },
-          ),
-        ]),
-      ],
-      settings: { editToolPartsExpanded: true },
-    })
-
-    const group = page.locator(`[data-timeline-part-ids="${first},${second}"]`)
-    await expect(group.locator('[data-slot="basic-tool-tool-title"]')).toContainText("Edit")
-    await expect(group.getByText("1 file", { exact: true })).toBeVisible()
-    await expect(group.locator('[data-slot="apply-patch-filename"]')).toHaveText(["first.ts"])
-    await expect(group.locator('[data-scope="apply-patch"] button')).toHaveAttribute("aria-expanded", "true")
-  })
-
   test("projects gaps, dividers, assistant parts, and errors together", async ({ page }) => {
     const firstUser = userMessage(
       [
@@ -236,25 +122,6 @@ test.describe("session timeline projection", () => {
     await expect(page.locator('[data-timeline-row="TurnGap"]')).toBeVisible()
   })
 
-  test("renders interruption independently when the turn is not compacted", async ({ page }) => {
-    const user = userMessage()
-    const before = assistantMessage([{ id: "prt_before", type: "text", text: "Before" }], {
-      id: "msg_1001_before",
-      error: { type: "MessageAbortedError", message: "Stopped" },
-    })
-    const after = assistantMessage([{ id: "prt_after", type: "text", text: "After" }], {
-      id: "msg_1002_after",
-      created: 1700000003000,
-    })
-    await setupTimeline(page, { messages: [user, before, after] })
-
-    await expect(page.getByText("Interrupted", { exact: true })).toBeVisible()
-    const rows = await page
-      .locator('[data-timeline-row="AssistantPart"], [data-timeline-row="TurnDivider"]')
-      .evaluateAll((elements) => elements.map((element) => element.getAttribute("data-timeline-row")))
-    expect(rows).toEqual(["AssistantPart", "TurnDivider", "AssistantPart"])
-  })
-
   test("renders aliased and long custom model notices", async ({ page }) => {
     const shortName = "GPT-5.4 nano"
     const longName = "Company Gateway Extra Long Context Model for Narrow Timeline Layouts"
@@ -291,77 +158,8 @@ test.describe("session timeline projection", () => {
     await expect(longNotice.locator("[title]")).toHaveAttribute("title", `Switched to ${longName}`)
     await expect.poll(() => longNotice.evaluate((element) => element.scrollWidth <= element.clientWidth)).toBe(true)
   })
-
-  test("renders user image, file attachment, file reference, and agent reference", async ({ page }) => {
-    const text = "Use @explore with @src/a.ts and inspect the attachments"
-    const parts: PartSeed<"user">[] = [
-      userText(text, { id: "prt_user_rich" }),
-      {
-        id: "prt_user_image",
-        type: "file",
-        mime: "image/png",
-        filename: "pixel.png",
-        url: "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg==",
-      },
-      {
-        id: "prt_user_attachment",
-        type: "file",
-        mime: "application/json",
-        filename: "tsconfig.json",
-        url: "data:application/json;base64,e30=",
-      },
-      {
-        id: "prt_user_reference",
-        type: "file",
-        mime: "text/plain",
-        filename: "a.ts",
-        url: "src/a.ts",
-        source: { type: "file", path: "src/a.ts", text: { value: "@src/a.ts", start: 18, end: 27 } },
-      },
-      {
-        id: "prt_user_agent",
-        type: "agent",
-        name: "explore",
-        source: { value: "@explore", start: 4, end: 12 },
-      },
-    ]
-    await setupTimeline(page, { messages: [userMessage(parts), assistantMessage()] })
-
-    await expect(page.getByAltText("pixel.png")).toBeVisible()
-    await expect(page.getByText("tsconfig.json")).toBeVisible()
-    await expect(page.getByText("@src/a.ts", { exact: true })).toBeVisible()
-    await expect(page.getByText("@explore", { exact: true })).toBeVisible()
-  })
 })
 
-function editPart(id: string) {
-  return toolPart(
-    id,
-    "edit",
-    "completed",
-    { path: "src/a.ts", oldString: "export const value = 1", newString: "export const value = 2" },
-    {
-      metadata: {
-        files: [patchFile("src/a.ts", "modified")],
-      },
-    },
-  )
-}
-
-function patchPart(id: string) {
-  return toolPart(
-    id,
-    "patch",
-    "completed",
-    { patchText: "Update the projected files" },
-    {
-      metadata: {
-        files: [patchFile("src/a.ts", "modified")],
-      },
-    },
-  )
-}
-
 function patchFile(file: string, status: "added" | "modified" | "deleted") {
   return {
     file,

packages/app/e2e/regression/session-timeline-reducer-projection.spec.ts

@@ -7,33 +7,25 @@ import {
   renderedPartID,
   setupTimeline,
   shell,
+  toolPart,
   status,
   textPart,
-  toolPart,
   userMessage,
 } from "../performance/timeline-stability/fixture"
 
-test("groups every collapsed tool until visible text separates the stack", async ({ page }) => {
-  const parts = [
-    toolPart("prt_boundary_01_read", "read", "completed", { path: "src/a.ts" }),
-    textPart("prt_boundary_02_text", "Boundary text"),
-    toolPart("prt_boundary_03_glob", "glob", "completed", { path: ".", pattern: "**/*.ts" }),
-    toolPart("prt_boundary_04_grep", "grep", "completed", { path: ".", pattern: "stable" }),
-    shell("prt_boundary_05_shell", "completed", "done"),
-    toolPart("prt_boundary_06_list", "list", "completed", { path: "src" }),
-  ]
-  await setupTimeline(page, { messages: [userMessage(), assistantMessage(parts)] })
+test("reducer-hardening: converges when idle arrives before final part and message completion", async ({ page }) => {
+  const textID = "prt_event_order_text"
+  const assistant = assistantMessage([textPart(textID, "Partial")], { completed: false })
+  const timeline = await setupTimeline(page, { messages: [userMessage(), assistant] })
+  await timeline.send(status("busy"), 100)
+  await timeline.send(status("idle"), 100)
+  await timeline.send(partUpdated(textPart(textID, "Final after early idle")), 120)
+  await timeline.send(messageUpdated(completedAssistantInfo(assistant)), 250)
 
-  await expect(page.locator('[data-timeline-part-ids="prt_boundary_01_read"]')).toBeVisible()
-  const group = page.locator(
-    '[data-timeline-part-ids="prt_boundary_03_glob,prt_boundary_04_grep,prt_boundary_05_shell,prt_boundary_06_list"]',
+  await expect(page.locator('[data-timeline-row="Thinking"]')).toHaveCount(0)
+  await expect(page.locator(`[data-timeline-part-id="${renderedPartID(textID)}"]`)).toContainText(
+    "Final after early idle",
   )
-  await expect(group).toBeVisible()
-  await expect(group.getByRole("button")).toHaveAccessibleName("Used Glob, Grep, Shell, List")
-  await expect(group.locator('[data-component="tag"]')).toHaveText("4")
-  await expect(page.locator('[data-timeline-row="AssistantPart"]')).toHaveCount(3)
-  await expect(page.locator('[data-timeline-spacing="content"]')).toHaveCount(2)
-  await expect(page.locator('[data-timeline-spacing="content"]').nth(0)).toHaveCSS("padding-top", "16px")
 })
 
 test("expands a mixed collapsed tool stack without expanding its individual calls", async ({ page }) => {
@@ -132,18 +124,3 @@ test("keeps failed search calls and their error cards inside the collapsed stack
     "Search timed out after 30 seconds",
   )
 })
-
-test("reducer-hardening: converges when idle arrives before final part and message completion", async ({ page }) => {
-  const textID = "prt_event_order_text"
-  const assistant = assistantMessage([textPart(textID, "Partial")], { completed: false })
-  const timeline = await setupTimeline(page, { messages: [userMessage(), assistant] })
-  await timeline.send(status("busy"), 100)
-  await timeline.send(status("idle"), 100)
-  await timeline.send(partUpdated(textPart(textID, "Final after early idle")), 120)
-  await timeline.send(messageUpdated(completedAssistantInfo(assistant)), 250)
-
-  await expect(page.locator('[data-timeline-row="Thinking"]')).toHaveCount(0)
-  await expect(page.locator(`[data-timeline-part-id="${renderedPartID(textID)}"]`)).toContainText(
-    "Final after early idle",
-  )
-})

packages/app/e2e/regression/session-timeline-tool-projection.spec.ts

@@ -7,31 +7,6 @@ import {
   userMessage,
 } from "../performance/timeline-stability/fixture"
 
-test("renders every tool error outcome without leaking hidden tools", async ({ page }) => {
-  const ordinary = ["shell", "edit", "write", "patch", "webfetch", "websearch", "subagent", "skill", "mcp_probe"]
-  const parts = ordinary.map((tool, index) =>
-    toolPart(`prt_error_${index}`, tool, "error", errorInput(tool), { error: `${tool} failed visibly` }),
-  )
-  parts.push(
-    toolPart("prt_question_dismissed", "question", "error", questionInput(), {
-      error: "The user dismissed this question",
-    }),
-    toolPart("prt_question_error", "question", "error", questionInput(), { error: "Question transport failed" }),
-    toolPart("prt_todo_error", "todowrite", "error", { todos: [] }, { error: "Hidden todo failure" }),
-  )
-  await setupTimeline(page, { messages: [userMessage(), assistantMessage(parts)] })
-
-  const group = page.locator(`[data-timeline-part-ids="${ordinary.map((_, index) => `prt_error_${index}`).join(",")}"]`)
-  await expect(group.locator('[data-component="tag"]')).toHaveText(String(ordinary.length))
-  await group.getByRole("button").click()
-  await expect(page.locator('[data-kind="tool-error-card"]')).toHaveCount(ordinary.length + 1)
-  await expect(page.getByText(/dismissed/i)).toBeVisible()
-  await expect(page.locator('[data-timeline-part-id="prt_todo_error"]')).toHaveCount(0)
-  for (let index = 0; index < ordinary.length; index++) {
-    await expect(page.locator(`[data-timeline-part-id="prt_error_${index}"]`)).toBeVisible()
-  }
-})
-
 test("transitions shell and question through running error outcomes", async ({ page }) => {
   const shellID = "prt_transition_error_shell"
   const questionID = "prt_transition_error_question"
@@ -138,62 +113,6 @@ test("preserves surviving grouped patch state when its first patch fails", async
     .toBeGreaterThanOrEqual(-0.5)
 })
 
-test("labels all web search provider variants", async ({ page }) => {
-  const parts = [
-    toolPart(
-      "prt_search_parallel",
-      "websearch",
-      "completed",
-      { query: "parallel" },
-      { metadata: { provider: "parallel" } },
-    ),
-    toolPart("prt_search_exa", "websearch", "completed", { query: "exa" }, { metadata: { provider: "exa" } }),
-    toolPart("prt_search_generic", "websearch", "completed", { query: "generic" }),
-  ]
-  await setupTimeline(page, { messages: [userMessage(), assistantMessage(parts)] })
-  await page.getByRole("button", { name: "Used Parallel Web Search, Exa Web Search, Web Search" }).click()
-
-  const tools = page.locator('[data-component="context-tool-group-list"]')
-  await expect(tools.getByRole("button", { name: /Parallel Web Search/ })).toBeVisible()
-  await expect(tools.getByRole("button", { name: /Exa Web Search/ })).toBeVisible()
-  await expect(tools.getByRole("button", { name: /^Web Search/ })).toBeVisible()
-})
-
-test("labels completed searches with result counts", async ({ page }) => {
-  const glob = "prt_glob_count"
-  const grep = "prt_grep_count"
-  await setupTimeline(page, {
-    messages: [
-      userMessage(),
-      assistantMessage([
-        toolPart(glob, "glob", "completed", { path: ".", pattern: "**/*.ts" }, { metadata: { count: 1 } }),
-        toolPart(grep, "grep", "completed", { path: ".", pattern: "value" }, { metadata: { matches: 12 } }),
-      ]),
-    ],
-  })
-
-  const group = page.locator(`[data-timeline-part-ids="${glob},${grep}"]`)
-  await group.locator('[data-slot="collapsible-trigger"]').click()
-  const rows = group.locator('[data-component="context-tool-group-list"] [data-component="tool-trigger"]')
-  await expect(rows.filter({ hasText: "Glob" })).toContainText("(1 match)")
-  await expect(rows.filter({ hasText: "Grep" })).toContainText("(12 matches)")
-})
-
-test("labels read tools from their path input", async ({ page }) => {
-  const id = "prt_read_path"
-  await setupTimeline(page, {
-    messages: [userMessage(), assistantMessage([toolPart(id, "read", "completed", { path: "src/a.ts" })])],
-  })
-
-  const group = page.locator(`[data-timeline-part-ids="${id}"]`)
-  await group.locator('[data-slot="collapsible-trigger"]').click()
-  await expect(
-    group
-      .locator('[data-component="context-tool-group-list"] [data-component="tool-trigger"]')
-      .filter({ hasText: "Read" }),
-  ).toContainText("a.ts")
-})
-
 test("groups instruction files loaded by the same read", async ({ page }) => {
   const id = "prt_read_instructions"
   await setupTimeline(page, {
@@ -221,36 +140,6 @@ test("groups instruction files loaded by the same read", async ({ page }) => {
   await expect(loaded.locator('[data-slot="tool-loaded-kind"]')).toHaveCount(0)
 })
 
-test("labels skill tools from IDs and result metadata", async ({ page }) => {
-  const pending = "prt_skill_id"
-  const completed = "prt_skill_name"
-  await setupTimeline(page, {
-    messages: [
-      userMessage(),
-      assistantMessage([
-        toolPart(pending, "skill", "running", { id: "frontend-design" }),
-        toolPart(completed, "skill", "completed", { id: "opencode" }, { metadata: { name: "OpenCode" } }),
-      ]),
-    ],
-  })
-
-  const group = page.locator(`[data-timeline-part-ids="${pending},${completed}"]`)
-  await expect(group.getByRole("button")).toHaveAccessibleName("Used Skill")
-  await expect(group.locator('[data-component="tag"]')).toHaveText("2")
-  await group.getByRole("button").click()
-
-  const loaded = group.locator('[data-component="tool-loaded-item"]')
-  await expect(loaded).toHaveCount(1)
-  await expect(loaded).toHaveAttribute("aria-label", "Loaded frontend-design, OpenCode skills")
-  await expect(loaded).toHaveCSS("line-height", "16px")
-  await expect(loaded.locator('[data-slot="tool-loaded-label"]')).toHaveText("Loaded")
-  await expect(loaded.locator('[data-slot="tool-loaded-kind"]')).toHaveText("skills")
-  const names = loaded.locator('[data-component="text-shimmer"]')
-  await expect(names).toHaveCount(2)
-  await expect(names.nth(0)).toHaveAttribute("aria-label", "frontend-design")
-  await expect(names.nth(1)).toHaveAttribute("aria-label", "OpenCode")
-})
-
 test("groups only consecutive successful skill tools", async ({ page }) => {
   const parts = [
     toolPart("prt_skill_first", "skill", "completed", { id: "ocpr" }),
@@ -273,14 +162,3 @@ test("groups only consecutive successful skill tools", async ({ page }) => {
 function questionInput() {
   return { questions: [{ header: "Stability", question: "Keep it stable?", options: [] }] }
 }
-
-function errorInput(tool: string) {
-  if (tool === "shell") return { command: "exit 1" }
-  if (["edit", "write"].includes(tool)) return { path: "src/error.ts", content: "" }
-  if (tool === "patch") return { patchText: "Update src/error.ts" }
-  if (tool === "webfetch") return { url: "https://example.com" }
-  if (tool === "websearch") return { query: "failure" }
-  if (tool === "subagent") return { description: "Fail subagent", agent: "explore", prompt: "Inspect the failure." }
-  if (tool === "skill") return { name: "failure" }
-  return { target: "failure" }
-}

packages/app/e2e/regression/terminal-composer-focus.spec.ts

@@ -87,6 +87,7 @@ test("clears the terminal line with Command+Delete", async ({ page }) => {
   const terminal = page.locator('[data-component="terminal"]')
   await page.keyboard.press("Control+Backquote")
   await expect(terminal.locator("textarea")).toHaveCount(1)
+  await expect.poll(() => sendPtyOutput).toBeDefined()
 
   await page.keyboard.press("Meta+Backspace")
 

packages/app/e2e/regression/terminal-hidden.spec.ts

@@ -127,7 +127,6 @@ test("animates review and terminal panels while caching hidden terminal content"
   await expectStackedGeometry(page)
   await expectPanelGapHeld(page)
 
-  await resetTerminalTopMotion(page)
   await resetTerminalBottomMotion(page)
   await resetTerminalAnchorGaps(page)
   await resetPanelGaps(page)
@@ -139,7 +138,6 @@ test("animates review and terminal panels while caching hidden terminal content"
   await expect(panel).toBeVisible()
   await expectHeightMotions(page, "session-side-region", 2)
   await expectHeightMotions(page, "session-side-terminal-region", 2)
-  await expectTerminalTopMotion(page)
   await expectTerminalBottomFixed(page)
   await expectTerminalTopAnchored(page)
   await expectPanelGapHeld(page)
@@ -225,7 +223,6 @@ type MotionProbe = {
   terminalAnchorGaps: number[]
   resetAnchorOnMotion: boolean
   panelGaps: number[]
-  terminalTops: number[]
   terminalBottoms: number[]
   heights: string[]
   animations: string[]
@@ -243,7 +240,6 @@ async function installMotionProbe(page: Page) {
       terminalAnchorGaps: [],
       resetAnchorOnMotion: false,
       panelGaps: [],
-      terminalTops: [],
       terminalBottoms: [],
       heights: [],
       animations: [],
@@ -270,7 +266,6 @@ async function installMotionProbe(page: Page) {
         const terminalContent = document.querySelector<HTMLElement>('[data-slot="terminal-panel-content"]')
         const panelGap = document.querySelector<HTMLElement>('[data-slot="session-side-panel-gap"]')
         if (!terminal || !terminalContent) return
-        probe.terminalTops.push(terminal.getBoundingClientRect().top)
         probe.terminalBottoms.push(terminal.getBoundingClientRect().bottom)
         probe.terminalContentSizes.push({
           width: terminalContent.getBoundingClientRect().width,
@@ -446,13 +441,6 @@ async function expectStackPainted(page: Page) {
   expect(Math.max(...gaps.map((gap) => gap.terminalSurface)), JSON.stringify(gaps)).toBeLessThanOrEqual(1)
 }
 
-async function resetTerminalTopMotion(page: Page) {
-  await page.evaluate(() => {
-    const probe = (window as Window & { __panelMotion?: MotionProbe }).__panelMotion
-    if (probe) probe.terminalTops = []
-  })
-}
-
 async function resetTerminalBottomMotion(page: Page) {
   await page.evaluate(() => {
     const probe = (window as Window & { __panelMotion?: MotionProbe }).__panelMotion
@@ -516,17 +504,6 @@ async function expectTerminalContentCachedSize(page: Page) {
   expect(Math.min(...sizes.map((size) => size.height))).toBeGreaterThan(100)
 }
 
-async function expectTerminalTopMotion(page: Page) {
-  const tops = await page.evaluate(
-    () => (window as Window & { __panelMotion?: MotionProbe }).__panelMotion?.terminalTops.map(Math.round) ?? [],
-  )
-  const unique = [...new Set(tops)]
-  const range = Math.max(...unique) - Math.min(...unique)
-  const maxDelta = Math.max(...unique.slice(1).map((value, index) => Math.abs(value - unique[index])))
-  expect(unique.length, JSON.stringify(unique)).toBeGreaterThan(6)
-  expect(maxDelta, JSON.stringify({ unique, range, maxDelta })).toBeLessThan(range * 0.3)
-}
-
 async function expectHeightMotions(page: Page, slot: string, count: number) {
   await expect
     .poll(() =>

packages/app/e2e/service-worker/cache.spec.ts

@@ -0,0 +1,137 @@
+import { expect, test } from "@playwright/test"
+import { readFile } from "node:fs/promises"
+import { createServer } from "node:http"
+import { once } from "node:events"
+
+const legacy = `
+self.addEventListener("install", event => event.waitUntil(
+  caches.open("workbox-precache-v2-" + self.registration.scope).then(cache =>
+    cache.addAll(["/index.html", "/assets/app-old.js", "/assets/lazy-old.js"])
+  )
+))
+self.addEventListener("fetch", event => {
+  if (event.request.mode === "navigate") {
+    event.respondWith(caches.match("/index.html"))
+    return
+  }
+  event.respondWith(caches.match(event.request).then(response => response || fetch(event.request)))
+})
+`
+
+const fixture = test.extend<{ site: { url: string; upgrade: () => void; repair: () => void } }>({
+  site: async ({}, use) => {
+    const worker = await readFile(new URL("../../dist/sw.js", import.meta.url), "utf8")
+    const state = { version: "old", repaired: false }
+    const server = createServer((request, response) => {
+      const pathname = new URL(request.url ?? "/", "http://localhost").pathname
+      const prefix = state.version === "old" ? "/assets" : "/_assets"
+      response.setHeader("cache-control", "no-store")
+      if (pathname === "/sw.js") {
+        response.setHeader("content-type", "text/javascript")
+        response.end(state.version === "old" ? legacy : worker)
+        return
+      }
+      if (pathname === `${prefix}/app-${state.version}.js`) {
+        response.setHeader("content-type", "text/javascript")
+        response.end(`import "${prefix}/startup-${state.version}.js"`)
+        return
+      }
+      if (pathname === `${prefix}/startup-${state.version}.js`) {
+        response.setHeader("content-type", "text/javascript")
+        response.end(`
+          document.getElementById("root").innerHTML = '<h1>${state.version}</h1><label>Draft<input></label><button>Load older chunk</button><output></output>'
+          document.querySelector("button").onclick = () => import("/assets/lazy-old.js")
+        `)
+        return
+      }
+      if (
+        (pathname === "/assets/lazy-old.js" && state.version === "old") ||
+        (pathname === "/_assets/retry.js" && state.repaired)
+      ) {
+        response.setHeader("content-type", "text/javascript")
+        response.end('document.querySelector("output").textContent = "Older chunk loaded"')
+        return
+      }
+      // Deliberately retain the old server's fallback so the worker must reject HTML asset responses itself.
+      response.setHeader("content-type", "text/html")
+      response.end(`<div id="root"></div><script type="module" src="${prefix}/app-${state.version}.js"></script>`)
+    })
+    server.listen(0, "127.0.0.1")
+    await once(server, "listening")
+    const address = server.address()
+    if (!address || typeof address === "string") throw new Error("Expected a TCP address")
+    try {
+      await use({
+        url: `http://127.0.0.1:${address.port}`,
+        upgrade: () => (state.version = "new"),
+        repair: () => (state.repaired = true),
+      })
+    } finally {
+      server.closeAllConnections()
+      await new Promise<void>((resolve, reject) => server.close((error) => (error ? reject(error) : resolve())))
+    }
+  },
+})
+
+fixture("updates a legacy worker without reloading drafts or deleting old chunks", async ({ page, site }) => {
+  await page.goto(site.url)
+  await expect(page.getByRole("heading")).toHaveText("old")
+  await page.evaluate(async () => {
+    await navigator.serviceWorker.register("/sw.js")
+    await navigator.serviceWorker.ready
+  })
+  await page.goto(site.url)
+  await expect(page.getByRole("heading")).toHaveText("old")
+  await page.getByLabel("Draft").fill("Keep this unsent prompt")
+
+  site.upgrade()
+  await page.evaluate(async () => {
+    const cache = await caches.open("opencode-assets")
+    await cache.put(
+      "/_assets/startup-new.js",
+      new Response("<html>stale fallback</html>", {
+        headers: { "content-type": "text/html" },
+      }),
+    )
+    const changed = new Promise<void>((resolve) =>
+      navigator.serviceWorker.addEventListener("controllerchange", () => resolve(), { once: true }),
+    )
+    const registration = await navigator.serviceWorker.getRegistration()
+    if (!registration) throw new Error("Missing legacy worker")
+    await registration.update()
+    await changed
+  })
+
+  await expect(page.getByLabel("Draft")).toHaveValue("Keep this unsent prompt")
+  await page.getByRole("button", { name: "Load older chunk" }).click()
+  await expect(page.getByRole("status")).toHaveText("Older chunk loaded")
+
+  await page.goto(`${site.url}/workspace/example`)
+  await expect(page.getByRole("heading")).toHaveText("new")
+  await expect
+    .poll(() =>
+      page.evaluate(async () =>
+        (await (await caches.open("opencode-assets")).match("/_assets/startup-new.js"))?.headers.get("content-type"),
+      ),
+    )
+    .toBe("text/javascript")
+})
+
+fixture("does not cache HTML responses under asset URLs", async ({ page, site }) => {
+  site.upgrade()
+  await page.goto(site.url)
+  await expect(page.getByRole("heading")).toHaveText("new")
+  await page.evaluate(async () => {
+    await navigator.serviceWorker.register("/sw.js")
+    await navigator.serviceWorker.ready
+  })
+  await page.goto(site.url)
+  await expect(page.getByRole("heading")).toHaveText("new")
+  expect(await page.evaluate(async () => (await fetch("/_assets/retry.js")).headers.get("content-type"))).toBe(
+    "text/html",
+  )
+  site.repair()
+  expect(await page.evaluate(async () => (await fetch("/_assets/retry.js")).headers.get("content-type"))).toBe(
+    "text/javascript",
+  )
+})

packages/app/e2e/service-worker/playwright.config.ts

@@ -0,0 +1,9 @@
+import { defineConfig } from "@playwright/test"
+
+export default defineConfig({
+  testDir: ".",
+  testMatch: "*.spec.ts",
+  outputDir: "../test-results/service-worker",
+  timeout: 30_000,
+  use: { browserName: "chromium" },
+})

packages/app/e2e/tsconfig.json

@@ -7,5 +7,5 @@
     "rootDir": "..",
     "types": ["node", "bun"]
   },
-  "include": ["./**/*.ts", "./**/*.tsx", "../src/types.ts"]
+  "include": ["./**/*.ts", "./**/*.tsx", "../component-tests/**/*.ts", "../src/types.ts"]
 }

packages/app/package.json

@@ -26,8 +26,11 @@
     "test:unit:watch": "bun test --conditions=solid --watch --preload ./happydom.ts ./src",
     "test:e2e": "playwright test",
     "test:e2e:local": "playwright test",
+    "test:components": "playwright test --config playwright.components.config.ts",
+    "test:components:ui": "playwright test --config playwright.components.config.ts --ui",
     "test:e2e:ui": "playwright test --ui",
     "test:e2e:report": "playwright show-report e2e/playwright-report",
+    "test:service-worker": "bun run build && playwright test --config e2e/service-worker/playwright.config.ts",
     "test:stability": "bun test ./e2e/performance/unit/visual-stability.test.ts && playwright test --config e2e/performance/timeline-stability/playwright.config.ts",
     "test:bench": "bun test ./e2e/performance/unit && playwright test --config e2e/performance/playwright.config.ts",
     "test:bench:devex": "bun test ./e2e/performance/unit/desktop-startup.test.ts && playwright test --config e2e/performance/devex/playwright.config.ts"

packages/app/playwright.components.config.ts

@@ -0,0 +1,4 @@
+import { fileURLToPath } from "node:url"
+import { componentConfig } from "../storybook/playwright/config"
+
+export default componentConfig(fileURLToPath(new URL(".", import.meta.url)))

packages/app/playwright.config.ts

@@ -9,7 +9,10 @@ const reuse = !process.env.CI
 const workers = Number(process.env.PLAYWRIGHT_WORKERS ?? (process.env.CI ? 5 : 0)) || undefined
 export default defineConfig({
   testDir: "./e2e",
-  testIgnore: process.env.OPENCODE_PERFORMANCE === "1" ? "performance/**/*.test.ts" : "performance/**",
+  testIgnore: [
+    "service-worker/**",
+    process.env.OPENCODE_PERFORMANCE === "1" ? "performance/**/*.test.ts" : "performance/**",
+  ],
   outputDir: "./e2e/test-results",
   timeout: 60_000,
   expect: {

packages/app/public/_headers

@@ -1,10 +1,10 @@
-/assets/*.js
+/_assets/*.js
   Content-Type: application/javascript
 
-/assets/*.mjs
+/_assets/*.mjs
   Content-Type: application/javascript
 
-/assets/*.css
+/_assets/*.css
   Content-Type: text/css
 
 /*.js

packages/app/src/workspaces/files/model.tsx

@@ -212,7 +212,7 @@ export const { use: useFile, provider: FileProvider } = createSimpleContext({
           {
             location: { directory: sdk().directory },
             query,
-            type: dirs === "true" ? "directory" : "file",
+            type: dirs === "true" ? undefined : "file",
             limit: options?.limit,
           },
           { signal: options?.signal },

packages/app/tsconfig.json

@@ -21,6 +21,12 @@
       "@/*": ["./src/*"]
     }
   },
-  "include": ["src", "package.json"],
+  "include": [
+    "src",
+    "component-tests",
+    "playwright.components.config.ts",
+    "../storybook/playwright/*.ts",
+    "package.json"
+  ],
   "exclude": ["dist", "ts-dist"]
 }

packages/app/vite.config.ts

@@ -29,33 +29,44 @@ export default defineConfig({
       injectRegister: false,
       manifest: false,
       workbox: {
-        cleanupOutdatedCaches: true,
         clientsClaim: false,
-        skipWaiting: false,
+        skipWaiting: true,
         inlineWorkboxRuntime: true,
-        navigateFallback: "/index.html",
-        navigateFallbackDenylist: [/^\/api(?:\/|$)/],
-        globPatterns: [
-          "index.html",
-          "site.webmanifest",
-          "favicon*",
-          "apple-touch-icon*",
-          "web-app-manifest*",
-          "assets/index-*.{js,css}",
-          "assets/session-*.js",
-          "assets/IBMPlexMono-Text-*.woff2",
-          "assets/Inter.ttf",
-          "assets/JetBrainsMonoNerdFontMono-Regular.woff2",
-        ],
+        // Always fetch the current HTML. Precaching a partial build can strand it without its chunks after an upgrade.
+        navigateFallback: null,
+        globPatterns: [],
         runtimeCaching: [
           {
-            urlPattern: ({ url }) => url.origin === self.location.origin && url.pathname.startsWith("/assets/"),
+            urlPattern: ({ url }) =>
+              url.origin === self.location.origin &&
+              (url.pathname.startsWith("/_assets/") || url.pathname.startsWith("/assets/")),
             handler: "CacheFirst",
             options: {
               cacheName: "opencode-assets",
-              cacheableResponse: {
-                statuses: [200],
-              },
+              plugins: [
+                {
+                  cachedResponseWillBeUsed: async ({ request, cachedResponse }) => {
+                    if (
+                      cachedResponse?.status === 200 &&
+                      !/^(text\/html|application\/xhtml\+xml)\b/i.test(cachedResponse.headers.get("content-type") ?? "")
+                    )
+                      return cachedResponse
+                    // Keep old tabs' precached chunks usable without retaining their stale HTML navigation handler.
+                    const response = await caches.match(request, {
+                      cacheName: `workbox-precache-v2-${self.location.origin}/`,
+                    })
+                    return response?.status === 200 &&
+                      !/^(text\/html|application\/xhtml\+xml)\b/i.test(response.headers.get("content-type") ?? "")
+                      ? response
+                      : null
+                  },
+                  cacheWillUpdate: async ({ response }) =>
+                    response.status === 200 &&
+                    !/^(text\/html|application\/xhtml\+xml)\b/i.test(response.headers.get("content-type") ?? "")
+                      ? response
+                      : null,
+                },
+              ],
               expiration: {
                 maxEntries: 1000,
               },
@@ -72,6 +83,7 @@ export default defineConfig({
     port: 3000,
   },
   build: {
+    assetsDir: "_assets",
     target: "esnext",
     sourcemap: true,
   },

packages/cli/package.json

@@ -27,7 +27,7 @@
     "@effect/platform-node": "catalog:",
     "@opencode-ai/client": "workspace:*",
     "@opencode-ai/plugin": "workspace:*",
-    "@opencode-ai/pty": "0.1.8",
+    "@opencode-ai/pty": "0.1.10",
     "@opencode-ai/schema": "workspace:*",
     "@opencode-ai/server": "workspace:*",
     "@opencode-ai/tui": "workspace:*",

packages/cli/src/services/web-ui.ts

@@ -26,6 +26,8 @@ export const handler = Effect.fn("cli.web-ui.handler")(function* (options?: { re
 
 function serveUI(request: HttpServerRequest.HttpServerRequest, url: URL, assets: AssetMap) {
   const key = url.pathname.replace(/^\//, "")
+  if (key.startsWith("_assets/") && assets[key] === undefined)
+    return Effect.succeed(HttpServerResponse.empty({ status: 404, headers: { "cache-control": "no-store" } }))
   const name = assets[key] !== undefined ? key : "index.html"
   const file = assets[name]
   if (!file) return Effect.succeed(HttpServerResponse.empty({ status: 404 }))

packages/cli/test/web-ui.test.ts

@@ -19,7 +19,7 @@ describe("web UI", () => {
     await writeFile(asset, "console.log('embedded')")
     const assets = {
       "index.html": await Bun.file(index).text(),
-      "app.js": await Bun.file(asset).text(),
+      "_assets/app.js": await Bun.file(asset).text(),
       "sw.js": "service worker",
       "registerSW.js": "registration",
       "font.woff2": new Uint8Array([0, 1, 2, 255]),
@@ -53,7 +53,16 @@ describe("web UI", () => {
           expect(missing.status).toBe(404)
           expect(yield* Effect.promise(() => missing.text())).toBe("")
 
-          const script = yield* Effect.promise(() => fetch(`${origin}/app.js`))
+          yield* Effect.forEach(["/_assets/old.js", "/_assets/old.css", "/_assets/missing"], (pathname) =>
+            Effect.gen(function* () {
+              const missing = yield* Effect.promise(() => fetch(`${origin}${pathname}`))
+              expect(missing.status).toBe(404)
+              expect(missing.headers.get("cache-control")).toBe("no-store")
+              expect(yield* Effect.promise(() => missing.text())).toBe("")
+            }),
+          )
+
+          const script = yield* Effect.promise(() => fetch(`${origin}/_assets/app.js`))
           expect(yield* Effect.promise(() => script.text())).toBe("console.log('embedded')")
           expect(script.headers.get("content-type")).toContain("javascript")
           expect(script.headers.get("cache-control")).toBe("public, max-age=31536000, immutable")
@@ -74,6 +83,14 @@ describe("web UI", () => {
           expect(yield* Effect.promise(() => fallback.text())).toContain("embedded")
           expect(fallback.headers.get("content-security-policy")).toContain("default-src 'self'")
           expect(fallback.headers.get("content-security-policy")).toContain("connect-src * data: blob:")
+
+          const dotted = yield* Effect.promise(() => fetch(`${origin}/workspace/example.js`))
+          expect(dotted.status).toBe(200)
+          expect(yield* Effect.promise(() => dotted.text())).toContain("embedded")
+
+          const legacy = yield* Effect.promise(() => fetch(`${origin}/assets/missing.js`))
+          expect(legacy.status).toBe(200)
+          expect(yield* Effect.promise(() => legacy.text())).toContain("embedded")
         }),
       ).pipe(Effect.provide(NodeFileSystem.layer)),
     )

packages/client/src/effect/api/api.ts

@@ -663,6 +663,15 @@ export type SessionLogOutput =
             readonly snapshot?: (string & Brand.Brand<"Snapshot.ID">) | undefined
           }
         }
+      | {
+          readonly id: Event.ID
+          readonly created: number
+          readonly metadata?: { readonly [x: string]: unknown } | undefined
+          readonly type: "session.step.streamed"
+          readonly durable: { readonly aggregateID: string; readonly seq: Event.Seq; readonly version: Event.Version }
+          readonly location?: Location.Ref | undefined
+          readonly data: { readonly sessionID: Session.ID; readonly assistantMessageID: SessionMessage.ID }
+        }
       | {
           readonly id: Event.ID
           readonly created: number
@@ -974,6 +983,19 @@ export type SessionLogOutput =
           readonly location?: Location.Ref | undefined
           readonly data: { readonly sessionID: Session.ID; readonly to: SessionMessage.ID }
         }
+      | {
+          readonly id: Event.ID
+          readonly created: number
+          readonly metadata?: { readonly [x: string]: unknown } | undefined
+          readonly type: "session.message.content.updated"
+          readonly durable: { readonly aggregateID: string; readonly seq: Event.Seq; readonly version: Event.Version }
+          readonly location?: Location.Ref | undefined
+          readonly data: {
+            readonly sessionID: Session.ID
+            readonly messageID: SessionMessage.ID
+            readonly content: ReadonlyArray<SessionMessage.AssistantContentEncoded>
+          }
+        }
       | {
           readonly id: Event.ID
           readonly created: number
@@ -1013,6 +1035,18 @@ export type SessionMessageInput = { readonly sessionID: Session.ID; readonly mes
 export type SessionMessageOutput = SessionMessage.Info
 export type SessionMessageOperation<E = never> = (input: SessionMessageInput) => Effect.Effect<SessionMessageOutput, E>
 
+export type SessionMessageUpdateInput = {
+  readonly sessionID: Session.ID
+  readonly messageID: SessionMessage.ID
+  readonly content: ReadonlyArray<
+    SessionMessage.AssistantText | SessionMessage.AssistantReasoning | SessionMessage.AssistantTool
+  >
+}
+export type SessionMessageUpdateOutput = SessionMessage.Assistant
+export type SessionMessageUpdateOperation<E = never> = (
+  input: SessionMessageUpdateInput,
+) => Effect.Effect<SessionMessageUpdateOutput, E>
+
 export type SessionEnvironmentInput = {
   readonly sessionID: Session.ID
   readonly variables: { readonly [x: string]: string }
@@ -1071,6 +1105,7 @@ export interface SessionApi<E = never> {
   readonly interrupt: SessionInterruptOperation<E>
   readonly background: SessionBackgroundOperation<E>
   readonly message: SessionMessageOperation<E>
+  readonly messageUpdate: SessionMessageUpdateOperation<E>
   readonly environment: SessionEnvironmentOperation<E>
   readonly view: SessionViewOperation<E>
 }

packages/client/src/effect/generated/client.ts

@@ -86,6 +86,8 @@ import type {
   SessionBackgroundOutput,
   SessionMessageInput,
   SessionMessageOutput,
+  SessionMessageUpdateInput,
+  SessionMessageUpdateOutput,
   SessionEnvironmentInput,
   SessionEnvironmentOutput,
   SessionViewInput,
@@ -651,6 +653,17 @@ const EndpointSessionMessage = (raw: RawClient["server.session"]) => (input: Ses
     ),
   )
 
+const EndpointSessionMessageUpdate = (raw: RawClient["server.session"]) => (input: SessionMessageUpdateInput) =>
+  preserveEffect<SessionMessageUpdateOutput>()(
+    raw["session.messageUpdate"]({
+      params: { sessionID: input["sessionID"], messageID: input["messageID"] },
+      payload: { content: input["content"] },
+    }).pipe(
+      Effect.mapError(mapClientError),
+      Effect.map((value) => value.data),
+    ),
+  )
+
 const EndpointSessionEnvironment = (raw: RawClient["server.session"]) => (input: SessionEnvironmentInput) =>
   preserveEffect<SessionEnvironmentOutput>()(
     raw["session.environment"]({
@@ -711,6 +724,7 @@ const adaptGroupSession = (raw: RawClient["server.session"]) => ({
   interrupt: EndpointSessionInterrupt(raw),
   background: EndpointSessionBackground(raw),
   message: EndpointSessionMessage(raw),
+  messageUpdate: EndpointSessionMessageUpdate(raw),
   environment: EndpointSessionEnvironment(raw),
   view: EndpointSessionView(raw),
 })

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

@@ -80,6 +80,8 @@ import type {
   SessionBackgroundOutput,
   SessionMessageInput,
   SessionMessageOutput,
+  SessionMessageUpdateInput,
+  SessionMessageUpdateOutput,
   SessionEnvironmentInput,
   SessionEnvironmentOutput,
   SessionViewInput,
@@ -929,6 +931,18 @@ export function make(options: ClientOptions) {
           },
           requestOptions,
         ).then((value) => value.data),
+      messageUpdate: (input: SessionMessageUpdateInput, requestOptions?: RequestOptions) =>
+        request<{ readonly data: SessionMessageUpdateOutput }>(
+          {
+            method: "PATCH",
+            path: `/api/session/${encodeURIComponent(input.sessionID)}/message/${encodeURIComponent(input.messageID)}`,
+            body: { content: input["content"] },
+            successStatus: 200,
+            declaredStatuses: [404, 400, 409, 401],
+            empty: false,
+          },
+          requestOptions,
+        ).then((value) => value.data),
       environment: (input: SessionEnvironmentInput, requestOptions?: RequestOptions) =>
         request<SessionEnvironmentOutput>(
           {

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

@@ -174,6 +174,12 @@ export type SessionMessageProviderState1 = { [x: string]: any }
 
 export type ToolFileContent1 = { type: "file"; uri: string; mime: string; name?: string | undefined }
 
+export type SessionMessageToolStateRunning1 = {
+  status: "running"
+  input: { [x: string]: any }
+  metadata: { [x: string]: JsonValue }
+}
+
 export type EventLogSynced = { type: "log.synced"; aggregateID: string; seq?: number }
 
 export type SessionInterruptResponse = { interrupted: boolean }
@@ -714,6 +720,16 @@ export type SessionStepStarted = {
   data: { sessionID: string; assistantMessageID: string; agent: string; model: ModelRef; snapshot?: string }
 }
 
+export type SessionStepStreamed = {
+  id: string
+  created: number
+  metadata?: { [x: string]: any }
+  type: "session.step.streamed"
+  durable: { aggregateID: string; seq: number; version: 1 }
+  location?: LocationRef
+  data: { sessionID: string; assistantMessageID: string }
+}
+
 export type SessionTextStarted = {
   id: string
   created: number
@@ -1315,6 +1331,15 @@ export type SessionToolCalled = {
   }
 }
 
+export type SessionMessageAssistantText1 = { type: "text"; text: string; state?: SessionMessageProviderState1 }
+
+export type SessionMessageAssistantReasoning1 = {
+  type: "reasoning"
+  text: string
+  state?: SessionMessageProviderState1
+  time?: { created: number; completed?: number }
+}
+
 export type ToolContent1 = ToolTextContent | ToolFileContent1
 
 export type ModelCompatibility = {
@@ -1743,6 +1768,21 @@ export type SessionToolFailed = {
   }
 }
 
+export type SessionMessageToolStateCompleted1 = {
+  status: "completed"
+  input: { [x: string]: any }
+  content: [ToolContent1, ...Array<ToolContent1>]
+  metadata?: { [x: string]: JsonValue }
+}
+
+export type SessionMessageToolStateError1 = {
+  status: "error"
+  input: { [x: string]: any }
+  error: SessionStructuredError
+  content?: [ToolContent1, ...Array<ToolContent1>]
+  metadata?: { [x: string]: JsonValue }
+}
+
 export type ModelInfo = {
   id: string
   modelID: string
@@ -2009,6 +2049,21 @@ export type SessionMessageAssistantTool = {
   time: { created: number; ran?: number; completed?: number }
 }
 
+export type SessionMessageAssistantTool1 = {
+  type: "tool"
+  id: string
+  name: string
+  executed?: boolean
+  providerState?: SessionMessageProviderState1
+  providerResultState?: SessionMessageProviderState1
+  state:
+    | SessionMessageToolStateStreaming
+    | SessionMessageToolStateRunning1
+    | SessionMessageToolStateCompleted1
+    | SessionMessageToolStateError1
+  time: { created: number; ran?: number; completed?: number }
+}
+
 export type FormFields = [FormField, ...Array<FormField>]
 
 export type FormFields2 = [FormField1, ...Array<FormField1>]
@@ -2028,7 +2083,7 @@ export type SessionInboxEnqueued = {
 export type SessionMessageAssistant = {
   id: string
   metadata?: { [x: string]: JsonValue }
-  time: { created: number; completed?: number }
+  time: { created: number; streamed?: number; completed?: number }
   type: "assistant"
   agent: string
   model: ModelRef
@@ -2043,6 +2098,11 @@ export type SessionMessageAssistant = {
   retry?: SessionMessageAssistantRetry
 }
 
+export type SessionMessageAssistantContentEncoded =
+  | SessionMessageAssistantText1
+  | SessionMessageAssistantReasoning1
+  | SessionMessageAssistantTool1
+
 export type IntegrationOAuthMethod = { id: string; type: "oauth"; label: string; form?: FormFields }
 
 export type IntegrationKeyMethod = { type: "key"; label?: string; form?: FormFields }
@@ -2051,6 +2111,50 @@ export type FormInfo = { id: string; sessionID: string; title: string; metadata?
 
 export type FormInfo1 = { id: string; sessionID: string; title: string; metadata?: FormMetadata1; fields: FormFields2 }
 
+export type SessionMessageInfo =
+  | SessionMessageAgentSelected
+  | SessionMessageModelSelected
+  | SessionMessageLocationSwitched
+  | SessionMessageUser
+  | SessionMessageSynthetic
+  | SessionMessageSystem
+  | SessionMessageSkill
+  | SessionMessageShell
+  | SessionMessageAssistant
+  | SessionMessageCompaction
+
+export type SessionMessageContentUpdated = {
+  id: string
+  created: number
+  metadata?: { [x: string]: any }
+  type: "session.message.content.updated"
+  durable: { aggregateID: string; seq: number; version: 1 }
+  location?: LocationRef
+  data: { sessionID: string; messageID: string; content: Array<SessionMessageAssistantContentEncoded> }
+}
+
+export type IntegrationMethod =
+  | IntegrationOAuthMethod
+  | IntegrationCommandMethod
+  | IntegrationKeyMethod
+  | IntegrationEnvMethod
+
+export type FormCreated = {
+  id: string
+  created: number
+  metadata?: { [x: string]: any }
+  type: "form.created"
+  location?: LocationRef
+  data: { form: FormInfo1 }
+}
+
+export type SessionTransferData = { info: SessionInfo; messages: Array<SessionMessageInfo> }
+
+export type SessionMessagesResponse = {
+  data: Array<SessionMessageInfo>
+  cursor: { previous?: string | null; next?: string | null }
+}
+
 export type SessionEventDurable =
   | SessionCreated
   | SessionAgentSelected
@@ -2074,6 +2178,7 @@ export type SessionEventDurable =
   | SessionShellStarted
   | SessionShellEnded
   | SessionStepStarted
+  | SessionStepStreamed
   | SessionStepEnded
   | SessionStepFailed
   | SessionTextStarted
@@ -2092,44 +2197,9 @@ export type SessionEventDurable =
   | SessionRevertStaged
   | SessionRevertCleared
   | SessionRevertCommitted
+  | SessionMessageContentUpdated
   | SessionUsageRecorded
 
-export type SessionMessageInfo =
-  | SessionMessageAgentSelected
-  | SessionMessageModelSelected
-  | SessionMessageLocationSwitched
-  | SessionMessageUser
-  | SessionMessageSynthetic
-  | SessionMessageSystem
-  | SessionMessageSkill
-  | SessionMessageShell
-  | SessionMessageAssistant
-  | SessionMessageCompaction
-
-export type IntegrationMethod =
-  | IntegrationOAuthMethod
-  | IntegrationCommandMethod
-  | IntegrationKeyMethod
-  | IntegrationEnvMethod
-
-export type FormCreated = {
-  id: string
-  created: number
-  metadata?: { [x: string]: any }
-  type: "form.created"
-  location?: LocationRef
-  data: { form: FormInfo1 }
-}
-
-export type SessionLogItem = SessionEventDurable | EventLogSynced
-
-export type SessionTransferData = { info: SessionInfo; messages: Array<SessionMessageInfo> }
-
-export type SessionMessagesResponse = {
-  data: Array<SessionMessageInfo>
-  cursor: { previous?: string | null; next?: string | null }
-}
-
 export type IntegrationInfo = {
   id: string
   name: string
@@ -2168,6 +2238,7 @@ export type V2Event =
   | SessionShellStarted
   | SessionShellEnded
   | SessionStepStarted
+  | SessionStepStreamed
   | SessionStepEnded
   | SessionStepFailed
   | SessionTextStarted
@@ -2191,6 +2262,7 @@ export type V2Event =
   | SessionRevertStaged
   | SessionRevertCleared
   | SessionRevertCommitted
+  | SessionMessageContentUpdated
   | FilesystemChanged
   | ReferenceUpdated
   | PermissionAsked
@@ -2229,6 +2301,8 @@ export type V2Event =
   | McpResourcesChanged
   | V2EventServerConnected
 
+export type SessionLogItem = SessionEventDurable | EventLogSynced
+
 export type UnauthorizedError = { readonly _tag: "UnauthorizedError"; readonly message: string }
 export const isUnauthorizedError = (value: unknown): value is UnauthorizedError =>
   typeof value === "object" && value !== null && "_tag" in value && value["_tag"] === "UnauthorizedError"
@@ -2789,7 +2863,7 @@ export type SessionImportInput = {
       | {
           readonly id: string
           readonly metadata?: { readonly [x: string]: JsonValue }
-          readonly time: { readonly created: number; readonly completed?: number }
+          readonly time: { readonly created: number; readonly streamed?: number; readonly completed?: number }
           readonly type: "assistant"
           readonly agent: string
           readonly model: { readonly id: string; readonly providerID: string; readonly variant?: string }
@@ -3065,7 +3139,7 @@ export type SessionImportInput = {
       | {
           readonly id: string
           readonly metadata?: { readonly [x: string]: JsonValue }
-          readonly time: { readonly created: number; readonly completed?: number }
+          readonly time: { readonly created: number; readonly streamed?: number; readonly completed?: number }
           readonly type: "assistant"
           readonly agent: string
           readonly model: { readonly id: string; readonly providerID: string; readonly variant?: string }
@@ -3341,7 +3415,7 @@ export type SessionImportInput = {
       | {
           readonly id: string
           readonly metadata?: { readonly [x: string]: JsonValue }
-          readonly time: { readonly created: number; readonly completed?: number }
+          readonly time: { readonly created: number; readonly streamed?: number; readonly completed?: number }
           readonly type: "assistant"
           readonly agent: string
           readonly model: { readonly id: string; readonly providerID: string; readonly variant?: string }
@@ -4036,6 +4110,91 @@ export type SessionMessageInput = {
 
 export type SessionMessageOutput = { data: SessionMessageInfo }["data"]
 
+export type SessionMessageUpdateInput = {
+  readonly sessionID: { readonly sessionID: string; readonly messageID: string }["sessionID"]
+  readonly messageID: { readonly sessionID: string; readonly messageID: string }["messageID"]
+  readonly content: {
+    readonly content: ReadonlyArray<
+      | { readonly type: "text"; readonly text: string; readonly state?: { readonly [x: string]: JsonValue } }
+      | {
+          readonly type: "reasoning"
+          readonly text: string
+          readonly state?: { readonly [x: string]: JsonValue }
+          readonly time?: { readonly created: number; readonly completed?: number }
+        }
+      | {
+          readonly type: "tool"
+          readonly id: string
+          readonly name: string
+          readonly executed?: boolean
+          readonly providerState?: { readonly [x: string]: JsonValue }
+          readonly providerResultState?: { readonly [x: string]: JsonValue }
+          readonly state:
+            | { readonly status: "streaming"; readonly input: string }
+            | {
+                readonly status: "running"
+                readonly input: { readonly [x: string]: JsonValue }
+                readonly metadata: { readonly [x: string]: JsonValue }
+              }
+            | {
+                readonly status: "completed"
+                readonly input: { readonly [x: string]: JsonValue }
+                readonly content: readonly [
+                  (
+                    | { readonly type: "text"; readonly text: string }
+                    | {
+                        readonly type: "file"
+                        readonly uri: string
+                        readonly mime: string
+                        readonly name?: string | null
+                      }
+                  ),
+                  ...Array<
+                    | { readonly type: "text"; readonly text: string }
+                    | {
+                        readonly type: "file"
+                        readonly uri: string
+                        readonly mime: string
+                        readonly name?: string | null
+                      }
+                  >,
+                ]
+                readonly metadata?: { readonly [x: string]: JsonValue }
+              }
+            | {
+                readonly status: "error"
+                readonly input: { readonly [x: string]: JsonValue }
+                readonly error: { readonly type: string; readonly message: string; readonly status?: number }
+                readonly content?: readonly [
+                  (
+                    | { readonly type: "text"; readonly text: string }
+                    | {
+                        readonly type: "file"
+                        readonly uri: string
+                        readonly mime: string
+                        readonly name?: string | null
+                      }
+                  ),
+                  ...Array<
+                    | { readonly type: "text"; readonly text: string }
+                    | {
+                        readonly type: "file"
+                        readonly uri: string
+                        readonly mime: string
+                        readonly name?: string | null
+                      }
+                  >,
+                ]
+                readonly metadata?: { readonly [x: string]: JsonValue }
+              }
+          readonly time: { readonly created: number; readonly ran?: number; readonly completed?: number }
+        }
+    >
+  }["content"]
+}
+
+export type SessionMessageUpdateOutput = { data: SessionMessageAssistant }["data"]
+
 export type SessionEnvironmentInput = {
   readonly sessionID: { readonly sessionID: string }["sessionID"]
   readonly variables: { readonly variables: { readonly [x: string]: string } }["variables"]

packages/client/src/solid/data.ts

@@ -167,6 +167,10 @@ function createSync() {
     has(key: string) {
       return state.has(key)
     },
+    pending(key: string) {
+      const active = state.get(key)
+      return active !== undefined && active !== true
+    },
     invalidate(key?: string) {
       if (key) {
         const active = state.get(key)
@@ -723,6 +727,17 @@ export function createData(config: CreateDataInput) {
           match.time.completed = event.created
         })
         return
+      case "session.message.content.updated": {
+        if (store.session.message[event.data.sessionID])
+          message.update(event.data.sessionID, (draft, index) => {
+            const assistant = message.assistant(draft, index, event.data.messageID)
+            if (assistant) assistant.content = [...event.data.content]
+          })
+        if (!sync.pending(`session.message:${event.data.sessionID}`)) return
+        result.session.message.invalidate(event.data.sessionID)
+        void result.session.message.sync(event.data.sessionID)
+        return
+      }
       case "session.step.started":
         message.update(event.data.sessionID, (draft, index) => {
           const position = index.get(event.data.assistantMessageID)
@@ -735,6 +750,7 @@ export function createData(config: CreateDataInput) {
             existing.finish = undefined
             existing.rawFinish = undefined
             existing.providerState = undefined
+            existing.time.streamed = undefined
             existing.time.completed = undefined
             if (event.data.snapshot) existing.snapshot = { ...existing.snapshot, start: event.data.snapshot }
             return
@@ -756,6 +772,12 @@ export function createData(config: CreateDataInput) {
           })
         })
         return
+      case "session.step.streamed":
+        message.update(event.data.sessionID, (draft, index) => {
+          const currentAssistant = message.assistant(draft, index, event.data.assistantMessageID)
+          if (currentAssistant) currentAssistant.time.streamed = event.created
+        })
+        return
       case "session.step.ended": {
         message.update(event.data.sessionID, (draft, index) => {
           const currentAssistant = message.assistant(draft, index, event.data.assistantMessageID)

packages/client/src/solid/pty.ts

@@ -1,4 +1,4 @@
-import type { OpenCodeClient, PtyConnectTokenInput } from "../promise"
+import type { ExperimentalPersistentPtyConnectTokenInput, OpenCodeClient, PtyConnectTokenInput } from "../promise"
 
 export type PtyClientOptions = {
   readonly url: string
@@ -11,6 +11,13 @@ export type PtyConnectInput = {
   readonly cursor?: number
 }
 
+export type PersistentPtyConnectInput = {
+  readonly ptyID: ExperimentalPersistentPtyConnectTokenInput["ptyID"]
+  readonly cursor: number
+  readonly attachmentID: string
+  readonly takeover?: boolean
+}
+
 export function createPtyClient(api: OpenCodeClient, options: PtyClientOptions) {
   return {
     async connect(input: PtyConnectInput) {
@@ -32,3 +39,25 @@ export function createPtyClient(api: OpenCodeClient, options: PtyClientOptions)
     },
   }
 }
+
+export function createPersistentPtyClient(api: OpenCodeClient, options: PtyClientOptions) {
+  return {
+    async connect(input: PersistentPtyConnectInput) {
+      const token = await api.experimental.persistentPty.connectToken({
+        ptyID: input.ptyID,
+        "x-opencode-ticket": "1",
+      })
+      const url = new URL(`/api/experimental/persistent-pty/${encodeURIComponent(input.ptyID)}/connect`, options.url)
+      url.searchParams.set("ticket", token.ticket)
+      url.searchParams.set("cursor", String(input.cursor))
+      url.searchParams.set("attachment_id", input.attachmentID)
+      url.searchParams.set("takeover", String(input.takeover ?? false))
+      url.searchParams.set("input_protocol", "1")
+      url.protocol = url.protocol === "https:" ? "wss:" : "ws:"
+
+      const socket = options.openSocket?.(url) ?? new WebSocket(url)
+      socket.binaryType = "arraybuffer"
+      return socket
+    },
+  }
+}

packages/client/test/promise.test.ts

@@ -514,6 +514,115 @@ test("event.subscribe exposes the Promise event stream wire projection", async (
   expect(events[1]?.type === "session.model.selected" && events[1].created).toBe(1_717_171_717_000)
 })
 
+// Moved from packages/app/e2e/regression/session-timeline-transport.spec.ts
+test("event.subscribe keeps one request open while delivering multiple events", async () => {
+  const requests: Request[] = []
+  const events = [
+    { id: "evt_first", created: 1, type: "server.connected", data: {} },
+    { id: "evt_second", created: 2, type: "server.connected", data: {} },
+  ]
+  const client = OpenCode.make({
+    baseUrl: "http://localhost:3000",
+    fetch: async (input, init) => {
+      requests.push(input instanceof Request ? input : new Request(input, init))
+      return new Response(events.map((event) => `data: ${JSON.stringify(event)}\n\n`).join(""), {
+        headers: { "content-type": "text/event-stream" },
+      })
+    },
+  })
+  const received = []
+  for await (const event of client.event.subscribe()) received.push(event)
+  expect(received).toEqual(events)
+  expect(requests).toHaveLength(1)
+})
+
+// Moved from packages/app/e2e/regression/session-timeline-transport.spec.ts
+test("event.subscribe delivers every event from one stream chunk", async () => {
+  const events = Array.from({ length: 4 }, (_, index) => ({
+    id: `evt_burst_${index}`,
+    created: index,
+    type: "server.connected",
+    data: {},
+  }))
+  const client = OpenCode.make({
+    baseUrl: "http://localhost:3000",
+    fetch: async () =>
+      new Response(new TextEncoder().encode(events.map((event) => `data: ${JSON.stringify(event)}\n\n`).join("")), {
+        headers: { "content-type": "text/event-stream" },
+      }),
+  })
+  const received = []
+  for await (const event of client.event.subscribe()) received.push(event)
+  expect(received).toEqual(events)
+  expect(new Set(received.map((event) => event.id)).size).toBe(4)
+})
+
+// Moved from packages/app/e2e/regression/session-timeline-transport.spec.ts
+test("event.subscribe parses split JSON and a split multibyte code point", async () => {
+  const event = {
+    id: "evt_split",
+    created: 1,
+    type: "server.connected",
+    data: { text: "split snowman \u2603\u2603\u2603" },
+  }
+  const encoded = new TextEncoder().encode(`data: ${JSON.stringify(event)}\n\n`)
+  const multibyte = encoded.indexOf(new TextEncoder().encode("\u2603")[0]!)
+  const boundaries = [9, multibyte + 1, multibyte + 2, encoded.length]
+  const client = OpenCode.make({
+    baseUrl: "http://localhost:3000",
+    fetch: async () =>
+      new Response(
+        new ReadableStream({
+          start(controller) {
+            boundaries.forEach((end, index) =>
+              controller.enqueue(encoded.slice(index ? boundaries[index - 1] : 0, end)),
+            )
+            controller.close()
+          },
+        }),
+        { headers: { "content-type": "text/event-stream" } },
+      ),
+  })
+  await expect(client.event.subscribe()[Symbol.asyncIterator]().next()).resolves.toEqual({ done: false, value: event })
+})
+
+// Moved from packages/app/e2e/regression/session-timeline-transport.spec.ts
+test("event.subscribe ignores server heartbeat comments", async () => {
+  const event = { id: "evt_sentinel", created: 1, type: "server.connected", data: {} }
+  const client = OpenCode.make({
+    baseUrl: "http://localhost:3000",
+    fetch: async () =>
+      new Response(`: heartbeat\n\ndata: ${JSON.stringify(event)}\n\n: heartbeat\n\n`, {
+        headers: { "content-type": "text/event-stream" },
+      }),
+  })
+  const received = []
+  for await (const item of client.event.subscribe()) received.push(item)
+  expect(received).toEqual([event])
+})
+
+// Moved from packages/app/e2e/regression/session-timeline-transport.spec.ts
+test("event transport passes through ordinary health requests", async () => {
+  const requests: string[] = []
+  const event = { id: "evt_connected", created: 1, type: "server.connected", data: {} }
+  const client = OpenCode.make({
+    baseUrl: "http://localhost:3000",
+    fetch: async (input, init) => {
+      const request = input instanceof Request ? input : new Request(input, init)
+      requests.push(new URL(request.url).pathname)
+      if (new URL(request.url).pathname === "/api/event") {
+        return new Response(`data: ${JSON.stringify(event)}\n\n`, {
+          headers: { "content-type": "text/event-stream" },
+        })
+      }
+      return Response.json({ healthy: true, version: "2.0.0", pid: 1 })
+    },
+  })
+  await expect(client.event.subscribe()[Symbol.asyncIterator]().next()).resolves.toEqual({ done: false, value: event })
+  await expect(client.health.get()).resolves.toEqual({ healthy: true, version: "2.0.0", pid: 1 })
+  expect(requests).toEqual(["/api/event", "/api/health"])
+})
+
 test("event.subscribe terminates on malformed Promise SSE data", async () => {
   const client = OpenCode.make({
     baseUrl: "http://localhost:3000",

packages/client/test/solid-data.test.ts

@@ -414,6 +414,77 @@ test("loads bounded message pages", async () => {
   }
 })
 
+test("preserves assistant content replacement events across an active message read", async () => {
+  const listeners = new Set<Parameters<CreateDataInput["event"]["listen"]>[0]>()
+  const release = Promise.withResolvers<void>()
+  let requests = 0
+  const content = [
+    { type: "text" as const, text: "replacement" },
+    { type: "reasoning" as const, text: "reasoning", time: { created: 3 } },
+  ]
+  const api = OpenCode.make({
+    baseUrl: "http://opencode.local",
+    fetch: async () => {
+      const current = ++requests
+      if (current === 2) await release.promise
+      return Response.json({
+        data: [
+          {
+            id: "msg_assistant",
+            type: "assistant",
+            agent: "build",
+            model: { id: "model", providerID: "provider" },
+            content: current === 3 ? content : [{ type: "text", text: "original" }],
+            time: { created: 1, completed: 2 },
+          },
+        ],
+        cursor: {},
+      })
+    },
+  })
+  const setup = createRoot((dispose) => ({
+    data: createData({
+      api: () => api,
+      directory: "/project",
+      event: {
+        on: () => () => {},
+        listen(handler) {
+          listeners.add(handler)
+          return () => listeners.delete(handler)
+        },
+      },
+    }),
+    dispose,
+  }))
+
+  try {
+    await setup.data.session.message.sync("ses_refresh")
+    setup.data.session.message.invalidate("ses_refresh")
+    const stale = setup.data.session.message.sync("ses_refresh")
+    await wait(() => requests === 2)
+    const updated: OpenCodeEvent = {
+      id: "evt_message_updated",
+      created: 3,
+      type: "session.message.content.updated",
+      durable: { aggregateID: "ses_refresh", seq: 3, version: 1 },
+      data: {
+        sessionID: "ses_refresh",
+        messageID: "msg_assistant",
+        content,
+      },
+    }
+    listeners.forEach((listener) => listener({ name: updated.type, details: updated }))
+
+    expect(setup.data.session.message.list("ses_refresh")[0]).toMatchObject({ content })
+    release.resolve()
+    await stale
+    await wait(() => requests === 3)
+    expect(setup.data.session.message.list("ses_refresh")[0]).toMatchObject({ content })
+  } finally {
+    setup.dispose()
+  }
+})
+
 async function wait(check: () => boolean) {
   const started = Date.now()
   while (!check()) {

packages/client/test/solid-pty.test.ts

@@ -1,6 +1,6 @@
 import { describe, expect, test } from "bun:test"
 import { OpenCode } from "../src/promise"
-import { createPtyClient } from "../src/solid"
+import { createPersistentPtyClient, createPtyClient } from "../src/solid"
 
 describe("createPtyClient", () => {
   test("mints an authenticated ticket before opening the terminal socket", async () => {
@@ -66,3 +66,36 @@ describe("createPtyClient", () => {
     expect(opened).toBe(false)
   })
 })
+
+describe("createPersistentPtyClient", () => {
+  test("mints an authenticated ticket before opening the persistent terminal socket", async () => {
+    let request: Request | undefined
+    let socketURL: URL | undefined
+    const socket = { binaryType: "blob" } as unknown as WebSocket
+    const api = OpenCode.make({
+      baseUrl: "https://server.example/base",
+      headers: { Authorization: "Basic credential" },
+      fetch: async (input, init) => {
+        request = input instanceof Request ? input : new Request(input, init)
+        return Response.json({ data: { ticket: "persistent-ticket", expires_in: 60 } })
+      },
+    })
+    const pty = createPersistentPtyClient(api, {
+      url: "https://server.example/base",
+      openSocket(url) {
+        socketURL = url
+        return socket
+      },
+    })
+
+    expect(await pty.connect({ ptyID: "pty_1", cursor: 42, attachmentID: "attachment_1", takeover: true })).toBe(socket)
+    expect(request?.method).toBe("POST")
+    expect(request?.url).toBe("https://server.example/api/experimental/persistent-pty/pty_1/connect-token")
+    expect(request?.headers.get("authorization")).toBe("Basic credential")
+    expect(request?.headers.get("x-opencode-ticket")).toBe("1")
+    expect(socketURL?.toString()).toBe(
+      "wss://server.example/api/experimental/persistent-pty/pty_1/connect?ticket=persistent-ticket&cursor=42&attachment_id=attachment_1&takeover=true&input_protocol=1",
+    )
+    expect(socket.binaryType).toBe("arraybuffer")
+  })
+})

packages/core/package.json

@@ -105,7 +105,6 @@
     "@ai-sdk/cohere": "3.0.27",
     "@ai-sdk/gateway": "3.0.104",
     "@ai-sdk/google-vertex": "4.0.128",
-    "@ai-sdk/groq": "3.0.31",
     "@ai-sdk/mistral": "3.0.51",
     "@ai-sdk/openai-compatible": "2.0.41",
     "@ai-sdk/perplexity": "3.0.26",
@@ -119,7 +118,7 @@
     "@ff-labs/fff-node": "0.10.5",
     "@opencode-ai/codemode": "workspace:*",
     "@opencode-ai/ai": "workspace:*",
-    "@opencode-ai/pty": "0.1.8",
+    "@opencode-ai/pty": "0.1.10",
     "@opencode-ai/schema": "workspace:*",
     "@opencode-ai/plugin": "workspace:*",
     "@opencode-ai/util": "workspace:*",

packages/core/src/aisdk-native.ts

@@ -55,6 +55,7 @@ export function map(input: MapInput): Mapping | undefined {
       }
     case "@ai-sdk/cerebras":
     case "@ai-sdk/deepinfra":
+    case "@ai-sdk/groq":
     case "@ai-sdk/togetherai":
       return {
         package: `@opencode-ai/ai/providers/${input.packageName.slice("@ai-sdk/".length)}`,

packages/core/src/job.ts

@@ -1,11 +1,41 @@
 export * as Job from "./job.js"
 
-import { Cause, Clock, Context, Deferred, Effect, Exit, Layer, Scope, SynchronizedRef } from "effect"
+import { Array, Cause, Clock, Context, Deferred, Effect, Exit, Layer, Schema, Scope, SynchronizedRef } from "effect"
 import { makeGlobalNode } from "@opencode-ai/util/effect/app-node"
 import { Identifier } from "./id/id.js"
+import { KV } from "./kv.js"
+import { SessionMessage } from "./session/message.js"
 import { SessionSchema } from "./session/schema.js"
 
-export type Status = "running" | "completed" | "error" | "cancelled"
+const Background = Schema.Struct({
+  id: Schema.String,
+  notificationID: SessionMessage.ID,
+  recovery: Schema.Union([
+    Schema.Struct({
+      kind: Schema.Literal("shell"),
+      sessionID: SessionSchema.ID,
+      shellID: Schema.String,
+      command: Schema.String,
+    }),
+    Schema.Struct({
+      kind: Schema.Literal("subagent"),
+      parentSessionID: SessionSchema.ID,
+      childSessionID: SessionSchema.ID,
+      agent: Schema.String,
+      description: Schema.String,
+    }),
+  ]),
+  status: Schema.Literals(["running", "completed", "error", "cancelled"]),
+  output: Schema.optionalKey(Schema.String),
+  error: Schema.optionalKey(Schema.String),
+})
+
+export type Background = typeof Background.Type
+export type Recovery = Background["recovery"]
+export type Status = Background["status"]
+
+const decodeBackground = Schema.decodeUnknownResult(Background)
+const backgroundPrefix = "job.background/"
 
 export type Info = {
   id: string
@@ -17,6 +47,7 @@ export type Info = {
   output?: string
   error?: string
   metadata?: Record<string, unknown>
+  notificationID?: SessionMessage.ID
 }
 
 type Active = {
@@ -27,6 +58,7 @@ type Active = {
   token: object
   blockingSessions: Map<SessionSchema.ID, number>
   isBackgrounded: boolean
+  recovery?: Recovery
 }
 
 type State = {
@@ -63,6 +95,8 @@ export type StartInput = {
   type: string
   title?: string
   metadata?: Record<string, unknown>
+  recovery?: Recovery
+  notificationID?: SessionMessage.ID
   run: Effect.Effect<string, unknown>
 }
 
@@ -96,6 +130,8 @@ export interface Interface {
   readonly background: (id: string) => Effect.Effect<Info | undefined>
   readonly backgroundAll: (input: BackgroundAllInput) => Effect.Effect<Info[]>
   readonly cancel: (id: string) => Effect.Effect<Info | undefined>
+  readonly pendingBackground: Effect.Effect<readonly Background[]>
+  readonly completeBackground: (notificationID: SessionMessage.ID) => Effect.Effect<void>
 }
 
 export class Service extends Context.Service<Service, Interface>()("@opencode/Job") {}
@@ -126,43 +162,57 @@ function decrementSession(input: Map<SessionSchema.ID, number>, sessionID: Sessi
 }
 
 /**
- * Makes one scoped, process-local registry. Entries are intentionally not
- * durable: process restart or owner-scope closure loses status and interrupts
- * live work. Persisted observation, restart recovery, and remote workers need a
- * separate durable ownership slice rather than pretending this registry has
- * those semantics.
+ * Makes one scoped, process-local registry. Explicitly recoverable background
+ * work also owns a durable notification marker until its notification is admitted.
  */
 export const make = Effect.gen(function* () {
+  const kv = yield* KV.Service
   const state: State = {
     jobs: yield* SynchronizedRef.make(new Map()),
     scope: yield* Scope.Scope,
   }
 
+  const persistBackground = Effect.fnUntraced(function* (job: Active) {
+    if (!job.recovery || !job.info.notificationID) return
+    yield* kv.set(`${backgroundPrefix}${job.info.notificationID}`, {
+      id: job.info.id,
+      notificationID: job.info.notificationID,
+      recovery: job.recovery,
+      status: job.info.status,
+      ...(job.info.output !== undefined ? { output: job.info.output } : {}),
+      ...(job.info.error !== undefined ? { error: job.info.error } : {}),
+    })
+  })
+
   const settle = Effect.fnUntraced(function* (id: string, token: object, exit: Exit.Exit<string, unknown>) {
     const completed_at = yield* Clock.currentTimeMillis
-    const result = yield* SynchronizedRef.modify(state.jobs, (jobs): readonly [FinishResult, Map<string, Active>] => {
-      const job = jobs.get(id)
-      if (!job) return [{}, jobs]
-      if (job.token !== token) return [{}, jobs]
-      if (job.info.status !== "running") return [{ info: snapshot(job) }, jobs]
-      const status: Exclude<Status, "running"> = Exit.isSuccess(exit)
-        ? "completed"
-        : Cause.hasInterruptsOnly(exit.cause)
-          ? "cancelled"
-          : "error"
-      const next = {
-        ...job,
-        blockingSessions: new Map<SessionSchema.ID, number>(),
-        info: {
-          ...job.info,
-          status,
-          completed_at,
-          ...(Exit.isSuccess(exit) ? { output: exit.value } : {}),
-          ...(Exit.isFailure(exit) ? { error: errorText(Cause.squash(exit.cause)) } : {}),
-        },
-      }
-      return [{ info: snapshot(next), done: job.done, scope: job.scope }, new Map(jobs).set(id, next)]
-    })
+    const result = yield* SynchronizedRef.modifyEffect(
+      state.jobs,
+      Effect.fnUntraced(function* (jobs): Effect.fn.Return<readonly [FinishResult, Map<string, Active>]> {
+        const job = jobs.get(id)
+        if (!job) return [{}, jobs]
+        if (job.token !== token) return [{}, jobs]
+        if (job.info.status !== "running") return [{ info: snapshot(job) }, jobs]
+        const status: Exclude<Status, "running"> = Exit.isSuccess(exit)
+          ? "completed"
+          : Cause.hasInterruptsOnly(exit.cause)
+            ? "cancelled"
+            : "error"
+        const next = {
+          ...job,
+          blockingSessions: new Map<SessionSchema.ID, number>(),
+          info: {
+            ...job.info,
+            status,
+            completed_at,
+            ...(Exit.isSuccess(exit) ? { output: exit.value } : {}),
+            ...(Exit.isFailure(exit) ? { error: errorText(Cause.squash(exit.cause)) } : {}),
+          },
+        }
+        if (status !== "cancelled") yield* persistBackground(next)
+        return [{ info: snapshot(next), done: job.done, scope: job.scope }, new Map(jobs).set(id, next)]
+      }),
+    )
     if (result.info && result.done) yield* Deferred.succeed(result.done, result.info).pipe(Effect.ignore)
     if (result.scope) {
       yield* Scope.close(result.scope, Exit.void).pipe(Effect.forkIn(state.scope, { startImmediately: true }))
@@ -170,22 +220,6 @@ export const make = Effect.gen(function* () {
     return result.info
   })
 
-  const fork = Effect.fnUntraced(function* (
-    scope: Scope.Scope,
-    id: string,
-    token: object,
-    run: Effect.Effect<string, unknown>,
-  ) {
-    return yield* run.pipe(
-      Effect.matchCauseEffect({
-        onSuccess: (output) => settle(id, token, Exit.succeed(output)),
-        onFailure: (cause) => settle(id, token, Exit.failCause(cause)),
-      }),
-      Effect.asVoid,
-      Effect.forkIn(scope, { startImmediately: true }),
-    )
-  })
-
   const get: Interface["get"] = Effect.fn("Job.get")(function* (id) {
     const job = (yield* SynchronizedRef.get(state.jobs)).get(id)
     if (!job) return undefined
@@ -201,10 +235,10 @@ export const make = Effect.gen(function* () {
         const backgrounded = yield* Deferred.make<Info>()
         const result = yield* SynchronizedRef.modifyEffect(
           state.jobs,
-          Effect.fnUntraced(function* (jobs) {
+          Effect.fnUntraced(function* (jobs): Effect.fn.Return<readonly [StartResult, Map<string, Active>]> {
             const existing = jobs.get(id)
             if (existing?.info.status === "running") {
-              return [{ info: snapshot(existing) }, jobs] as readonly [StartResult, Map<string, Active>]
+              return [{ info: snapshot(existing) }, jobs]
             }
             const scope = yield* Scope.fork(state.scope, "parallel")
             const token = {}
@@ -216,6 +250,7 @@ export const make = Effect.gen(function* () {
                 status: "running" as const,
                 started_at,
                 metadata: input.metadata,
+                ...(input.notificationID ? { notificationID: input.notificationID } : {}),
               },
               done,
               backgrounded,
@@ -223,14 +258,18 @@ export const make = Effect.gen(function* () {
               token,
               blockingSessions: new Map<SessionSchema.ID, number>(),
               isBackgrounded: false,
+              recovery: input.recovery,
             }
-            return [{ info: snapshot(job), scope, token }, new Map(jobs).set(id, job)] as readonly [
-              StartResult,
-              Map<string, Active>,
-            ]
+            return [{ info: snapshot(job), scope, token }, new Map(jobs).set(id, job)]
           }),
         )
-        if ("scope" in result) yield* fork(result.scope, id, result.token, restore(input.run))
+        if ("scope" in result)
+          yield* restore(input.run).pipe(
+            Effect.exit,
+            Effect.flatMap((exit) => settle(id, result.token, exit)),
+            Effect.asVoid,
+            Effect.forkIn(result.scope, { startImmediately: true }),
+          )
         return result.info
       }),
     )
@@ -281,20 +320,31 @@ export const make = Effect.gen(function* () {
     ).pipe(Effect.ensuring(removeBlock(input)))
   })
 
-  const background: Interface["background"] = Effect.fn("Job.background")(function* (id) {
-    const result = yield* SynchronizedRef.modify(
-      state.jobs,
-      (jobs): readonly [BackgroundResult, Map<string, Active>] => {
-        const job = jobs.get(id)
-        if (!job || job.info.status !== "running") return [{}, jobs]
-        if (job.isBackgrounded) return [{ info: snapshot(job) }, jobs]
-        const next = {
-          ...job,
-          isBackgrounded: true,
-          blockingSessions: new Map<SessionSchema.ID, number>(),
-        }
-        return [{ info: snapshot(next), backgrounded: job.backgrounded }, new Map(jobs).set(id, next)]
+  const markBackground = Effect.fnUntraced(function* (job: Active) {
+    const next = {
+      ...job,
+      isBackgrounded: true,
+      blockingSessions: new Map<SessionSchema.ID, number>(),
+      info: {
+        ...job.info,
+        ...(job.recovery ? { notificationID: job.info.notificationID ?? SessionMessage.ID.create() } : {}),
       },
+    }
+    yield* persistBackground(next)
+    return next
+  })
+
+  const background: Interface["background"] = Effect.fn("Job.background")(function* (id) {
+    const result = yield* SynchronizedRef.modifyEffect(
+      state.jobs,
+      Effect.fnUntraced(function* (jobs): Effect.fn.Return<readonly [BackgroundResult, Map<string, Active>]> {
+        const job = jobs.get(id)
+        // Recoverable work may finish before the caller backgrounds it.
+        if (!job || (job.info.status !== "running" && !job.recovery)) return [{}, jobs]
+        if (job.isBackgrounded) return [{ info: snapshot(job) }, jobs]
+        const next = yield* markBackground(job)
+        return [{ info: snapshot(next), backgrounded: job.backgrounded }, new Map(jobs).set(id, next)]
+      }),
     )
     if (result.info && result.backgrounded)
       yield* Deferred.succeed(result.backgrounded, result.info).pipe(Effect.ignore)
@@ -302,60 +352,83 @@ export const make = Effect.gen(function* () {
   })
 
   const backgroundAll: Interface["backgroundAll"] = Effect.fn("Job.backgroundAll")(function* (input) {
-    const result = yield* SynchronizedRef.modify(
+    const result = yield* SynchronizedRef.modifyEffect(
       state.jobs,
-      (jobs): readonly [BackgroundResult[], Map<string, Active>] => {
-        const results: BackgroundResult[] = []
+      Effect.fnUntraced(function* (jobs): Effect.fn.Return<
+        readonly [Required<BackgroundResult>[], Map<string, Active>]
+      > {
+        const results: Required<BackgroundResult>[] = []
         const next = new Map(jobs)
         for (const [id, job] of jobs) {
           if (job.info.status !== "running") continue
           if (job.isBackgrounded) continue
           if (input.type !== undefined && job.info.type !== input.type) continue
           if (!job.blockingSessions.has(input.sessionID)) continue
-          const updated = {
-            ...job,
-            isBackgrounded: true,
-            blockingSessions: new Map<SessionSchema.ID, number>(),
-          }
+          const updated = yield* markBackground(job)
           results.push({ info: snapshot(updated), backgrounded: job.backgrounded })
           next.set(id, updated)
         }
         return [results, next]
-      },
+      }),
     )
-    yield* Effect.forEach(
-      result,
-      (item) => (item.info && item.backgrounded ? Deferred.succeed(item.backgrounded, item.info) : Effect.void),
-      { discard: true },
-    )
-    return result.flatMap((item) => (item.info ? [item.info] : []))
+    yield* Effect.forEach(result, (item) => Deferred.succeed(item.backgrounded, item.info), { discard: true })
+    return result.map((item) => item.info)
   })
 
   const cancel: Interface["cancel"] = Effect.fn("Job.cancel")(function* (id) {
     const completed_at = yield* Clock.currentTimeMillis
-    const result = yield* SynchronizedRef.modify(state.jobs, (jobs): readonly [FinishResult, Map<string, Active>] => {
-      const job = jobs.get(id)
-      if (!job) return [{}, jobs]
-      if (job.info.status !== "running") return [{ info: snapshot(job) }, jobs]
-      const next = {
-        ...job,
-        blockingSessions: new Map<SessionSchema.ID, number>(),
-        info: {
-          ...job.info,
-          status: "cancelled" as const,
-          completed_at,
-        },
-      }
-      return [{ info: snapshot(next), done: job.done, scope: job.scope }, new Map(jobs).set(id, next)]
-    })
+    const result = yield* SynchronizedRef.modifyEffect(
+      state.jobs,
+      Effect.fnUntraced(function* (jobs): Effect.fn.Return<readonly [FinishResult, Map<string, Active>]> {
+        const job = jobs.get(id)
+        if (!job) return [{}, jobs]
+        if (job.info.status !== "running") return [{ info: snapshot(job) }, jobs]
+        const next = {
+          ...job,
+          blockingSessions: new Map<SessionSchema.ID, number>(),
+          info: {
+            ...job.info,
+            status: "cancelled" as const,
+            completed_at,
+          },
+        }
+        yield* persistBackground(next)
+        return [{ info: snapshot(next), done: job.done, scope: job.scope }, new Map(jobs).set(id, next)]
+      }),
+    )
     if (result.info && result.done) yield* Deferred.succeed(result.done, result.info).pipe(Effect.ignore)
     if (result.scope) yield* Scope.close(result.scope, Exit.void)
     return result.info
   })
 
-  return Service.of({ get, start, wait, block, background, backgroundAll, cancel })
+  const pendingBackground: Interface["pendingBackground"] = Effect.gen(function* () {
+    const recovered: Background[] = []
+    let after: string | undefined
+    do {
+      const page = yield* kv.scan({ prefix: backgroundPrefix, after })
+      recovered.push(...Array.filterMap(page.entries, (entry) => decodeBackground(entry.value)))
+      after = page.next
+    } while (after)
+    return recovered
+  }).pipe(Effect.withSpan("Job.pendingBackground"))
+
+  const completeBackground: Interface["completeBackground"] = Effect.fn("Job.completeBackground")((notificationID) =>
+    kv.remove(`${backgroundPrefix}${notificationID}`),
+  )
+
+  return Service.of({
+    get,
+    start,
+    wait,
+    block,
+    background,
+    backgroundAll,
+    cancel,
+    pendingBackground,
+    completeBackground,
+  })
 })
 
 const layer = Layer.effect(Service, make)
 
-export const node = makeGlobalNode({ service: Service, layer, deps: [] })
+export const node = makeGlobalNode({ service: Service, layer, deps: [KV.node] })

packages/core/src/model-resolver.ts

@@ -337,6 +337,7 @@ function usesAPIKeyAuth(packageName: string | undefined) {
     name === "@ai-sdk/deepinfra" ||
     name === "@ai-sdk/openai-compatible" ||
     name === "@ai-sdk/google" ||
+    name === "@ai-sdk/groq" ||
     name === "@ai-sdk/togetherai" ||
     name === "@ai-sdk/xai" ||
     name === "@openrouter/ai-sdk-provider" ||
@@ -349,6 +350,7 @@ function usesAPIKeyAuth(packageName: string | undefined) {
     name === "@opencode-ai/ai/providers/deepinfra" ||
     name === "@opencode-ai/ai/providers/openai-compatible" ||
     name === "@opencode-ai/ai/providers/google" ||
+    name === "@opencode-ai/ai/providers/groq" ||
     name === "@opencode-ai/ai/providers/togetherai" ||
     name === "@opencode-ai/ai/providers/xai" ||
     name === "@opencode-ai/ai/providers/openrouter" ||

packages/core/src/plugin.ts

@@ -161,7 +161,7 @@ const layer = Layer.effect(
     yield* Effect.addFinalizer((exit) =>
       Effect.gen(function* () {
         active.clear()
-        yield* State.batch(Scope.close(scope, exit))
+        yield* State.batch(Scope.close(scope, exit), { flush: false })
       }),
     )
 

packages/core/src/plugin/host.ts

@@ -384,7 +384,7 @@ export const make = Effect.fn("PluginHost.make")(function* (plugin: Interface, p
               add: (tool) => draft.add(tool),
             }),
           )
-          .pipe(Effect.orDie, Effect.as({ dispose: Effect.void })),
+          .pipe(Effect.as({ dispose: Effect.void })),
       hook: (name, callback) => hooks.register("tool", name, callback),
     },
     vcs: {

packages/core/src/plugin/provider.ts

@@ -11,7 +11,6 @@ import { GatewayPlugin } from "./provider/gateway.js"
 import { GithubCopilotPlugin } from "./provider/github-copilot.js"
 import { GitLabPlugin } from "./provider/gitlab.js"
 import { GoogleVertexPlugin } from "./provider/google-vertex.js"
-import { GroqPlugin } from "./provider/groq.js"
 import { KiloPlugin } from "./provider/kilo.js"
 import { LLMGatewayPlugin } from "./provider/llmgateway.js"
 import { LMStudioPlugin } from "./provider/lmstudio.js"
@@ -45,7 +44,6 @@ export const ProviderPlugins: PluginInternal.InternalPlugin[] = [
   GithubCopilotPlugin,
   GitLabPlugin,
   GoogleVertexPlugin,
-  GroqPlugin,
   KiloPlugin,
   LLMGatewayPlugin,
   LMStudioPlugin,

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

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

packages/core/src/plugin/runtime.ts

@@ -29,7 +29,7 @@ export interface Interface {
     | "wait"
     | "context"
   >
-  readonly job: Pick<Job.Interface, "start" | "wait" | "block" | "background" | "cancel">
+  readonly job: Pick<Job.Interface, "start" | "wait" | "block" | "background" | "cancel" | "completeBackground">
   readonly location: {
     readonly agent: {
       readonly list: (
@@ -92,6 +92,8 @@ export const layerWithCell = (cell: Cell) =>
         block: (input) => require(cell, (runtime) => runtime.job.block(input)),
         background: (id) => require(cell, (runtime) => runtime.job.background(id)),
         cancel: (id) => require(cell, (runtime) => runtime.job.cancel(id)),
+        completeBackground: (notificationID) =>
+          require(cell, (runtime) => runtime.job.completeBackground(notificationID)),
       },
       location: {
         agent: {

packages/core/src/provider.ts

@@ -62,6 +62,7 @@ const builtins = new Map<string, () => Promise<unknown>>([
     "@opencode-ai/ai/providers/google-vertex/messages",
     () => import("@opencode-ai/ai/providers/google-vertex/messages"),
   ],
+  ["@opencode-ai/ai/providers/groq", () => import("@opencode-ai/ai/providers/groq")],
   ["@opencode-ai/ai/providers/openai", () => import("@opencode-ai/ai/providers/openai")],
   ["@opencode-ai/ai/providers/openai/chat", () => import("@opencode-ai/ai/providers/openai/chat")],
   ["@opencode-ai/ai/providers/openai/responses", () => import("@opencode-ai/ai/providers/openai/responses")],

packages/core/src/session.ts

@@ -139,6 +139,27 @@ export class CompactionConflictError extends Schema.TaggedError<CompactionConfli
 export class BusyError extends Schema.TaggedError<BusyError>()("Session.BusyError", {
   sessionID: SessionSchema.ID,
 }) {}
+export class MessageNotAssistantError extends Schema.TaggedError<MessageNotAssistantError>()(
+  "Session.MessageNotAssistantError",
+  {
+    sessionID: SessionSchema.ID,
+    messageID: SessionMessage.ID,
+  },
+) {}
+export class MessageIncompleteError extends Schema.TaggedError<MessageIncompleteError>()(
+  "Session.MessageIncompleteError",
+  {
+    sessionID: SessionSchema.ID,
+    messageID: SessionMessage.ID,
+  },
+) {}
+export class MessageToolIncompleteError extends Schema.TaggedError<MessageToolIncompleteError>()(
+  "Session.MessageToolIncompleteError",
+  {
+    sessionID: SessionSchema.ID,
+    messageID: SessionMessage.ID,
+  },
+) {}
 export class InboxConflictError extends Schema.TaggedError<InboxConflictError>()("Session.InboxConflictError", {
   sessionID: SessionSchema.ID,
   inboxID: SessionMessage.ID,
@@ -193,6 +214,19 @@ export interface Interface {
     sessionID: SessionSchema.ID
     messageID: SessionMessage.ID
   }) => Effect.Effect<SessionMessage.Info | undefined>
+  readonly updateMessage: (input: {
+    readonly sessionID: SessionSchema.ID
+    readonly messageID: SessionMessage.ID
+    readonly content: readonly SessionMessage.AssistantContent[]
+  }) => Effect.Effect<
+    SessionMessage.Assistant,
+    | NotFoundError
+    | MessageNotFoundError
+    | BusyError
+    | MessageNotAssistantError
+    | MessageIncompleteError
+    | MessageToolIncompleteError
+  >
   readonly context: (
     sessionID: SessionSchema.ID,
   ) => Effect.Effect<SessionMessage.Info[], NotFoundError | MessageDecodeError>
@@ -560,6 +594,29 @@ const layer = Layer.effect(
         const stored = yield* store.message(input.messageID)
         return stored?.sessionID === input.sessionID ? stored.message : undefined
       }),
+      updateMessage: Effect.fn("Session.updateMessage")(function* (input) {
+        const ref = { sessionID: input.sessionID, messageID: input.messageID }
+        yield* result.get(ref.sessionID)
+        if ((yield* execution.active).has(ref.sessionID)) return yield* new BusyError({ sessionID: ref.sessionID })
+        const message = yield* result.message(ref)
+        if (!message) return yield* new MessageNotFoundError(ref)
+        if (message.type !== "assistant") return yield* new MessageNotAssistantError(ref)
+        if (!message.time.completed) return yield* new MessageIncompleteError(ref)
+        if (
+          input.content.some(
+            (content) =>
+              content.type === "tool" && (content.state.status === "streaming" || content.state.status === "running"),
+          )
+        )
+          return yield* new MessageToolIncompleteError(ref)
+        yield* bus.publish(SessionEvent.MessageContentUpdated, {
+          ...ref,
+          content: Schema.encodeSync(Schema.Array(SessionMessage.AssistantContent))(input.content),
+        })
+        const updated = yield* result.message(ref)
+        if (updated?.type !== "assistant") return yield* new MessageNotFoundError(ref)
+        return updated
+      }),
       context: Effect.fn("Session.context")(function* (sessionID) {
         yield* result.get(sessionID)
         return yield* store.context(sessionID)

packages/core/src/session/execution.ts

@@ -3,6 +3,7 @@ export * as SessionExecution from "./execution.js"
 import { Cause, Context, Effect, Exit, Layer } from "effect"
 import { Bus } from "../bus.js"
 import { Database } from "../database/database.js"
+import { Job } from "../job.js"
 import { LocationServiceMap } from "../location-service-map.js"
 import { makeGlobalNode } from "@opencode-ai/util/effect/app-node"
 import { SessionEvent } from "./event.js"
@@ -52,6 +53,7 @@ export const layer = Layer.effect(
     const store = yield* SessionStore.Service
     const locations = yield* LocationServiceMap.Service
     const bus = yield* Bus.Service
+    const jobs = yield* Job.Service
     const db = (yield* Database.Service).db
     const reportLifecycle = <A>(sessionID: SessionSchema.ID, effect: Effect.Effect<A>) =>
       effect.pipe(
@@ -94,7 +96,7 @@ export const layer = Layer.effect(
               : Effect.logError("Failed to drain Session", cause).pipe(Effect.annotateLogs({ sessionID })),
           ),
         )
-        if (result.type === "complete") return
+        if (result._tag === "Complete") return
         return yield* drain(sessionID, false, result.continuation, promotable)
       })
     }
@@ -118,6 +120,7 @@ export const layer = Layer.effect(
             if (outcome.type === "interrupted") {
               // A user cancel releases the claim: the turn must not resurrect at the next
               // boot. Shutdown interruption keeps it for restart continuity.
+              if (outcome.reason === "user") yield* jobs.cancel(sessionID)
               yield* bus.publish(
                 SessionEvent.Execution.Interrupted,
                 { sessionID, reason: outcome.reason },
@@ -167,7 +170,7 @@ export const layer = Layer.effect(
 export const node = makeGlobalNode({
   service: Service,
   layer,
-  deps: [SessionStore.node, LocationServiceMap.node, Bus.node, Database.node],
+  deps: [SessionStore.node, LocationServiceMap.node, Bus.node, Database.node, Job.node],
 })
 
 /** Low-level compatibility layer for callers that only need durable Session recording. */

packages/core/src/session/execution/restart.ts

@@ -3,6 +3,8 @@ export * as SessionRestart from "./restart.js"
 import { Context, Effect, Layer } from "effect"
 import { makeGlobalNode } from "@opencode-ai/util/effect/app-node"
 import { Bus } from "../../bus.js"
+import { Job } from "../../job.js"
+import { Session } from "../../session.js"
 import { SessionEvent } from "../event.js"
 import { SessionExecution } from "../execution.js"
 import { SessionSchema } from "../schema.js"
@@ -45,6 +47,9 @@ export interface Interface {
  * process: crash, SIGKILL, isolate eviction, and graceful restart all leave
  * the same durable signature.
  *
+ * Recovery is at-least-once: local coordination prevents concurrent drains,
+ * not repeated external side effects after a crash.
+ *
  * The sweep assumes every orphaned claim's owner is dead. The managed-server
  * protocol guarantees this: a successor is only spawned after the previous
  * process is confirmed dead (client service `kill`/`evict` poll the PID), the
@@ -62,14 +67,16 @@ export const layer = (options?: Options) =>
       const store = yield* SessionStore.Service
       const execution = yield* SessionExecution.Service
       const bus = yield* Bus.Service
+      const jobs = yield* Job.Service
+      const sessions = yield* Session.Service
       const scope = yield* Effect.scope
       const maxAttempts = options?.maxAttempts ?? DEFAULT_MAX_ATTEMPTS
 
-      const resumeOne = Effect.fnUntraced(function* (sessionID: SessionSchema.ID) {
+      const prepareResume = Effect.fnUntraced(function* (sessionID: SessionSchema.ID) {
         // Durable before the resume runs, so a crash inside the resumed turn is
         // counted by the next sweep and the budget cannot be dodged.
         const attempts = yield* store.countResume(sessionID)
-        if (attempts === undefined) return // the Session was deleted since listing
+        if (attempts === undefined) return false
         if (attempts > maxAttempts) {
           // Terminalize instead: the release hook clears the claim and resets the
           // counter atomically with the terminal event.
@@ -78,31 +85,166 @@ export const layer = (options?: Options) =>
             { sessionID, error: RESUME_EXHAUSTED },
             { commit: () => store.release(sessionID) },
           )
-          return
+          return false
         }
         yield* bus.publish(SessionEvent.Synthetic, {
           sessionID,
           text: CONTINUE_AFTER_SERVER_RESTART,
           description: "Continuing after restart",
         })
-        // Forked into the service scope so boot never waits on resumed turns;
-        // resuming an already-live Session joins its execution. Drain failures
-        // are logged and durably recorded by the execution layer.
-        yield* execution.resume(sessionID).pipe(Effect.ignore, Effect.forkIn(scope))
+        return true
+      })
+
+      const recoverShell = Effect.fnUntraced(function* (
+        background: Job.Background,
+        recovery: Extract<Job.Recovery, { kind: "shell" }>,
+      ) {
+        const state = background.status === "running" ? "cancelled" : background.status
+        const text =
+          background.status === "running"
+            ? "Command cancelled because the server restarted"
+            : state === "completed"
+              ? (background.output ?? "Command completed")
+              : state === "error"
+                ? (background.error ?? "Command failed")
+                : "Command cancelled"
+
+        yield* sessions
+          .synthetic({
+            id: background.notificationID,
+            sessionID: recovery.sessionID,
+            description: recovery.command,
+            text: `<shell id="${background.id}" state="${state}" command="${recovery.command}">\n${text}\n</shell>`,
+            metadata: {
+              source: "shell",
+              jobID: background.id,
+              shellID: recovery.shellID,
+              state,
+            },
+            resume: false,
+          })
+          .pipe(
+            Effect.catchTag("Session.NotFoundError", () => Effect.void),
+            Effect.orDie,
+          )
+        yield* jobs.completeBackground(background.notificationID)
+      })
+
+      const recoverSubagent = Effect.fnUntraced(function* (
+        background: Job.Background,
+        recovery: Extract<Job.Recovery, { kind: "subagent" }>,
+        suspended: ReadonlySet<SessionSchema.ID>,
+      ) {
+        const child = yield* store.get(recovery.childSessionID)
+        if (!child || child.parentID !== recovery.parentSessionID || !(yield* store.get(recovery.parentSessionID))) {
+          yield* jobs.completeBackground(background.notificationID)
+          return
+        }
+
+        const notify = Effect.fnUntraced(function* (result: Pick<Job.Background, "status" | "output" | "error">) {
+          if (result.status === "running") return
+          const text =
+            result.status === "completed"
+              ? (result.output ?? "Subagent completed without a text response.")
+              : result.status === "error"
+                ? (result.error ?? "Subagent failed")
+                : "Subagent cancelled"
+          yield* sessions
+            .synthetic({
+              id: background.notificationID,
+              sessionID: recovery.parentSessionID,
+              ...(suspended.has(recovery.parentSessionID) ? { resume: false } : {}),
+              description: recovery.description,
+              text: `<subagent sessionID="${recovery.childSessionID}" state="${result.status}" description="${recovery.description}">\n${text}\n</subagent>`,
+              metadata: {
+                source: "subagent",
+                childID: recovery.childSessionID,
+                agent: recovery.agent,
+                state: result.status,
+              },
+            })
+            .pipe(Effect.orDie)
+          yield* jobs.completeBackground(background.notificationID)
+        })
+
+        if (background.status !== "running") {
+          yield* notify(background)
+          return
+        }
+        if ((yield* execution.active).has(recovery.childSessionID)) return
+        if (!(yield* prepareResume(recovery.childSessionID))) {
+          yield* notify({ status: "error", error: RESUME_EXHAUSTED.message })
+          return
+        }
+
+        yield* jobs.start({
+          id: background.id,
+          type: "subagent",
+          title: recovery.description,
+          notificationID: background.notificationID,
+          recovery,
+          run: execution.resume(recovery.childSessionID).pipe(
+            Effect.andThen(store.context(recovery.childSessionID)),
+            Effect.map((messages) => {
+              const assistant = messages.findLast(
+                (message) =>
+                  message.type === "assistant" && message.time.completed !== undefined && message.error === undefined,
+              )
+              if (assistant?.type !== "assistant") return "Subagent completed without a text response."
+              return (
+                assistant.content
+                  .filter((part) => part.type === "text")
+                  .map((part) => part.text)
+                  .join("") || "Subagent completed without a text response."
+              )
+            }),
+          ),
+        })
+        yield* jobs.background(background.id)
+        yield* jobs.wait({ id: background.id }).pipe(
+          Effect.flatMap((result) => (result.info ? notify(result.info) : Effect.void)),
+          Effect.ignore,
+          Effect.forkIn(scope),
+        )
       })
 
       return Service.of({
         resumeSuspendedSessions: Effect.gen(function* () {
-          // Child claims never drive recovery (children are not resumed), so a
-          // dead child's claim is noise no terminal will ever release. Clearing
-          // is safe even against a live child: claims are recovery markers, not
-          // locks, and children are excluded from that recovery.
-          yield* store.releaseChildClaims
           const active = yield* execution.active
-          // Sessions already draining in this process keep their claim; resuming
-          // them would only inject a stray continuation into a live turn.
-          const orphaned = (yield* store.listSuspended()).filter((sessionID) => !active.has(sessionID))
-          yield* Effect.forEach(orphaned, resumeOne, { concurrency: "unbounded", discard: true })
+          // Early notices wait for root recovery's accounting, including roots that exhaust their budget.
+          const suspended = new Set((yield* store.listSuspended()).filter((sessionID) => !active.has(sessionID)))
+          const pending = yield* jobs.pendingBackground
+          yield* store.releaseChildClaims(
+            pending.flatMap((background) =>
+              background.status === "running" && background.recovery.kind === "subagent"
+                ? [background.recovery.childSessionID]
+                : [],
+            ),
+          )
+          yield* Effect.forEach(
+            pending,
+            Effect.fnUntraced(function* (background) {
+              if ((yield* jobs.get(background.id))?.status === "running") return
+              const recovery = background.recovery
+              yield* recovery.kind === "shell"
+                ? recoverShell(background, recovery)
+                : recoverSubagent(background, recovery, suspended)
+            }),
+            { discard: true },
+          )
+
+          // Background completion can wake a parent, so inspect local ownership only after recovery.
+          const resumed = yield* execution.active
+          yield* Effect.forEach(
+            (yield* store.listSuspended()).filter((sessionID) => !resumed.has(sessionID)),
+            (sessionID) =>
+              execution
+                .resume(sessionID)
+                .pipe(Effect.ignore, Effect.forkIn(scope), Effect.when(prepareResume(sessionID))),
+            { concurrency: "unbounded", discard: true },
+          )
+          // Async observers consult this set at delivery; later completions wake parents normally.
+          suspended.clear()
         }),
       })
     }),
@@ -111,5 +253,5 @@ export const layer = (options?: Options) =>
 export const node = makeGlobalNode({
   service: Service,
   layer: layer(),
-  deps: [SessionStore.node, SessionExecution.node, Bus.node],
+  deps: [SessionStore.node, SessionExecution.node, Bus.node, Job.node, Session.node],
 })

packages/core/src/session/message-updater.ts

@@ -1,5 +1,5 @@
 import { castDraft, produce, type WritableDraft } from "immer"
-import { DateTime, Effect, Match, pipe } from "effect"
+import { DateTime, Effect, Match, pipe, Schema } from "effect"
 import { SessionEvent } from "./event.js"
 import { SessionMessage } from "./message.js"
 
@@ -71,6 +71,12 @@ export function update(adapter: Adapter, event: SessionEvent.DurableEvent) {
     Match.discriminatorsExhaustive("type")({
       "session.created": () => Effect.void,
       "session.viewed": () => Effect.void,
+      "session.message.content.updated": (event) =>
+        updateOwnedAssistant(event.data.messageID, (draft) => {
+          draft.content = castDraft(
+            Schema.decodeUnknownSync(Schema.Array(SessionMessage.AssistantContent))(event.data.content),
+          )
+        }),
       "session.usage.recorded": () => Effect.void,
       "session.agent.selected": (event) =>
         Effect.gen(function* () {
@@ -204,6 +210,7 @@ export function update(adapter: Adapter, event: SessionEvent.DurableEvent) {
                 draft.finish = undefined
                 draft.rawFinish = undefined
                 draft.providerState = undefined
+                draft.time.streamed = undefined
                 draft.time.completed = undefined
                 if (event.data.snapshot) draft.snapshot = { ...draft.snapshot, start: event.data.snapshot }
               }),
@@ -232,6 +239,11 @@ export function update(adapter: Adapter, event: SessionEvent.DurableEvent) {
             }),
           )
         }),
+      "session.step.streamed": (event) => {
+        return updateOwnedAssistant(event.data.assistantMessageID, (draft) => {
+          draft.time.streamed = created
+        })
+      },
       "session.step.ended": (event) => {
         return updateOwnedAssistant(event.data.assistantMessageID, (draft) => {
           draft.time.completed = created

packages/core/src/session/model-request.ts

@@ -45,7 +45,7 @@ const declineDefect = (cause: Cause.Cause<Tool.Error>) => {
   return decline ? Result.succeed(decline) : Result.fail(cause)
 }
 
-interface Prepared {
+export interface Prepared {
   readonly request: LLMRequest
   readonly options: StreamOptions
   /**

packages/core/src/session/projector.ts

@@ -581,6 +581,7 @@ const layer = Layer.effectDiscard(
         .run()
         .pipe(Effect.orDie)
     })
+    yield* bus.project(SessionEvent.MessageContentUpdated, (event) => run(db, event))
     yield* bus.project(SessionEvent.UsageRecorded, (event) => applyUsage(db, event.data.sessionID, event.data))
     yield* bus.project(SessionEvent.Forked, (event) => projectFork(db, event))
     yield* bus.project(SessionEvent.InboxDelivered, (event) =>
@@ -659,6 +660,7 @@ const layer = Layer.effectDiscard(
     yield* bus.project(SessionEvent.Shell.Started, (event) => run(db, event))
     yield* bus.project(SessionEvent.Shell.Ended, (event) => run(db, event))
     yield* bus.project(SessionEvent.Step.Started, (event) => run(db, event))
+    yield* bus.project(SessionEvent.Step.Streamed, (event) => run(db, event))
     yield* bus.project(SessionEvent.Step.Ended, (event) =>
       Effect.gen(function* () {
         yield* run(db, event)

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

@@ -1,7 +1,7 @@
 export * as SessionRunner from "./index.js"
 
 import type { AIError } from "@opencode-ai/ai"
-import { Context, Effect } from "effect"
+import { Context, Data, Effect } from "effect"
 import { SessionSchema } from "../schema.js"
 import type { Promotable } from "../inbox.js"
 import type { AgentNotFoundError, MessageDecodeError, StepFailedError, UserInterruptedError } from "../error.js"
@@ -19,9 +19,11 @@ export type RunError =
 
 export type Continuation = { readonly step: number }
 
-export type DrainResult =
-  | { readonly type: "complete" }
-  | { readonly type: "moved"; readonly continuation?: Continuation }
+export type DrainResult = Data.TaggedEnum<{
+  Complete: {}
+  Moved: { readonly continuation?: Continuation }
+}>
+export const DrainResult = Data.taggedEnum<DrainResult>()
 
 /** Runs one local continuation from already-recorded Session history. */
 export interface Interface {

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

@@ -1,20 +1,9 @@
 export * as SessionRunnerLLM from "./llm.js"
 
-import {
-  LLMClient,
-  AIError,
-  InvalidProviderOutputReason,
-  LLMEvent,
-  Message,
-  isContextOverflowFailure,
-  type ProviderErrorEvent,
-  type ToolCall,
-} from "@opencode-ai/ai"
-import { Cause, Config, Data, Effect, Exit, Fiber, FiberMap, Layer, Option, Pull, Schedule, Stream } from "effect"
+import { Message } from "@opencode-ai/ai"
+import { Cause, Config, Effect, Exit, FiberMap, Layer, Pull, Schedule } from "effect"
 import { Database } from "../../database/database.js"
 import { Bus } from "../../bus.js"
-import { Permission } from "../../permission.js"
-import { QuestionTool } from "../../tool/plugin/question.js"
 import { InstructionState } from "../instruction-state.js"
 import { SessionCompaction } from "../compaction.js"
 import { SessionContext } from "../context.js"
@@ -26,100 +15,18 @@ import { SessionMessage } from "../message.js"
 import { SessionSchema } from "../schema.js"
 import { SessionStore } from "../store.js"
 import { SessionTitle } from "../title.js"
-import { Service, type Continuation } from "./index.js"
-import { createLLMEventPublisher, type StepRecord } from "./publish-llm-event.js"
+import { DrainResult, Service, type Continuation } from "./index.js"
 import { Snapshot } from "../../snapshot.js"
 import { makeLocationNode } from "@opencode-ai/util/effect/app-node"
 import { llmClient } from "../../effect/app-node-platform.js"
 import { StepFailedError } from "../error.js"
-import { toSessionError } from "../to-session-error.js"
 import { SessionRunnerRetry } from "./retry.js"
-import { SessionUsage } from "../usage.js"
+import { SessionStep } from "./step.js"
 import { ToolOutput } from "../../tool-output.js"
 import { PluginSupervisor } from "../../plugin/supervisor.js"
-import { Tool } from "../../tool.js"
 import { PromptCacheDiagnostics } from "../prompt-cache-diagnostics.js"
 import { MAX_STEPS_PROMPT } from "./max-steps.js"
 
-/** How one model call ended: settled, awaiting retry/recovery, or restarted by compaction. */
-type CallOutcome = Data.TaggedEnum<{
-  Completed: { readonly needsContinuation: boolean; readonly step: number }
-  Retry: { readonly step: number }
-  Continue: {
-    readonly cause: AIError
-    readonly error: SessionRunnerRetry.RetryableFailure["error"]
-    readonly step: number
-  }
-  RecoverFull: { readonly step: number }
-  Restart: { readonly step: number; readonly recoveredOverflow: boolean }
-}>
-const CallOutcome = Data.taggedEnum<CallOutcome>()
-
-// Declining an interactive prompt halts the drain instead of becoming model-facing tool output.
-const isDecline = (
-  error: SessionModelRequest.ExecuteError,
-): error is Permission.DeclinedError | QuestionTool.CancelledError =>
-  error._tag === "Permission.DeclinedError" || error._tag === "QuestionTool.CancelledError"
-
-const isInterruptedStream = (failure: AIError) => {
-  if (failure.reason._tag === "InvalidProviderOutput")
-    return failure.reason.classification === "incomplete-stream"
-  if (failure.reason._tag === "Transport") return failure.reason.operation === "read"
-  return false
-}
-
-/**
- * Classifies how the owned tool fibers ended. Interrupts abort the step; a user decline
- * settles its own call and then aborts the step; a defect from a tool implementation
- * becomes a failed tool call the model can read; a typed infrastructure failure must
- * fail the assistant and then the drain.
- */
-const classifyToolExits = (
-  settled: Exit.Exit<Array<Exit.Exit<void, SessionModelRequest.ExecuteError>>, never>,
-  calls: ReadonlyArray<ToolCall>,
-) => {
-  // Exits align with calls by construction: one owned fiber per accepted local call.
-  const exits = settled._tag === "Success" ? settled.value : []
-  const declines = exits.flatMap((exit, index) =>
-    exit._tag === "Failure"
-      ? exit.cause.reasons.flatMap((reason) =>
-          Cause.isFailReason(reason) && isDecline(reason.error) ? [{ call: calls[index], reason: reason.error }] : [],
-        )
-      : [],
-  )
-  const causes =
-    settled._tag === "Failure"
-      ? [settled.cause]
-      : exits.flatMap((exit) => (exit._tag === "Failure" ? [exit.cause] : []))
-  // The first non-interrupt, non-decline failure, rebuilt without decline reasons so the
-  // drain's error channel never carries a decline.
-  const failure = causes
-    .flatMap((cause) => {
-      if (Cause.hasInterrupts(cause)) return []
-      const reasons = cause.reasons.flatMap(
-        (reason): Array<Cause.Reason<never>> =>
-          Cause.isFailReason(reason)
-            ? isDecline(reason.error)
-              ? []
-              : // A typed failure here broke the ExecuteError contract (the per-fiber
-                // `catchTag("Tool.Error")` consumes honest ones). Surfacing it as a defect
-                // keeps it from being dropped, which would leave its call unsettled forever.
-                [Cause.makeDieReason(reason.error)]
-            : [reason],
-      )
-      return reasons.length > 0 ? [Cause.fromReasons(reasons)] : []
-    })
-    .at(0)
-  return {
-    interrupted: causes.some(Cause.hasInterrupts),
-    declines,
-    failure,
-  }
-}
-
-const TOOLS_INTERRUPTED = { type: "aborted", message: "Tool execution interrupted" } as const
-const STEP_INTERRUPTED = { type: "aborted", message: "Step interrupted" } as const
-const RESULT_MISSING = { type: "tool.result-missing", message: "Provider did not return a tool result" } as const
 const CONTINUE_AFTER_INCOMPLETE_STREAM =
   "The previous response was interrupted. Continue from where you left off without repeating completed content."
 
@@ -127,17 +34,15 @@ const layer = Layer.effect(
   Service,
   Effect.gen(function* () {
     const bus = yield* Bus.Service
-    const llm = yield* LLMClient.Service
     const store = yield* SessionStore.Service
     const context = yield* SessionContext.Service
     const modelRequests = yield* SessionModelRequest.Service
     const modelTransport = yield* SessionModelTransport.Service
-    const snapshots = yield* Snapshot.Service
     const db = (yield* Database.Service).db
     const compaction = yield* SessionCompaction.Service
     const plugins = yield* PluginSupervisor.Service
     const title = yield* SessionTitle.Service
-    const toolOutput = yield* ToolOutput.Service
+    const steps = yield* SessionStep.make
     const diagnostics = yield* Config.boolean("OPENCODE_PROMPT_CACHE_DIAGNOSTICS").pipe(
       Config.withDefault(false),
       Effect.orDie,
@@ -166,539 +71,218 @@ const layer = Layer.effect(
     })
     // Title generation starts once input is visible and must not delay model execution.
     const titles = yield* FiberMap.make<SessionSchema.ID, void, never>()
-    /**
-     * Drains eligible manual compaction and user input until the Session becomes idle.
-     * Execution lifecycle is published per busy period by SessionExecution, not here.
-     */
+
     const drain = Effect.fn("SessionRunner.drain")(function* (input: {
       readonly sessionID: SessionSchema.ID
       readonly force: boolean
       readonly continuation?: Continuation
       readonly promotable?: SessionInbox.Promotable
     }) {
+      const sessionID = input.sessionID
       let force = input.force
-      let continuation = input.continuation
+      let continuing = input.continuation !== undefined
+      let step = input.continuation?.step ?? 1
+      let entering = true
       const promotable = input.promotable ?? "input"
-      if (!force && !continuation && !(yield* eligible(input.sessionID, promotable)))
-        return { type: "complete" as const }
-      yield* plugins.flush
-      yield* settleStaleToolCalls(input.sessionID)
-      while (true) {
-        // Between-turn control items run under any drain scope: scope gates which user
-        // input may promote, not whether admitted housekeeping runs. Steered control
-        // items go ahead of any queued input; only a queue-delivered control item
-        // parked behind a queued prompt is not the next eligible item.
-        if (yield* runPendingCompaction(input.sessionID, "input")) {
-          force = false
-          continue
-        }
-        if (yield* runPendingMove(input.sessionID, "input")) return { type: "moved" as const }
-        if (!force && !continuation && !(yield* SessionInbox.has(db, input.sessionID, promotable)))
-          return { type: "complete" as const }
-        const result = yield* runSteps(input.sessionID, continuation, promotable)
-        if (result.type === "moved") return result
-        force = false
-        continuation = undefined
-      }
-    })
-
-    /** Work this drain may perform: scoped input, or a between-turn control item next in line. */
-    const eligible = Effect.fnUntraced(function* (sessionID: SessionSchema.ID, promotable: SessionInbox.Promotable) {
-      if (yield* SessionInbox.has(db, sessionID, promotable)) return true
-      if (promotable === "input") return false
-      const next = yield* SessionInbox.nextPromotable(db, sessionID, "input")
-      return next?.type === "compaction" || next?.type === "move"
-    })
-
-    /**
-     * Runs logical steps until no tool result or newly admitted steer requires another
-     * model call. Queued inputs remain pending until the current model work reaches idle.
-     */
-    const runSteps = Effect.fn("SessionRunner.runSteps")(function* (
-      sessionID: SessionSchema.ID,
-      continuation: Continuation | undefined,
-      drainPromotable: SessionInbox.Promotable,
-    ) {
-      // Fresh work may promote queued input; resumed turns and later steps absorb steers only.
-      let promotable: SessionInbox.Promotable = continuation ? "steer" : drainPromotable
-      let step = continuation?.step ?? 1
-      let next = continuation
-      // The drain admitted this work, so the first step always runs — even after a
-      // control item consumed at this boundary (unlike drain's one-shot force).
-      let first = true
-      // Every boundary has the same shape: control items first, then one exit decision,
-      // then the model. The turn continues only while the first step, a continuation, or
-      // steer input is owed. Deciding after control items means consuming the last
-      // steered compaction ends the turn instead of issuing an input-free model call.
-      while (true) {
-        if (yield* runPendingCompaction(sessionID, "steer")) continue
-        if (yield* runPendingMove(sessionID, "steer")) return { type: "moved" as const, continuation: next }
-        if (!first && !next && !(yield* SessionInbox.has(db, sessionID, "steer")))
-          return { type: "complete" as const }
-        const result = yield* runStep(sessionID, promotable, step)
-        first = false
-        promotable = "steer"
-        step = result.step + 1
-        next = result.needsContinuation ? { step } : undefined
-      }
-    })
-
-    /** Completes one logical model step, transparently retrying or rebuilding after compaction. */
-    const runStep = Effect.fnUntraced(function* (
-      sessionID: SessionSchema.ID,
-      promotable: SessionInbox.Promotable,
-      step: number,
-    ) {
-      // Minting message identity before any attempt lets retries resume the same durable
-      // message. A compaction restart re-mints: the old message is stranded behind the new
-      // compaction boundary, so the rebuilt step needs identity inside the new epoch.
-      let assistantMessageID = SessionMessage.ID.create()
-      const retry = yield* Schedule.toStepWithSleep(
-        SessionRunnerRetry.schedule(bus, sessionID, () => assistantMessageID),
-      )
-      /**
-       * Consumes one retry allowance: sleeps the scheduled backoff, or publishes
-       * Step.Failed and fails once attempts are exhausted. The step loop performs
-       * the retry itself on the next iteration.
-       */
-      const waitForRetry = (failure: SessionRunnerRetry.RetryableFailure) =>
-        retry(failure).pipe(
-          Effect.as(CallOutcome.Retry({ step: failure.step })),
-          Pull.catchDone(() =>
-            bus
-              .publish(SessionEvent.Step.Failed, {
-                sessionID,
-                assistantMessageID,
-                error: failure.error,
-              })
-              .pipe(Effect.andThen(Effect.fail(failure.cause))),
-          ),
+      if (!force && !continuing) {
+        const pending = yield* SessionInbox.nextPromotable(db, sessionID, "input")
+        if (
+          !pending ||
+          (pending.delivery === "queue" &&
+            promotable === "steer" &&
+            pending.type !== "compaction" &&
+            pending.type !== "move")
         )
-      let currentPromotable: SessionInbox.Promotable | undefined = promotable
-      let currentStep = step
-      // Overflow recovery is one-shot: a call after recovery must not recover another overflow.
-      let recoverOverflow = true
-      // Continuation rejection permits one immediate full-context Physical Attempt without generic backoff.
-      let recoverContinuation = true
-      while (true) {
-        const outcome = yield* callModel(
-          sessionID,
-          currentPromotable,
-          currentStep,
-          recoverOverflow,
-          recoverContinuation,
-          assistantMessageID,
-        ).pipe(Effect.catchTag("SessionRunner.RetryableFailure", waitForRetry))
-        if (outcome._tag === "Completed") return { needsContinuation: outcome.needsContinuation, step: outcome.step }
-        if (outcome._tag === "Continue") {
-          yield* retry(
-            new SessionRunnerRetry.RetryableFailure({
-              cause: outcome.cause,
-              error: outcome.error,
-              step: outcome.step,
-            }),
-          ).pipe(Pull.catchDone(() => Effect.fail(outcome.cause)))
-          yield* bus.publish(SessionEvent.Synthetic, {
-            sessionID,
-            text: CONTINUE_AFTER_INCOMPLETE_STREAM,
-          })
-          assistantMessageID = SessionMessage.ID.create()
-        }
-        if (outcome._tag === "Restart") {
-          if (outcome.recoveredOverflow) recoverOverflow = false
-          assistantMessageID = SessionMessage.ID.create()
-        }
-        if (outcome._tag === "RecoverFull") recoverContinuation = false
-        // Neither a retry nor a compaction restart re-promotes input.
-        currentPromotable = undefined
-        currentStep = outcome.step
+          return DrainResult.Complete()
       }
-    })
+      yield* plugins.flush
+      yield* settleStaleToolCalls(sessionID)
 
-    /**
-     * Prepares and runs at most one model call, executes its local tools, and durably
-     * settles the step. Compaction may instead request that the logical step restart.
-     */
-    const callModel = Effect.fn("SessionRunner.callModel")(function* (
-      sessionID: SessionSchema.ID,
-      promotable: SessionInbox.Promotable | undefined,
-      step: number,
-      recoverOverflow: boolean,
-      recoverContinuation: boolean,
-      assistantMessageID: SessionMessage.ID,
-    ) {
-      const selected = yield* context.select(sessionID)
-      // Establish what the model knows before admitting what the user said, so
-      // a blocked first step leaves pending inputs untouched.
-      yield* InstructionState.prepare(db, bus, selected.instructions, selected.session.id)
-      const promoted = promotable ? yield* SessionInbox.promote(db, bus, selected.session.id, promotable) : 0
-      if (promoted > 0 && !selected.session.parentID && SessionTitle.isUntitled(selected.session))
-        yield* FiberMap.run(titles, sessionID, title.generate(sessionID).pipe(Effect.ignore), {
-          onlyIfMissing: true,
-        })
-      // Promoted input opens a fresh step allowance.
-      const currentStep = promoted > 0 ? 1 : step
-      const loaded = yield* context.load(selected)
-      const { session, agent } = loaded
-      const resolved = loaded.model
-      // Make room: history must fit the context window before the call. A pending manual
-      // compaction owns this instead; the runner executes it between steps.
-      const compactionInput = { session, messages: loaded.messages, resolved }
-      if (compaction.required(compactionInput)) {
-        const compacted = yield* compaction.compact(compactionInput)
-        if (compacted.status === "completed")
-          return CallOutcome.Restart({ step: currentStep, recoveredOverflow: false })
-        return yield* new StepFailedError({ error: compacted.error })
-      }
-      const stepLimitReached = agent.info.steps !== undefined && currentStep >= agent.info.steps
-      const transcript = SessionModelRequest.baseTranscript({
-        agent: agent.info,
-        model: resolved,
-        tools: loaded.tools,
-        initial: loaded.initial,
-        messages: loaded.messages,
-      })
-      const prepared = yield* modelRequests.prepare({
-        scope: { session, agentID: agent.id, model: resolved, tools: loaded.tools },
-        transcript: {
-          system: transcript.system,
-          messages: stepLimitReached
-            ? [...transcript.messages, Message.assistant(MAX_STEPS_PROMPT)]
-            : transcript.messages,
-        },
-        // The final Step keeps definitions available to protocols with native "none",
-        // preserving their prompt cache prefix. Calls are still rejected at execution.
-        toolChoice: stepLimitReached ? "none" : undefined,
-        webSocket: "session",
-      })
-      yield* diagnosePromptCache(session.id, prepared.request)
-      const executeTool = (input: Parameters<typeof prepared.executeTool>[0]) => {
-        if (stepLimitReached) return new Tool.Error({ message: "Tools are disabled after the maximum agent steps" })
-        return prepared.executeTool(input)
-      }
-      // Every local tool call forked here is owned until it reaches one durable settlement.
-      const toolRuns: Array<{
-        readonly call: ToolCall
-        readonly fiber: Fiber.Fiber<void, SessionModelRequest.ExecuteError>
-      }> = []
-      const interruptTools = Effect.suspend(() => Fiber.interruptAll(toolRuns.map((run) => run.fiber)))
-      const startSnapshot = yield* snapshots.capture()
-      const publisher = createLLMEventPublisher(bus, {
-        sessionID: session.id,
-        agent: agent.id,
-        // The selected catalog identity, not model.id: route-level ids are provider API
-        // model ids (for example gpt-5.5-fast resolves to api id gpt-5.5).
-        model: resolved.ref,
-        providerMetadataKey: transcript.providerMetadataKey,
-        snapshot: startSnapshot,
-        assistantMessageID,
-      })
-      const stepUsage = (finish: NonNullable<StepRecord["finish"]>) => ({
-        cost: SessionUsage.calculateCost(resolved.cost, finish.tokens),
-        tokens: finish.tokens,
-      })
-
-      const captureStepEnd = Effect.fnUntraced(function* () {
-        const snapshot = yield* snapshots.capture()
-        const files =
-          startSnapshot && snapshot
-            ? startSnapshot === snapshot
-              ? []
-              : yield* snapshots
-                  .files({ from: startSnapshot, to: snapshot })
-                  .pipe(Effect.orElseSucceed(() => undefined))
-            : undefined
-        return { snapshot, files }
-      })
-
-      const publishStepEnd = (finish: NonNullable<StepRecord["finish"]>) =>
-        Effect.gen(function* () {
-          const end = yield* captureStepEnd()
-          yield* bus.publish(SessionEvent.Step.Ended, {
-            sessionID: session.id,
-            assistantMessageID: yield* publisher.startAssistant(),
-            finish: finish.finish,
-            rawFinish: finish.rawFinish,
-            providerState: finish.providerState,
-            ...stepUsage(finish),
-            ...end,
-          })
-        })
-
-      // Concurrent writers, no lock: the provider loop and each tool fiber publish
-      // durable events unserialized. This is safe because every publisher method commits
-      // its state marks synchronously before its first await (see publish-llm-event.ts),
-      // every required event order is per-source (each source is one sequential fiber),
-      // and a fiber's events are causally after its own Tool.Called: the fork happens
-      // below that publish. Cross-source order is unconstrained; either interleaving is
-      // a truthful history of concurrent work.
-      //
-      // The stream is defined here but runs inside the settlement mask below: publish each
-      // event durably, fork one fiber per local tool call, and hold back a virgin
-      // context-overflow provider error so settlement may recover it via compaction.
-      let overflowFailure: ProviderErrorEvent | undefined
-      const providerStream = llm.stream(prepared.request, prepared.options).pipe(
-        Stream.runForEach((event) =>
+      const advanceToStep = Effect.fn("SessionRunner.advanceToStep")(() =>
+        Effect.uninterruptibleMask((restore) =>
           Effect.gen(function* () {
-            if (overflowFailure || publisher.hasProviderError()) return
-            if (
-              LLMEvent.is.providerError(event) &&
-              isContextOverflowFailure(event) &&
-              !publisher.record().outputStarted
-            ) {
-              overflowFailure = event
-              return
-            }
-            yield* publisher.publish(event)
-            if (event.type !== "tool-call" || event.providerExecuted) return
-            const assistantMessageID = yield* publisher.assistantMessageID(event.id)
-            toolRuns.push({
-              call: event,
-              fiber: yield* Effect.uninterruptibleMask((restore) =>
-                restore(
-                  executeTool({
-                    sessionID: session.id,
-                    agent: agent.id,
-                    messageID: assistantMessageID,
-                    call: event,
-                    // Progress is ephemeral, not durable history: nothing to order.
-                    progress: (update) => publisher.progress(event.id, update),
+            while (true) {
+              // Location entry and idle boundaries allow queued controls, not necessarily queued prompts.
+              const pending = yield* SessionInbox.serialized(
+                sessionID,
+                Effect.gen(function* () {
+                  const next = yield* SessionInbox.nextPromotable(
+                    db,
+                    sessionID,
+                    entering || !continuing ? "input" : "steer",
+                  )
+                  if (next?.type === "compaction")
+                    yield* bus.publishAll([
+                      [SessionEvent.InboxDelivered, { sessionID, inboxID: next.id }],
+                      [SessionEvent.Compaction.Started, { sessionID, reason: "manual", recent: "", inputID: next.id }],
+                    ])
+                  if (next?.type === "move")
+                    yield* restore(
+                      Effect.gen(function* () {
+                        yield* modelTransport.close(sessionID)
+                        yield* bus.publishAll([
+                          [SessionEvent.InboxDelivered, { sessionID, inboxID: next.id }],
+                          [SessionEvent.Moved, { sessionID, ...next.payload }],
+                        ])
+                      }),
+                    )
+                  return next
+                }),
+              )
+              if (!continuing && pending?.delivery !== "steer") {
+                entering = true
+                step = 1
+              }
+              if (pending?.type === "move")
+                return DrainResult.Moved({ continuation: !entering && continuing ? { step } : undefined })
+              if (pending?.type === "compaction") {
+                const session = yield* store.get(sessionID)
+                if (!session) return yield* Effect.die(new Error(`Session not found: ${sessionID}`))
+                const compacted = yield* restore(
+                  Effect.gen(function* () {
+                    return yield* compaction.compactManual({
+                      session,
+                      messages: yield* store.context(sessionID),
+                      inputID: pending.id,
+                      started: true,
+                    })
                   }),
-                ).pipe(
-                  // The fiber owns its call: it publishes its own completion, masked so a
-                  // finished execution always reaches its durable settlement.
-                  Effect.flatMap(toolOutput.truncate),
-                  Effect.flatMap((outcome) => publisher.toolExecution(event.id, event.name, outcome)),
-                  Effect.catchTag("Tool.Error", (error) =>
-                    publisher.failTool(event.id, toSessionError(error), error.metadata).pipe(Effect.asVoid),
-                  ),
-                ),
-              ).pipe(Effect.forkScoped),
-            })
+                ).pipe(Effect.exit)
+                if (Exit.isFailure(compacted)) {
+                  yield* bus.publish(SessionEvent.Compaction.Failed, {
+                    sessionID,
+                    reason: "manual",
+                    error: Cause.hasInterruptsOnly(compacted.cause)
+                      ? { type: "aborted", message: "Compaction cancelled" }
+                      : { type: "compaction.failed", message: Cause.pretty(compacted.cause) },
+                    inputID: pending.id,
+                  })
+                  return yield* Effect.failCause(compacted.cause)
+                }
+                force = false
+                continue
+              }
+              if (!force && !continuing && (!pending || (pending.delivery === "queue" && promotable === "steer")))
+                return DrainResult.Complete()
+              return yield* restore(
+                Effect.gen(function* () {
+                  const selected = yield* prepareContext(sessionID)
+                  const promoted = yield* SessionInbox.promote(
+                    db,
+                    bus,
+                    sessionID,
+                    entering && !continuing ? promotable : "steer",
+                  )
+                  if (promoted > 0 && !selected.session.parentID && SessionTitle.isUntitled(selected.session))
+                    yield* FiberMap.run(titles, sessionID, title.generate(sessionID).pipe(Effect.ignore), {
+                      onlyIfMissing: true,
+                    })
+                  if (promoted > 0) step = 1
+                  return { _tag: "Ready" as const, context: yield* context.load(selected) }
+                }),
+              )
+            }
           }),
         ),
-        Effect.ensuring(publisher.flush()),
       )
 
-      // Settle: only the stream and the fiber joins are interruptible (restore); every
-      // other line is protected so a started call always reaches one durable outcome.
-      return yield* Effect.uninterruptibleMask((restore) =>
-        Effect.gen(function* () {
-          const stream = yield* restore(providerStream).pipe(Effect.exit)
-          const streamFailure = Option.getOrUndefined(Exit.findErrorOption(stream))
-          // Note: Exit.hasInterrupts is a type guard whose false branch unsoundly narrows
-          // away non-interrupt failures, so both interrupt checks stay Cause-based.
-          const streamInterrupted = stream._tag === "Failure" && Cause.hasInterrupts(stream.cause)
-
-          // Join every owned tool run first: await all exits, not just the first failure.
-          // Afterwards no fiber is alive, settlement is the only writer, and the record
-          // is final. A failed join means the waiting itself was interrupted, so the runs
-          // we abandoned are interrupted before settlement closes them out.
-          if (streamInterrupted) yield* interruptTools
-          const joined = yield* restore(
-            Effect.forEach(toolRuns, (run) => Fiber.await(run.fiber), { concurrency: "unbounded" }),
-          ).pipe(Effect.exit)
-          if (joined._tag === "Failure") yield* interruptTools
-          const tools = classifyToolExits(
-            joined,
-            toolRuns.map((run) => run.call),
-          )
-
-          // A context overflow before any assistant output is recoverable: compact and
-          // restart the step instead of surfacing the provider error.
-          if (
-            recoverOverflow &&
-            compaction.enabled() &&
-            !publisher.record().outputStarted &&
-            isContextOverflowFailure(overflowFailure ?? streamFailure) &&
-            (yield* restore(compaction.compact(compactionInput))).status === "completed"
-          )
-            return CallOutcome.Restart({ step: currentStep, recoveredOverflow: true })
-
-          // An unrecovered held-back overflow becomes the step's durable provider error.
-          if (overflowFailure) yield* publisher.publish(overflowFailure)
-          // A thrown LLM failure not already recorded as the provider error either
-          // escapes as a scheduled retry or fails the assistant durably.
-          const unknownFinish =
-            stream._tag === "Success" && publisher.record().finish?.finish === "unknown"
-              ? new AIError({
-                  module: "session",
-                  method: "stream",
-                  reason: new InvalidProviderOutputReason({
-                    classification: "incomplete-stream",
-                    message: "The provider response ended with an unknown finish reason.",
-                  }),
-                })
-              : undefined
-          const llmFailure = streamFailure instanceof AIError ? streamFailure : unknownFinish
-          const llmError = llmFailure && !publisher.record().providerFailed ? toSessionError(llmFailure) : undefined
-          if (
-            recoverContinuation &&
-            llmFailure?.reason._tag === "Transport" &&
-            (llmFailure.reason.recovery === "retry-full" || llmFailure.reason.recovery === "rotate-and-retry-full") &&
-            !publisher.record().outputStarted
-          )
-            return CallOutcome.RecoverFull({ step: currentStep })
-          if (
-            llmFailure &&
-            llmError &&
-            SessionRunnerRetry.isRetryable(llmFailure) &&
-            !publisher.record().outputStarted
-          ) {
-            // RetryScheduled and Step.Failed fold onto an existing assistant message, so
-            // Step.Started must be durable before the failure escapes.
-            yield* publisher.startAssistant()
-            return yield* new SessionRunnerRetry.RetryableFailure({
-              cause: llmFailure,
-              error: llmError,
-              step: currentStep,
-            })
-          }
-          if (llmError) yield* publisher.failAssistant(llmError)
-
-          // Close every unsettled call with the reason it could not settle truthfully,
-          // and fail the assistant when the step itself cannot complete. A declined call
-          // settles with its own reason before the generic sweeps.
-          for (const decline of tools.declines)
-            yield* publisher.failTool(decline.call.id, {
-              type: "aborted",
-              message:
-                decline.reason._tag === "QuestionTool.CancelledError"
-                  ? decline.reason.message
-                  : "The user declined this tool call",
-            })
-          if (tools.declines.length > 0 || streamInterrupted || tools.interrupted) {
-            yield* publisher.failUnsettledTools(TOOLS_INTERRUPTED)
-            yield* publisher.failAssistant(STEP_INTERRUPTED)
-          }
-          if (tools.failure !== undefined) {
-            const error = toSessionError(Cause.squash(tools.failure))
-            yield* publisher.failUnsettledTools(error)
-          }
-          // Local calls have joined, so the remaining sweeps only close hosted calls the
-          // provider promised but never resolved.
-          if (publisher.record().providerFailed) yield* publisher.failUnsettledTools(TOOLS_INTERRUPTED)
-          if (llmError) yield* publisher.failUnsettledTools(RESULT_MISSING, "hosted")
-          // A clean stream that still left hosted calls unresolved fails the step itself.
-          if (stream._tag === "Success" && !publisher.record().providerFailed) {
-            const hostedResultMissing = yield* publisher.failUnsettledTools(RESULT_MISSING, "hosted")
-            if (hostedResultMissing && !publisher.record().finish) yield* publisher.failAssistant(RESULT_MISSING)
-          }
-
-          // One terminal event: Step.Ended on a clean finish, Step.Failed otherwise.
-          const record = publisher.record()
-          if (record.finish && !record.failure) yield* publishStepEnd(record.finish)
-          if (record.failure) {
-            const end = yield* captureStepEnd()
-            yield* publisher.publishStepFailure({
-              ...(record.finish ? stepUsage(record.finish) : {}),
-              ...end,
-            })
-          }
-
-          if (
-            llmFailure &&
-            llmError &&
-            isInterruptedStream(llmFailure) &&
-            record.outputStarted &&
-            tools.declines.length === 0 &&
-            !tools.interrupted
-          )
-            return CallOutcome.Continue({
-              cause: llmFailure,
-              error: llmError,
-              step: currentStep,
-            })
-
-          if (stream._tag === "Failure") return yield* Effect.failCause(stream.cause)
-          if (tools.declines.length > 0) return yield* Effect.interrupt
-          if (tools.interrupted && tools.failure) return yield* Effect.failCause(tools.failure)
-          if (tools.interrupted && joined._tag === "Failure") return yield* Effect.failCause(joined.cause)
-          if (record.failure) return yield* new StepFailedError({ error: record.failure })
-          return CallOutcome.Completed({
-            // A local call or malformed tool input requires another model step, unless
-            // this step already exhausted the agent's allowance.
-            needsContinuation:
-              !stepLimitReached && record.calls.some((call) => !call.providerExecuted && (call.called || call.settled)),
-            step: currentStep,
-          })
-        }),
-      )
-    }, Effect.scoped)
-
-    /** Executes a previously admitted manual compaction request, if one is pending. */
-    const runPendingCompaction = Effect.fn("SessionRunner.runPendingCompaction")(function* (
-      sessionID: SessionSchema.ID,
-      promotable: SessionInbox.Promotable,
-    ) {
-      return yield* Effect.uninterruptibleMask((restore) =>
-        Effect.gen(function* () {
-          const pending = yield* SessionInbox.serialized(
-            sessionID,
-            Effect.gen(function* () {
-              const selected = yield* SessionInbox.nextPromotable(db, sessionID, promotable)
-              if (selected?.type !== "compaction") return
-              yield* bus.publishAll([
-                [SessionEvent.InboxDelivered, { sessionID, inboxID: selected.id }],
-                [SessionEvent.Compaction.Started, { sessionID, reason: "manual", recent: "", inputID: selected.id }],
-              ])
-              return selected
-            }),
-          )
-          if (pending?.type !== "compaction") return false
-          const session = yield* getSession(sessionID)
-          const compacted = yield* restore(
-            Effect.gen(function* () {
-              return yield* compaction.compactManual({
-                session,
-                messages: yield* store.context(sessionID),
-                inputID: pending.id,
-                started: true,
-              })
-            }),
-          ).pipe(Effect.exit)
-          if (Exit.isSuccess(compacted)) return true
-          yield* bus.publish(SessionEvent.Compaction.Failed, {
-            sessionID,
-            reason: "manual",
-            error: Cause.hasInterruptsOnly(compacted.cause)
-              ? { type: "aborted", message: "Compaction cancelled" }
-              : { type: "compaction.failed", message: Cause.pretty(compacted.cause) },
-            inputID: pending.id,
-          })
-          return yield* Effect.failCause(compacted.cause)
-        }),
-      )
+      while (true) {
+        const next = yield* advanceToStep()
+        if (next._tag !== "Ready") return next
+        continuing = yield* runStep(next.context, step)
+        step++
+        force = false
+        entering = false
+      }
     })
 
-    const runPendingMove = Effect.fn("SessionRunner.runPendingMove")(function* (
-      sessionID: SessionSchema.ID,
-      promotable: SessionInbox.Promotable,
-    ) {
-      return yield* SessionInbox.serialized(
-        sessionID,
-        Effect.gen(function* () {
-          const pending = yield* SessionInbox.nextPromotable(db, sessionID, promotable)
-          if (pending?.type !== "move") return false
-          yield* modelTransport.close(sessionID)
-          yield* bus.publishAll([
-            [SessionEvent.InboxDelivered, { sessionID, inboxID: pending.id }],
-            [
-              SessionEvent.Moved,
-              {
-                sessionID,
-                location: pending.payload.location,
-                projectID: pending.payload.projectID,
-                subpath: pending.payload.subpath,
-              },
-            ],
-          ])
-          return true
-        }),
-      )
+    const prepareContext = Effect.fn("SessionRunner.prepareContext")(function* (sessionID: SessionSchema.ID) {
+      const selected = yield* context.select(sessionID)
+      // A blocked initial instruction baseline must leave admitted input pending.
+      yield* InstructionState.prepare(db, bus, selected.instructions, sessionID)
+      return selected
+    })
+
+    /** Owns logical Step policy; each attempt owns its streaming, tools, and durable settlement. */
+    const runStep = Effect.fn("SessionRunner.runStep")(function* (first: SessionContext.Loaded, step: number) {
+      const sessionID = first.session.id
+      let assistantMessageID = SessionMessage.ID.create()
+      const retry = yield* Schedule.toStepWithSleep(SessionRunnerRetry.schedule(bus, sessionID))
+      let initial: SessionContext.Loaded | undefined = first
+      let recoverOverflow = true
+      let recoverContinuation = true
+      while (true) {
+        // Reuse boundary preparation once; retries refresh context without delivering more input.
+        const loaded = initial ?? (yield* prepareContext(sessionID).pipe(Effect.flatMap(context.load)))
+        initial = undefined
+        const compactionInput = { session: loaded.session, messages: loaded.messages, resolved: loaded.model }
+        if (compaction.required(compactionInput)) {
+          const compacted = yield* compaction.compact(compactionInput)
+          if (compacted.status !== "completed") return yield* new StepFailedError({ error: compacted.error })
+          assistantMessageID = SessionMessage.ID.create()
+          continue
+        }
+        const stepLimitReached = loaded.agent.info.steps !== undefined && step >= loaded.agent.info.steps
+        const transcript = SessionModelRequest.baseTranscript({
+          agent: loaded.agent.info,
+          model: loaded.model,
+          tools: loaded.tools,
+          initial: loaded.initial,
+          messages: loaded.messages,
+        })
+        const prepared = yield* modelRequests.prepare({
+          scope: { session: loaded.session, agentID: loaded.agent.id, model: loaded.model, tools: loaded.tools },
+          transcript: {
+            system: transcript.system,
+            messages: stepLimitReached
+              ? [...transcript.messages, Message.assistant(MAX_STEPS_PROMPT)]
+              : transcript.messages,
+          },
+          // Keep tool definitions on the final Step to preserve the provider's cached prefix.
+          toolChoice: stepLimitReached ? "none" : undefined,
+          webSocket: "session",
+        })
+        yield* diagnosePromptCache(sessionID, prepared.request)
+        const outcome = yield* steps.attempt({
+          sessionID,
+          assistantMessageID,
+          agent: loaded.agent.id,
+          model: loaded.model,
+          prepared,
+          toolsDisabled: stepLimitReached,
+          recoverContinuation,
+          recoverOverflow: Effect.suspend(() =>
+            recoverOverflow && compaction.enabled()
+              ? compaction.compact(compactionInput).pipe(Effect.map((result) => result.status === "completed"))
+              : Effect.succeed(false),
+          ),
+        })
+        if (outcome._tag === "Completed") return outcome.needsContinuation
+        if (outcome._tag === "Retry" || outcome._tag === "Continue") {
+          yield* retry({ cause: outcome.cause, error: outcome.error, assistantMessageID }).pipe(
+            Pull.catchDone(() =>
+              Effect.gen(function* () {
+                if (outcome._tag === "Retry")
+                  yield* bus.publish(SessionEvent.Step.Failed, { sessionID, assistantMessageID, error: outcome.error })
+                return yield* outcome.cause
+              }),
+            ),
+          )
+          if (outcome._tag === "Continue") {
+            yield* bus.publish(SessionEvent.Synthetic, { sessionID, text: CONTINUE_AFTER_INCOMPLETE_STREAM })
+            assistantMessageID = SessionMessage.ID.create()
+          }
+          continue
+        }
+        if (outcome._tag === "Compacted") {
+          recoverOverflow = false
+          assistantMessageID = SessionMessage.ID.create()
+          continue
+        }
+        recoverContinuation = false
+      }
     })
 
-    /** Closes stale tool calls left active by an earlier interrupted drain. */
     const settleStaleToolCalls = Effect.fn("SessionRunner.settleStaleToolCalls")(function* (
       sessionID: SessionSchema.ID,
     ) {
@@ -706,23 +290,24 @@ const layer = Layer.effect(
         if (message.type !== "assistant") continue
         for (const tool of message.content) {
           if (tool.type !== "tool" || (tool.state.status !== "streaming" && tool.state.status !== "running")) continue
+          const metadata = tool.state.status === "running" ? tool.state.metadata : undefined
+          const childID =
+            tool.name === "subagent" && typeof metadata?.sessionID === "string" ? metadata.sessionID : undefined
           yield* bus.publish(SessionEvent.Tool.Failed, {
             sessionID,
             assistantMessageID: message.id,
             id: tool.id,
-            error: { type: "aborted", message: `Tool execution interrupted: ${tool.name}` },
+            error: {
+              type: "aborted",
+              message: `Tool execution interrupted: ${tool.name}${childID ? ` (sessionID: ${childID})` : ""}`,
+            },
+            ...(metadata && Object.keys(metadata).length > 0 ? { metadata } : {}),
             executed: tool.executed === true,
           })
         }
       }
     })
 
-    const getSession = Effect.fn("SessionRunner.getSession")(function* (sessionID: SessionSchema.ID) {
-      const session = yield* store.get(sessionID)
-      if (!session) return yield* Effect.die(new Error(`Session not found: ${sessionID}`))
-      return session
-    })
-
     return Service.of({ drain })
   }),
 )