feat(ai): support server-side response compaction

.changeset/unify-request-envelope.md

@@ -0,0 +1,5 @@
+---
+"@opencode-ai/core": patch
+---
+
+Title generation and compaction summaries now build their model requests through the shared session request boundary, gaining unsupported-media filtering and image bounds while explicitly opting out of session context hooks: plugins that shape the agent conversation do not observe title or compaction requests. Title requests gain the fork-aware session prompt cache key, and compaction summaries in forked sessions reuse the fork root's prompt cache key instead of the fork's own.

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

@@ -69,6 +69,13 @@ const OpenResponsesReasoningItem = Schema.Struct({
   encrypted_content: optionalNull(Schema.String),
 })
 
+const OpenResponsesCompactionItem = Schema.Struct({
+  type: Schema.tag("compaction"),
+  id: Schema.optionalKey(Schema.String),
+  encrypted_content: Schema.String,
+})
+type OpenResponsesCompactionItem = Schema.Schema.Type<typeof OpenResponsesCompactionItem>
+
 const OpenResponsesItemReference = Schema.Struct({
   type: Schema.tag("item_reference"),
   id: Schema.String,
@@ -100,6 +107,7 @@ export const InputItem = Schema.Union([
     phase: Schema.optionalKey(MessagePhase),
   }),
   OpenResponsesReasoningItem,
+  OpenResponsesCompactionItem,
   OpenResponsesItemReference,
   Schema.Struct({
     type: Schema.tag("function_call"),
@@ -339,6 +347,7 @@ export interface ParserState {
   readonly messagePhase: (value: unknown) => MessagePhase | null | undefined
   readonly messagePhases: Readonly<Record<string, MessagePhase | null>>
   readonly reasoningItems: Readonly<Record<string, ReasoningStreamItem>>
+  readonly compactionItems: ReadonlyArray<OpenResponsesCompactionItem>
   readonly store: boolean | undefined
 }
 
@@ -418,6 +427,12 @@ const hostedToolItemID = (part: ToolResultPart, providerMetadataKey: string) =>
   return itemID(part.providerMetadata, providerMetadataKey)
 }
 
+const compactionItems = (message: LLMRequest["messages"][number], providerMetadataKey: string) => {
+  const native = message.native?.[providerMetadataKey]
+  if (!ProviderShared.isRecord(native) || !Array.isArray(native.compactionItems)) return []
+  return native.compactionItems.filter(Schema.is(OpenResponsesCompactionItem))
+}
+
 const lowerMedia = Effect.fn("OpenResponses.lowerMedia")(function* (
   part: MediaPart,
   request: LLMRequest,
@@ -499,6 +514,7 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
     }
 
     if (message.role === "assistant") {
+      input.push(...compactionItems(message, providerMetadataKey))
       const content: TextPart[] = []
       const reasoningItems: Record<string, OpenResponsesReasoningInput> = {}
       const reasoningReferences = new Set<string>()
@@ -1029,6 +1045,21 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
     ] satisfies StepResult
   }
 
+  if (item.type === "compaction") {
+    if (!item.id || typeof item.encrypted_content !== "string")
+      return yield* ProviderShared.eventError(state.id, "Open Responses compaction item is malformed")
+    return [
+      {
+        ...state,
+        compactionItems: [
+          ...state.compactionItems,
+          { type: "compaction", id: item.id, encrypted_content: item.encrypted_content },
+        ],
+      },
+      NO_EVENTS,
+    ] satisfies StepResult
+  }
+
   return [state, NO_EVENTS] satisfies StepResult
 })
 
@@ -1049,10 +1080,11 @@ const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (
     },
     usage: mapUsage(event.response?.usage, state.providerMetadataKey),
     providerMetadata:
-      event.response?.id || event.response?.service_tier
+      event.response?.id || event.response?.service_tier || state.compactionItems.length > 0
         ? providerMetadata(state, {
-            responseId: event.response.id,
-            serviceTier: event.response.service_tier,
+            responseId: event.response?.id,
+            serviceTier: event.response?.service_tier,
+            ...(state.compactionItems.length > 0 ? { compactionItems: state.compactionItems } : {}),
           })
         : undefined,
   })
@@ -1162,6 +1194,7 @@ export const initial = (request: LLMRequest, extension: Extension = BASE): Parse
   messagePhase: (value) => messagePhase(value, extension),
   messagePhases: {},
   reasoningItems: {},
+  compactionItems: [],
   store: OpenResponsesOptions.resolve(request).store,
 })
 

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

@@ -13,6 +13,7 @@ import { OpenAIImage } from "./utils/openai-image.js"
 import { ToolSchemaProjection } from "./utils/tool-schema.js"
 import { OpenResponsesChannel } from "./open-responses-channel.js"
 import { OpenAIResponsesChannel } from "./openai-responses-channel.js"
+import { OpenAIOptions } from "./utils/openai-options.js"
 
 const ADAPTER = "openai-responses"
 const NAME = "OpenAI Responses"
@@ -56,6 +57,14 @@ const OpenAIResponsesCoreFields = {
   input: Schema.Array(OpenAIResponsesInputItem),
   tools: optionalArray(OpenAIResponsesTools),
   tool_choice: Schema.optional(OpenAIResponsesToolChoice),
+  context_management: Schema.optional(
+    Schema.Array(
+      Schema.Struct({
+        type: Schema.tag("compaction"),
+        compact_threshold: Schema.optional(Schema.Int.check(Schema.isGreaterThan(0))),
+      }),
+    ),
+  ),
 }
 
 const OpenAIResponsesBody = Schema.Struct({
@@ -115,6 +124,7 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
     extension,
   )
   const toolSchemaCompatibility = request.model.compatibility?.toolSchema
+  const contextManagement = OpenAIOptions.resolve(request).contextManagement
   return {
     ...body,
     tools:
@@ -125,6 +135,10 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
           ),
     tool_choice:
       body.tool_choice ?? (request.toolChoice ? yield* lowerToolChoice(request.toolChoice, request.tools) : undefined),
+    context_management: contextManagement?.map((item) => ({
+      type: item.type,
+      compact_threshold: item.compactThreshold,
+    })),
   } satisfies OpenAIResponsesBody
 })
 

packages/ai/src/protocols/utils/openai-options.ts

@@ -1,3 +1,5 @@
+import { Option, Schema } from "effect"
+import type { LLMRequest } from "../../schema/index.js"
 import { OpenResponsesOptions } from "./open-responses-options.js"
 
 export const OpenAIReasoningEfforts = OpenResponsesOptions.ReasoningEfforts
@@ -19,6 +21,22 @@ export const OpenAIServiceTier = OpenResponsesOptions.ServiceTierSchema
 
 export const isReasoningEffort = (effort: unknown): effort is OpenAIReasoningEffort => typeof effort === "string"
 
-export const resolve = OpenResponsesOptions.resolve
+export const ContextManagement = Schema.Array(
+  Schema.Struct({
+    type: Schema.tag("compaction"),
+    compactThreshold: Schema.optional(Schema.Int.check(Schema.isGreaterThan(0))),
+  }),
+)
+export type ContextManagement = typeof ContextManagement.Type
+
+const Options = Schema.Struct({
+  contextManagement: Schema.optional(ContextManagement),
+})
+const decodeOptions = Schema.decodeUnknownOption(Options)
+
+export const resolve = (request: LLMRequest) => ({
+  ...OpenResponsesOptions.resolve(request),
+  ...Option.getOrElse(decodeOptions(request.providerOptions), () => ({})),
+})
 
 export * as OpenAIOptions from "./openai-options.js"

packages/ai/src/providers/openai-options.ts

@@ -1,9 +1,12 @@
 import { mergeProviderOptions, type ProviderOptions } from "../schema/index.js"
 import type { OpenResponsesOptionsInput } from "./open-responses-options.js"
+import type { ContextManagement } from "../protocols/utils/openai-options.js"
 
 export type { OpenAIResponseIncludable, OpenAIServiceTier } from "../protocols/utils/openai-options.js"
 
-export type OpenAIOptionsInput = OpenResponsesOptionsInput
+export type OpenAIOptionsInput = OpenResponsesOptionsInput & {
+  readonly contextManagement?: ContextManagement
+}
 
 export type OpenAIProviderOptionsInput = OpenAIOptionsInput
 

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

@@ -0,0 +1,100 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "openai",
+    "protocol": "openai-responses",
+    "transport": "websocket",
+    "model": "gpt-5.5",
+    "tags": [
+      "prefix:openai-responses-websocket",
+      "provider:openai",
+      "protocol:openai-responses",
+      "transport:websocket",
+      "tool",
+      "continuation"
+    ],
+    "name": "openai-responses-websocket/continues-a-tool-call-over-one-socket",
+    "recordedAt": "2026-08-20T00:00:00.000Z"
+  },
+  "interactions": [
+    {
+      "transport": "websocket",
+      "connection": {
+        "sequence": 0,
+        "url": "wss://api.openai.com/v1/responses",
+        "protocols": [],
+        "close": {
+          "code": 1000,
+          "reason": ""
+        }
+      },
+      "events": [
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Call get_weather once, then reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_tool_1\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"function_call\",\"id\":\"fc_ws_weather\",\"call_id\":\"call_ws_weather\",\"name\":\"get_weather\",\"arguments\":\"\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.function_call_arguments.delta\",\"item_id\":\"fc_ws_weather\",\"delta\":\"{\\\"city\\\":\\\"Paris\\\"}\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"function_call\",\"id\":\"fc_ws_weather\",\"call_id\":\"call_ws_weather\",\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_tool_1\"}}"
+        },
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"type\":\"function_call_output\",\"call_id\":\"call_ws_weather\",\"output\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}],\"store\":false,\"max_output_tokens\":50,\"previous_response_id\":\"resp_ws_tool_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_tool_2\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_tool_2\",\"role\":\"assistant\",\"content\":[]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"item_id\":\"msg_ws_tool_2\",\"delta\":\"Paris is sunny.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.done\",\"item_id\":\"msg_ws_tool_2\",\"text\":\"Paris is sunny.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_tool_2\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Paris is sunny.\"}]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_tool_2\"}}"
+        }
+      ]
+    }
+  ]
+}

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

@@ -0,0 +1,119 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "openai",
+    "protocol": "openai-responses",
+    "transport": "websocket",
+    "model": "gpt-5.5",
+    "tags": [
+      "prefix:openai-responses-websocket",
+      "provider:openai",
+      "protocol:openai-responses",
+      "transport:websocket",
+      "reconnect",
+      "full-context"
+    ],
+    "name": "openai-responses-websocket/reconstructs-full-context-after-reconnect",
+    "recordedAt": "2026-08-20T00:00:00.000Z"
+  },
+  "interactions": [
+    {
+      "transport": "websocket",
+      "connection": {
+        "sequence": 0,
+        "url": "wss://api.openai.com/v1/responses",
+        "protocols": [],
+        "close": {
+          "code": 1000,
+          "reason": ""
+        }
+      },
+      "events": [
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_reconnect_1\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_reconnect_1\",\"role\":\"assistant\",\"content\":[]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"item_id\":\"msg_ws_reconnect_1\",\"delta\":\"Alpha.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.done\",\"item_id\":\"msg_ws_reconnect_1\",\"text\":\"Alpha.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_reconnect_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Alpha.\"}]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_reconnect_1\"}}"
+        }
+      ]
+    },
+    {
+      "transport": "websocket",
+      "connection": {
+        "sequence": 1,
+        "url": "wss://api.openai.com/v1/responses",
+        "protocols": [],
+        "close": {
+          "code": 1000,
+          "reason": ""
+        }
+      },
+      "events": [
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Alpha.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_reconnect_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Alpha.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Beta.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_reconnect_2\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_reconnect_2\",\"role\":\"assistant\",\"content\":[]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"item_id\":\"msg_ws_reconnect_2\",\"delta\":\"Beta.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.done\",\"item_id\":\"msg_ws_reconnect_2\",\"text\":\"Beta.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_reconnect_2\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Beta.\"}]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_reconnect_2\"}}"
+        }
+      ]
+    }
+  ]
+}

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

@@ -0,0 +1,129 @@
+{
+  "version": 1,
+  "metadata": {
+    "provider": "openai",
+    "protocol": "openai-responses",
+    "transport": "websocket",
+    "model": "gpt-5.5",
+    "tags": [
+      "prefix:openai-responses-websocket",
+      "provider:openai",
+      "protocol:openai-responses",
+      "transport:websocket",
+      "continuation",
+      "recovery"
+    ],
+    "name": "openai-responses-websocket/recovers-from-explicit-continuation-rejection",
+    "recordedAt": "2026-08-20T00:00:00.000Z"
+  },
+  "interactions": [
+    {
+      "transport": "websocket",
+      "connection": {
+        "sequence": 0,
+        "url": "wss://api.openai.com/v1/responses",
+        "protocols": [],
+        "close": {
+          "code": 1000,
+          "reason": ""
+        }
+      },
+      "events": [
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_rejection_1\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_rejection_1\",\"role\":\"assistant\",\"content\":[]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"item_id\":\"msg_ws_rejection_1\",\"delta\":\"Ready.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.done\",\"item_id\":\"msg_ws_rejection_1\",\"text\":\"Ready.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_rejection_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Ready.\"}]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_rejection_1\"}}"
+        }
+      ]
+    },
+    {
+      "transport": "websocket",
+      "connection": {
+        "sequence": 1,
+        "url": "wss://api.openai.com/v1/responses",
+        "protocols": [],
+        "close": {
+          "code": 1000,
+          "reason": ""
+        }
+      },
+      "events": [
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"previous_response_id\":\"resp_ws_rejection_1\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"error\",\"error\":{\"code\":\"previous_response_not_found\",\"message\":\"Previous response not found\"}}"
+        },
+        {
+          "direction": "client",
+          "kind": "text",
+          "body": "{\"type\":\"response.create\",\"model\":\"gpt-5.5\",\"input\":[{\"role\":\"system\",\"content\":\"Follow the user's exact reply instruction.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Ready.\"}]},{\"type\":\"message\",\"id\":\"msg_ws_rejection_1\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Ready.\"}]},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"Reply exactly: Recovered.\"}]}],\"store\":false,\"max_output_tokens\":30,\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"text\":{\"verbosity\":\"low\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.created\",\"response\":{\"id\":\"resp_ws_rejection_2\"}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.added\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_rejection_2\",\"role\":\"assistant\",\"content\":[]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.delta\",\"item_id\":\"msg_ws_rejection_2\",\"delta\":\"Recovered.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_text.done\",\"item_id\":\"msg_ws_rejection_2\",\"text\":\"Recovered.\"}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_ws_rejection_2\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"Recovered.\"}]}}"
+        },
+        {
+          "direction": "server",
+          "kind": "text",
+          "body": "{\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ws_rejection_2\"}}"
+        }
+      ]
+    }
+  ]
+}

packages/ai/test/provider-options/openai.types.ts

@@ -10,6 +10,11 @@ LLM.request({ model: selected, prompt: "Hello", providerOptions: { textVerbosity
 LLM.request({ model: selected, prompt: "Hello", providerOptions: { textVerbosity: "verbose" } })
 LLM.request({ model: chat, prompt: "Hello", providerOptions: { reasoningEffort: "max" } })
 LLM.request({ model: chat, prompt: "Hello", providerOptions: { reasoningEffort: "experimental" } })
+LLM.request({
+  model: selected,
+  prompt: "Hello",
+  providerOptions: { contextManagement: [{ type: "compaction", compactThreshold: 100_000 }] },
+})
 
 LLM.request({
   model: selected,

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

@@ -0,0 +1,216 @@
+import { describe, expect } from "bun:test"
+import { Effect, Stream } from "effect"
+import { Socket } from "effect/unstable/socket"
+import { LLM, LLMRequest, Message, ToolRuntime } from "../../src/index.js"
+import {
+  LLMClient,
+  WebSocketTransport,
+  type ChannelCheckpoint,
+  type ChannelObservation,
+  type WebSocketChannelExchange,
+  type WebSocketChannelExecutor,
+  type WebSocketConnection,
+} from "../../src/route.js"
+import { configure } from "../../src/providers/openai.js"
+import { decodeJson } from "../../src/protocols/shared.js"
+import { weatherRuntimeTool, weatherTool, weatherToolName } from "../recorded-scenarios.js"
+import { recordedTests } from "../recorded-test.js"
+
+const model = configure({ apiKey: process.env.OPENAI_API_KEY ?? "fixture" }).responses("gpt-5.5")
+const recorded = recordedTests({
+  prefix: "openai-responses-websocket",
+  provider: "openai",
+  protocol: "openai-responses",
+  requires: ["OPENAI_API_KEY"],
+  tags: ["transport:websocket"],
+  metadata: { transport: "websocket", model: model.id },
+})
+
+const observationFrame = (observation: ChannelObservation) => {
+  if (observation.type === "frame" || observation.type === "completed" || observation.type === "incomplete")
+    return Effect.succeed(observation.frame)
+  return Effect.fail(observation.error)
+}
+
+const terminal = (observation: ChannelObservation) => observation.type !== "frame"
+
+// This deliberately models only sequential test traffic. Core owns production connection pooling and recovery.
+const makeChannel = Effect.gen(function* () {
+  const constructor = yield* Socket.WebSocketConstructor
+  let connection: WebSocketConnection | undefined
+  let checkpoint: ChannelCheckpoint | undefined
+  let pending: ChannelCheckpoint | undefined
+  let opens = 0
+  const sent: unknown[] = []
+
+  const close = Effect.suspend(() => {
+    const current = connection
+    connection = undefined
+    return current ? current.close : Effect.void
+  })
+  yield* Effect.addFinalizer(() => close)
+
+  const executor: WebSocketChannelExecutor = {
+    execute: (exchange: WebSocketChannelExchange) =>
+      Effect.gen(function* () {
+        if (!connection) {
+          connection = yield* WebSocketTransport.open(exchange.connect).pipe(
+            Effect.provideService(Socket.WebSocketConstructor, constructor),
+          )
+          opens += 1
+        }
+        const current = connection
+        const create = yield* exchange.driver.create(checkpoint)
+        if (create.mode === "full") checkpoint = undefined
+        pending = undefined
+        sent.push(decodeJson(create.message))
+        yield* current.sendText(create.message)
+        const decoder = new TextDecoder()
+        return {
+          frames: current.messages.pipe(
+            Stream.map((message) => WebSocketTransport.messageText(message, decoder)),
+            Stream.mapEffect((frame) => exchange.driver.observe(create, frame)),
+            Stream.tap((observation) =>
+              Effect.sync(() => {
+                if (!terminal(observation)) return
+                pending = observation.type === "completed" ? observation.checkpoint : undefined
+                if (observation.type !== "completed") checkpoint = undefined
+              }),
+            ),
+            Stream.takeUntil(terminal),
+            Stream.mapEffect(observationFrame),
+          ),
+          complete: Effect.sync(() => {
+            checkpoint = pending
+            pending = undefined
+          }),
+        }
+      }),
+  }
+
+  return {
+    executor,
+    sent,
+    opens: () => opens,
+    reconnect: (preserveCheckpoint = false) =>
+      close.pipe(
+        Effect.andThen(
+          Effect.sync(() => {
+            pending = undefined
+            if (!preserveCheckpoint) checkpoint = undefined
+          }),
+        ),
+      ),
+  }
+})
+
+describe("OpenAI Responses WebSocket recorded", () => {
+  recorded.effect.with("continues a tool call over one socket", { tags: ["tool", "continuation"] }, () =>
+    Effect.gen(function* () {
+      const channel = yield* makeChannel
+      const request = LLM.request({
+        id: "recorded_openai_responses_websocket_tool",
+        model,
+        system: "Call get_weather once, then reply exactly: Paris is sunny.",
+        prompt: "What is the weather in Paris?",
+        tools: [weatherTool],
+        generation: { maxTokens: 50 },
+        cache: "none",
+      })
+      const first = yield* LLMClient.generate(request, { webSocket: channel.executor })
+      const call = first.toolCalls[0]
+      if (!call) yield* Effect.die("Expected get_weather tool call")
+      const result = yield* ToolRuntime.dispatch({ [weatherToolName]: weatherRuntimeTool }, call)
+      const second = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          messages: [
+            ...request.messages,
+            first.message,
+            Message.tool({ id: call.id, name: call.name, result: result.result }),
+          ],
+        }),
+        { webSocket: channel.executor },
+      )
+
+      expect(second.text).toBe("Paris is sunny.")
+      expect(channel.opens()).toBe(1)
+      expect(channel.sent).toHaveLength(2)
+      expect(channel.sent[1]).toMatchObject({
+        previous_response_id: expect.any(String),
+        input: [{ type: "function_call_output", call_id: call.id, output: expect.any(String) }],
+      })
+    }),
+  )
+
+  recorded.effect.with("reconstructs full context after reconnect", { tags: ["reconnect", "full-context"] }, () =>
+    Effect.gen(function* () {
+      const channel = yield* makeChannel
+      const request = LLM.request({
+        id: "recorded_openai_responses_websocket_reconnect",
+        model,
+        system: "Follow the user's exact reply instruction.",
+        prompt: "Reply exactly: Alpha.",
+        generation: { maxTokens: 30 },
+        cache: "none",
+      })
+      const first = yield* LLMClient.generate(request, { webSocket: channel.executor })
+      yield* channel.reconnect()
+      const second = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          messages: [...request.messages, first.message, Message.user("Reply exactly: Beta.")],
+        }),
+        { webSocket: channel.executor },
+      )
+
+      expect(first.text).toBe("Alpha.")
+      expect(second.text).toBe("Beta.")
+      expect(channel.opens()).toBe(2)
+      expect(channel.sent[1]).not.toHaveProperty("previous_response_id")
+      expect(channel.sent[1]).toMatchObject({
+        input: [
+          { role: "system", content: "Follow the user's exact reply instruction." },
+          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Alpha." }] },
+          { role: "assistant", content: [{ type: "output_text", text: "Alpha." }] },
+          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Beta." }] },
+        ],
+      })
+    }),
+  )
+
+  recorded.effect.with("recovers from explicit continuation rejection", { tags: ["continuation", "recovery"] }, () =>
+    Effect.gen(function* () {
+      const channel = yield* makeChannel
+      const request = LLM.request({
+        id: "recorded_openai_responses_websocket_rejection",
+        model,
+        system: "Follow the user's exact reply instruction.",
+        prompt: "Reply exactly: Ready.",
+        generation: { maxTokens: 30 },
+        cache: "none",
+      })
+      const first = yield* LLMClient.generate(request, { webSocket: channel.executor })
+      const continuation = LLMRequest.update(request, {
+        messages: [...request.messages, first.message, Message.user("Reply exactly: Recovered.")],
+      })
+      yield* channel.reconnect(true)
+      const rejected = yield* LLMClient.generate(continuation, { webSocket: channel.executor }).pipe(Effect.flip)
+      const recovered = yield* LLMClient.generate(continuation, { webSocket: channel.executor })
+
+      expect(rejected).toMatchObject({
+        reason: { _tag: "Transport", delivery: "rejected", recovery: "retry-full" },
+      })
+      expect(recovered.text).toBe("Recovered.")
+      expect(channel.opens()).toBe(2)
+      expect(channel.sent[1]).toHaveProperty("previous_response_id", expect.any(String))
+      expect(channel.sent[2]).not.toHaveProperty("previous_response_id")
+      expect(channel.sent[2]).toMatchObject({
+        input: [
+          { role: "system", content: "Follow the user's exact reply instruction." },
+          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Ready." }] },
+          { role: "assistant", content: [{ type: "output_text", text: "Ready." }] },
+          { role: "user", content: [{ type: "input_text", text: "Reply exactly: Recovered." }] },
+        ],
+      })
+    }),
+  )
+})

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

@@ -168,6 +168,50 @@ describe("OpenAI Responses route", () => {
     }),
   )
 
+  it.effect("enables server-side compaction", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLMRequest.update(request, {
+          providerOptions: {
+            contextManagement: [{ type: "compaction", compactThreshold: 100_000 }],
+          },
+        }),
+      )
+
+      expect(prepared.body.context_management).toEqual([{ type: "compaction", compact_threshold: 100_000 }])
+    }),
+  )
+
+  it.effect("replays durable server-side compaction items", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.make({
+              role: "assistant",
+              content: "After compaction",
+              native: {
+                openai: {
+                  compactionItems: [{ type: "compaction", id: "cmp_1", encrypted_content: "opaque-state" }],
+                },
+              },
+            }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        { type: "compaction", id: "cmp_1", encrypted_content: "opaque-state" },
+        {
+          type: "message",
+          role: "assistant",
+          content: [{ type: "output_text", text: "After compaction" }],
+        },
+      ])
+    }),
+  )
+
   it.effect("passes through custom OpenAI reasoning effort strings", () =>
     Effect.gen(function* () {
       const prepared = yield* compileRequest(
@@ -1506,6 +1550,45 @@ describe("OpenAI Responses route", () => {
     }),
   )
 
+  it.effect("retains server-side compaction output for continuation", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.done",
+                item: {
+                  type: "compaction",
+                  id: "cmp_1",
+                  encrypted_content: "opaque-state",
+                  status: "completed",
+                },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter(LLMEvent.is.stepFinish)).toEqual([
+        {
+          type: "step-finish",
+          index: 0,
+          reason: { normalized: "stop", raw: undefined },
+          providerMetadata: {
+            openai: {
+              responseId: "resp_1",
+              serviceTier: undefined,
+              compactionItems: [{ type: "compaction", id: "cmp_1", encrypted_content: "opaque-state" }],
+            },
+          },
+          usage: undefined,
+        },
+      ])
+    }),
+  )
+
   it.effect("preserves standard refusal content as ordinary assistant text", () =>
     Effect.gen(function* () {
       const response = yield* LLMClient.generate(request).pipe(

packages/ai/test/recorded-test.ts

@@ -1,5 +1,7 @@
 import { HttpRecorder } from "@opencode-ai/http-recorder"
+import { NodeSocket } from "@effect/platform-node"
 import { Layer } from "effect"
+import { Socket } from "effect/unstable/socket"
 import * as path from "node:path"
 import { fileURLToPath } from "node:url"
 import { LLMClient, RequestExecutor } from "../src/route.js"
@@ -16,7 +18,7 @@ import {
 const __dirname = path.dirname(fileURLToPath(import.meta.url))
 const FIXTURES_DIR = path.resolve(__dirname, "fixtures", "recordings")
 
-type RecordedEnv = RequestExecutorService | LLMClientService | ImageClientService
+type RecordedEnv = RequestExecutorService | LLMClientService | ImageClientService | Socket.WebSocketConstructor
 
 type RecordedTestsOptions = RecordedGroupOptions & {
   readonly options?: HttpRecorder.RecorderOptions
@@ -69,7 +71,7 @@ export const recordedTests = (options: RecordedTestsOptions) =>
         ...metadata,
       }
       if (recording) {
-        if (process.env.CI !== undefined) throw new Error("Unset CI before recording HTTP cassettes")
+        if (process.env.CI !== undefined) throw new Error("Unset CI before recording cassettes")
         HttpRecorder.removeCassetteSync(cassette, { directory: FIXTURES_DIR })
       }
       const requestExecutor = RequestExecutor.layer.pipe(
@@ -81,10 +83,16 @@ export const recordedTests = (options: RecordedTestsOptions) =>
           }),
         ),
       )
+      const webSocket = HttpRecorder.layerWebSocketConstructor(cassette, {
+        ...recorderOptions,
+        directory: FIXTURES_DIR,
+        metadata: recorderMetadata,
+      }).pipe(Layer.provide(NodeSocket.layerWebSocketConstructorWS))
       return Layer.mergeAll(
         requestExecutor,
         LLMClient.layer.pipe(Layer.provide(requestExecutor)),
         ImageClient.layer.pipe(Layer.provide(requestExecutor)),
+        webSocket,
       )
     },
   })

packages/core/src/mcp/index.ts

@@ -5,7 +5,7 @@ import { McpEvent } from "@opencode-ai/schema/mcp-event"
 import { Command } from "@opencode-ai/schema/command"
 import { createHash } from "node:crypto"
 import { isDeepStrictEqual } from "node:util"
-import { Cause, Context, Deferred, Effect, Exit, FiberSet, Layer, Schema, Scope, Stream, Types } from "effect"
+import { Cause, Context, Effect, Exit, FiberSet, Latch, Layer, Schema, Scope, Stream, Types } from "effect"
 import { makeLocationNode } from "@opencode-ai/util/effect/app-node"
 import { Credential } from "../credential.js"
 import { Bus } from "../bus.js"
@@ -111,7 +111,7 @@ export class ToolCallError extends Schema.TaggedError<ToolCallError>()("MCP.Tool
 type ServerEntry = {
   readonly config: Mcp.ServerConfig
   status: Status
-  readonly startup: Deferred.Deferred<void>
+  readonly startup: Latch.Latch
   scope?: Scope.Closeable
   client?: MCPClient.Connection
   tools?: ReadonlyArray<Tool>
@@ -535,7 +535,7 @@ export const layer = (options?: Options) =>
               : { status: "failed", error: error instanceof Error ? error.message : String(error) }
           yield* Effect.logWarning("mcp connect failed", { server: name, status: entry.status })
           yield* bus.publish(McpEvent.StatusChanged, { server: name }).pipe(Effect.ignore)
-        }).pipe(Effect.ensuring(Deferred.succeed(entry.startup, undefined)))
+        }).pipe(Effect.ensuring(entry.startup.open))
 
       const stopServer = Effect.fnUntraced(function* (name: ServerName, entry: ServerEntry) {
         const scope = entry.scope
@@ -562,7 +562,7 @@ export const layer = (options?: Options) =>
         const entry: ServerEntry = {
           config: serverConfig,
           status: { status: "pending" },
-          startup: Deferred.makeUnsafe<void>(),
+          startup: Latch.makeUnsafe(),
         }
         entries.set(name, entry)
         yield* Effect.gen(function* () {
@@ -575,7 +575,7 @@ export const layer = (options?: Options) =>
           yield* startServer(name, entry)
         }).pipe(
           // Settle startup even when registration fails or replacement is interrupted, so readers cannot hang.
-          Effect.ensuring(Effect.sync(() => Deferred.doneUnsafe(entry.startup, Exit.void))),
+          Effect.ensuring(entry.startup.open),
         )
       })
 
@@ -597,7 +597,7 @@ export const layer = (options?: Options) =>
             entries.set(name, {
               config: server,
               status: { status: "pending" },
-              startup: Deferred.makeUnsafe<void>(),
+              startup: Latch.makeUnsafe(),
             })
           }
           yield* Effect.forEach(entries, ([name, entry]) => register(name, entry), { discard: true })
@@ -607,7 +607,7 @@ export const layer = (options?: Options) =>
           for (const [name, entry] of entries) {
             if (entry.config.disabled) {
               entry.status = { status: "disabled" }
-              Deferred.doneUnsafe(entry.startup, Exit.void)
+              entry.startup.openUnsafe()
               yield* bus.publish(McpEvent.StatusChanged, { server: name }).pipe(Effect.ignore)
               continue
             }
@@ -685,7 +685,7 @@ export const layer = (options?: Options) =>
 
       // Suspend so each await sees current entries; a bare Map iterator is exhausted after one run.
       const whenAllReady = Effect.suspend(() =>
-        Effect.forEach(Array.from(entries.values()), (entry) => Deferred.await(entry.startup), {
+        Effect.forEach(Array.from(entries.values()), (entry) => entry.startup.await, {
           concurrency: "unbounded",
           discard: true,
         }),
@@ -734,7 +734,7 @@ export const layer = (options?: Options) =>
         }),
         callTool: Effect.fn("MCP.callTool")(function* (input) {
           const target = yield* requireServer(input.server)
-          yield* Deferred.await(target.entry.startup)
+          yield* target.entry.startup.await
           if (!target.entry.client)
             return yield* new ToolCallError({
               server: target.name,
@@ -773,7 +773,7 @@ export const layer = (options?: Options) =>
         }),
         prompt: Effect.fn("MCP.prompt")(function* (input) {
           const target = yield* requireServer(input.server)
-          yield* Deferred.await(target.entry.startup)
+          yield* target.entry.startup.await
           if (!target.entry.client) return undefined
           const result = yield* target.entry.client
             .prompt({ name: input.name, args: input.args })
@@ -827,7 +827,7 @@ export const layer = (options?: Options) =>
         }),
         readResource: Effect.fn("MCP.readResource")(function* (input) {
           const target = yield* requireServer(input.server)
-          yield* Deferred.await(target.entry.startup)
+          yield* target.entry.startup.await
           if (!target.entry.client) return undefined
           const result = yield* target.entry.client
             .readResource({ uri: input.uri })

packages/core/src/plugin/supervisor.ts

@@ -3,7 +3,7 @@ export { Service, type Interface } from "./supervisor-service.js"
 
 import type { Plugin as PluginDefinition } from "@opencode-ai/plugin/effect/plugin"
 import { Event } from "@opencode-ai/schema/config"
-import { Cause, Deferred, Effect, Layer, Schema, Stream } from "effect"
+import { Cause, Effect, Latch, Layer, Schema, Stream } from "effect"
 import path from "path"
 import { pathToFileURL } from "url"
 import { ConfigPluginSource } from "../config/plugin/source.js"
@@ -137,7 +137,7 @@ export const layer = Layer.effect(
     const sdk = yield* SdkPlugins.Service
     const sources = yield* ConfigPluginSource.Service
     const bus = yield* Bus.Service
-    const ready = { current: yield* Deferred.make<void>() }
+    const ready = yield* Latch.make()
     let observed = 0
 
     const activate = Effect.fn("PluginSupervisor.activate")(function* () {
@@ -164,7 +164,7 @@ export const layer = Layer.effect(
       Stream.mapEffect(() =>
         Effect.gen(function* () {
           observed++
-          if (yield* Deferred.isDone(ready.current)) ready.current = yield* Deferred.make<void>()
+          yield* ready.close
           return observed
         }),
       ),
@@ -176,12 +176,12 @@ export const layer = Layer.effect(
       Stream.runForEach((target) =>
         Effect.gen(function* () {
           yield* activate()
-          if (observed === target) yield* Deferred.succeed(ready.current, undefined)
+          if (observed === target) yield* ready.open
         }).pipe(Effect.catchCause((cause) => Effect.logError("failed to reload plugins", { cause }))),
       ),
       Effect.forkScoped({ startImmediately: true }),
     )
-    return Service.of({ flush: Effect.suspend(() => Deferred.await(ready.current)) })
+    return Service.of({ flush: ready.await })
   }),
 )
 

packages/core/src/session/compaction.ts

@@ -1,6 +1,6 @@
 export * as SessionCompaction from "./compaction.js"
 
-import { LLM, LLMClient, AIError, LLMEvent, Message, type LLMRequest } from "@opencode-ai/ai"
+import { LLMClient, AIError, LLMEvent, Message, type LLMRequest } from "@opencode-ai/ai"
 import type { StreamOptions } from "@opencode-ai/ai/route"
 import { SessionError } from "@opencode-ai/schema/session-error"
 import { Context, Effect, Layer, Stream } from "effect"
@@ -9,17 +9,12 @@ import { makeLocationNode } from "@opencode-ai/util/effect/app-node"
 import { llmClient } from "../effect/app-node-platform.js"
 import { SessionEvent } from "./event.js"
 import type { SessionMessage } from "./message.js"
-import { SessionModelHeaders } from "./model-headers.js"
-import { SessionModelHook } from "./model-hook.js"
-import { SessionModelHttp } from "./model-http.js"
-import { SessionPromptCacheKey } from "./prompt-cache-key.js"
-import { App } from "../app.js"
+import { SessionModelRequest } from "./model-request.js"
 import { SessionRunnerModel } from "./runner/model.js"
 import { SessionSchema } from "./schema.js"
 import { toSessionError } from "./to-session-error.js"
 import { Token } from "../util/token.js"
 import { SessionUsage } from "./usage.js"
-import { PluginHooks } from "../plugin/hooks.js"
 import { Agent } from "../agent.js"
 import { State } from "../state.js"
 
@@ -70,13 +65,12 @@ export type Draft = {
 }
 
 type Dependencies = {
-  readonly app: App.Info
   readonly bus: Bus.Interface
   readonly llm: {
     readonly stream: (request: LLMRequest, options?: StreamOptions) => Stream.Stream<LLMEvent, AIError>
   }
   readonly models: SessionRunnerModel.Interface
-  readonly hooks: PluginHooks.Interface
+  readonly modelRequests: SessionModelRequest.Interface
 }
 
 export type AutoInput = {
@@ -85,6 +79,8 @@ export type AutoInput = {
   readonly resolved: SessionRunnerModel.Resolved
 }
 
+type RequiredInput = Pick<AutoInput, "messages" | "resolved">
+
 export type ManualInput = {
   readonly session: SessionSchema.Info
   readonly messages: readonly SessionMessage.Info[]
@@ -92,8 +88,6 @@ export type ManualInput = {
   readonly started?: boolean
 }
 
-type RequiredInput = Pick<AutoInput, "messages" | "resolved">
-
 type Plan = {
   readonly session: SessionSchema.Info
   readonly resolved: SessionRunnerModel.Resolved
@@ -116,8 +110,19 @@ export interface Interface extends State.Transformable<Draft> {
 
 export class Service extends Context.Service<Service, Interface>()("@opencode/SessionCompaction") {}
 
-const truncate = (value: string) =>
-  value.length <= TOOL_OUTPUT_MAX_CHARS ? value : `${value.slice(0, TOOL_OUTPUT_MAX_CHARS)}\n[truncated]`
+export const truncateToolOutput = (value: string) => {
+  if (value.length <= TOOL_OUTPUT_MAX_CHARS) return value
+  let end = 0
+  for (let count = 0; count < TOOL_OUTPUT_MAX_CHARS && end < value.length; count++) {
+    const code = value.charCodeAt(end)
+    end +=
+      code >= 0xd800 && code <= 0xdbff && value.charCodeAt(end + 1) >= 0xdc00 && value.charCodeAt(end + 1) <= 0xdfff
+        ? 2
+        : 1
+  }
+  if (end === value.length) return value
+  return `${value.slice(0, end)}\n[truncated]`
+}
 
 export const serializeToolContent = (content: SessionMessage.ToolStateCompleted["content"]) =>
   content
@@ -146,7 +151,7 @@ const serialize = (message: SessionMessage.Info) => {
         if (part.state.status === "completed")
           return [
             `[Assistant tool call]: ${part.name}(${input})`,
-            `[Tool result]: ${truncate(serializeToolContent(part.state.content))}`,
+            `[Tool result]: ${truncateToolOutput(serializeToolContent(part.state.content))}`,
           ]
         if (part.state.status === "error")
           return [`[Assistant tool call]: ${part.name}(${input})`, `[Tool error]: ${part.state.error.message}`]
@@ -157,7 +162,8 @@ const serialize = (message: SessionMessage.Info) => {
   if (message.type === "system") return `[System update]: ${message.text}`
   if (message.type === "synthetic") return `[Synthetic context]: ${message.text}`
   if (message.type === "skill") return `[Skill activated: ${message.name}]\n${message.text}`
-  if (message.type === "shell") return `[Shell]: ${message.command}\n${truncate(message.output?.output ?? "")}`
+  if (message.type === "shell")
+    return `[Shell]: ${message.command}\n${truncateToolOutput(message.output?.output ?? "")}`
   return ""
 }
 
@@ -266,65 +272,51 @@ const make = (dependencies: Dependencies) => {
           })
         : Effect.void,
     )
-    const request = yield* SessionModelHook.apply(
-      dependencies.hooks,
-      { sessionID: plan.session.id, agent: Agent.ID.make("compaction"), model: plan.resolved.ref },
-      LLM.request({
-        model: plan.resolved.model,
-        promptCacheKey: SessionPromptCacheKey.make(plan.session.id),
-        http: { headers: SessionModelHeaders.make(plan.session, dependencies.app) },
-        messages: [Message.user(plan.prompt)],
-        tools: [],
+    const prepared = yield* dependencies.modelRequests.prepare({
+      scope: { session: plan.session, agentID: Agent.ID.make("compaction"), model: plan.resolved },
+      transcript: { system: [], messages: [Message.user(plan.prompt)] },
+      contextHooks: false,
+    })
+    yield* dependencies.llm.stream(prepared.request, prepared.options).pipe(
+      Stream.runForEach((event) => {
+        if (LLMEvent.is.providerError(event))
+          failure = {
+            type: event.classification === "context-overflow" ? "provider.invalid-request" : "provider.error",
+            message: event.message,
+          }
+        if (LLMEvent.is.textDelta(event)) {
+          chunks.push(event.text)
+          return dependencies.bus.publish(SessionEvent.Compaction.Delta, {
+            sessionID: plan.session.id,
+            text: event.text,
+          })
+        }
+        if (LLMEvent.is.stepFinish(event)) {
+          const step = SessionUsage.record(event.usage, plan.resolved.cost)
+          usage = usage ? SessionUsage.add(usage, step) : step
+        }
+        return Effect.void
       }),
-    )
-    yield* dependencies.llm
-      .stream(request, {
-        http: SessionModelHttp.middleware(dependencies.hooks, {
-          sessionID: plan.session.id,
-          agent: Agent.ID.make("compaction"),
-          model: plan.resolved.ref,
+      Effect.catchTag("AI.Error", (error) =>
+        Effect.sync(() => {
+          failure = toSessionError(error)
         }),
-      })
-      .pipe(
-        Stream.runForEach((event) => {
-          if (LLMEvent.is.providerError(event))
-            failure = {
-              type: event.classification === "context-overflow" ? "provider.invalid-request" : "provider.error",
-              message: event.message,
-            }
-          if (LLMEvent.is.textDelta(event)) {
-            chunks.push(event.text)
-            return dependencies.bus.publish(SessionEvent.Compaction.Delta, {
-              sessionID: plan.session.id,
-              text: event.text,
-            })
-          }
-          if (LLMEvent.is.stepFinish(event)) {
-            const step = SessionUsage.record(event.usage, plan.resolved.cost)
-            usage = usage ? SessionUsage.add(usage, step) : step
-          }
-          return Effect.void
-        }),
-        Effect.catchTag("AI.Error", (error) =>
-          Effect.sync(() => {
-            failure = toSessionError(error)
-          }),
-        ),
-        Effect.onInterrupt(() =>
-          recordUsage.pipe(
-            Effect.andThen(
-              plan.reason === "auto"
-                ? failed({
-                    sessionID: plan.session.id,
-                    reason: plan.reason,
-                    error: { type: "compaction.interrupted", message: "Compaction was interrupted" },
-                    inputID: plan.inputID,
-                  }).pipe(Effect.asVoid)
-                : Effect.void,
-            ),
+      ),
+      Effect.onInterrupt(() =>
+        recordUsage.pipe(
+          Effect.andThen(
+            plan.reason === "auto"
+              ? failed({
+                  sessionID: plan.session.id,
+                  reason: plan.reason,
+                  error: { type: "compaction.interrupted", message: "Compaction was interrupted" },
+                  inputID: plan.inputID,
+                }).pipe(Effect.asVoid)
+              : Effect.void,
           ),
         ),
-      )
+      ),
+    )
     yield* recordUsage
     const summary = chunks.join("")
     if (failure || !summary.trim()) {
@@ -425,14 +417,13 @@ export const layer = Layer.effect(
     const bus = yield* Bus.Service
     const llm = yield* LLMClient.Service
     const models = yield* SessionRunnerModel.Service
-    const app = yield* App.Metadata
-    const hooks = yield* PluginHooks.Service
-    return make({ bus, llm, models, app, hooks })
+    const modelRequests = yield* SessionModelRequest.Service
+    return make({ bus, llm, models, modelRequests })
   }),
 )
 
 export const node = makeLocationNode({
   service: Service,
   layer,
-  deps: [Bus.node, llmClient, SessionRunnerModel.node, App.node, PluginHooks.node],
+  deps: [Bus.node, llmClient, SessionRunnerModel.node, SessionModelRequest.node],
 })

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

@@ -61,13 +61,20 @@ interface PrepareInput {
     readonly session: SessionSchema.Info
     readonly agentID: Agent.ID
     readonly model: SessionRunnerModel.Resolved
-    readonly tools: Tool.Snapshot
+    /** Omitted for requests that carry no tools (title, compaction). */
+    readonly tools?: Tool.Snapshot
   }
   readonly transcript: {
     readonly system: Array<SystemPart>
     readonly messages: Array<Message>
   }
   readonly toolChoice?: LLM.RequestInput["toolChoice"]
+  /**
+   * Session context hooks shape the agent conversation. Requests that are not
+   * part of the conversation (title, compaction) opt out: their transcripts
+   * pass through unchanged.
+   */
+  readonly contextHooks?: false
   /** Stateful Session WebSocket channels require an explicit durable-runner opt-in. */
   readonly webSocket?: "session"
 }
@@ -209,7 +216,10 @@ export const layer = Layer.effect(
       const session = input.scope.session
       const resolved = input.scope.model
       const model = resolved.model
-      const tools = input.scope.tools
+      const tools = input.scope.tools ?? {
+        definitions: [],
+        execute: () => new Tool.Error({ message: "Tools are not available for this request" }),
+      }
       const registry = new Map(tools.definitions.map((tool) => [tool.name, tool]))
       // The definition objects we hand to hooks, mapped back to their tools. Hooks rename a
       // tool by moving its definition to a new key; recognizing the object recovers the tool.
@@ -219,14 +229,18 @@ export const layer = Layer.effect(
         ),
       )
       // Hooks mutate this record in place: edit descriptions and schemas, rename, or remove.
-      const context = yield* hooks.trigger("session", "context", {
-        sessionID: session.id,
-        agent: input.scope.agentID,
-        model: resolved.ref,
-        system: input.transcript.system,
-        messages: input.transcript.messages,
-        tools: Object.fromEntries(Array.from(given, ([definition, tool]) => [tool.name, definition])),
-      })
+      const definitions = Object.fromEntries(Array.from(given, ([definition, tool]) => [tool.name, definition]))
+      const context =
+        input.contextHooks === false
+          ? { system: input.transcript.system, messages: input.transcript.messages, tools: definitions }
+          : yield* hooks.trigger("session", "context", {
+              sessionID: session.id,
+              agent: input.scope.agentID,
+              model: resolved.ref,
+              system: input.transcript.system,
+              messages: input.transcript.messages,
+              tools: definitions,
+            })
       // Match each surviving entry back to its tool, by recognizing a moved definition or
       // by key. Identity wins so a definition moved onto another tool's name still executes
       // the tool it describes. Entries matching neither were invented by a hook and dropped.

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

@@ -89,7 +89,15 @@ const classifyToolExits = (
     .flatMap((cause) => {
       if (Cause.hasInterrupts(cause)) return []
       const reasons = cause.reasons.flatMap(
-        (reason): Array<Cause.Reason<never>> => (Cause.isFailReason(reason) ? [] : [reason]),
+        (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)] : []
     })
@@ -315,7 +323,6 @@ const layer = Layer.effect(
       const loaded = yield* context.load(selected)
       const { session, agent } = loaded
       const resolved = loaded.model
-      const model = resolved.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 }

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

@@ -146,6 +146,10 @@ const assistant = (message: SessionMessage.Assistant, model: Model.Ref, provider
   const sameProvider = String(message.model.providerID) === String(model.providerID)
   const sameModel = sameProvider && String(message.model.id) === String(model.id)
   const reuseProviderMetadata = sameModel && message.error === undefined
+  const native =
+    reuseProviderMetadata && Array.isArray(message.providerState?.compactionItems)
+      ? { [providerMetadataKey]: message.providerState }
+      : undefined
   const content = message.content.flatMap((item): ContentPart[] => {
     if (item.type === "text")
       return [
@@ -204,9 +208,15 @@ const assistant = (message: SessionMessage.Assistant, model: Model.Ref, provider
     )
     .filter((message) => message !== undefined)
     .map(Message.tool)
-  if (meaningful.length === 0) return results
+  if (meaningful.length === 0 && native === undefined) return results
   return [
-    Message.make({ id: message.id, role: "assistant", content: meaningful, metadata: message.metadata }),
+    Message.make({
+      id: message.id,
+      role: "assistant",
+      content: meaningful,
+      metadata: message.metadata,
+      native,
+    }),
     ...results,
   ]
 }

packages/core/src/session/title.ts

@@ -1,6 +1,6 @@
 export * as SessionTitle from "./title.js"
 
-import { LLM, LLMClient, AIError, LLMEvent, Message, type LLMRequest } from "@opencode-ai/ai"
+import { LLMClient, AIError, LLMEvent, Message, SystemPart, type LLMRequest } from "@opencode-ai/ai"
 import type { StreamOptions } from "@opencode-ai/ai/route"
 import { Context, DateTime, Effect, Layer, Stream } from "effect"
 import { Agent } from "../agent.js"
@@ -8,14 +8,10 @@ import { Database } from "../database/database.js"
 import { Bus } from "../bus.js"
 import { makeLocationNode } from "@opencode-ai/util/effect/app-node"
 import { isExactRootFallback } from "@opencode-ai/util/session-title-fallback"
-import { App } from "../app.js"
 import { llmClient } from "../effect/app-node-platform.js"
-import { PluginHooks } from "../plugin/hooks.js"
 import { SessionEvent } from "./event.js"
 import { SessionHistory } from "./history.js"
-import { SessionModelHeaders } from "./model-headers.js"
-import { SessionModelHook } from "./model-hook.js"
-import { SessionModelHttp } from "./model-http.js"
+import { SessionModelRequest } from "./model-request.js"
 import { SessionRunnerModel } from "./runner/model.js"
 import { SessionSchema } from "./schema.js"
 import { SessionUsage } from "./usage.js"
@@ -25,15 +21,14 @@ const MAX_LENGTH = 100
 const titleChanged = Symbol("Session title changed")
 
 type Dependencies = {
-  readonly app: App.Info
   readonly bus: Bus.Interface
   readonly llm: {
     readonly stream: (request: LLMRequest, options?: StreamOptions) => Stream.Stream<LLMEvent, AIError>
   }
   readonly agents: Agent.Interface
   readonly models: SessionRunnerModel.Interface
+  readonly modelRequests: SessionModelRequest.Interface
   readonly store: SessionStore.Interface
-  readonly hooks: PluginHooks.Interface
 }
 
 export interface Interface {
@@ -81,39 +76,28 @@ const make = (dependencies: Dependencies) => {
           })
         : Effect.void,
     )
-    const request = yield* SessionModelHook.apply(
-      dependencies.hooks,
-      { sessionID: session.id, agent: agent.id, model: resolved.ref },
-      LLM.request({
-        model: resolved.model,
-        http: { headers: SessionModelHeaders.make(session, dependencies.app) },
-        system: agent.system,
+    const prepared = yield* dependencies.modelRequests.prepare({
+      scope: { session, agentID: agent.id, model: resolved },
+      transcript: {
+        system: agent.system ? [SystemPart.make(agent.system)] : [],
         messages: [Message.user(firstUser.text)],
-        tools: [],
+      },
+      contextHooks: false,
+    })
+    const streamed = yield* dependencies.llm.stream(prepared.request, prepared.options).pipe(
+      Stream.runForEach((event) => {
+        if (LLMEvent.is.providerError(event)) failed = true
+        if (LLMEvent.is.textDelta(event)) chunks.push(event.text)
+        if (LLMEvent.is.stepFinish(event)) {
+          const step = SessionUsage.record(event.usage, resolved.cost)
+          usage = usage ? SessionUsage.add(usage, step) : step
+        }
+        return Effect.void
       }),
+      Effect.as(true),
+      Effect.catchTag("AI.Error", () => Effect.succeed(false)),
+      Effect.onInterrupt(() => recordUsage.pipe(Effect.asVoid)),
     )
-    const streamed = yield* dependencies.llm
-      .stream(request, {
-        http: SessionModelHttp.middleware(dependencies.hooks, {
-          sessionID: session.id,
-          agent: agent.id,
-          model: resolved.ref,
-        }),
-      })
-      .pipe(
-        Stream.runForEach((event) => {
-          if (LLMEvent.is.providerError(event)) failed = true
-          if (LLMEvent.is.textDelta(event)) chunks.push(event.text)
-          if (LLMEvent.is.stepFinish(event)) {
-            const step = SessionUsage.record(event.usage, resolved.cost)
-            usage = usage ? SessionUsage.add(usage, step) : step
-          }
-          return Effect.void
-        }),
-        Effect.as(true),
-        Effect.catchTag("AI.Error", () => Effect.succeed(false)),
-        Effect.onInterrupt(() => recordUsage.pipe(Effect.asVoid)),
-      )
     yield* recordUsage
     if (!streamed || failed) return
     const title = chunks
@@ -146,11 +130,10 @@ export const layer = Layer.effect(
     const llm = yield* LLMClient.Service
     const agents = yield* Agent.Service
     const models = yield* SessionRunnerModel.Service
+    const modelRequests = yield* SessionModelRequest.Service
     const store = yield* SessionStore.Service
     const database = yield* Database.Service
-    const app = yield* App.Metadata
-    const hooks = yield* PluginHooks.Service
-    const title = make({ bus, llm, agents, models, store, app, hooks })
+    const title = make({ bus, llm, agents, models, modelRequests, store })
     return Service.of({
       generateForFirstPrompt: (sessionID) => title.generateForFirstPrompt(database.db, sessionID),
     })
@@ -165,9 +148,8 @@ export const node = makeLocationNode({
     llmClient,
     Agent.node,
     SessionRunnerModel.node,
+    SessionModelRequest.node,
     SessionStore.node,
     Database.node,
-    App.node,
-    PluginHooks.node,
   ],
 })

packages/core/src/shell.ts

@@ -1,7 +1,7 @@
 export * as Shell from "./shell.js"
 
 import path from "path"
-import { Context, Deferred, Duration, Effect, Fiber, Layer, Schema, Schedule, Stream } from "effect"
+import { Context, Deferred, Duration, Effect, Fiber, Latch, Layer, Schema, Schedule, Stream } from "effect"
 import { ChildProcess } from "effect/unstable/process"
 import { produce } from "immer"
 import { Shell } from "@opencode-ai/schema/shell"
@@ -286,7 +286,7 @@ const layer = () =>
               sessions.set(id, session)
 
               const stream = createWriteStream(file)
-              const outputDone = Deferred.makeUnsafe<void>()
+              const outputDone = Latch.makeUnsafe()
               const pump = handle.all.pipe(
                 Stream.runForEach((chunk: Uint8Array) =>
                   Effect.sync(() => {
@@ -304,8 +304,8 @@ const layer = () =>
                         stream.end(() => resolve())
                       }),
                   )
-                  yield* Deferred.succeed(outputDone, undefined)
-                }).pipe(Effect.catch(() => Deferred.succeed(outputDone, undefined))),
+                  yield* outputDone.open
+                }).pipe(Effect.catch(() => outputDone.open)),
               )
               yield* Effect.promise(
                 () =>
@@ -324,7 +324,7 @@ const layer = () =>
                     draft.time.completed = Date.now()
                   })
                   yield* beforeWait
-                  yield* Deferred.await(outputDone)
+                  yield* outputDone.await
                   // Resolve waiters with the terminal Info before any retention eviction, so an evicted
                   // session still reports success rather than the removal NotFoundError. This runs before
                   // the timeout-fiber interrupt below, which on the timeout path would otherwise cancel

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

@@ -2,7 +2,6 @@ export * as ShellTool from "./shell.js"
 
 import path from "path"
 import { ToolFailure } from "@opencode-ai/ai"
-import type { Content } from "@opencode-ai/schema/tool"
 import type { Context as PluginContext } from "@opencode-ai/plugin/effect/plugin"
 import { Deferred, Effect, Schema, Scope } from "effect"
 import { Config } from "../../config.js"
@@ -72,10 +71,27 @@ const Output = Schema.Struct({
 
 type Output = typeof Output.Type
 
-const modelOutput = (output: Output): string | undefined => {
-  if (output.status === "running") return BACKGROUND_INSTRUCTION
-  if (output.timeout) return "Command timed out before completion."
-  return `Command exited with code ${output.exit}.`
+const resultMessages = (output: Output) => {
+  const notice = (() => {
+    if (output.status === "running") return BACKGROUND_INSTRUCTION
+    if (output.timeout) return "Command timed out before completion."
+    if (output.exit !== undefined) return `Command exited with code ${output.exit}.`
+  })()
+  return [output.output, ...(notice ? [notice] : [])]
+}
+
+const toolResult = (output: Output) => {
+  return {
+    output,
+    content: resultMessages(output).map((text) => ({ type: "text" as const, text })),
+    metadata: {
+      status: output.status,
+      truncated: output.truncated,
+      ...(output.exit !== undefined ? { exit: output.exit } : {}),
+      ...(output.shellID !== undefined ? { shellID: output.shellID } : {}),
+      ...(output.timeout !== undefined ? { timeout: output.timeout } : {}),
+    },
+  }
 }
 
 export const Plugin = {
@@ -92,32 +108,50 @@ export const Plugin = {
     const notifyWhenDone = Effect.fn("ShellTool.notifyWhenDone")(function* (
       sessionID: SessionSchema.ID,
       id: string,
+      shellID: string,
       command: string,
+      settled: Deferred.Deferred<Output>,
     ) {
       yield* runtime.job.wait({ id: id }).pipe(
-        Effect.flatMap((result) => {
-          const state =
-            result.info?.status === "completed"
-              ? "completed"
-              : result.info?.status === "error"
-                ? "error"
-                : result.info?.status === "cancelled"
-                  ? "cancelled"
-                  : undefined
-          if (state === undefined) return Effect.void
-          const text =
-            state === "completed"
-              ? (result.info!.output ?? "")
+        Effect.flatMap((result) =>
+          Effect.gen(function* () {
+            const info = result.info
+            if (!info) return
+            const state =
+              info.status === "completed"
+                ? "completed"
+                : info.status === "error"
+                  ? "error"
+                  : info.status === "cancelled"
+                    ? "cancelled"
+                    : undefined
+            if (state === undefined) return
+            const output = state === "completed" ? yield* Deferred.await(settled) : undefined
+            const text = output
+              ? resultMessages(output).join("\n\n")
               : state === "error"
-                ? (result.info!.error ?? "Command failed")
+                ? (info.error ?? "Command failed")
                 : "Command cancelled"
-          return runtime.session.synthetic({
-            sessionID,
-            text: `<shell id="${id}" state="${state}" command="${command}">\n${text}\n</shell>`,
-            description: command,
-            metadata: { source: "shell", jobID: id, state },
-          })
-        }),
+            yield* runtime.session.synthetic({
+              sessionID,
+              text: `<shell id="${id}" state="${state}" command="${command}">\n${text}\n</shell>`,
+              description: command,
+              metadata: {
+                source: "shell",
+                jobID: id,
+                shellID,
+                state,
+                ...(output
+                  ? {
+                      truncated: output.truncated,
+                      ...(output.exit !== undefined ? { exit: output.exit } : {}),
+                      ...(output.timeout !== undefined ? { timeout: output.timeout } : {}),
+                    }
+                  : {}),
+              },
+            })
+          }),
+        ),
         Effect.forkIn(scope, { startImmediately: true }),
       )
     })
@@ -268,7 +302,7 @@ export const Plugin = {
 
               if (input.background === true) {
                 yield* runtime.job.background(job.id)
-                yield* notifyWhenDone(context.sessionID, context.id, info.command)
+                yield* notifyWhenDone(context.sessionID, context.id, info.id, info.command, settled)
                 return {
                   output: BACKGROUND_STARTED,
                   shellID: info.id,
@@ -282,7 +316,7 @@ export const Plugin = {
                 .pipe(Effect.onInterrupt(() => runtime.job.cancel(job.id).pipe(Effect.ignore)))
               if (result?.type === "backgrounded") {
                 yield* shell.timeout(info.id, 0)
-                yield* notifyWhenDone(context.sessionID, context.id, info.command)
+                yield* notifyWhenDone(context.sessionID, context.id, info.id, info.command, settled)
                 return {
                   output: BACKGROUND_STARTED,
                   shellID: info.id,
@@ -296,22 +330,7 @@ export const Plugin = {
 
               return yield* Deferred.await(settled)
             }).pipe(
-              Effect.map((output) => {
-                const content: Array<Content> = [{ type: "text", text: output.output }]
-                const model = modelOutput(output)
-                if (model) content.push({ type: "text", text: model })
-                return {
-                  output,
-                  content,
-                  metadata: {
-                    status: output.status,
-                    truncated: output.truncated,
-                    ...("exit" in output && output.exit !== undefined ? { exit: output.exit } : {}),
-                    ...("shellID" in output && output.shellID !== undefined ? { shellID: output.shellID } : {}),
-                    ...("timeout" in output && output.timeout !== undefined ? { timeout: output.timeout } : {}),
-                  },
-                }
-              }),
+              Effect.map(toolResult),
               Effect.mapError(
                 (error) => new ToolFailure({ message: `Unable to execute command: ${input.command}`, error }),
               ),

packages/core/src/tool/runtime.ts

@@ -13,7 +13,20 @@ export const definition = (tool: Tool.Info<any, any>): ToolDefinition => ({
 export const execute = (tool: Tool.Info<any, any>, input: unknown, context: Tool.Context) =>
   Effect.gen(function* () {
     const decoded = yield* decodeInput(tool.input, input)
-    const result = yield* tool.execute(decoded, context)
+    // Tool implementations declare `Tool.Error` but plugins can fail with anything at
+    // runtime. A foreign typed failure would slip past every `catchTag("Tool.Error")`
+    // downstream and leave its call permanently unsettled, so the declared contract is
+    // enforced here at the untrusted boundary. Declines tunnel through as defects and
+    // interrupts are not errors; neither is touched.
+    const result = yield* tool.execute(decoded, context).pipe(
+      Effect.mapError((error: unknown) =>
+        error instanceof Tool.Error
+          ? error
+          : new Tool.Error({
+              message: error instanceof globalThis.Error ? error.message : String(error),
+            }),
+      ),
+    )
     if (tool.output === undefined) {
       if ("output" in result) return yield* Effect.die("Tool result declared output without an output schema")
       return {

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

@@ -1,5 +1,5 @@
 import { expect, test } from "bun:test"
-import { LLMClient, LLMEvent, LanguageModel, type LLMRequest } from "@opencode-ai/ai"
+import { LLMClient, LLMEvent, LanguageModel, SystemPart, type LLMRequest } from "@opencode-ai/ai"
 import { OpenAIChat } from "@opencode-ai/ai/protocols"
 import { Database } from "@opencode-ai/core/database/database"
 import { AppNodeBuilder } from "@opencode-ai/core/effect/app-node-builder"
@@ -19,6 +19,7 @@ import { Project } from "@opencode-ai/core/project"
 import { ProjectTable } from "@opencode-ai/core/project/sql"
 import { App } from "@opencode-ai/core/app"
 import { Agent } from "@opencode-ai/core/agent"
+import { PluginHooks } from "@opencode-ai/core/plugin/hooks"
 import { Location } from "@opencode-ai/core/location"
 import { AbsolutePath } from "@opencode-ai/core/schema"
 import { Money } from "@opencode-ai/schema/money"
@@ -76,7 +77,14 @@ const models = Layer.mock(SessionRunnerModel.Service)({
 })
 const it = testEffect(
   AppNodeBuilder.build(
-    LayerNode.group([Database.node, Bus.node, SessionProjector.node, SessionStore.node, SessionCompaction.node]),
+    LayerNode.group([
+      Database.node,
+      Bus.node,
+      SessionProjector.node,
+      SessionStore.node,
+      PluginHooks.node,
+      SessionCompaction.node,
+    ]),
     [
       [Bus.node, Bus.configured({ persist: true })],
       [llmClient, client],
@@ -110,6 +118,13 @@ test("compaction describes tool media without embedding base64", () => {
   expect(serialized).not.toContain(base64)
 })
 
+test("compaction truncation does not split surrogate pairs", () => {
+  const prefix = "a".repeat(1_999)
+
+  expect(SessionCompaction.truncateToolOutput(`${prefix}😀suffix`)).toBe(`${prefix}😀\n[truncated]`)
+  expect(SessionCompaction.truncateToolOutput("😀".repeat(2_000))).toBe("😀".repeat(2_000))
+})
+
 test("compaction prompt requires the checkpoint headings in order", () => {
   const prompt = SessionCompaction.buildPrompt({ context: ["Conversation history"] })
   expect(prompt.match(/^#{2,3} .+$/gm)).toEqual([
@@ -174,6 +189,35 @@ it.effect("auto compaction reserves a buffer below the prompt ceiling", () =>
   }),
 )
 
+/** Seeds the global project plus one session row, returning the projected session. */
+const insertSession = (id: Session.ID, overrides?: Partial<typeof SessionTable.$inferInsert>) =>
+  Effect.gen(function* () {
+    const db = (yield* Database.Service).db
+    yield* db
+      .insert(ProjectTable)
+      .values({ id: Project.ID.global, worktree: AbsolutePath.make("/project"), sandboxes: [] })
+      .onConflictDoNothing()
+      .run()
+      .pipe(Effect.orDie)
+    yield* db
+      .insert(SessionTable)
+      .values({
+        id,
+        project_id: Project.ID.global,
+        slug: id,
+        directory: "/project",
+        title: id,
+        version: "test",
+        ...overrides,
+      })
+      .run()
+      .pipe(Effect.orDie)
+    const store = yield* SessionStore.Service
+    return yield* store
+      .get(id)
+      .pipe(Effect.flatMap((session) => (session ? Effect.succeed(session) : Effect.die(`session missing: ${id}`))))
+  })
+
 it.effect("manual compaction summarizes short context instead of no-op", () =>
   Effect.gen(function* () {
     requests = []
@@ -189,33 +233,7 @@ it.effect("manual compaction summarizes short context instead of no-op", () =>
       text: "Manual compaction should include this short conversation.",
       time: { created: DateTime.makeUnsafe(0) },
     }
-    yield* db
-      .insert(ProjectTable)
-      .values({ id: Project.ID.global, worktree: AbsolutePath.make("/project"), sandboxes: [] })
-      .onConflictDoNothing()
-      .run()
-      .pipe(Effect.orDie)
-    yield* db
-      .insert(SessionTable)
-      .values({
-        id: sessionID,
-        project_id: Project.ID.global,
-        parent_id: parentID,
-        slug: "manual-compaction",
-        directory: "/project",
-        title: "Manual compaction",
-        version: "test",
-      })
-      .run()
-      .pipe(Effect.orDie)
-
-    const session = yield* store
-      .get(sessionID)
-      .pipe(
-        Effect.flatMap((session) =>
-          session ? Effect.succeed(session) : Effect.die("manual compaction test session missing"),
-        ),
-      )
+    const session = yield* insertSession(sessionID, { parent_id: parentID })
 
     const delta = yield* bus
       .subscribe(SessionEvent.Compaction.Delta)
@@ -265,3 +283,66 @@ it.effect("manual compaction summarizes short context instead of no-op", () =>
     ])
   }),
 )
+
+it.effect("forked session compaction reuses the fork root prompt cache key", () =>
+  Effect.gen(function* () {
+    requests = []
+    const compaction = yield* SessionCompaction.Service
+    const sessionID = Session.ID.make("ses_fork_compaction")
+    const rootID = Session.ID.make("ses_fork_compaction_root")
+    const session = yield* insertSession(sessionID, {
+      fork_session_id: rootID,
+      fork_boundary: { type: "before", messageID: SessionMessage.ID.create() },
+    })
+    expect(
+      yield* compaction.compactManual({
+        session,
+        messages: [
+          {
+            id: SessionMessage.ID.create(),
+            type: "user",
+            text: "Summarize the forked conversation.",
+            time: { created: DateTime.makeUnsafe(0) },
+          },
+        ],
+        inputID: SessionMessage.ID.make("msg_fork_compaction"),
+      }),
+    ).toEqual({ status: "completed" })
+
+    expect(requests).toHaveLength(1)
+    expect(requests[0]?.promptCacheKey).toBe(rootID)
+  }),
+)
+
+it.effect("keeps session context hooks away from compaction requests", () =>
+  Effect.gen(function* () {
+    requests = []
+    const compaction = yield* SessionCompaction.Service
+    // Context hooks shape the agent conversation; compaction is not part of it,
+    // so it opts out and the transcript passes through unchanged.
+    const hooks = yield* PluginHooks.Service
+    yield* hooks.register("session", "context", (event) =>
+      Effect.sync(() => {
+        event.system.push(SystemPart.make("Injected conversation context"))
+      }),
+    )
+    const session = yield* insertSession(Session.ID.make("ses_hook_compaction"))
+    expect(
+      yield* compaction.compactManual({
+        session,
+        messages: [
+          {
+            id: SessionMessage.ID.create(),
+            type: "user",
+            text: "Summarize this conversation.",
+            time: { created: DateTime.makeUnsafe(0) },
+          },
+        ],
+        inputID: SessionMessage.ID.make("msg_hook_compaction"),
+      }),
+    ).toEqual({ status: "completed" })
+
+    expect(requests).toHaveLength(1)
+    expect(requests[0]?.system).toEqual([])
+  }),
+)

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

@@ -1019,6 +1019,40 @@ Recent work
     ])
   })
 
+  test("carries same-model server compaction state as native message data", () => {
+    const messages = toLLMMessages(
+      [
+        SessionMessage.Assistant.make({
+          id: id("assistant-compaction"),
+          type: "assistant",
+          agent: build,
+          model,
+          content: [],
+          providerState: {
+            responseId: "resp_1",
+            compactionItems: [{ type: "compaction", id: "cmp_1", encrypted_content: "opaque-state" }],
+          },
+          time: { created, completed: created },
+        }),
+      ],
+      model,
+    )
+
+    expect(messages).toEqual([
+      Message.make({
+        id: id("assistant-compaction"),
+        role: "assistant",
+        content: [],
+        native: {
+          provider: {
+            responseId: "resp_1",
+            compactionItems: [{ type: "compaction", id: "cmp_1", encrypted_content: "opaque-state" }],
+          },
+        },
+      }),
+    ])
+  })
+
   test("preserves assistant text provider state across same-provider model changes and failures", () => {
     const messages = toLLMMessages(
       [

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

@@ -1,5 +1,5 @@
 import { expect } from "bun:test"
-import { LLMClient, LLMEvent, LanguageModel, type LLMRequest } from "@opencode-ai/ai"
+import { LLMClient, LLMEvent, LanguageModel, SystemPart, type LLMRequest } from "@opencode-ai/ai"
 import { OpenAIChat } from "@opencode-ai/ai/protocols"
 import { Agent } from "@opencode-ai/core/agent"
 import { Database } from "@opencode-ai/core/database/database"
@@ -14,6 +14,7 @@ import { SessionRunnerModel } from "@opencode-ai/core/session/runner/model"
 import { SessionTable } from "@opencode-ai/core/session/sql"
 import { SessionStore } from "@opencode-ai/core/session/store"
 import { SessionTitle } from "@opencode-ai/core/session/title"
+import { PluginHooks } from "@opencode-ai/core/plugin/hooks"
 import { Session } from "@opencode-ai/core/session"
 import { Project } from "@opencode-ai/core/project"
 import { ProjectTable } from "@opencode-ai/core/project/sql"
@@ -78,7 +79,15 @@ const models = Layer.mock(SessionRunnerModel.Service)({
 })
 const it = testEffect(
   AppNodeBuilder.build(
-    LayerNode.group([Database.node, Bus.node, SessionProjector.node, SessionStore.node, Agent.node, SessionTitle.node]),
+    LayerNode.group([
+      Database.node,
+      Bus.node,
+      SessionProjector.node,
+      SessionStore.node,
+      Agent.node,
+      PluginHooks.node,
+      SessionTitle.node,
+    ]),
     [
       [llmClient, client],
       [SessionRunnerModel.node, models],
@@ -155,6 +164,9 @@ it.effect("generates a title from the sole user message and renames the session"
       "x-opencode-session": sessionID,
       "x-opencode-client": "opencode",
     })
+    expect(requests[0]?.promptCacheKey).toBe(sessionID)
+    expect(requests[0]?.tools).toEqual([])
+    expect(requests[0]?.system.map((part) => part.text)).toEqual(["You are a title generator."])
     expect(JSON.stringify(requests[0]?.messages)).toContain("Help me debug the failing build")
     const renamed = yield* store.get(sessionID)
     expect(renamed?.title).toBe("Generated Title")
@@ -323,6 +335,38 @@ it.effect("retries after a failed title request", () =>
   }),
 )
 
+it.effect("keeps session context hooks away from title requests", () =>
+  Effect.gen(function* () {
+    requests = []
+    titleStream = successfulTitle
+    const agentService = yield* Agent.Service
+    yield* agentService.transform((editor) => {
+      editor.update(Agent.ID.make("title"), (agent) => {
+        agent.mode = "primary"
+        agent.hidden = true
+        agent.system = "You are a title generator."
+      })
+    })
+    // Context hooks shape the agent conversation; title generation is not part of
+    // it, so it opts out and the transcript passes through unchanged.
+    const hooks = yield* PluginHooks.Service
+    yield* hooks.register("session", "context", (event) =>
+      Effect.sync(() => {
+        event.system.push(SystemPart.make("Keep titles in sentence case."))
+      }),
+    )
+    const sessionID = Session.ID.make("ses_title_context_hook")
+    yield* insertSession(sessionID)
+    yield* prompt(sessionID, "Hook this title request")
+
+    const title = yield* SessionTitle.Service
+    yield* title.generateForFirstPrompt(sessionID)
+
+    expect(requests).toHaveLength(1)
+    expect(requests[0]?.system.map((part) => part.text)).toEqual(["You are a title generator."])
+  }),
+)
+
 it.effect("preserves a manual rename completed while generation is in flight", () =>
   Effect.gen(function* () {
     requests = []

packages/core/test/tool-execute.test.ts

@@ -94,6 +94,24 @@ test("declared outputs cannot bypass validation and raw outputs stay JSON-compat
   )
 })
 
+test("foreign typed failures settle as Tool.Error at the untrusted boundary", async () => {
+  class ForeignFailure extends Schema.TaggedError<ForeignFailure>()("Plugin.ForeignFailure", {
+    message: Schema.String,
+  }) {}
+  const lying: Info = {
+    name: "lying",
+    description: "Fails with a non-Tool.Error typed failure",
+    input: Schema.Struct({}),
+    execute: () => new ForeignFailure({ message: "transport died" }) as never,
+  }
+
+  const exit = await Effect.runPromiseExit(execute(lying, {}, context))
+  expect(exit._tag).toBe("Failure")
+  const error = exit._tag === "Failure" ? exit.cause.reasons.find((reason) => "error" in reason)?.error : undefined
+  expect(error).toBeInstanceOf(Tool.Error)
+  expect((error as Tool.Error).message).toBe("transport died")
+})
+
 test("execute supports callable namespace tools", async () => {
   const callable: Info = {
     name: "admin",

packages/core/test/tool-shell.test.ts

@@ -760,10 +760,52 @@ describe("ShellTool", () => {
             expect((yield* shell.list()).map((info) => info.id)).toContain(id)
             expect((yield* shell.wait(id)).status).toBe("timeout")
             expect((yield* Fiber.join(admitted)).valueOrUndefined?.data.item.payload).toMatchObject({
+              text: expect.stringContaining("Command timed out before completion."),
               description: idleCommand,
               metadata: {
                 source: "shell",
+                shellID,
                 state: "completed",
+                timeout: true,
+                truncated: false,
+              },
+            })
+          }),
+        )
+      },
+      (tmp) => Effect.promise(() => tmp[Symbol.asyncDispose]().then(() => undefined)),
+    ),
+  )
+
+  it.live("preserves a background command's non-zero exit", () =>
+    Effect.acquireUseRelease(
+      Effect.promise(() => tmpdir()),
+      (tmp) => {
+        reset()
+        return withSession(tmp.path, (registry) =>
+          Effect.gen(function* () {
+            const bus = yield* Bus.Service
+            const admitted = yield* bus.subscribe(SessionEvent.InboxEnqueued).pipe(
+              Stream.filter((event) => event.data.sessionID === sessionID && event.data.item.type === "synthetic"),
+              Stream.runHead,
+              Effect.forkScoped({ startImmediately: true }),
+            )
+            const settled = yield* executeTool(
+              registry,
+              call({ command: bodyExitCommand, background: true }, "call-background-nonzero"),
+            )
+            const shellID = settled.metadata?.shellID
+            expect(typeof shellID).toBe("string")
+            expect((yield* Fiber.join(admitted)).valueOrUndefined?.data.item.payload).toMatchObject({
+              text: expect.stringContaining("Command exited with code 7."),
+              description: bodyExitCommand,
+              metadata: {
+                source: "shell",
+                jobID: "call-background-nonzero",
+                shellID,
+                state: "completed",
+                exit: 7,
+                truncated: false,
               },
             })
           }),

packages/http-recorder/src/websocket/recorder.ts

@@ -31,8 +31,11 @@ interface PendingRecordings {
 }
 type Frame = string | Uint8Array
 
-const normalizeProtocols = (protocols?: string | Array<string>): Array<string> =>
-  protocols === undefined ? [] : typeof protocols === "string" ? [protocols] : [...protocols]
+const normalizeProtocols = (protocols: unknown): Array<string> => {
+  if (typeof protocols === "string") return [protocols]
+  if (Array.isArray(protocols)) return protocols.filter((protocol): protocol is string => typeof protocol === "string")
+  return []
+}
 const frameFromWebSocketData = async (data: unknown): Promise<Frame> => {
   if (typeof data === "string") return data
   if (data instanceof Blob) return new Uint8Array(await data.arrayBuffer())
@@ -371,7 +374,7 @@ const makeRecordingWebSocketConstructor = (
   return (url, protocols) => {
     const sequence = nextSequence++
     const requestedProtocols = normalizeProtocols(protocols)
-    const native = upstream(url, requestedProtocols)
+    const native = Reflect.apply(upstream, undefined, [url, protocols])
     const events: WebSocketEvent[] = []
     let opened = false
     let failed = false

packages/http-recorder/test/websocket.test.ts

@@ -80,6 +80,38 @@ describe("WebSocket", () => {
     ])
   })
 
+  test("constructor recording forwards handshake options", async () => {
+    using directory = tempDirectory("http-recorder-websocket-constructor-")
+    let received: unknown
+    const recorder = HttpRecorder.layerWebSocketConstructor("websocket/constructor-options", {
+      directory: directory.path,
+    }).pipe(
+      Layer.provide(
+        Layer.succeed(Socket.WebSocketConstructor, (url, options) => {
+          received = options
+          // oxlint-disable-next-line typescript-eslint/no-unsafe-type-assertion -- the fixture implements the WebSocket surface used by the recorder.
+          return new EchoWebSocket(url) as unknown as globalThis.WebSocket
+        }),
+      ),
+    )
+
+    await Effect.runPromise(
+      Effect.gen(function* () {
+        const constructor = yield* Socket.WebSocketConstructor
+        const options = { headers: { authorization: "Bearer fixture" } }
+        const socket = Reflect.apply(constructor, undefined, ["wss://echo.example.test/options", options])
+        yield* Effect.callback<void>((resume) => {
+          socket.addEventListener("open", () => {
+            socket.close()
+            resume(Effect.void)
+          })
+        })
+      }).pipe(Effect.scoped, Effect.provide(recorder)),
+    )
+
+    expect(received).toEqual({ headers: { authorization: "Bearer fixture" } })
+  })
+
   test("constructor replay validates dynamic URLs and protocols without opening a live socket", async () => {
     using directory = tempDirectory("http-recorder-websocket-constructor-")
     await seedCassetteDirectory(directory.path, "websocket/constructor", [

packages/server/src/process.ts

@@ -3,7 +3,7 @@ export * as ServerProcess from "./process"
 import { NodeHttpServer } from "@effect/platform-node"
 import { SessionRestart } from "@opencode-ai/core/session/execution/restart"
 import { hasPtyConnectTicketURL } from "@opencode-ai/protocol/groups/pty"
-import { Cause, Context, Deferred, Effect, Exit, Layer, Option, Ref, Scope } from "effect"
+import { Cause, Context, Effect, Exit, Latch, Layer, Option, Ref, Scope } from "effect"
 import { HttpMiddleware, HttpRouter, HttpServer, HttpServerRequest, HttpServerResponse } from "effect/unstable/http"
 import { createServer } from "node:http"
 import { ServerAuth } from "./auth"
@@ -47,7 +47,7 @@ export const start = Effect.fn("ServerProcess.start")(function* <E, R>(
   if (!password) return yield* Effect.fail(new Error("Missing server password"))
   const hostname = options.hostname ?? "127.0.0.1"
   const port = Option.fromNullishOr(options.port)
-  const shutdown = yield* Deferred.make<void>()
+  const shutdown = yield* Latch.make()
   const status = yield* Status.make()
   const bound = yield* listen({ hostname, port })
   const application = yield* Ref.make(Option.none<App>())
@@ -61,7 +61,7 @@ export const start = Effect.fn("ServerProcess.start")(function* <E, R>(
     )
     .pipe(withoutParentSpan)
   if (lifecycle)
-    yield* lifecycle.onListen(bound.http.address, Deferred.succeed(shutdown, undefined).pipe(Effect.asVoid)).pipe(
+    yield* lifecycle.onListen(bound.http.address, shutdown.open.pipe(Effect.asVoid)).pipe(
       Effect.flatMap((cleanup) =>
         Effect.addFinalizer(() => Scope.close(bound.scope, Exit.void).pipe(Effect.andThen(cleanup))),
       ),
@@ -101,7 +101,7 @@ export const start = Effect.fn("ServerProcess.start")(function* <E, R>(
     const app = Context.get(context, HttpRouter.HttpRouter).asHttpEffect()
     yield* Ref.set(application, Option.some(transform ? transform(app) : app))
     yield* status.ready
-    return { address: bound.http.address, shutdown: Deferred.await(shutdown) }
+    return { address: bound.http.address, shutdown: shutdown.await }
   }).pipe(
     Effect.catchCause((cause) => {
       if (!lifecycle || Cause.hasInterruptsOnly(cause)) return Effect.failCause(cause)
@@ -119,7 +119,7 @@ export const start = Effect.fn("ServerProcess.start")(function* <E, R>(
     }),
   )
   if (!lifecycle) return yield* boot
-  return yield* Effect.raceFirst(boot, Deferred.await(shutdown).pipe(Effect.andThen(Effect.interrupt)))
+  return yield* Effect.raceFirst(boot, shutdown.await.pipe(Effect.andThen(Effect.interrupt)))
 })
 
 function listen(options: { readonly hostname: string; readonly port: Option.Option<number> }) {

packages/tui/src/app.tsx

@@ -1,6 +1,6 @@
 import { render, useRenderer, useTerminalDimensions } from "@opentui/solid"
 import { registerOpencodeSpinner } from "./component/register-spinner"
-import { Deferred, Effect } from "effect"
+import { Effect, Latch } from "effect"
 import { Service, type Endpoint } from "@opencode-ai/client/effect/service"
 import { OpenCode, type SessionInfo } from "@opencode-ai/client"
 import { Global } from "@opencode-ai/util/global"
@@ -274,13 +274,13 @@ export const run = Effect.fn("Tui.run")(function* (input: TuiInput) {
             .forEach((result) => log("error", "Failed to dispose TUI resource", { error: result.reason }))
         }),
       )
-      const shutdown = yield* Deferred.make<unknown>()
+      const shutdown = yield* Latch.make()
       const onSighup = () => destroyRenderer(renderer)
       yield* Effect.acquireRelease(
         Effect.sync(() => process.on("SIGHUP", onSighup)),
         () => Effect.sync(() => process.off("SIGHUP", onSighup)),
       )
-      renderer.once("destroy", () => Deferred.doneUnsafe(shutdown, Effect.void))
+      renderer.once("destroy", () => shutdown.openUnsafe())
       yield* Effect.tryPromise(async () => {
         // Prewarm palette before ThemeProvider mounts so `system` theme avoids a first-paint fallback flash.
         void renderer.getPalette({ size: 16 }).catch(() => undefined)
@@ -443,7 +443,7 @@ export const run = Effect.fn("Tui.run")(function* (input: TuiInput) {
           renderer.requestRender()
         }
       })
-      yield* Deferred.await(shutdown)
+      yield* shutdown.await
       return { epilogue: exit.epilogue, reason: exit.reason }
     }),
   )

packages/tui/src/routes/session/index.tsx

@@ -121,7 +121,14 @@ const TRANSCRIPT_BACKFILL_CHUNK = 60
 type PendingAction = "steer" | "queue" | "cancel"
 
 const context = createContext<{
+  /** Content width: terminal width minus vertical tabs, sidebar, and padding. */
   width: number
+  /**
+   * Shared reactive terminal size. Transcript-row components must read this
+   * instead of calling useTerminalDimensions(), which registers one renderer
+   * resize listener per mounted component and grows with transcript length.
+   */
+  terminal: { width: number; height: number }
   sessionID: string
   thinkingMode: () => ThinkingMode
   showThinking: () => boolean
@@ -1124,12 +1131,25 @@ export function Session(props: { verticalTabsWidth: number }) {
     ),
   )
 
+  // Memoized per axis so width readers do not re-run on height-only resizes
+  // (dimensions() is one object signal with identity equality) and vice versa.
+  const terminalWidth = createMemo(() => dimensions().width)
+  const terminalHeight = createMemo(() => dimensions().height)
+
   return (
     <context.Provider
       value={{
         get width() {
           return contentWidth()
         },
+        terminal: {
+          get width() {
+            return terminalWidth()
+          },
+          get height() {
+            return terminalHeight()
+          },
+        },
         sessionID: route.sessionID,
         thinkingMode,
         showThinking,
@@ -1805,7 +1825,6 @@ function AssistantFooter(props: { message: SessionMessageAssistant }) {
   const ctx = use()
   const data = useData()
   const local = useLocal()
-  const dimensions = useTerminalDimensions()
   const theme = useTheme("elevated")
   const model = createMemo(
     () =>
@@ -1829,10 +1848,10 @@ function AssistantFooter(props: { message: SessionMessageAssistant }) {
           <span style={{ fg: props.message.error ? theme.text.subdued : local.agent.color(props.message.agent) }}>
             {Locale.titlecase(props.message.agent)}
           </span>
-          <Show when={dimensions().width >= 28}>
+          <Show when={ctx.terminal.width >= 28}>
             <span style={{ fg: theme.text.subdued }}> · {model()}</span>
           </Show>
-          <Show when={duration() && (dimensions().width < 28 || dimensions().width >= 36)}>
+          <Show when={duration() && (ctx.terminal.width < 28 || ctx.terminal.width >= 36)}>
             <span style={{ fg: theme.text.subdued }}> · {Locale.duration(duration())}</span>
           </Show>
           <Show when={interrupted()}>
@@ -2521,9 +2540,8 @@ function ToolImages(props: { parts: readonly SessionMessageAssistantTool[] }) {
 function SessionImages(props: { images: readonly { uri: string }[]; paddingLeft?: number }) {
   const ctx = use()
   const dialog = useDialog()
-  const dimensions = useTerminalDimensions()
   const images = createMemo(() => (ctx.config.session?.image_preview ? props.images : []))
-  const height = createMemo(() => Math.max(4, Math.min(8, Math.floor(dimensions().height / 4))))
+  const height = createMemo(() => Math.max(4, Math.min(8, Math.floor(ctx.terminal.height / 4))))
   const visible = createMemo(() => images().slice(0, 3))
 
   return (