fix(ai): batch Anthropic parallel tool results (#41948)

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

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

@@ -573,7 +573,10 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
         cache_control: cacheControl(breakpoints, part.cache),
       })
     }
-    messages.push({ role: "user", content })
+    const previous = messages.at(-1)
+    if (previous?.role === "user" && previous.content.every((block) => block.type === "tool_result"))
+      messages[messages.length - 1] = { role: "user", content: [...previous.content, ...content] }
+    else messages.push({ role: "user", content })
   }
 
   return messages

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

@@ -157,7 +157,6 @@ const OpenAIChatUsage = Schema.StructWithRest(
     prompt_tokens: optionalNull(Schema.Number),
     completion_tokens: optionalNull(Schema.Number),
     total_tokens: optionalNull(Schema.Number),
-    cost: optionalNull(Schema.Number),
     prompt_tokens_details: optionalNull(
       Schema.StructWithRest(
         Schema.Struct({
@@ -596,7 +595,6 @@ const mapUsage = (usage: OpenAIChatEvent["usage"]): Usage | undefined => {
     cacheWriteInputTokens: cacheWrite,
     reasoningTokens: reasoning,
     totalTokens: ProviderShared.totalTokens(input, output, usage.total_tokens ?? undefined),
-    cost: usage.cost ?? undefined,
     providerMetadata: { openai: usage },
   })
 }

packages/ai/src/route/executor.ts

@@ -1,4 +1,4 @@
-import { Cause, Context, Effect, Layer, Option, Schema } from "effect"
+import { Cause, Context, Effect, Layer, Option, Schema, Stream } from "effect"
 import {
   FetchHttpClient,
   Headers,
@@ -297,44 +297,86 @@ export const classifyHttpFailure = (input: {
   })
 }
 
-const toHttpError = (redactedNames: ReadonlyArray<string | RegExp>) => (error: unknown) => {
-  const transportError = (input: {
-    readonly message: string
-    readonly kind?: string | undefined
-    readonly request?: HttpClientRequest.HttpClientRequest | undefined
-  }) =>
+type HttpOperation = "request" | "read"
+
+const NativeTransportFailure = Schema.Struct({
+  message: Schema.String,
+  code: Schema.optionalKey(Schema.String),
+  cause: Schema.optionalKey(Schema.Unknown),
+})
+const decodeNativeTransportFailure = Schema.decodeUnknownOption(NativeTransportFailure)
+
+const nativeTransportFailure = (error: unknown) => {
+  const failure = Option.getOrUndefined(decodeNativeTransportFailure(error))
+  if (!failure) return undefined
+  if (failure.code !== undefined) return failure
+  const cause = Option.getOrUndefined(decodeNativeTransportFailure(failure.cause))
+  if (cause?.code !== undefined) return cause
+  return failure
+}
+
+const httpError = (input: {
+  readonly error: unknown
+  readonly request: HttpClientRequest.HttpClientRequest
+  readonly operation: HttpOperation
+  readonly redactedNames: ReadonlyArray<string | RegExp>
+}) => {
+  const request = HttpClientError.isHttpClientError(input.error) ? input.error.request : input.request
+  const transportError = (failure: { readonly message: string; readonly code?: string | undefined }) =>
     new AIError({
       module: "RequestExecutor",
-      method: "execute",
+      method: input.operation,
       reason: new TransportReason({
-        message: input.message,
-        kind: input.kind,
-        url: input.request ? redactUrl(input.request.url) : undefined,
-        http: input.request ? new HttpContext({ request: requestDetails(input.request, redactedNames) }) : undefined,
+        message: failure.message,
+        transport: "http",
+        operation: input.operation,
+        code: failure.code,
+        url: redactUrl(request.url),
+        http: new HttpContext({ request: requestDetails(request, input.redactedNames) }),
       }),
     })
 
-  if (Cause.isTimeoutError(error)) {
-    return transportError({ message: error.message, kind: "Timeout" })
-  }
-  if (!HttpClientError.isHttpClientError(error)) {
-    return transportError({ message: error instanceof Error ? error.message : "HTTP transport failed" })
-  }
-  const request = "request" in error ? error.request : undefined
-  if (error.reason._tag === "TransportError") {
+  const source =
+    HttpClientError.isHttpClientError(input.error) && "cause" in input.error.reason
+      ? input.error.reason.cause
+      : input.error
+  const native = nativeTransportFailure(source)
+  const code = native?.code
+  const raw = native?.message ?? (input.error instanceof Error ? input.error.message : undefined)
+  const detail = raw ? redactBody(raw, secretValues(request)) : undefined
+  const message = code && detail && !detail.includes(code) ? `${code}: ${detail}` : detail
+
+  if (Cause.isTimeoutError(input.error) || Cause.isTimeoutError(source))
+    return transportError({ message: message ?? "HTTP transport timed out", code: code ?? "Timeout" })
+  if (!HttpClientError.isHttpClientError(input.error))
+    return transportError({ message: message ?? "HTTP transport failed", code })
+  if (input.error.reason._tag === "TransportError") {
     return transportError({
-      message: error.reason.description ?? "HTTP transport failed",
-      kind: error.reason._tag,
-      request,
+      message: message ?? input.error.reason.description ?? "HTTP transport failed",
+      code: code ?? input.error.reason._tag,
     })
   }
   return transportError({
-    message: `HTTP transport failed: ${error.reason._tag}`,
-    kind: error.reason._tag,
-    request,
+    message: message ?? `HTTP transport failed: ${input.error.reason._tag}`,
+    code: code ?? input.error.reason._tag,
   })
 }
 
+export const stream = (
+  executor: Interface,
+  request: HttpClientRequest.HttpClientRequest,
+  middleware?: HttpMiddleware,
+): Stream.Stream<Uint8Array, AIError> =>
+  Stream.unwrap(
+    Effect.gen(function* () {
+      const redactedNames = yield* Headers.CurrentRedactedNames
+      const response = yield* executor.execute(request, middleware)
+      return response.stream.pipe(
+        Stream.mapError((error) => httpError({ error, request: response.request, operation: "read", redactedNames })),
+      )
+    }),
+  )
+
 export const layer: Layer.Layer<Service, never, HttpClient.HttpClient> = Layer.effect(
   Service,
   Effect.gen(function* () {
@@ -343,15 +385,16 @@ export const layer: Layer.Layer<Service, never, HttpClient.HttpClient> = Layer.e
       Effect.gen(function* () {
         const redactedNames = yield* Headers.CurrentRedactedNames
         if (!middleware)
-          return yield* http
-            .execute(request)
-            .pipe(Effect.mapError(toHttpError(redactedNames)), Effect.flatMap(statusError(request, redactedNames)))
+          return yield* http.execute(request).pipe(
+            Effect.mapError((error) => httpError({ error, request, operation: "request", redactedNames })),
+            Effect.flatMap(statusError(request, redactedNames)),
+          )
 
         const response = yield* middleware(request, (input) =>
           http
             .execute(input)
             .pipe(Effect.mapError((cause) => (cause instanceof Error ? cause : new Error(String(cause))))),
-        ).pipe(Effect.mapError(toHttpError(redactedNames)))
+        ).pipe(Effect.mapError((error) => httpError({ error, request, operation: "request", redactedNames })))
         return yield* statusError(response.request, redactedNames)(response)
       })
     return Service.of({

packages/ai/src/route/transport/http.ts

@@ -1,4 +1,4 @@
-import { Effect, Stream } from "effect"
+import { Effect } from "effect"
 import { Headers, HttpClientRequest } from "effect/unstable/http"
 import { Auth } from "../auth.js"
 import { render as renderEndpoint } from "../endpoint.js"
@@ -6,6 +6,7 @@ import { Framing } from "../framing.js"
 import type { HttpMiddleware, Transport, TransportPrepareInput } from "./index.js"
 import * as ProviderShared from "../../protocols/shared.js"
 import { mergeJsonRecords, type LLMRequest } from "../../schema/index.js"
+import { RequestExecutor } from "../executor.js"
 
 export type JsonRequestInput<Body> = TransportPrepareInput<Body>
 
@@ -86,26 +87,8 @@ export const httpJson = <Body, Frame>(input: HttpJsonInput<Body, Frame>): HttpJs
         middleware: prepareInput.middleware,
       }
     }),
-  frames: (prepared, request, runtime) =>
-    Stream.unwrap(
-      runtime.http
-        .execute(prepared.request, prepared.middleware)
-        .pipe(
-          Effect.map((response) =>
-            prepared.framing.frame(
-              response.stream.pipe(
-                Stream.mapError((error) =>
-                  ProviderShared.eventError(
-                    `${request.model.provider}/${request.model.route.id}`,
-                    `Failed to read ${request.model.provider}/${request.model.route.id} stream`,
-                    ProviderShared.errorText(error),
-                  ),
-                ),
-              ),
-            ),
-          ),
-        ),
-    ),
+  frames: (prepared, _request, runtime) =>
+    prepared.framing.frame(RequestExecutor.stream(runtime.http, prepared.request, prepared.middleware)),
 })
 
 export const sseJson = {

packages/ai/src/route/transport/websocket.ts

@@ -1,6 +1,6 @@
 import { Cause, Context, Effect, Layer, Queue, Stream } from "effect"
 import { Headers } from "effect/unstable/http"
-import { AIError, TransportReason } from "../../schema/index.js"
+import { AIError, TransportReason, type TransportOperation } from "../../schema/index.js"
 import * as HttpTransport from "./http.js"
 import type { Transport } from "./index.js"
 
@@ -29,12 +29,18 @@ export class Service extends Context.Service<Service, Interface>()("@opencode/AI
 const transportError = (
   method: string,
   message: string,
-  input: { readonly url?: string; readonly kind?: string } = {},
+  input: { readonly operation: TransportOperation; readonly url?: string; readonly code?: string },
 ) =>
   new AIError({
     module: "WebSocketExecutor",
     method,
-    reason: new TransportReason({ message, url: input.url, kind: input.kind }),
+    reason: new TransportReason({
+      message,
+      transport: "websocket",
+      operation: input.operation,
+      url: input.url,
+      code: input.code,
+    }),
   })
 
 const eventMessage = (event: Event) => {
@@ -55,7 +61,8 @@ const waitOpen = (ws: globalThis.WebSocket, input: WebSocketRequest) => {
     return Effect.fail(
       transportError("open", `WebSocket closed before opening (state ${ws.readyState})`, {
         url: input.url,
-        kind: "open",
+        operation: "request",
+        code: "closed",
       }),
     )
   }
@@ -79,7 +86,10 @@ const waitOpen = (ws: globalThis.WebSocket, input: WebSocketRequest) => {
       cleanup()
       resume(
         Effect.fail(
-          transportError("open", `Failed to open WebSocket: ${eventMessage(event)}`, { url: input.url, kind: "open" }),
+          transportError("open", `Failed to open WebSocket: ${eventMessage(event)}`, {
+            url: input.url,
+            operation: "request",
+          }),
         ),
       )
     }
@@ -89,7 +99,8 @@ const waitOpen = (ws: globalThis.WebSocket, input: WebSocketRequest) => {
         Effect.fail(
           transportError("open", `WebSocket closed before opening with code ${event.code}`, {
             url: input.url,
-            kind: "open",
+            operation: "request",
+            code: String(event.code),
           }),
         ),
       )
@@ -118,7 +129,8 @@ const webSocketUrl = (value: string) =>
     catch: (error) =>
       transportError("prepare", error instanceof Error ? error.message : "Invalid WebSocket URL", {
         url: value,
-        kind: "websocket",
+        operation: "request",
+        code: "invalid-url",
       }),
   })
 
@@ -129,7 +141,7 @@ export const open = (input: WebSocketRequest) =>
     catch: (error) =>
       transportError("open", error instanceof Error ? error.message : "Failed to construct WebSocket", {
         url: input.url,
-        kind: "open",
+        operation: "request",
       }),
   }).pipe(Effect.flatMap((ws) => fromWebSocket(ws, input)))
 
@@ -150,7 +162,10 @@ export const fromWebSocket = (
       Queue.failCauseUnsafe(
         messages,
         Cause.fail(
-          transportError("message", "Unsupported WebSocket message payload", { url: input.url, kind: "message" }),
+          transportError("message", "Unsupported WebSocket message payload", {
+            url: input.url,
+            operation: "read",
+          }),
         ),
       )
     }
@@ -158,7 +173,10 @@ export const fromWebSocket = (
       Queue.failCauseUnsafe(
         messages,
         Cause.fail(
-          transportError("message", `WebSocket error: ${eventMessage(event)}`, { url: input.url, kind: "message" }),
+          transportError("message", `WebSocket error: ${eventMessage(event)}`, {
+            url: input.url,
+            operation: "read",
+          }),
         ),
       )
     }
@@ -167,7 +185,11 @@ export const fromWebSocket = (
       Queue.failCauseUnsafe(
         messages,
         Cause.fail(
-          transportError("message", `WebSocket closed with code ${event.code}`, { url: input.url, kind: "close" }),
+          transportError("message", `WebSocket closed with code ${event.code}`, {
+            url: input.url,
+            operation: "read",
+            code: String(event.code),
+          }),
         ),
       )
     }
@@ -188,7 +210,7 @@ export const fromWebSocket = (
           catch: (error) =>
             transportError("sendText", error instanceof Error ? error.message : "Failed to send WebSocket message", {
               url: input.url,
-              kind: "write",
+              operation: "write",
             }),
         }),
       messages: Stream.fromQueue(messages),
@@ -243,7 +265,8 @@ export const json = <Body, Message>(input: JsonInput<Body, Message>): JsonTransp
       return Stream.fail(
         transportError("json", "WebSocket JSON transport requires WebSocketExecutor.Service", {
           url: prepared.url,
-          kind: "websocket",
+          operation: "request",
+          code: "unavailable",
         }),
       )
     }

packages/ai/src/schema/errors.ts

@@ -92,10 +92,18 @@ export class ProviderInternalReason extends Schema.Class<ProviderInternalReason>
   http: Schema.optional(HttpContext),
 }) {}
 
+export const TransportType = Schema.Literals(["http", "websocket"])
+export type TransportType = typeof TransportType.Type
+
+export const TransportOperation = Schema.Literals(["request", "read", "write"])
+export type TransportOperation = typeof TransportOperation.Type
+
 export class TransportReason extends Schema.Class<TransportReason>("AI.Error.Transport")({
   _tag: Schema.tag("Transport"),
   message: Schema.String,
-  kind: Schema.optional(Schema.String),
+  transport: TransportType,
+  operation: TransportOperation,
+  code: Schema.optional(Schema.String),
   url: Schema.optional(Schema.String),
   http: Schema.optional(HttpContext),
 }) {}

packages/ai/src/schema/events.ts

@@ -56,8 +56,6 @@ export class Usage extends Schema.Class<Usage>("AI.Usage")({
   cacheWriteInputTokens: Schema.optional(Schema.Number),
   reasoningTokens: Schema.optional(Schema.Number),
   totalTokens: Schema.optional(Schema.Number),
-  /** Provider-reported cost for this physical request, normalized to USD. */
-  cost: Schema.optional(Schema.Number),
   providerMetadata: Schema.optional(ProviderMetadata),
 }) {
   /**

packages/ai/test/executor.test.ts

@@ -1,10 +1,10 @@
 import { describe, expect } from "bun:test"
-import { Effect, Layer, Ref } from "effect"
-import { Headers, HttpClient, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
+import { Effect, Layer, Ref, Stream } from "effect"
+import { Headers, HttpClient, HttpClientError, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
 import { LLM, AIError } from "../src/index.js"
 import { LLMClient, RequestExecutor } from "../src/route.js"
 import * as OpenAIChat from "../src/protocols/openai-chat.js"
-import { dynamicResponse } from "./lib/http.js"
+import { dynamicResponse, systemError } from "./lib/http.js"
 import { deltaChunk } from "./lib/openai-chunks.js"
 import { sseRaw } from "./lib/sse.js"
 import { it } from "./lib/effect.js"
@@ -67,6 +67,62 @@ const expectAIError = (error: unknown) => {
 const errorHttp = (error: AIError) => ("http" in error.reason ? error.reason.http : undefined)
 
 describe("RequestExecutor", () => {
+  it.effect("parses response body failures at the executor seam", () =>
+    Effect.gen(function* () {
+      const executor = yield* RequestExecutor.Service
+      const error = yield* RequestExecutor.stream(executor, secretRequest).pipe(Stream.runDrain, Effect.flip)
+
+      expectAIError(error)
+      expect(error.reason).toMatchObject({
+        _tag: "Transport",
+        message: "ECONNRESET: disconnected <redacted> <redacted>",
+        transport: "http",
+        operation: "read",
+        code: "ECONNRESET",
+        url: "https://provider.test/v1/chat?api_key=%3Credacted%3E&debug=1",
+      })
+    }).pipe(
+      Effect.provide(
+        responsesLayer([
+          new Response(
+            new ReadableStream({
+              start(controller) {
+                controller.error(systemError("ECONNRESET", "disconnected query-secret-123 header-secret-456"))
+              },
+            }),
+          ),
+        ]),
+      ),
+    ),
+  )
+
+  it.effect("unwraps native transport failure causes", () =>
+    Effect.gen(function* () {
+      const executor = yield* RequestExecutor.Service
+      const error = yield* RequestExecutor.stream(executor, secretRequest).pipe(Stream.runDrain, Effect.flip)
+
+      expectAIError(error)
+      expect(error.reason).toMatchObject({
+        _tag: "Transport",
+        message: "ECONNRESET: socket closed",
+        operation: "read",
+        code: "ECONNRESET",
+      })
+    }).pipe(
+      Effect.provide(
+        responsesLayer([
+          new Response(
+            new ReadableStream({
+              pull(controller) {
+                controller.error(new TypeError("fetch failed", { cause: systemError("ECONNRESET", "socket closed") }))
+              },
+            }),
+          ),
+        ]),
+      ),
+    ),
+  )
+
   it.effect("preserves middleware error messages", () =>
     Effect.gen(function* () {
       const executor = yield* RequestExecutor.Service
@@ -79,6 +135,48 @@ describe("RequestExecutor", () => {
     }).pipe(Effect.provide(responsesLayer([]))),
   )
 
+  it.effect("reports the request sent by middleware", () =>
+    Effect.gen(function* () {
+      const executor = yield* RequestExecutor.Service
+      const error = yield* executor
+        .execute(request, (original, handler) =>
+          handler(
+            original.pipe(
+              HttpClientRequest.setUrl("https://proxy.test/v1/chat?api_key=proxy-secret"),
+              HttpClientRequest.setHeader("authorization", "Bearer proxy-secret"),
+            ),
+          ),
+        )
+        .pipe(Effect.flip)
+
+      expectAIError(error)
+      expect(error.reason).toMatchObject({
+        _tag: "Transport",
+        message: "ECONNRESET: proxy disconnected <redacted>",
+        url: "https://proxy.test/v1/chat?api_key=%3Credacted%3E",
+        http: {
+          request: {
+            url: "https://proxy.test/v1/chat?api_key=%3Credacted%3E",
+            headers: { authorization: "<redacted>" },
+          },
+        },
+      })
+    }).pipe(
+      Effect.provide(
+        dynamicResponse((input) =>
+          Effect.fail(
+            new HttpClientError.HttpClientError({
+              reason: new HttpClientError.TransportError({
+                request: input.request,
+                cause: systemError("ECONNRESET", "proxy disconnected proxy-secret"),
+              }),
+            }),
+          ),
+        ),
+      ),
+    ),
+  )
+
   it.effect("classifies context overflow responses", () =>
     Effect.gen(function* () {
       const executor = yield* RequestExecutor.Service

packages/ai/test/lib/http.ts

@@ -1,5 +1,5 @@
 import { Effect, Layer, Ref } from "effect"
-import { HttpClient, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
+import { HttpClient, HttpClientError, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
 import { LLMClient, RequestExecutor, WebSocketExecutor } from "../../src/route.js"
 import type { Service as LLMClientService } from "../../src/route/client.js"
 import type { Service as RequestExecutorService } from "../../src/route/executor.js"
@@ -14,7 +14,9 @@ export type HandlerInput = {
   ) => HttpClientResponse.HttpClientResponse
 }
 
-export type Handler = (input: HandlerInput) => Effect.Effect<HttpClientResponse.HttpClientResponse>
+export type Handler = (
+  input: HandlerInput,
+) => Effect.Effect<HttpClientResponse.HttpClientResponse, HttpClientError.HttpClientError>
 
 const handlerLayer = (handler: Handler): Layer.Layer<HttpClient.HttpClient> =>
   Layer.succeed(
@@ -34,6 +36,12 @@ const handlerLayer = (handler: Handler): Layer.Layer<HttpClient.HttpClient> =>
 
 export type RuntimeEnv = RequestExecutorService | WebSocketExecutorService | LLMClientService
 
+export interface SystemError extends Error {
+  readonly code: string
+}
+
+export const systemError = (code: string, message: string): SystemError => Object.assign(new Error(message), { code })
+
 export const runtimeLayer = (layer: Layer.Layer<HttpClient.HttpClient>): Layer.Layer<RuntimeEnv> => {
   const requestExecutorLayer = RequestExecutor.layer.pipe(Layer.provide(layer))
   const deps = Layer.mergeAll(requestExecutorLayer, WebSocketExecutor.layer)
@@ -63,14 +71,20 @@ export const dynamicResponse = (handler: Handler) => runtimeLayer(handlerLayer(h
  * Layer that emits the supplied SSE chunks and then aborts mid-stream. Used to
  * exercise transport errors that surface during parsing.
  */
-export const truncatedStream = (chunks: ReadonlyArray<string>) =>
+export const truncatedStream = (chunks: ReadonlyArray<string>, error: Error = new Error("connection reset")) =>
   dynamicResponse((input) =>
     Effect.sync(() => {
       const encoder = new TextEncoder()
+      let index = 0
       const stream = new ReadableStream({
-        start(controller) {
-          for (const chunk of chunks) controller.enqueue(encoder.encode(chunk))
-          controller.error(new Error("connection reset"))
+        pull(controller) {
+          const chunk = chunks[index]
+          if (chunk !== undefined) {
+            index++
+            controller.enqueue(encoder.encode(chunk))
+            return
+          }
+          controller.error(error)
         },
       })
       return input.respond(stream, { headers: SSE_HEADERS })

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

@@ -271,6 +271,47 @@ describe("Anthropic Messages route", () => {
     }),
   )
 
+  it.effect("batches parallel tool results into one Anthropic user message", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user("Check both cities."),
+            Message.assistant([
+              { type: "text", text: "I'll check both." },
+              ToolCallPart.make({ id: "call_paris", name: "weather", input: { city: "Paris" } }),
+              ToolCallPart.make({ id: "call_london", name: "weather", input: { city: "London" } }),
+            ]),
+            Message.tool({ id: "call_paris", name: "weather", result: { temperature: 22 } }),
+            Message.tool({ id: "call_london", name: "weather", result: { temperature: 18 } }),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toMatchObject([
+        { role: "user", content: [{ type: "text", text: "Check both cities." }] },
+        {
+          role: "assistant",
+          content: [
+            { type: "text", text: "I'll check both." },
+            { type: "tool_use", id: "call_paris", name: "weather", input: { city: "Paris" } },
+            { type: "tool_use", id: "call_london", name: "weather", input: { city: "London" } },
+          ],
+        },
+        {
+          role: "user",
+          content: [
+            { type: "tool_result", tool_use_id: "call_paris", content: '{"temperature":22}' },
+            { type: "tool_result", tool_use_id: "call_london", content: '{"temperature":18}' },
+          ],
+        },
+      ])
+      expect(prepared.body.messages).toHaveLength(3)
+    }),
+  )
+
   it.effect("keeps tools and sends tool_choice none", () =>
     Effect.gen(function* () {
       const prepared = yield* compileRequest(
@@ -915,6 +956,54 @@ describe("Anthropic Messages route", () => {
     }),
   )
 
+  it.effect("assembles and persists multiple tool calls from one Anthropic response", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              {
+                type: "content_block_start",
+                index: 0,
+                content_block: { type: "tool_use", id: "call_paris", name: "weather", input: {} },
+              },
+              {
+                type: "content_block_delta",
+                index: 0,
+                delta: { type: "input_json_delta", partial_json: '{"city":"Paris"}' },
+              },
+              { type: "content_block_stop", index: 0 },
+              {
+                type: "content_block_start",
+                index: 1,
+                content_block: { type: "tool_use", id: "call_london", name: "weather", input: {} },
+              },
+              {
+                type: "content_block_delta",
+                index: 1,
+                delta: { type: "input_json_delta", partial_json: '{"city":"London"}' },
+              },
+              { type: "content_block_stop", index: 1 },
+              { type: "message_delta", delta: { stop_reason: "tool_use" }, usage: { output_tokens: 2 } },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.toolCalls).toMatchObject([
+        { id: "call_paris", name: "weather", input: { city: "Paris" } },
+        { id: "call_london", name: "weather", input: { city: "London" } },
+      ])
+      expect(response.message.content).toMatchObject([
+        { type: "tool-call", id: "call_paris", name: "weather", input: { city: "Paris" } },
+        { type: "tool-call", id: "call_london", name: "weather", input: { city: "London" } },
+      ])
+      expect(response.finishReason).toEqual({ normalized: "tool-calls", raw: "tool_use" })
+    }),
+  )
+
   it.effect("keeps malformed server tool input terminal", () =>
     Effect.gen(function* () {
       const body = sseEvents(

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

@@ -1,5 +1,5 @@
 import { describe, expect } from "bun:test"
-import { Effect, Schema, Stream } from "effect"
+import { Effect, Ref, Schema, Stream } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import {
   HttpOptions,
@@ -22,7 +22,7 @@ import { ProviderShared } from "../../src/protocols/shared.js"
 import { Auth, LLMClient } from "../../src/route.js"
 import { compileRequest } from "../../src/route/client.js"
 import { it } from "../lib/effect.js"
-import { dynamicResponse, fixedResponse, truncatedStream } from "../lib/http.js"
+import { dynamicResponse, fixedResponse, systemError, truncatedStream } from "../lib/http.js"
 import { deltaChunk, usageChunk } from "../lib/openai-chunks.js"
 import { sseEvents } from "../lib/sse.js"
 
@@ -1221,12 +1221,44 @@ describe("OpenAI Chat route", () => {
 
   it.effect("surfaces transport errors that occur mid-stream", () =>
     Effect.gen(function* () {
-      const layer = truncatedStream([
-        `data: ${JSON.stringify(deltaChunk({ role: "assistant", content: "Hello" }))}\n\n`,
-      ])
-      const error = yield* LLMClient.generate(request).pipe(Effect.provide(layer), Effect.flip)
+      const layer = truncatedStream(
+        [`data: ${JSON.stringify(deltaChunk({ role: "assistant", content: "Hello" }))}\n\n`],
+        systemError("ECONNRESET", "socket closed unexpectedly"),
+      )
+      const events = yield* Ref.make<ReadonlyArray<LLMEvent>>([])
+      const error = yield* LLMClient.stream(request).pipe(
+        Stream.tap((event) => Ref.update(events, (current) => [...current, event])),
+        Stream.runDrain,
+        Effect.provide(layer),
+        Effect.flip,
+      )
 
-      expect(error.message).toContain("Failed to read openai/openai-chat stream")
+      expect((yield* Ref.get(events)).some((event) => event.type === "text-delta")).toBeTrue()
+      expect(error.reason).toMatchObject({
+        _tag: "Transport",
+        message: "ECONNRESET: socket closed unexpectedly",
+        transport: "http",
+        operation: "read",
+        code: "ECONNRESET",
+        url: "https://api.openai.test/v1/chat/completions",
+      })
+    }),
+  )
+
+  it.effect("surfaces transport errors before the first stream frame", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(truncatedStream([], systemError("ECONNRESET", "socket closed before output"))),
+        Effect.flip,
+      )
+
+      expect(error.reason).toMatchObject({
+        _tag: "Transport",
+        message: "ECONNRESET: socket closed before output",
+        transport: "http",
+        operation: "read",
+        code: "ECONNRESET",
+      })
     }),
   )
 

packages/ai/test/provider/openrouter.test.ts

@@ -246,24 +246,6 @@ describe("OpenRouter", () => {
     }),
   )
 
-  it.effect("reports OpenRouter's streamed USD cost", () =>
-    Effect.gen(function* () {
-      const model = OpenRouter.configure({ apiKey: "test-key" }).model("openai/gpt-4o-mini")
-      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Say hello." })).pipe(
-        Effect.provide(
-          fixedResponse(
-            sseEvents({
-              choices: [{ delta: { content: "Hello" }, finish_reason: "stop" }],
-              usage: { prompt_tokens: 10, completion_tokens: 2, total_tokens: 12, cost: 0.00123 },
-            }),
-          ),
-        ),
-      )
-
-      expect(response.usage?.cost).toBe(0.00123)
-    }),
-  )
-
   it.effect("fails on a mid-stream provider error", () =>
     Effect.gen(function* () {
       const model = OpenRouter.configure({ apiKey: "test-key" }).model("openai/gpt-4o-mini")

packages/core/src/aisdk.ts

@@ -343,8 +343,7 @@ function modelFromLanguage(info: Info, language: LanguageModelV3) {
     model: (input) =>
       LanguageModel.make({ ...input, provider: "provider" in input ? input.provider : info.providerID, route }),
     prepareTransport: (body) => Effect.succeed(body),
-    streamPrepared: (prepared) =>
-      streamLanguage(language, prepared as LanguageModelV3CallOptions, info.providerID === Provider.ID.githubCopilot),
+    streamPrepared: (prepared) => streamLanguage(language, prepared as LanguageModelV3CallOptions),
   }
   return LanguageModel.make({
     id: info.modelID ?? info.id,
@@ -428,7 +427,6 @@ function callOptions(request: LLMRequest): LanguageModelV3CallOptions {
     toolChoice: toolChoice(request.toolChoice),
     headers: request.http?.headers,
     providerOptions: providerOptions(request.providerOptions),
-    includeRawChunks: request.model.provider === ProviderID.make(Provider.ID.githubCopilot),
   }
 }
 
@@ -549,15 +547,8 @@ function providerOptions(input: LLMRequest["providerOptions"]): SharedV3Provider
   return Object.fromEntries(Object.entries(input).map(([key, value]) => [key, jsonObject(value)]))
 }
 
-interface StreamState {
-  step: number
-  toolNames: Record<string, string>
-  copilot: boolean
-  cost?: number
-}
-
-function streamLanguage(language: LanguageModelV3, options: LanguageModelV3CallOptions, copilot: boolean) {
-  const state: StreamState = { step: 0, toolNames: {}, copilot }
+function streamLanguage(language: LanguageModelV3, options: LanguageModelV3CallOptions) {
+  const state = { step: 0, toolNames: {} as Record<string, string> }
   return Stream.concat(
     Stream.make(LLMEvent.stepStart({ index: state.step })),
     Stream.unwrap(
@@ -580,19 +571,17 @@ function streamLanguage(language: LanguageModelV3, options: LanguageModelV3CallO
 }
 
 function streamPartEvents(
-  state: StreamState,
+  state: { step: number; toolNames: Record<string, string> },
   event: LanguageModelV3StreamPart,
 ): Effect.Effect<ReadonlyArray<LLMEvent>, AIError> {
   switch (event.type) {
     case "stream-start":
     case "response-metadata":
+    case "raw":
     case "file":
     case "source":
     case "tool-approval-request":
       return Effect.succeed([])
-    case "raw":
-      if (state.copilot) state.cost = copilotCost(event.rawValue) ?? state.cost
-      return Effect.succeed([])
     case "text-start":
       return Effect.succeed([
         LLMEvent.textStart({ id: event.id, providerMetadata: providerMetadata(event.providerMetadata) }),
@@ -683,18 +672,16 @@ function streamPartEvents(
         }),
       ])
     case "finish":
-      const normalized = usage(event.usage, state.cost)
-      state.cost = undefined
       return Effect.succeed([
         LLMEvent.stepFinish({
           index: state.step++,
           reason: { normalized: finishReason(event.finishReason), raw: event.finishReason.raw },
-          usage: normalized,
+          usage: usage(event.usage),
           providerMetadata: providerMetadata(event.providerMetadata),
         }),
         LLMEvent.finish({
           reason: { normalized: finishReason(event.finishReason), raw: event.finishReason.raw },
-          usage: normalized,
+          usage: usage(event.usage),
           providerMetadata: providerMetadata(event.providerMetadata),
         }),
       ])
@@ -703,10 +690,7 @@ function streamPartEvents(
   }
 }
 
-function usage(
-  input: Extract<LanguageModelV3StreamPart, { type: "finish" }>["usage"],
-  cost?: number,
-): UsageInput | undefined {
+function usage(input: Extract<LanguageModelV3StreamPart, { type: "finish" }>["usage"]): UsageInput | undefined {
   const output = {
     inputTokens: input.inputTokens.total,
     nonCachedInputTokens: input.inputTokens.noCache,
@@ -718,22 +702,10 @@ function usage(
       input.inputTokens.total === undefined || input.outputTokens.total === undefined
         ? undefined
         : input.inputTokens.total + input.outputTokens.total,
-    cost,
   }
   return Object.values(output).some((value) => value !== undefined) ? output : undefined
 }
 
-function copilotCost(input: unknown): number | undefined {
-  if (!ProviderShared.isRecord(input)) return undefined
-  const raw = input
-  const response = ProviderShared.isRecord(raw.response) ? raw.response : undefined
-  const usage = raw.copilot_usage ?? response?.copilot_usage
-  if (!ProviderShared.isRecord(usage)) return undefined
-  const total = usage.total_nano_aiu
-  if (typeof total !== "number" || !Number.isFinite(total) || total < 0) return undefined
-  return total / 100_000_000_000
-}
-
 function finishReason(value: LanguageModelV3FinishReason): FinishReason {
   return value.unified === "other" ? "unknown" : value.unified
 }
@@ -791,7 +763,9 @@ function apiCallErrorReason(error: APICallError) {
   if (error.statusCode !== undefined || !error.isRetryable) return reason
   return new TransportReason({
     message: reason.message,
-    kind: error.name,
+    transport: "http",
+    operation: "request",
+    code: error.name,
     url: error.url,
     http: "http" in reason ? reason.http : undefined,
   })

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

@@ -272,8 +272,10 @@ const layer = Layer.effect(
         snapshot: startSnapshot,
         assistantMessageID,
       })
-      const stepUsage = (finish: NonNullable<StepRecord["finish"]>) =>
-        SessionUsage.record(finish.usage, resolved.cost)
+      const stepUsage = (finish: NonNullable<StepRecord["finish"]>) => ({
+        cost: SessionUsage.calculateCost(resolved.cost, finish.tokens),
+        tokens: finish.tokens,
+      })
 
       const captureStepEnd = Effect.fnUntraced(function* () {
         const snapshot = yield* snapshots.capture()

packages/core/src/session/runner/publish-llm-event.ts

@@ -35,7 +35,7 @@ export interface StepRecord {
   /** Present once the provider finished the step normally. */
   readonly finish?: {
     readonly finish: Extract<LLMEvent, { type: "step-finish" }>["reason"]["normalized"]
-    readonly usage: Extract<LLMEvent, { type: "step-finish" }>["usage"]
+    readonly tokens: ReturnType<typeof SessionUsage.tokens>
   }
   readonly calls: ReadonlyArray<{
     readonly id: string
@@ -495,7 +495,7 @@ export const createLLMEventPublisher = (bus: Pick<Bus.Interface, "publish">, inp
       case "step-finish":
         yield* flush()
         if (stepSettlement) return yield* Effect.die(new Error("Duplicate step finish"))
-        stepSettlement = { finish: event.reason.normalized, usage: event.usage }
+        stepSettlement = { finish: event.reason.normalized, tokens: SessionUsage.tokens(event.usage) }
         if (event.reason.normalized === "content-filter") {
           providerFailed = true
           yield* failAssistant({ type: "provider.content-filter", message: "Provider blocked the response" })

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

@@ -48,10 +48,12 @@ export const schedule = (
   assistantMessageID: () => SessionMessage.ID,
 ) =>
   Schedule.max([Schedule.exponential("2 seconds"), Schedule.recurs(4)]).pipe(
+    Schedule.jittered,
     Schedule.setInputType<RetryableFailure>(),
     Schedule.modifyDelay(({ input: failure, duration: delay }) => {
       const minimum = retryAfter(failure)
-      return Effect.succeed(minimum === undefined ? delay : Duration.max(delay, Duration.millis(minimum)))
+      const duration = minimum === undefined ? delay : Duration.max(delay, Duration.millis(minimum))
+      return Effect.succeed(Duration.millis(Math.ceil(Duration.toMillis(duration))))
     }),
     Schedule.tap((metadata) =>
       bus.publish(SessionEvent.RetryScheduled, {

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

@@ -1,5 +1,6 @@
 import { Message, ToolCallPart, ToolResultPart, type ContentPart, type ProviderMetadata } from "@opencode-ai/ai"
 import { Option, Schema } from "effect"
+import { fileURLToPath } from "url"
 import type { Model } from "../../model.js"
 import { SessionMessage } from "../message.js"
 import type { FileAttachment } from "@opencode-ai/schema/prompt"
@@ -14,6 +15,17 @@ const media = (file: FileAttachment): ContentPart => ({
   metadata: file.description === undefined ? undefined : { description: file.description },
 })
 
+const attachmentLocation = (file: FileAttachment) => {
+  if (file.source.type !== "uri") return undefined
+  const url = URL.parse(file.source.uri)
+  if (url?.protocol !== "file:") return undefined
+  try {
+    return fileURLToPath(url)
+  } catch {
+    return undefined
+  }
+}
+
 const textAttachment = (file: FileAttachment): ContentPart => ({
   type: "text",
   text: `\n\n${[
@@ -36,7 +48,7 @@ const textAttachment = (file: FileAttachment): ContentPart => ({
 const directoryAttachment = (file: FileAttachment): ContentPart => ({
   type: "text",
   text: `\n\n${[
-    `Attached directory: ${file.name ?? (file.source.type === "uri" ? file.source.uri : "directory")}`,
+    `Attached directory: ${attachmentLocation(file) ?? file.name ?? (file.source.type === "uri" ? file.source.uri : "directory")}`,
     file.description === undefined ? undefined : `Description: ${file.description}`,
     file.data.length === 0 ? undefined : "",
     file.data.length === 0 ? undefined : Buffer.from(file.data, "base64").toString("utf8"),
@@ -55,7 +67,10 @@ const directoryAttachment = (file: FileAttachment): ContentPart => ({
 const attachmentContent = (file: FileAttachment): ContentPart[] => {
   if (file.mime === "text/plain") return [textAttachment(file)]
   if (file.mime === "application/x-directory") return [directoryAttachment(file)]
-  if (imageMimes.has(file.mime)) return [media(file)]
+  if (imageMimes.has(file.mime)) {
+    const location = attachmentLocation(file)
+    return [...(location === undefined ? [] : [Message.text(`Attached file: ${location}`)]), media(file)]
+  }
   return []
 }
 

packages/core/src/session/usage.ts

@@ -17,6 +17,7 @@ export const tokens = (usage: Usage | undefined): TokenUsage.Info => ({
   },
 })
 
+// TODO(#35765): Use Copilot's reported billed amount once billing has a dedicated typed runtime contract.
 export function calculateCost(costs: Model.Info["cost"], usage: TokenUsage.Info) {
   const context = usage.input + usage.cache.read + usage.cache.write
   const tier = costs
@@ -37,14 +38,7 @@ export type Recorded = { readonly tokens: TokenUsage.Info; readonly cost: Money.
 
 export const record = (usage: Usage | undefined, costs: Model.Info["cost"]): Recorded => {
   const normalized = tokens(usage)
-  const reported = usage?.cost
-  return {
-    tokens: normalized,
-    cost:
-      reported !== undefined && Number.isFinite(reported) && reported >= 0
-        ? Money.USD.make(reported)
-        : calculateCost(costs, normalized),
-  }
+  return { tokens: normalized, cost: calculateCost(costs, normalized) }
 }
 
 export const add = (a: Recorded, b: Recorded): Recorded => ({

packages/core/test/aisdk.test.ts

@@ -1,5 +1,5 @@
 import { APICallError } from "@ai-sdk/provider"
-import type { LanguageModelV3, LanguageModelV3CallOptions, LanguageModelV3StreamPart } from "@ai-sdk/provider"
+import type { LanguageModelV3, LanguageModelV3StreamPart } from "@ai-sdk/provider"
 import { AISDK } from "@opencode-ai/core/aisdk"
 import { SessionRunnerRetry } from "@opencode-ai/core/session/runner/retry"
 import { toSessionError } from "@opencode-ai/core/session/to-session-error"
@@ -23,26 +23,21 @@ const model = (packageName: string, settings: Record<string, unknown> = {}) =>
     limit: { context: 100, output: 20 },
   })
 
-const streamModel = (
-  events: ReadonlyArray<LanguageModelV3StreamPart>,
-  inspect?: (options: LanguageModelV3CallOptions) => void,
-): LanguageModelV3 => ({
+const streamModel = (events: ReadonlyArray<LanguageModelV3StreamPart>): LanguageModelV3 => ({
   specificationVersion: "v3",
   provider: "test",
   modelId: "test",
   supportedUrls: {},
   doGenerate: () => Promise.reject(new Error("Unexpected non-streaming request")),
-  doStream: (options) => {
-    inspect?.(options)
-    return Promise.resolve({
+  doStream: () =>
+    Promise.resolve({
       stream: new ReadableStream({
         start(controller) {
           events.forEach((event) => controller.enqueue(event))
           controller.close()
         },
       }),
-    })
-  },
+    }),
 })
 
 const usage = {
@@ -54,7 +49,9 @@ const client = LLMClient.layer.pipe(
   Layer.provide(
     Layer.succeed(
       RequestExecutor.Service,
-      RequestExecutor.Service.of({ execute: () => Effect.die("Unexpected HTTP request") }),
+      RequestExecutor.Service.of({
+        execute: () => Effect.die("Unexpected HTTP request"),
+      }),
     ),
   ),
 )
@@ -380,41 +377,6 @@ it.effect("emits malformed AI SDK tool input without executing it", () =>
   }),
 )
 
-it.effect("normalizes Copilot billed usage to USD", () =>
-  Effect.gen(function* () {
-    const aisdk = yield* AISDK.Service
-    let options: LanguageModelV3CallOptions | undefined
-    yield* aisdk.hook.sdk((event) => {
-      event.sdk = {
-        languageModel: () =>
-          streamModel(
-            [
-              {
-                type: "raw",
-                rawValue: { type: "message_delta", copilot_usage: { total_nano_aiu: 4_473_525_000 } },
-              },
-              { type: "finish", finishReason: { unified: "stop", raw: "end_turn" }, usage },
-            ],
-            (input) => {
-              options = input
-            },
-          ),
-      }
-    })
-
-    const resolved = yield* aisdk.model({
-      ...model("@ai-sdk/github-copilot"),
-      providerID: Provider.ID.githubCopilot,
-    })
-    const response = yield* LLMClient.generate(LLM.request({ model: resolved, prompt: "Hello" })).pipe(
-      Effect.provide(client),
-    )
-
-    expect(options?.includeRawChunks).toBeTrue()
-    expect(response.usage?.cost).toBeCloseTo(0.04473525)
-  }),
-)
-
 it.effect("keeps malformed provider-executed AI SDK input terminal", () =>
   Effect.gen(function* () {
     const aisdk = yield* AISDK.Service
@@ -582,7 +544,12 @@ it.effect("retries status-less AI SDK transport failures", () =>
         isRetryable: true,
       }),
     )
-    expect(error.reason).toMatchObject({ _tag: "Transport", kind: "AI_APICallError" })
+    expect(error.reason).toMatchObject({
+      _tag: "Transport",
+      transport: "http",
+      operation: "request",
+      code: "AI_APICallError",
+    })
     expect(SessionRunnerRetry.isRetryable(error)).toBeTrue()
     expect("http" in error.reason ? error.reason.http?.request.url : undefined).toBe("https://api.example.com/chat")
   }),

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

@@ -39,7 +39,9 @@ describe("toSessionError", () => {
     )
     expect(toSessionError(llm(new QuotaExceededReason({ message: "quota" }))).type).toBe("provider.quota")
     expect(toSessionError(llm(new ContentPolicyReason({ message: "blocked" }))).type).toBe("provider.content-filter")
-    expect(toSessionError(llm(new TransportReason({ message: "transport" }))).type).toBe("provider.transport")
+    expect(
+      toSessionError(llm(new TransportReason({ message: "transport", transport: "http", operation: "request" }))).type,
+    ).toBe("provider.transport")
     expect(toSessionError(llm(new ProviderInternalReason({ message: "internal", status: 500 }))).type).toBe(
       "provider.internal",
     )
@@ -111,7 +113,7 @@ describe("toSessionError", () => {
     const eligible = [
       llm(new RateLimitReason({ message: "rate" })),
       llm(new ProviderInternalReason({ message: "internal", status: 500 })),
-      llm(new TransportReason({ message: "transport" })),
+      llm(new TransportReason({ message: "transport", transport: "http", operation: "request" })),
     ]
     const ineligible = [
       llm(new AuthenticationReason({ message: "auth", kind: "invalid" })),

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

@@ -32,7 +32,7 @@ describe("SessionExecution lifecycle", () => {
           new AIError({
             module: "test",
             method: "stream",
-            reason: new TransportReason({ message: "Disconnected" }),
+            reason: new TransportReason({ message: "Disconnected", transport: "http", operation: "request" }),
           }),
         ),
       ),

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

@@ -11,6 +11,8 @@ import { Shell } from "@opencode-ai/schema/shell"
 import { Location } from "@opencode-ai/schema/location"
 import { AbsolutePath } from "@opencode-ai/schema/schema"
 import { DateTime } from "effect"
+import path from "path"
+import { pathToFileURL } from "url"
 
 const created = DateTime.makeUnsafe(0)
 const id = (value: string) => SessionMessage.ID.make(`msg_${value}`)
@@ -267,12 +269,13 @@ Recent work
     ])
   })
 
-  test("lowers directory attachments as directory context", () => {
+  test("exposes admitted reference directory source paths in model context", () => {
+    const location = path.resolve("/references/harness-engineering")
     const directory = FileAttachment.make({
       data: Base64.make(Buffer.from("lib/\nindex.ts").toString("base64")),
       mime: "application/x-directory",
-      source: { type: "uri", uri: "file:///project/src" },
-      name: "src/",
+      source: { type: "uri", uri: pathToFileURL(location).href },
+      name: "harness-engineering",
     })
     const messages = toLLMMessages(
       [
@@ -295,14 +298,15 @@ Recent work
         { type: "text", text: "Review this directory" },
         {
           type: "text",
-          text: "\n\nAttached directory: src/\n\nlib/\nindex.ts",
-          metadata: { attachment: { source: directory.source, name: "src/" } },
+          text: `\n\nAttached directory: ${location}\n\nlib/\nindex.ts`,
+          metadata: { attachment: { source: directory.source, name: "harness-engineering" } },
         },
       ],
     })
   })
 
   test("preserves attachment order after the prompt", () => {
+    const directory = path.resolve("/project/src")
     const messages = toLLMMessages(
       [
         SessionMessage.User.make({
@@ -313,7 +317,7 @@ Recent work
             FileAttachment.make({
               data: Base64.make(Buffer.from("index.ts").toString("base64")),
               mime: "application/x-directory",
-              source: { type: "uri", uri: "file:///project/src" },
+              source: { type: "uri", uri: pathToFileURL(directory).href },
               name: "src/",
             }),
             FileAttachment.make({
@@ -332,12 +336,13 @@ Recent work
     expect(messages).toHaveLength(1)
     expect(messages[0]?.content.map((part) => (part.type === "text" ? part.text : part.type))).toEqual([
       "Review these attachments",
-      "\n\nAttached directory: src/\n\nindex.ts",
+      `\n\nAttached directory: ${directory}\n\nindex.ts`,
       "\n\nAttached file: main.ts\n\nexport const value = 1",
     ])
   })
 
   test("omits empty prompt text before an attachment", () => {
+    const directory = path.resolve("/project/src")
     const messages = toLLMMessages(
       [
         SessionMessage.User.make({
@@ -348,7 +353,7 @@ Recent work
             FileAttachment.make({
               data: Base64.make(Buffer.from("index.ts").toString("base64")),
               mime: "application/x-directory",
-              source: { type: "uri", uri: "file:///project/src" },
+              source: { type: "uri", uri: pathToFileURL(directory).href },
               name: "src/",
             }),
           ],
@@ -359,7 +364,9 @@ Recent work
     )
 
     expect(messages).toHaveLength(1)
-    expect(messages[0]?.content).toMatchObject([{ type: "text", text: "\n\nAttached directory: src/\n\nindex.ts" }])
+    expect(messages[0]?.content).toMatchObject([
+      { type: "text", text: `\n\nAttached directory: ${directory}\n\nindex.ts` },
+    ])
   })
 
   test("uses materialized image data as provider media and drops unsupported attachments", () => {
@@ -391,6 +398,108 @@ Recent work
     ])
   })
 
+  test("exposes admitted local image source paths before provider media", () => {
+    const data = Base64.make("AAECAw==")
+    const location = path.resolve("/project/IMG_3480.JPG")
+    const image = FileAttachment.make({
+      data,
+      mime: "image/png",
+      source: { type: "uri", uri: pathToFileURL(location).href },
+      name: "IMG_3480.JPG",
+    })
+
+    const messages = toLLMMessages(
+      [
+        SessionMessage.User.make({
+          id: id("user-local-image-path"),
+          type: "user",
+          text: "Inspect this image",
+          files: [image],
+          time: { created },
+        }),
+      ],
+      model,
+    )
+
+    expect(messages[0]?.content).toEqual([
+      { type: "text", text: "Inspect this image" },
+      { type: "text", text: `Attached file: ${location}` },
+      { type: "media", mediaType: "image/png", data, filename: "IMG_3480.JPG" },
+    ])
+  })
+
+  test("falls back to attachment names for invalid local source paths", () => {
+    const data = Base64.make("AAECAw==")
+    const messages = toLLMMessages(
+      [
+        SessionMessage.User.make({
+          id: id("user-invalid-local-paths"),
+          type: "user",
+          text: "Inspect these attachments",
+          files: [
+            FileAttachment.make({
+              data: Base64.make(Buffer.from("index.ts").toString("base64")),
+              mime: "application/x-directory",
+              source: { type: "uri", uri: "file:///project/src%2Flib" },
+              name: "src/",
+            }),
+            FileAttachment.make({
+              data,
+              mime: "image/png",
+              source: { type: "uri", uri: "file:///project/image%2Fpreview.png" },
+              name: "preview.png",
+            }),
+          ],
+          time: { created },
+        }),
+      ],
+      model,
+    )
+
+    expect(messages[0]?.content).toEqual([
+      { type: "text", text: "Inspect these attachments" },
+      {
+        type: "text",
+        text: "\n\nAttached directory: src/\n\nindex.ts",
+        metadata: {
+          attachment: {
+            source: { type: "uri", uri: "file:///project/src%2Flib" },
+            name: "src/",
+          },
+        },
+      },
+      { type: "media", mediaType: "image/png", data, filename: "preview.png" },
+    ])
+  })
+
+  test("does not add attachment location text for non-local provider media", () => {
+    const data = Base64.make("AAECAw==")
+    const messages = toLLMMessages(
+      [
+        SessionMessage.User.make({
+          id: id("user-remote-image"),
+          type: "user",
+          text: "Inspect this image",
+          files: [
+            FileAttachment.make({
+              data,
+              mime: "image/png",
+              source: { type: "uri", uri: "https://example.com/image.png" },
+              name: "image.png",
+            }),
+          ],
+          time: { created },
+        }),
+      ],
+      model,
+    )
+
+    expect(messages[0]?.content).toEqual([
+      { type: "text", text: "Inspect this image" },
+      { type: "media", mediaType: "image/png", data, filename: "image.png" },
+    ])
+  })
+
   test("deduplicates provider media while preserving durable attachment references", () => {
     const data = Base64.make("AAECAw==")
     const messages = toLLMMessages(
@@ -468,7 +577,7 @@ Recent work
             FileAttachment.make({
               data,
               mime: "image/png",
-              source: { type: "uri", uri: "file:///project/image.png" },
+              source: { type: "uri", uri: pathToFileURL(path.resolve("/project/image.png")).href },
               name: "image.png",
               mention: { start: 0, end: 9, text: "[Image 1]" },
             }),

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

@@ -1,6 +1,7 @@
 import { expect, test } from "bun:test"
 import { Cause, Effect, Exit, Schema } from "effect"
 import { LLMEvent } from "@opencode-ai/ai"
+import { Money } from "@opencode-ai/schema/money"
 import { Bus } from "@opencode-ai/core/bus"
 import { Event } from "@opencode-ai/schema/event"
 import { Agent } from "@opencode-ai/core/agent"
@@ -12,7 +13,6 @@ import { Provider } from "@opencode-ai/core/provider"
 import { RelativePath } from "@opencode-ai/core/schema"
 import { Snapshot } from "@opencode-ai/core/snapshot"
 import { createLLMEventPublisher } from "@opencode-ai/core/session/runner/publish-llm-event"
-import { SessionUsage } from "@opencode-ai/core/session/usage"
 
 const sessionID = Session.ID.make("ses_tool_event_test")
 const base64 = "iVBORw0KGgoAAAANSUhEUgAAAAEAAAAB"
@@ -280,7 +280,6 @@ test("content-filter finish retains failure evidence until step closeout", async
           nonCachedInputTokens: 8,
           outputTokens: 3,
           reasoningTokens: 1,
-          cost: 1.25,
         },
       }),
     ),
@@ -290,13 +289,13 @@ test("content-filter finish retains failure evidence until step closeout", async
   const settlement = publisher.record().finish
   expect(settlement).toMatchObject({
     finish: "content-filter",
-    usage: { nonCachedInputTokens: 8, outputTokens: 3, reasoningTokens: 1, cost: 1.25 },
+    tokens: { input: 8, output: 2, reasoning: 1 },
   })
   if (!settlement) throw new Error("Expected content-filter settlement")
-  const recorded = SessionUsage.record(settlement.usage, [])
   await Effect.runPromise(
     publisher.publishStepFailure({
-      ...recorded,
+      cost: Money.USD.make(1.25),
+      tokens: settlement.tokens,
       snapshot: Snapshot.ID.make("tree-end"),
       files: [RelativePath.make("src/changed.ts")],
     }),

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

@@ -515,7 +515,11 @@ const providerUnavailable = () =>
   new AIError({
     module: "test",
     method: "stream",
-    reason: new TransportReason({ message: "Provider unavailable" }),
+    reason: new TransportReason({
+      message: "Provider unavailable",
+      transport: "http",
+      operation: "request",
+    }),
   })
 
 const incompleteStream = () =>
@@ -3947,7 +3951,7 @@ describe("SessionRunnerLLM", () => {
     }),
   )
 
-  it.effect("retries eligible pre-output failures after exponential backoff", () =>
+  it.effect("bounds jittered exponential backoff for eligible pre-output failures", () =>
     Effect.gen(function* () {
       const session = yield* setup
       yield* admit(session, "Retry transport")
@@ -3956,9 +3960,9 @@ describe("SessionRunnerLLM", () => {
 
       const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
       yield* TestLLM.wait(1)
-      yield* TestClock.adjust("1999 millis")
+      yield* TestClock.adjust("1599 millis")
       expect(requests).toHaveLength(1)
-      yield* TestClock.adjust("1 millis")
+      yield* TestClock.adjust("801 millis")
       yield* Fiber.join(run)
 
       expect(requests).toHaveLength(2)
@@ -3983,7 +3987,7 @@ describe("SessionRunnerLLM", () => {
 
       const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
       yield* TestLLM.wait(1)
-      yield* TestClock.adjust("2 seconds")
+      yield* TestClock.adjust("2400 millis")
       yield* Fiber.join(run)
 
       expect(requests).toHaveLength(2)
@@ -4028,7 +4032,7 @@ describe("SessionRunnerLLM", () => {
 
       const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
       yield* TestLLM.wait(1)
-      yield* TestClock.adjust("2 seconds")
+      yield* TestClock.adjust("2400 millis")
       yield* Fiber.join(run)
 
       expect(requests).toHaveLength(2)
@@ -4085,7 +4089,7 @@ describe("SessionRunnerLLM", () => {
 
       const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
       yield* TestLLM.wait(1)
-      yield* TestClock.adjust("2 seconds")
+      yield* TestClock.adjust("2400 millis")
       yield* Fiber.join(run)
 
       expect(requests[1]?.messages.at(-2)).toMatchObject({
@@ -4126,7 +4130,7 @@ describe("SessionRunnerLLM", () => {
       const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
       yield* TestLLM.wait(1)
       while (!(yield* recordedEventTypes(sessionID)).includes("session.retry.scheduled.1")) yield* Effect.yieldNow
-      yield* TestClock.adjust("2 seconds")
+      yield* TestClock.adjust("2400 millis")
       yield* Fiber.join(run)
 
       expect(executions).toEqual(["settled"])
@@ -4165,7 +4169,7 @@ describe("SessionRunnerLLM", () => {
       const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
       yield* TestLLM.wait(1)
       while (!(yield* recordedEventTypes(sessionID)).includes("session.retry.scheduled.1")) yield* Effect.yieldNow
-      yield* TestClock.adjust("2 seconds")
+      yield* TestClock.adjust("2400 millis")
       yield* Fiber.join(run)
 
       expect(messageRoles(requests[1])).toEqual(["user", "assistant", "tool", "user"])
@@ -4203,7 +4207,7 @@ describe("SessionRunnerLLM", () => {
 
       const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
       yield* TestLLM.wait(1)
-      for (const [index, delay] of [2_000, 4_000, 8_000, 16_000].entries()) {
+      for (const [index, delay] of [2_400, 4_800, 9_600, 19_200].entries()) {
         yield* TestClock.adjust(delay)
         yield* TestLLM.wait(index + 2)
       }
@@ -4224,7 +4228,7 @@ describe("SessionRunnerLLM", () => {
 
       const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
       yield* TestLLM.wait(1)
-      for (const [index, delay] of [2_000, 4_000, 8_000, 16_000].entries()) {
+      for (const [index, delay] of [2_400, 4_800, 9_600, 19_200].entries()) {
         yield* TestClock.adjust(delay)
         yield* TestLLM.wait(index + 2)
       }
@@ -4239,12 +4243,15 @@ describe("SessionRunnerLLM", () => {
         .orderBy(asc(EventTable.seq))
         .all()
         .pipe(Effect.orDie)
-      expect(retries.map((event) => event.data)).toMatchObject([
-        { attempt: 2, at: 2_000 },
-        { attempt: 3, at: 6_000 },
-        { attempt: 4, at: 14_000 },
-        { attempt: 5, at: 30_000 },
-      ])
+      for (const [index, range] of [
+        [1_600, 2_400],
+        [4_800, 7_200],
+        [11_200, 16_800],
+        [24_000, 36_000],
+      ].entries()) {
+        expect(retries[index]?.data.at).toBeGreaterThanOrEqual(range[0]!)
+        expect(retries[index]?.data.at).toBeLessThanOrEqual(range[1]!)
+      }
       expect((yield* recordedEventTypes(sessionID)).filter((type) => type === "session.step.started.1")).toHaveLength(5)
       const assistant = requireAssistant(yield* session.context(sessionID))
       expect(yield* recordedStepSettlementEvents(sessionID, assistant.id)).toMatchObject([
@@ -4274,7 +4281,7 @@ describe("SessionRunnerLLM", () => {
 
       const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
       yield* TestLLM.wait(1)
-      yield* TestClock.adjust("2 seconds")
+      yield* TestClock.adjust("2400 millis")
       yield* Fiber.join(run)
 
       expect(requests).toHaveLength(3)

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

@@ -1,28 +0,0 @@
-import { expect, test } from "bun:test"
-import { Usage } from "@opencode-ai/ai"
-import { Money } from "@opencode-ai/schema/money"
-import { SessionUsage } from "@opencode-ai/core/session/usage"
-
-const costs = [
-  {
-    input: Money.USDPerMillionTokens.make(1),
-    output: Money.USDPerMillionTokens.make(2),
-    cache: { read: Money.USDPerMillionTokens.zero, write: Money.USDPerMillionTokens.zero },
-  },
-]
-
-test("prefers provider-reported cost", () => {
-  expect(SessionUsage.record(new Usage({ nonCachedInputTokens: 1_000_000, cost: 0.25 }), costs).cost).toBe(
-    Money.USD.make(0.25),
-  )
-  expect(SessionUsage.record(new Usage({ nonCachedInputTokens: 1_000_000, cost: 0 }), costs).cost).toBe(Money.USD.zero)
-})
-
-test("falls back to catalog pricing for invalid reported cost", () => {
-  expect(SessionUsage.record(new Usage({ nonCachedInputTokens: 1_000_000, cost: Number.NaN }), costs).cost).toBe(
-    Money.USD.make(1),
-  )
-  expect(SessionUsage.record(new Usage({ nonCachedInputTokens: 1_000_000, cost: -1 }), costs).cost).toBe(
-    Money.USD.make(1),
-  )
-})

packages/server/src/fetch.ts

@@ -7,15 +7,6 @@ import { isAllowedCorsOrigin } from "./cors"
 import { createRoutes } from "./routes"
 import type { ServerOptions } from "./options"
 
-export interface BootOptions {
-  /**
-   * Resumes execution-journaled Sessions once the application layer boots. Pair with
-   * `SessionExecution.configured({ suspendOnStart: true })` on runtimes that can die without
-   * teardown, so turns orphaned by a hard death replay on the next boot.
-   */
-  readonly resumeSuspendedSessions?: boolean
-}
-
 /**
  * Builds a web-standard fetch handler — `(request: Request) => Promise<Response>` — serving the
  * same HttpApi routes as the Node server process without binding a port, owning a listener, or
@@ -32,13 +23,17 @@ export interface BootOptions {
  * Auth follows `createRoutes` semantics: `options.password` enforces Basic auth; omitting it
  * serves unauthenticated, so an embedder without a password must front the handler with its own
  * access control.
+ *
+ * Sessions whose execution claim was never released resume once the layer is built, exactly as
+ * the Node server process does: a runtime that dies without teardown — an evicted Durable
+ * Object leaves the same durable signature as a killed process — replays orphaned turns on the
+ * next boot, and the sweep is a no-op when nothing is suspended.
  */
-export const make = Effect.fn("ServerFetch.make")(function* (options: ServerOptions = {}, boot: BootOptions = {}) {
+export const make = Effect.fn("ServerFetch.make")(function* (options: ServerOptions = {}) {
   const context = yield* Layer.build(createRoutes(options, () => []).pipe(Layer.provide(HttpServer.layerServices)))
   // Forked so the returned handler is never delayed; resumed drains are already
   // logged and durably recorded by the execution layer.
-  if (boot.resumeSuspendedSessions)
-    yield* Effect.forkDetach(Context.get(context, SessionRestart.Service).resumeSuspendedSessions)
+  yield* Effect.forkDetach(Context.get(context, SessionRestart.Service).resumeSuspendedSessions)
   return Context.get(context, HttpRouter.HttpRouter)
     .asHttpEffect()
     .pipe(

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

@@ -1814,7 +1814,7 @@ function CompactionMessage(props: { message: Extract<SessionMessageInfo, { type:
             streaming={true}
             internalBlockMode="top-level"
             content={content()}
-            tableOptions={{ style: "grid" }}
+            tableOptions={{ style: "grid", cellPaddingX: 1 }}
             conceal={ctx.markdownMode() === "rendered"}
             fg={theme.markdown.text}
             bg={theme.background.default}
@@ -2264,7 +2264,7 @@ function TextPart(props: { last: boolean; part: SessionMessageAssistantText }) {
           streaming={true}
           internalBlockMode="top-level"
           content={props.part.text.trim()}
-          tableOptions={{ style: "grid" }}
+          tableOptions={{ style: "grid", cellPaddingX: 1 }}
           conceal={ctx.markdownMode() === "rendered"}
           fg={theme.markdown.text}
           bg={theme.background.default}