@neurocode-ai/core 1.18.8

package/test/session-runner.test.ts

@@ -0,0 +1,3365 @@
+import { describe, expect } from "bun:test"
+import {
+  LLMClient,
+  LLMError,
+  LLMEvent,
+  Model,
+  TransportReason,
+  InvalidRequestReason,
+  type LLMClientShape,
+  type LLMRequest,
+} from "@neurocode-ai/llm"
+import * as OpenAIChat from "@neurocode-ai/llm/protocols/openai-chat"
+import { Database } from "@neurocode-ai/core/database/database"
+import { makeLocationNode } from "@neurocode-ai/core/effect/app-node"
+import { AppNodeBuilder } from "@neurocode-ai/core/effect/app-node-builder"
+import { LayerNodePlatform } from "@neurocode-ai/core/effect/app-node-platform"
+import { LayerNode } from "@neurocode-ai/core/effect/layer-node"
+import { EventV2 } from "@neurocode-ai/core/event"
+import { PermissionV2 } from "@neurocode-ai/core/permission"
+import { EventTable } from "@neurocode-ai/core/event/sql"
+import { Project } from "@neurocode-ai/core/project"
+import { ProjectTable } from "@neurocode-ai/core/project/sql"
+import { QuestionV2 } from "@neurocode-ai/core/question"
+import { AbsolutePath } from "@neurocode-ai/core/schema"
+import { SessionV2 } from "@neurocode-ai/core/session"
+import { Snapshot } from "@neurocode-ai/core/snapshot"
+import { ContextSnapshotDecodeError } from "@neurocode-ai/core/session/error"
+import { SessionEvent } from "@neurocode-ai/core/session/event"
+import { SessionInput } from "@neurocode-ai/core/session/input"
+import { SessionMessage } from "@neurocode-ai/core/session/message"
+import { Prompt } from "@neurocode-ai/core/session/prompt"
+import { SessionProjector } from "@neurocode-ai/core/session/projector"
+import { SessionExecution } from "@neurocode-ai/core/session/execution"
+import { SessionRunCoordinator } from "@neurocode-ai/core/session/run-coordinator"
+import { SessionRunner } from "@neurocode-ai/core/session/runner"
+import * as SessionRunnerLLM from "@neurocode-ai/core/session/runner/llm"
+import { SessionRunnerModel } from "@neurocode-ai/core/session/runner/model"
+import { ToolRegistry } from "@neurocode-ai/core/tool/registry"
+import { ApplicationTools } from "@neurocode-ai/core/tool/application-tools"
+import { AgentV2 } from "@neurocode-ai/core/agent"
+import { Config } from "@neurocode-ai/core/config"
+import { ConfigCompaction } from "@neurocode-ai/core/config/compaction"
+import { Tool } from "@neurocode-ai/core/tool/tool"
+import {
+  SessionContextEpochTable,
+  SessionInputTable,
+  SessionMessageTable,
+  SessionTable,
+} from "@neurocode-ai/core/session/sql"
+import { SessionStore } from "@neurocode-ai/core/session/store"
+import { SystemContext } from "@neurocode-ai/core/system-context"
+import { SystemContextRegistry } from "@neurocode-ai/core/system-context/registry"
+import { SkillGuidance } from "@neurocode-ai/core/skill/guidance"
+import { ReferenceGuidance } from "@neurocode-ai/core/reference/guidance"
+import { ModelV2 } from "@neurocode-ai/core/model"
+import { Location } from "@neurocode-ai/core/location"
+import { ProviderV2 } from "@neurocode-ai/core/provider"
+import { Cause, DateTime, Deferred, Effect, Exit, Fiber, Layer, Schema, Stream } from "effect"
+import { asc, eq } from "drizzle-orm"
+import { testEffect } from "./lib/effect"
+
+const requests: LLMRequest[] = []
+let response: LLMEvent[] = []
+let responses: LLMEvent[][] | undefined
+let responseStream: Stream.Stream<LLMEvent, LLMError> | undefined
+let streamGate: Deferred.Deferred<void> | undefined
+let streamStarted: Deferred.Deferred<void> | undefined
+let streamFailure: LLMError | undefined
+let toolExecutionGate: Deferred.Deferred<void> | undefined
+let toolExecutionsStarted: Deferred.Deferred<void> | undefined
+let toolExecutionsReady = 5
+let activeToolExecutions = 0
+let maxActiveToolExecutions = 0
+const client = Layer.succeed(
+  LLMClient.Service,
+  LLMClient.Service.of({
+    prepare: () => Effect.die("unused"),
+    stream: ((request: LLMRequest) => {
+      requests.push(request)
+      if (responseStream) {
+        const stream = responseStream
+        responseStream = undefined
+        return stream
+      }
+      const events = streamFailure
+        ? Stream.fail(streamFailure)
+        : Stream.fromIterable(responses === undefined ? response : (responses.shift() ?? []))
+      if (!streamGate) return events
+      return Stream.unwrap(
+        (streamStarted ? Deferred.succeed(streamStarted, undefined) : Effect.void).pipe(
+          Effect.andThen(Deferred.await(streamGate)),
+          Effect.as(events),
+        ),
+      )
+    }) as unknown as LLMClientShape["stream"],
+    generate: () => Effect.die("unused"),
+  }),
+)
+const model = Model.make({ id: "fake-model", provider: "fake", route: OpenAIChat.route })
+const replacementModel = Model.make({ id: "replacement", provider: "fake", route: OpenAIChat.route })
+const compactModel = Model.make({
+  id: "compact",
+  provider: "fake",
+  route: OpenAIChat.route.with({ limits: { context: 4_000, output: 50 } }),
+})
+const recoveryModel = Model.make({
+  id: "recovery",
+  provider: "fake",
+  route: OpenAIChat.route.with({ limits: { context: 20_000, output: 1_000 } }),
+})
+const authorizations: Tool.Context[] = []
+const executions: string[] = []
+const permission = Layer.succeed(
+  PermissionV2.Service,
+  PermissionV2.Service.of({
+    assert: () => Effect.die("unused"),
+    ask: () => Effect.die("unused"),
+    reply: () => Effect.die("unused"),
+    get: () => Effect.die("unused"),
+    forSession: () => Effect.die("unused"),
+    list: () => Effect.die("unused"),
+  }),
+)
+const echo = Layer.effectDiscard(
+  ToolRegistry.Service.use((registry) =>
+    registry.register({
+      echo: Tool.make({
+        description: "Echo text",
+        input: Schema.Struct({ text: Schema.String }),
+        output: Schema.Struct({ text: Schema.String }),
+        toModelOutput: ({ output }) => [{ type: "text", text: output.text }],
+        execute: ({ text }, context) =>
+          Effect.gen(function* () {
+            authorizations.push(context)
+            executions.push(text)
+            activeToolExecutions++
+            maxActiveToolExecutions = Math.max(maxActiveToolExecutions, activeToolExecutions)
+            if (activeToolExecutions === toolExecutionsReady && toolExecutionsStarted) {
+              yield* Deferred.succeed(toolExecutionsStarted, undefined)
+            }
+            if (toolExecutionGate) yield* Deferred.await(toolExecutionGate)
+            return { text }
+          }).pipe(Effect.ensuring(Effect.sync(() => activeToolExecutions--))),
+      }),
+      defect: Tool.make({
+        description: "Fail unexpectedly",
+        input: Schema.Struct({}),
+        output: Schema.Struct({}),
+        execute: () => Effect.die("unexpected tool defect"),
+      }),
+    }),
+  ),
+)
+const echoNode = makeLocationNode({ name: "test/session-runner-tools", layer: echo, deps: [ToolRegistry.node] })
+let modelResolveHook = Effect.void
+let currentModel = model
+const models = SessionRunnerModel.layerWith((session) =>
+  modelResolveHook.pipe(Effect.as(session.model?.id === "replacement" ? replacementModel : currentModel)),
+)
+const systemContextKey = SystemContext.Key.make("test/context")
+let systemBaseline = "Initial context"
+let systemRemoved = false
+let systemUnavailable = false
+let systemLoadHook = Effect.void
+const skillBaselines = new Map<AgentV2.ID, string>()
+const systemContext = Layer.effectDiscard(
+  SystemContextRegistry.Service.pipe(
+    Effect.flatMap((registry) =>
+      registry.register({
+        key: systemContextKey,
+        load: Effect.sync(() =>
+          SystemContext.combine(
+            systemRemoved
+              ? []
+              : [
+                  SystemContext.make({
+                    key: systemContextKey,
+                    codec: Schema.toCodecJson(Schema.String),
+                    load: systemLoadHook.pipe(
+                      Effect.andThen(
+                        Effect.sync(() => (systemUnavailable ? SystemContext.unavailable : systemBaseline)),
+                      ),
+                    ),
+                    baseline: String,
+                    update: (_previous, current) => current,
+                    removed: () => "System context source removed: test/context",
+                  }),
+                ],
+          ),
+        ),
+      }),
+    ),
+  ),
+).pipe(Layer.provideMerge(AppNodeBuilder.build(SystemContextRegistry.node)))
+const skillGuidance = Layer.mock(SkillGuidance.Service, {
+  load: (agent) =>
+    Effect.succeed(
+      skillBaselines.has(agent.id)
+        ? SystemContext.make({
+            key: SystemContext.Key.make("test/skill-guidance"),
+            codec: Schema.toCodecJson(Schema.String),
+            load: Effect.succeed(skillBaselines.get(agent.id)!),
+            baseline: String,
+            update: (_previous, current) => current,
+            removed: () => "Skill guidance removed",
+          })
+        : SystemContext.empty,
+    ),
+})
+const referenceGuidance = Layer.mock(ReferenceGuidance.Service, { load: () => Effect.succeed(SystemContext.empty) })
+const config = Layer.succeed(
+  Config.Service,
+  Config.Service.of({
+    entries: () =>
+      Effect.succeed([
+        new Config.Document({
+          type: "document",
+          info: new Config.Info({
+            compaction: new ConfigCompaction.Info({
+              buffer: 3_000,
+              keep: new ConfigCompaction.Keep({ tokens: 1_000 }),
+            }),
+          }),
+        }),
+      ]),
+  }),
+)
+const runnerLayer = AppNodeBuilder.build(SessionRunnerLLM.node, [
+  [Snapshot.node, Snapshot.noopLayer],
+  [LayerNodePlatform.llmClient, client],
+  [SessionRunnerModel.node, models],
+  [SystemContextRegistry.node, systemContext],
+  [Location.node, Location.boundNode({ directory: AbsolutePath.make("/project") })],
+  [SkillGuidance.node, skillGuidance],
+  [ReferenceGuidance.node, referenceGuidance],
+  [PermissionV2.node, permission],
+  [Config.node, config],
+])
+const execution = Layer.effect(
+  SessionExecution.Service,
+  Effect.gen(function* () {
+    const sessionRunner = yield* SessionRunner.Service
+    const coordinator = yield* SessionRunCoordinator.make<SessionV2.ID, SessionRunner.RunError>({
+      drain: (sessionID, force) => sessionRunner.run({ sessionID, force }),
+    })
+    return SessionExecution.Service.of({
+      active: coordinator.active,
+      resume: coordinator.run,
+      wake: coordinator.wake,
+      interrupt: coordinator.interrupt,
+    })
+  }),
+).pipe(Layer.provide(runnerLayer))
+const it = testEffect(
+  AppNodeBuilder.build(
+    LayerNode.group([
+      Database.node,
+      EventV2.node,
+      QuestionV2.node,
+      SessionProjector.node,
+      SessionStore.node,
+      ApplicationTools.node,
+      AgentV2.node,
+      ToolRegistry.node,
+      ToolRegistry.toolsNode,
+      echoNode,
+      SessionRunnerModel.node,
+      SystemContextRegistry.node,
+      SkillGuidance.node,
+      ReferenceGuidance.node,
+      Config.node,
+      Snapshot.node,
+      SessionRunnerLLM.node,
+      SessionExecution.node,
+      SessionV2.node,
+    ]),
+    [
+      [LayerNodePlatform.llmClient, client],
+      [PermissionV2.node, permission],
+      [SessionRunnerModel.node, models],
+      [SystemContextRegistry.node, systemContext],
+      [Location.node, Location.boundNode({ directory: AbsolutePath.make("/project") })],
+      [SkillGuidance.node, skillGuidance],
+      [ReferenceGuidance.node, referenceGuidance],
+      [Snapshot.node, Snapshot.noopLayer],
+      [SessionExecution.node, execution],
+      [Config.node, config],
+    ],
+  ),
+)
+const sessionID = SessionV2.ID.make("ses_runner_test")
+const otherSessionID = SessionV2.ID.make("ses_runner_other")
+
+const insertSession = (id: SessionV2.ID) =>
+  Effect.gen(function* () {
+    const { db } = yield* Database.Service
+    yield* db
+      .insert(SessionTable)
+      .values({
+        id,
+        project_id: Project.ID.global,
+        slug: id,
+        directory: "/project",
+        title: "test",
+        version: "test",
+      })
+      .onConflictDoNothing()
+      .run()
+      .pipe(Effect.orDie)
+  })
+
+const setup = Effect.gen(function* () {
+  const { db } = yield* Database.Service
+  response = []
+  systemBaseline = "Initial context"
+  systemRemoved = false
+  systemUnavailable = false
+  systemLoadHook = Effect.void
+  modelResolveHook = Effect.void
+  currentModel = model
+  skillBaselines.clear()
+  responses = undefined
+  streamFailure = undefined
+  responseStream = undefined
+  streamGate = undefined
+  streamStarted = undefined
+  toolExecutionGate = undefined
+  toolExecutionsStarted = undefined
+  toolExecutionsReady = 5
+  activeToolExecutions = 0
+  maxActiveToolExecutions = 0
+  yield* db
+    .insert(ProjectTable)
+    .values({ id: Project.ID.global, worktree: AbsolutePath.make("/project"), sandboxes: [] })
+    .onConflictDoNothing()
+    .run()
+    .pipe(Effect.orDie)
+  yield* insertSession(sessionID)
+})
+
+const providerUnavailable = () =>
+  new LLMError({
+    module: "test",
+    method: "stream",
+    reason: new TransportReason({ message: "Provider unavailable" }),
+  })
+
+const setupOverflowRecovery = Effect.gen(function* () {
+  yield* setup
+  const session = yield* SessionV2.Service
+  response = fragmentFixture("text", "text-earlier", ["Earlier answer"]).completeEvents
+  yield* session.prompt({
+    sessionID,
+    prompt: Prompt.make({ text: "Earlier question ".repeat(700) }),
+    resume: false,
+  })
+  yield* session.resume(sessionID)
+  currentModel = recoveryModel
+  requests.length = 0
+  return session
+})
+
+const messageTexts = (request: LLMRequest, role: "user" | "system") =>
+  request.messages.flatMap((message) =>
+    message.role === role ? message.content.flatMap((content) => (content.type === "text" ? [content.text] : [])) : [],
+  )
+const userTexts = (request: LLMRequest) => messageTexts(request, "user")
+const systemTexts = (request: LLMRequest) => messageTexts(request, "system")
+
+const replaySessionProjection = (id: SessionV2.ID) =>
+  Effect.gen(function* () {
+    const { db } = yield* Database.Service
+    const events = yield* EventV2.Service
+    const recorded = yield* db
+      .select()
+      .from(EventTable)
+      .where(eq(EventTable.aggregate_id, id))
+      .orderBy(asc(EventTable.seq))
+      .all()
+      .pipe(Effect.orDie)
+
+    yield* events.remove(id)
+    yield* db.delete(SessionInputTable).where(eq(SessionInputTable.session_id, id)).run().pipe(Effect.orDie)
+    yield* db.delete(SessionMessageTable).where(eq(SessionMessageTable.session_id, id)).run().pipe(Effect.orDie)
+    yield* events.replayAll(
+      recorded.map((event) => ({
+        id: event.id,
+        aggregateID: event.aggregate_id,
+        seq: event.seq,
+        type: event.type,
+        data: event.data,
+      })),
+    )
+  })
+
+type FragmentKind = "text" | "reasoning" | "tool input"
+
+type FragmentFixture = {
+  readonly delta: EventV2.Definition
+  readonly completeEvents: LLMEvent[]
+  readonly partialEvents: LLMEvent[]
+  readonly expectedAssistant: unknown
+  readonly expectedContent: unknown
+}
+
+const fragmentKinds: readonly FragmentKind[] = ["text", "reasoning", "tool input"]
+
+const fragmentID = (kind: FragmentKind, suffix: string) => `${kind === "tool input" ? "call" : kind}-${suffix}`
+
+const fragmentFixture = (kind: FragmentKind, id: string, chunks: readonly string[]): FragmentFixture => {
+  const text = chunks.join("")
+  switch (kind) {
+    case "text": {
+      const partialEvents = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.textStart({ id }),
+        ...chunks.map((text) => LLMEvent.textDelta({ id, text })),
+      ]
+      const expectedContent = { type: "text", id, text }
+      return {
+        delta: SessionEvent.Text.Delta,
+        partialEvents,
+        completeEvents: [
+          ...partialEvents,
+          LLMEvent.textEnd({ id }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+        expectedAssistant: { type: "assistant", finish: "stop", content: [expectedContent] },
+        expectedContent,
+      }
+    }
+    case "reasoning": {
+      const partialEvents = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.reasoningStart({ id }),
+        ...chunks.map((text) => LLMEvent.reasoningDelta({ id, text })),
+      ]
+      const expectedContent = { type: "reasoning", id, text }
+      return {
+        delta: SessionEvent.Reasoning.Delta,
+        partialEvents,
+        completeEvents: [
+          ...partialEvents,
+          LLMEvent.reasoningEnd({ id }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+        expectedAssistant: { type: "assistant", finish: "stop", content: [expectedContent] },
+        expectedContent,
+      }
+    }
+    case "tool input": {
+      const partialEvents = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolInputStart({ id, name: "echo" }),
+        ...chunks.map((text) => LLMEvent.toolInputDelta({ id, name: "echo", text })),
+      ]
+      const expectedContent = { type: "tool", id, state: { status: "pending", input: text } }
+      return {
+        delta: SessionEvent.Tool.Input.Delta,
+        partialEvents,
+        completeEvents: [...partialEvents, LLMEvent.toolInputEnd({ id, name: "echo" })],
+        expectedAssistant: { type: "assistant", content: [expectedContent] },
+        expectedContent,
+      }
+    }
+  }
+}
+
+const verifyEphemeralDeltas = (kind: FragmentKind) =>
+  Effect.gen(function* () {
+    yield* setup
+    const session = yield* SessionV2.Service
+    const prompt = `Stream ${kind}`
+    const chunks = Array.from({ length: 32 }, (_, index) => `${index},`)
+    const fixture = fragmentFixture(kind, fragmentID(kind, "many"), chunks)
+    const expectedContext = [{ type: "user", text: prompt }, fixture.expectedAssistant]
+    yield* session.prompt({ sessionID, prompt: Prompt.make({ text: prompt }), resume: false })
+    const events = yield* EventV2.Service
+    const live = yield* events.subscribe(fixture.delta).pipe(Stream.take(32), Stream.runCollect, Effect.forkScoped)
+    yield* Effect.yieldNow
+    response = fixture.completeEvents
+
+    yield* session.resume(sessionID)
+
+    const { db } = yield* Database.Service
+    const deltas = yield* db
+      .select({ type: EventTable.type })
+      .from(EventTable)
+      .where(eq(EventTable.type, EventV2.versionedType(fixture.delta.type, 1)))
+      .all()
+      .pipe(Effect.orDie)
+    expect(Array.from(yield* Fiber.join(live))).toHaveLength(32)
+    expect(deltas).toHaveLength(0)
+    expect(yield* session.context(sessionID)).toMatchObject(expectedContext)
+
+    yield* replaySessionProjection(sessionID)
+
+    expect(yield* session.context(sessionID)).toMatchObject(expectedContext)
+  })
+
+const verifyPartialFlushOnFailure = (kind: FragmentKind) =>
+  Effect.gen(function* () {
+    yield* setup
+    const session = yield* SessionV2.Service
+    const prompt = `Fail after ${kind}`
+    const fixture = fragmentFixture(kind, fragmentID(kind, "partial"), ["Partial"])
+    const failure = providerUnavailable()
+    yield* session.prompt({ sessionID, prompt: Prompt.make({ text: prompt }), resume: false })
+    responseStream = Stream.concat(Stream.fromIterable(fixture.partialEvents), Stream.fail(failure))
+
+    expect(yield* session.resume(sessionID).pipe(Effect.flip)).toBe(failure)
+    expect(yield* session.context(sessionID)).toMatchObject([
+      { type: "user", text: prompt },
+      {
+        type: "assistant",
+        finish: "error",
+        error: { type: "unknown", message: "Provider unavailable" },
+        content: [fixture.expectedContent],
+      },
+    ])
+  })
+
+const verifyPartialFlushOnInterruption = (kind: FragmentKind) =>
+  Effect.gen(function* () {
+    yield* setup
+    const session = yield* SessionV2.Service
+    const prompt = `Interrupt after ${kind}`
+    const fixture = fragmentFixture(kind, fragmentID(kind, "interrupted"), ["Partial"])
+    const streamed = yield* Deferred.make<void>()
+    yield* session.prompt({ sessionID, prompt: Prompt.make({ text: prompt }), resume: false })
+    responseStream = Stream.concat(
+      Stream.fromIterable(fixture.partialEvents),
+      Stream.fromEffect(Deferred.succeed(streamed, undefined)).pipe(Stream.flatMap(() => Stream.never)),
+    )
+
+    const runner = yield* SessionRunner.Service
+    const fiber = yield* runner.run({ sessionID, force: true }).pipe(Effect.forkChild)
+    yield* Deferred.await(streamed)
+    yield* Fiber.interrupt(fiber)
+    expect(yield* session.context(sessionID)).toMatchObject([
+      { type: "user", text: prompt },
+      {
+        type: "assistant",
+        finish: "error",
+        error: { type: "unknown", message: "Provider turn interrupted" },
+        content: [
+          kind === "tool input"
+            ? { type: "tool", id: fragmentID(kind, "interrupted"), state: { status: "error" } }
+            : fixture.expectedContent,
+        ],
+      },
+    ])
+  })
+
+describe("SessionRunnerLLM", () => {
+  it.effect("advertises and executes a globally attached application tool", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const applicationTools = yield* ApplicationTools.Service
+      const session = yield* SessionV2.Service
+      const contexts: Tool.Context[] = []
+      yield* applicationTools.register({
+        application_context: Tool.make({
+          description: "Read application context",
+          input: Schema.Struct({ query: Schema.String }),
+          output: Schema.Struct({ answer: Schema.String }),
+          execute: ({ query }, context) =>
+            Effect.sync(() => {
+              contexts.push(context)
+              return { answer: query.toUpperCase() }
+            }),
+        }),
+      })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Use application context" }), resume: false })
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-application", name: "application_context", input: { query: "hello" } }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [],
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(requests[0]?.tools.map((tool) => tool.name)).toContain("application_context")
+      expect(contexts).toEqual([
+        {
+          sessionID,
+          agent: AgentV2.ID.make("build"),
+          assistantMessageID: expect.stringMatching(/^msg_/),
+          toolCallID: "call-application",
+        },
+      ])
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Use application context" },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "call-application",
+              state: { status: "completed", structured: { answer: "HELLO" } },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("starts a real runner turn after default prompt recording", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      requests.length = 0
+      responses = undefined
+      streamGate = undefined
+      streamStarted = undefined
+      response = []
+
+      const message = yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Run automatically" }) })
+
+      expect(requests).toHaveLength(1)
+      expect(yield* session.messages({ sessionID })).toMatchObject([
+        { id: message.id, type: "user", text: "Run automatically" },
+      ])
+    }),
+  )
+
+  it.effect("streams one request with registry definitions from chronological V2 user history", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Second" }), resume: false })
+
+      requests.length = 0
+      responses = undefined
+      streamGate = undefined
+      streamStarted = undefined
+      response = []
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(requests[0]?.model).toBe(model)
+      expect(requests[0]?.tools.map((tool) => tool.name)).toEqual(["echo", "defect"])
+      expect(requests[0]?.messages.map((message) => ({ role: message.role, content: message.content }))).toEqual([
+        { role: "user", content: [{ type: "text", text: "First" }] },
+        { role: "user", content: [{ type: "text", text: "Second" }] },
+      ])
+      expect(yield* session.messages({ sessionID })).toHaveLength(2)
+    }),
+  )
+
+  it.effect("retries the first provider turn after system context becomes available", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const { db } = yield* Database.Service
+      const messageID = SessionMessage.ID.create()
+      systemUnavailable = true
+      yield* session.prompt({ id: messageID, sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+      requests.length = 0
+
+      const exit = yield* session.resume(sessionID).pipe(Effect.exit)
+
+      expect(Exit.isFailure(exit)).toBe(true)
+      if (Exit.isFailure(exit)) expect(Cause.squash(exit.cause)).toBeInstanceOf(SystemContext.InitializationBlocked)
+      expect(requests).toHaveLength(0)
+      expect(yield* SessionInput.hasPending(db, sessionID, "steer")).toBe(true)
+      expect(
+        yield* db
+          .select()
+          .from(SessionContextEpochTable)
+          .where(eq(SessionContextEpochTable.session_id, sessionID))
+          .get(),
+      ).toBeUndefined()
+
+      systemUnavailable = false
+      yield* session.prompt({ id: messageID, sessionID, prompt: Prompt.make({ text: "First" }) })
+
+      expect(requests).toHaveLength(1)
+      expect(requests[0]?.messages.map((message) => message.role)).toEqual(["user"])
+    }),
+  )
+
+  it.effect("interrupts a source Location runner after a Session moves", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      const { db } = yield* Database.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+
+      yield* events.publish(SessionEvent.Moved, {
+        sessionID,
+        timestamp: DateTime.makeUnsafe(1),
+        location: Location.Ref.make({ directory: AbsolutePath.make("/moved") }),
+      })
+      expect(
+        yield* db
+          .select()
+          .from(SessionContextEpochTable)
+          .where(eq(SessionContextEpochTable.session_id, sessionID))
+          .get(),
+      ).toBeUndefined()
+
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Second" }), resume: false })
+      const exit = yield* session.resume(sessionID).pipe(Effect.exit)
+
+      expect(Exit.isFailure(exit) && Cause.hasInterruptsOnly(exit.cause)).toBe(true)
+      expect(requests).toHaveLength(1)
+      expect(yield* SessionInput.hasPending(db, sessionID, "steer")).toBe(true)
+    }),
+  )
+
+  it.effect("fails gracefully when a stored context snapshot cannot be decoded", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const { db } = yield* Database.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+      response = []
+      yield* session.resume(sessionID)
+      yield* db
+        .update(SessionContextEpochTable)
+        .set({ snapshot: { invalid: { value: "bad" } } })
+        .where(eq(SessionContextEpochTable.session_id, sessionID))
+        .run()
+        .pipe(Effect.orDie)
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Second" }), resume: false })
+      requests.length = 0
+
+      const exit = yield* session.resume(sessionID).pipe(Effect.exit)
+
+      expect(Exit.isFailure(exit)).toBe(true)
+      if (Exit.isFailure(exit)) expect(Cause.squash(exit.cause)).toBeInstanceOf(ContextSnapshotDecodeError)
+      expect(requests).toHaveLength(0)
+    }),
+  )
+
+  it.effect("reuses one durable baseline after the context producer changes", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+      systemBaseline = "Changed context"
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Second" }), resume: false })
+      yield* session.resume(sessionID)
+
+      expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
+        ["Initial context"],
+        ["Initial context"],
+      ])
+      expect(requests[1]?.messages.map((message) => message.role)).toEqual(["user", "user", "system"])
+      expect(requests[1]?.messages.at(-1)?.content).toEqual([{ type: "text", text: "Changed context" }])
+      expect(yield* session.messages({ sessionID })).toHaveLength(3)
+      const { db } = yield* Database.Service
+      expect(
+        yield* db
+          .select({ id: EventTable.id })
+          .from(EventTable)
+          .where(eq(EventTable.type, "session.next.context.updated.1"))
+          .all()
+          .pipe(Effect.orDie),
+      ).toHaveLength(1)
+      yield* replaySessionProjection(sessionID)
+      expect(yield* session.messages({ sessionID })).toHaveLength(3)
+    }),
+  )
+
+  it.effect("includes the effective default agent system before durable context", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const agent = yield* AgentV2.Service
+      yield* agent.transform((editor) =>
+        editor.update(AgentV2.ID.make("build"), (agent) => {
+          agent.system = "Build agent instructions"
+          agent.mode = "primary"
+        }),
+      )
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = fragmentFixture("text", "text-build", ["Done"]).completeEvents
+      yield* session.resume(sessionID)
+
+      expect(requests.at(-1)?.system.map((part) => part.text)).toEqual(["Build agent instructions", "Initial context"])
+    }),
+  )
+
+  it.effect("uses the configured default agent system for omitted-agent sessions", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const agent = yield* AgentV2.Service
+      yield* agent.transform((editor) => {
+        editor.update(AgentV2.ID.make("build"), (agent) => {
+          agent.system = "Build agent instructions"
+          agent.mode = "primary"
+        })
+        editor.update(AgentV2.ID.make("reviewer"), (agent) => {
+          agent.system = "Reviewer instructions"
+          agent.mode = "primary"
+        })
+        editor.default(AgentV2.ID.make("reviewer"))
+      })
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = fragmentFixture("text", "text-reviewer", ["Done"]).completeEvents
+      yield* session.resume(sessionID)
+
+      expect(requests.at(-1)?.system.map((part) => part.text)).toEqual(["Reviewer instructions", "Initial context"])
+      expect((yield* session.messages({ sessionID }))[0]).toMatchObject({ type: "assistant", agent: "reviewer" })
+    }),
+  )
+
+  it.effect("uses an explicitly selected non-build agent system", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const { db } = yield* Database.Service
+      const agent = yield* AgentV2.Service
+      yield* agent.transform((editor) =>
+        editor.update(AgentV2.ID.make("reviewer"), (agent) => {
+          agent.system = "Reviewer instructions"
+          agent.mode = "primary"
+        }),
+      )
+      yield* db
+        .update(SessionTable)
+        .set({ agent: "reviewer" })
+        .where(eq(SessionTable.id, sessionID))
+        .run()
+        .pipe(Effect.orDie)
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = fragmentFixture("text", "text-selected", ["Done"]).completeEvents
+      yield* session.resume(sessionID)
+
+      expect(requests.at(-1)?.system.map((part) => part.text)).toEqual(["Reviewer instructions", "Initial context"])
+      expect((yield* session.messages({ sessionID }))[0]).toMatchObject({ type: "assistant", agent: "reviewer" })
+    }),
+  )
+
+  it.effect("updates selected-agent skill guidance after an agent switch", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      skillBaselines.set(AgentV2.ID.make("build"), "Build skills")
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+      skillBaselines.set(AgentV2.ID.make("reviewer"), "Reviewer skills")
+      yield* events.publish(SessionEvent.AgentSwitched, {
+        sessionID,
+        messageID: SessionMessage.ID.create(),
+        timestamp: DateTime.makeUnsafe(1),
+        agent: "reviewer",
+      })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Second" }), resume: false })
+      yield* session.resume(sessionID)
+
+      expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
+        ["Initial context\n\nBuild skills"],
+        ["Initial context\n\nBuild skills"],
+      ])
+      expect(systemTexts(requests[1]!)).toContainEqual(expect.stringContaining("Reviewer skills"))
+    }),
+  )
+
+  it.effect("keeps the sampled agent when selection changes during observation", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      skillBaselines.set(AgentV2.ID.make("build"), "Build skills")
+      skillBaselines.set(AgentV2.ID.make("reviewer"), "Reviewer skills")
+      let switched = false
+      systemLoadHook = Effect.suspend(() => {
+        if (switched) return Effect.void
+        switched = true
+        return events
+          .publish(SessionEvent.AgentSwitched, {
+            sessionID,
+            messageID: SessionMessage.ID.create(),
+            timestamp: DateTime.makeUnsafe(1),
+            agent: "reviewer",
+          })
+          .pipe(Effect.asVoid)
+      })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+
+      expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
+        ["Initial context\n\nBuild skills"],
+      ])
+    }),
+  )
+
+  it.effect("keeps the sampled model when selection changes during model resolution", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      let switched = false
+      modelResolveHook = Effect.suspend(() => {
+        if (switched) return Effect.void
+        switched = true
+        return events
+          .publish(SessionEvent.ModelSwitched, {
+            sessionID,
+            messageID: SessionMessage.ID.create(),
+            timestamp: DateTime.makeUnsafe(1),
+            model: { id: ModelV2.ID.make("replacement"), providerID: ProviderV2.ID.make("fake") },
+          })
+          .pipe(Effect.asVoid)
+      })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+      expect(requests.map((request) => request.model)).toEqual([model])
+      expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([["Initial context"]])
+    }),
+  )
+
+  it.effect("admits removed context as a chronological System message", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+      systemRemoved = true
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Second" }), resume: false })
+      yield* session.resume(sessionID)
+
+      expect(requests[1]?.messages.map((message) => message.role)).toEqual(["user", "user", "system"])
+      expect(requests[1]?.messages.at(-1)?.content).toEqual([
+        { type: "text", text: "System context source removed: test/context" },
+      ])
+      expect(yield* session.messages({ sessionID })).toHaveLength(3)
+    }),
+  )
+
+  it.effect("keeps the baseline and chronological System updates after a model switch", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+      systemBaseline = "Changed context"
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Second" }), resume: false })
+      yield* session.resume(sessionID)
+      yield* events.publish(SessionEvent.ModelSwitched, {
+        sessionID,
+        messageID: SessionMessage.ID.create(),
+        timestamp: DateTime.makeUnsafe(1),
+        model: { id: ModelV2.ID.make("replacement"), providerID: ProviderV2.ID.make("fake") },
+      })
+      systemBaseline = "Replacement context"
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Third" }), resume: false })
+      yield* session.resume(sessionID)
+
+      expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
+        ["Initial context"],
+        ["Initial context"],
+        ["Initial context"],
+      ])
+      expect(requests[1]?.messages.map((message) => message.role)).toEqual(["user", "user", "system"])
+      expect(requests[2]?.messages.filter((message) => message.role === "system")).toHaveLength(2)
+      expect((yield* session.context(sessionID)).map((message) => message.type)).toEqual([
+        "user",
+        "user",
+        "system",
+        "model-switched",
+        "user",
+        "system",
+      ])
+      yield* replaySessionProjection(sessionID)
+      expect(yield* session.messages({ sessionID })).toHaveLength(6)
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Fourth" }), resume: false })
+      yield* session.resume(sessionID)
+    }),
+  )
+
+  it.effect("preserves the baseline while context is temporarily unavailable", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+      yield* events.publish(SessionEvent.ModelSwitched, {
+        sessionID,
+        messageID: SessionMessage.ID.create(),
+        timestamp: DateTime.makeUnsafe(1),
+        model: { id: ModelV2.ID.make("replacement"), providerID: ProviderV2.ID.make("fake") },
+      })
+      systemUnavailable = true
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Second" }), resume: false })
+      yield* session.resume(sessionID)
+      systemUnavailable = false
+      systemBaseline = "Replacement context"
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Third" }), resume: false })
+      yield* session.resume(sessionID)
+
+      expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
+        ["Initial context"],
+        ["Initial context"],
+        ["Initial context"],
+      ])
+    }),
+  )
+
+  it.effect("rebuilds the baseline directly after completed compaction", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+      const compactionID = SessionMessage.ID.create()
+      yield* events.publish(SessionEvent.Compaction.Started, {
+        sessionID,
+        messageID: compactionID,
+        timestamp: DateTime.makeUnsafe(1),
+        reason: "manual",
+      })
+      yield* events.publish(SessionEvent.Compaction.Ended, {
+        sessionID,
+        messageID: compactionID,
+        timestamp: DateTime.makeUnsafe(2),
+        reason: "manual",
+        text: "summary",
+        recent: "",
+      })
+      systemBaseline = "Replacement context"
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Second" }), resume: false })
+      yield* session.resume(sessionID)
+
+      expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
+        ["Initial context"],
+        ["Replacement context"],
+      ])
+      yield* replaySessionProjection(sessionID)
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Third" }), resume: false })
+      yield* session.resume(sessionID)
+    }),
+  )
+
+  it.effect("automatically compacts into a completed summary and retained recent turn", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      response = fragmentFixture("text", "text-first", ["Earlier answer"]).completeEvents
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Earlier question ".repeat(180) }),
+        resume: false,
+      })
+      yield* session.resume(sessionID)
+
+      currentModel = compactModel
+      requests.length = 0
+      responses = [
+        fragmentFixture("text", "text-summary", ["## Objective\n- Preserve the task"]).completeEvents,
+        fragmentFixture("text", "text-final", ["Continued"]).completeEvents,
+      ]
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Recent exact request ".repeat(180) }),
+        resume: false,
+      })
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      expect(userTexts(requests[0])[0]).toContain("## Objective")
+      expect(userTexts(requests[1])).toHaveLength(1)
+      expect(userTexts(requests[1])[0]).toContain("<summary>\n## Objective\n- Preserve the task\n</summary>")
+      expect(userTexts(requests[1])[0]).toContain(`[User]: ${"Recent exact request ".repeat(180)}`)
+
+      const context = yield* (yield* SessionStore.Service).context(sessionID)
+      expect(context.map((message) => message.type)).toEqual(["compaction", "assistant"])
+      expect(context[0]).toMatchObject({
+        type: "compaction",
+        summary: "## Objective\n- Preserve the task",
+      })
+
+      requests.length = 0
+      executions.length = 0
+      responses = [
+        fragmentFixture("text", "text-summary-2", ["## Objective\n- Preserve the updated task"]).completeEvents,
+        fragmentFixture("text", "text-final-2", ["Continued again"]).completeEvents,
+      ]
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Newest exact request ".repeat(180) }),
+        resume: false,
+      })
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      expect(userTexts(requests[0])[0]).toContain(
+        "<previous-summary>\n## Objective\n- Preserve the task\n</previous-summary>",
+      )
+      expect(userTexts(requests[0])[0]).toContain("Recent exact request")
+      expect((yield* (yield* SessionStore.Service).context(sessionID))[0]).toMatchObject({
+        type: "compaction",
+        summary: "## Objective\n- Preserve the updated task",
+      })
+    }),
+  )
+
+  it.effect("forces one compaction and retries after provider context overflow", () =>
+    Effect.gen(function* () {
+      const session = yield* setupOverflowRecovery
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" }),
+        ],
+        fragmentFixture("text", "text-summary", ["## Objective\n- Recover overflow"]).completeEvents,
+        fragmentFixture("text", "text-final", ["Recovered"]).completeEvents,
+      ]
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Continue" }), resume: false })
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(3)
+      expect(userTexts(requests[1])[0]).toContain("## Objective")
+      expect(userTexts(requests[2])[0]).toContain("<summary>\n## Objective\n- Recover overflow\n</summary>")
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "compaction", summary: "## Objective\n- Recover overflow" },
+        { type: "assistant", finish: "stop" },
+      ])
+      yield* replaySessionProjection(sessionID)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "compaction" },
+        { type: "assistant", finish: "stop" },
+      ])
+    }),
+  )
+
+  it.effect("persists a second context overflow after one recovery", () =>
+    Effect.gen(function* () {
+      const session = yield* setupOverflowRecovery
+      const overflow = () => [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" }),
+      ]
+      responses = [
+        overflow(),
+        fragmentFixture("text", "text-summary", ["## Objective\n- Recover once"]).completeEvents,
+        overflow(),
+      ]
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Continue" }), resume: false })
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(3)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "compaction" },
+        { type: "assistant", finish: "error", error: { message: "prompt too long" } },
+      ])
+    }),
+  )
+
+  it.effect("recovers once from a raw context overflow failure", () =>
+    Effect.gen(function* () {
+      const session = yield* setupOverflowRecovery
+      responseStream = Stream.fail(
+        new LLMError({
+          module: "test",
+          method: "stream",
+          reason: new InvalidRequestReason({
+            message: "prompt too long",
+            classification: "context-overflow",
+          }),
+        }),
+      )
+      responses = [
+        fragmentFixture("text", "text-summary", ["## Objective\n- Recover raw overflow"]).completeEvents,
+        fragmentFixture("text", "text-final", ["Recovered"]).completeEvents,
+      ]
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Continue" }), resume: false })
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(3)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "compaction", summary: "## Objective\n- Recover raw overflow" },
+        { type: "assistant", finish: "stop" },
+      ])
+    }),
+  )
+
+  it.effect("publishes the original overflow when recovery summarization fails", () =>
+    Effect.gen(function* () {
+      const session = yield* setupOverflowRecovery
+      responses = [
+        [LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" })],
+        [LLMEvent.providerError({ message: "summary unavailable" })],
+      ]
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Continue" }), resume: false })
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      const context = yield* session.context(sessionID)
+      expect(context.some((message) => message.type === "compaction")).toBe(false)
+      expect(context.slice(-2)).toMatchObject([
+        { type: "user", text: "Continue" },
+        { type: "assistant", finish: "error", error: { message: "prompt too long" } },
+      ])
+    }),
+  )
+
+  it.effect("interrupts overflow recovery while the summary provider is running", () =>
+    Effect.gen(function* () {
+      const session = yield* setupOverflowRecovery
+      responses = [
+        [LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" })],
+        fragmentFixture("text", "text-summary", ["## Objective\n- Interrupted"]).completeEvents,
+      ]
+      const firstGate = yield* Deferred.make<void>()
+      const summaryGate = yield* Deferred.make<void>()
+      streamGate = firstGate
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Continue" }), resume: false })
+      const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      while (requests.length < 1) yield* Effect.yieldNow
+      streamGate = summaryGate
+      yield* Deferred.succeed(firstGate, undefined)
+      while (requests.length < 2) yield* Effect.yieldNow
+
+      yield* session.interrupt(sessionID)
+      expect(yield* Fiber.await(run)).toMatchObject({ _tag: "Failure" })
+      streamGate = undefined
+      expect(requests).toHaveLength(2)
+      expect((yield* session.context(sessionID)).some((message) => message.type === "compaction")).toBe(false)
+    }),
+  )
+
+  it.effect("preserves effective System updates while compaction rebaseline is blocked", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First" }), resume: false })
+
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+      systemBaseline = "Changed context"
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Second" }), resume: false })
+      yield* session.resume(sessionID)
+      const compactionID = SessionMessage.ID.create()
+      yield* events.publish(SessionEvent.Compaction.Started, {
+        sessionID,
+        messageID: compactionID,
+        timestamp: DateTime.makeUnsafe(1),
+        reason: "manual",
+      })
+      yield* events.publish(SessionEvent.Compaction.Ended, {
+        sessionID,
+        messageID: compactionID,
+        timestamp: DateTime.makeUnsafe(2),
+        reason: "manual",
+        text: "summary",
+        recent: "",
+      })
+      systemUnavailable = true
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Third" }), resume: false })
+      yield* session.resume(sessionID)
+
+      expect(requests.at(-1)?.system.map((part) => part.text)).toEqual(["Initial context"])
+      expect(systemTexts(requests.at(-1)!)).toContain("Changed context")
+    }),
+  )
+
+  it.effect("projects reasoning and tool events without executing or continuing tools", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Use tools" }), resume: false })
+
+      requests.length = 0
+      responses = undefined
+      streamGate = undefined
+      streamStarted = undefined
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.reasoningStart({ id: "reasoning-1" }),
+        LLMEvent.reasoningDelta({ id: "reasoning-1", text: "Think" }),
+        LLMEvent.reasoningEnd({ id: "reasoning-1" }),
+        LLMEvent.toolInputStart({ id: "call-error", name: "write" }),
+        LLMEvent.toolInputDelta({ id: "call-error", name: "write", text: '{"path":"README.md"}' }),
+        LLMEvent.toolInputEnd({ id: "call-error", name: "write" }),
+        LLMEvent.toolCall({ id: "call-error", name: "write", input: { path: "README.md" }, providerExecuted: true }),
+        LLMEvent.toolError({ id: "call-error", name: "write", message: "Denied" }),
+        LLMEvent.toolResult({ id: "call-error", name: "write", result: { type: "error", value: "Denied" } }),
+        LLMEvent.toolCall({
+          id: "call-provider",
+          name: "web_search",
+          input: { query: "hello" },
+          providerExecuted: true,
+          providerMetadata: { fake: { source: "provider" } },
+        }),
+        LLMEvent.toolResult({
+          id: "call-provider",
+          name: "web_search",
+          result: {
+            type: "content",
+            value: [
+              { type: "text", text: "Hello" },
+              { type: "file", uri: "data:image/png;base64,aGVsbG8=", mime: "image/png", name: "hello.png" },
+            ],
+          },
+          providerExecuted: true,
+          providerMetadata: { fake: { source: "provider" } },
+        }),
+        LLMEvent.stepFinish({
+          index: 0,
+          reason: "tool-calls",
+          usage: {
+            inputTokens: 10,
+            nonCachedInputTokens: 8,
+            outputTokens: 4,
+            reasoningTokens: 1,
+            cacheReadInputTokens: 2,
+          },
+        }),
+        LLMEvent.finish({ reason: "tool-calls" }),
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(requests[0]?.tools.map((tool) => tool.name)).toEqual(["echo", "defect"])
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Use tools" },
+        {
+          type: "assistant",
+          finish: "tool-calls",
+          tokens: { input: 8, output: 3, reasoning: 1, cache: { read: 2, write: 0 } },
+          content: [
+            { type: "reasoning", id: "reasoning-1", text: "Think" },
+            {
+              type: "tool",
+              id: "call-error",
+              name: "write",
+              state: {
+                status: "error",
+                input: { path: "README.md" },
+                error: { type: "unknown", message: "Denied" },
+              },
+            },
+            {
+              type: "tool",
+              id: "call-provider",
+              name: "web_search",
+              provider: { executed: true, metadata: { fake: { source: "provider" } } },
+              state: {
+                status: "completed",
+                input: { query: "hello" },
+                structured: {},
+                content: [
+                  { type: "text", text: "Hello" },
+                  { type: "file", mime: "image/png", uri: "data:image/png;base64,aGVsbG8=", name: "hello.png" },
+                ],
+              },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("continues with reloaded history after durably settling one local tool call", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Echo this" }), resume: false })
+
+      requests.length = 0
+      authorizations.length = 0
+      executions.length = 0
+      streamGate = undefined
+      streamStarted = undefined
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-echo", name: "echo", input: { text: "hello" } }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.textStart({ id: "text-final" }),
+          LLMEvent.textDelta({ id: "text-final", text: "Done" }),
+          LLMEvent.textEnd({ id: "text-final" }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      expect(requests[1]?.messages.map((message) => message.role)).toEqual(["user", "assistant", "tool"])
+      expect(authorizations).toMatchObject([{ sessionID, toolCallID: "call-echo" }])
+      expect(executions).toEqual(["hello"])
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Echo this" },
+        {
+          type: "assistant",
+          finish: "tool-calls",
+          content: [
+            {
+              type: "tool",
+              id: "call-echo",
+              name: "echo",
+              state: {
+                status: "completed",
+                input: { text: "hello" },
+                structured: { text: "hello" },
+                content: [{ type: "text", text: "hello" }],
+              },
+            },
+          ],
+        },
+        { type: "assistant", finish: "stop", content: [{ type: "text", id: "text-final", text: "Done" }] },
+      ])
+    }),
+  )
+
+  it.effect("reloads a model switch before a tool-driven continuation turn", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Echo this" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-echo", name: "echo", input: { text: "hello" } }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+      toolExecutionGate = yield* Deferred.make<void>()
+      toolExecutionsStarted = yield* Deferred.make<void>()
+      toolExecutionsReady = 1
+      const run = yield* Effect.forkChild(session.resume(sessionID))
+      yield* Deferred.await(toolExecutionsStarted)
+      yield* events.publish(SessionEvent.ModelSwitched, {
+        sessionID,
+        messageID: SessionMessage.ID.create(),
+        timestamp: DateTime.makeUnsafe(1),
+        model: { id: ModelV2.ID.make("replacement"), providerID: ProviderV2.ID.make("fake") },
+      })
+      systemBaseline = "Replacement context"
+      yield* Deferred.succeed(toolExecutionGate, undefined)
+      yield* Fiber.join(run)
+
+      expect(requests.map((request) => request.model)).toEqual([model, replacementModel])
+      expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
+        ["Initial context"],
+        ["Initial context"],
+      ])
+      expect(systemTexts(requests[1]!)).toContain("Replacement context")
+    }),
+  )
+
+  it.effect("restores durable reasoning provider metadata in a second-turn request", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Think first" }), resume: false })
+
+      requests.length = 0
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.reasoningStart({ id: "reasoning-anthropic" }),
+        LLMEvent.reasoningDelta({ id: "reasoning-anthropic", text: "Signed thought" }),
+        LLMEvent.reasoningEnd({ id: "reasoning-anthropic", providerMetadata: { anthropic: { signature: "sig_1" } } }),
+        LLMEvent.reasoningStart({
+          id: "reasoning-openai",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: null } },
+        }),
+        LLMEvent.reasoningDelta({ id: "reasoning-openai", text: "Encrypted thought" }),
+        LLMEvent.reasoningEnd({
+          id: "reasoning-openai",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
+        }),
+        LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+        LLMEvent.finish({ reason: "stop" }),
+      ]
+      yield* session.resume(sessionID)
+      yield* replaySessionProjection(sessionID)
+
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Think first" },
+        {
+          type: "assistant",
+          content: [
+            { type: "reasoning", text: "Signed thought", providerMetadata: { anthropic: { signature: "sig_1" } } },
+            {
+              type: "reasoning",
+              text: "Encrypted thought",
+              providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
+            },
+          ],
+        },
+      ])
+
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Continue" }), resume: false })
+      response = []
+      yield* session.resume(sessionID)
+
+      expect(requests[1]?.messages[1]?.content).toEqual([
+        { type: "reasoning", text: "Signed thought", providerMetadata: { anthropic: { signature: "sig_1" } } },
+        {
+          type: "reasoning",
+          text: "Encrypted thought",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
+        },
+      ])
+    }),
+  )
+
+  it.effect("replays durable provider-executed tool results inline in a second-turn request", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Search first" }), resume: false })
+
+      requests.length = 0
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolCall({
+          id: "hosted-search",
+          name: "web_search",
+          input: { query: "Effect" },
+          providerExecuted: true,
+          providerMetadata: { openai: { itemId: "hosted-search" } },
+        }),
+        LLMEvent.toolResult({
+          id: "hosted-search",
+          name: "web_search",
+          result: { type: "json", value: [{ title: "Effect" }] },
+          providerExecuted: true,
+          providerMetadata: { anthropic: { blockType: "web_search_tool_result" } },
+        }),
+        LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+        LLMEvent.finish({ reason: "stop" }),
+      ]
+      yield* session.resume(sessionID)
+      yield* replaySessionProjection(sessionID)
+
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Continue" }), resume: false })
+      response = []
+      yield* session.resume(sessionID)
+
+      expect(requests[1]?.messages.map((message) => message.role)).toEqual(["user", "assistant", "user"])
+      expect(requests[1]?.messages[1]?.content).toMatchObject([
+        {
+          type: "tool-call",
+          id: "hosted-search",
+          name: "web_search",
+          input: { query: "Effect" },
+          providerExecuted: true,
+          providerMetadata: { openai: { itemId: "hosted-search" } },
+        },
+        {
+          type: "tool-result",
+          id: "hosted-search",
+          name: "web_search",
+          result: { type: "json", value: [{ title: "Effect" }] },
+          providerExecuted: true,
+          providerMetadata: { anthropic: { blockType: "web_search_tool_result" } },
+        },
+      ])
+    }),
+  )
+
+  it.effect("starts recorded local tools eagerly and awaits settlement before continuing", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Echo five times" }), resume: false })
+
+      requests.length = 0
+      executions.length = 0
+      toolExecutionGate = yield* Deferred.make<void>()
+      toolExecutionsStarted = yield* Deferred.make<void>()
+      const providerGate = yield* Deferred.make<void>()
+      response = []
+      responses = undefined
+      const initial = Stream.fromIterable([
+        LLMEvent.stepStart({ index: 0 }),
+        ...Array.from({ length: 5 }, (_, index) =>
+          LLMEvent.toolCall({ id: `call-echo-${index}`, name: "echo", input: { text: `${index}` } }),
+        ),
+      ])
+      const final = Stream.fromIterable([
+        LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+        LLMEvent.finish({ reason: "tool-calls" }),
+      ])
+      streamGate = undefined
+      responseStream = Stream.concat(
+        initial,
+        Stream.fromEffect(Deferred.await(providerGate)).pipe(Stream.flatMap(() => final)),
+      )
+
+      const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(toolExecutionsStarted)
+
+      expect(executions).toHaveLength(5)
+      expect(maxActiveToolExecutions).toBe(5)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Echo five times" },
+        {
+          type: "assistant",
+          content: Array.from({ length: 5 }, (_, index) => ({
+            type: "tool",
+            id: `call-echo-${index}`,
+            state: { status: "running", input: { text: `${index}` } },
+          })),
+        },
+      ])
+
+      yield* Deferred.succeed(providerGate, undefined)
+      yield* Effect.yieldNow
+      expect(requests).toHaveLength(1)
+
+      yield* Deferred.succeed(toolExecutionGate, undefined)
+      yield* Fiber.join(run)
+      toolExecutionGate = undefined
+      toolExecutionsStarted = undefined
+
+      expect(executions).toHaveLength(5)
+      expect(maxActiveToolExecutions).toBe(5)
+      expect(requests).toHaveLength(2)
+    }),
+  )
+
+  it.effect("settles repeated provider-local tool call IDs against their owning assistant messages", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Echo twice" }), resume: false })
+
+      requests.length = 0
+      executions.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "tool_0", name: "echo", input: { text: "first" } }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "tool_0", name: "echo", input: { text: "second" } }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [],
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(executions).toEqual(["first", "second"])
+      expect(requests).toHaveLength(3)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Echo twice" },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "tool_0",
+              state: { status: "completed", structured: { text: "first" }, content: [{ type: "text", text: "first" }] },
+            },
+          ],
+        },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "tool_0",
+              state: {
+                status: "completed",
+                structured: { text: "second" },
+                content: [{ type: "text", text: "second" }],
+              },
+            },
+          ],
+        },
+      ])
+
+      yield* replaySessionProjection(sessionID)
+
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Echo twice" },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "tool_0",
+              state: { status: "completed", structured: { text: "first" }, content: [{ type: "text", text: "first" }] },
+            },
+          ],
+        },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "tool_0",
+              state: {
+                status: "completed",
+                structured: { text: "second" },
+                content: [{ type: "text", text: "second" }],
+              },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("joins concurrent resume calls into one active provider run", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Run once" }), resume: false })
+
+      requests.length = 0
+      responses = undefined
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.textStart({ id: "text-once" }),
+        LLMEvent.textDelta({ id: "text-once", text: "Once" }),
+        LLMEvent.textEnd({ id: "text-once" }),
+        LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+        LLMEvent.finish({ reason: "stop" }),
+      ]
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      const second = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Effect.yieldNow
+
+      expect(requests).toHaveLength(1)
+      yield* Deferred.succeed(streamGate, undefined)
+      yield* Fiber.join(first)
+      yield* Fiber.join(second)
+      streamGate = undefined
+      streamStarted = undefined
+
+      expect(requests).toHaveLength(1)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Run once" },
+        { type: "assistant", finish: "stop", content: [{ type: "text", id: "text-once", text: "Once" }] },
+      ])
+    }),
+  )
+
+  it.effect("steers an active provider turn with newly recorded prompts", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Start working" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Change direction" }) })
+      yield* Deferred.succeed(streamGate, undefined)
+      yield* Fiber.join(first)
+      streamGate = undefined
+      streamStarted = undefined
+      yield* Effect.yieldNow
+
+      expect(requests).toHaveLength(2)
+      expect(userTexts(requests[0]!)).toEqual(["Start working"])
+      expect(userTexts(requests[1]!)).toEqual(["Start working", "Change direction"])
+      expect((yield* session.context(sessionID)).map((message) => message.type)).toEqual([
+        "user",
+        "assistant",
+        "user",
+        "assistant",
+      ])
+    }),
+  )
+
+  it.effect("promotes queued input after continuation ends", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Start working" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-echo", name: "echo", input: { text: "hello" } }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Wait until continuation ends" }),
+        delivery: "queue",
+      })
+      yield* Deferred.succeed(streamGate, undefined)
+      yield* Fiber.join(first)
+      streamGate = undefined
+      streamStarted = undefined
+
+      expect(requests).toHaveLength(3)
+      expect(userTexts(requests[0]!)).toEqual(["Start working"])
+      expect(userTexts(requests[1]!)).toEqual(["Start working"])
+      expect(userTexts(requests[2]!)).toEqual(["Start working", "Wait until continuation ends"])
+    }),
+  )
+
+  it.effect("preserves durable queued input for a later wake after interruption", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const { db } = yield* Database.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Interrupt current work" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Run after interrupt" }),
+        delivery: "queue",
+      })
+      yield* session.interrupt(sessionID)
+      expect(yield* Fiber.await(run)).toMatchObject({ _tag: "Failure" })
+      expect(requests).toHaveLength(1)
+      expect(yield* SessionInput.hasPending(db, sessionID, "queue")).toBe(true)
+      const resumed = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      while (requests.length < 2) yield* Effect.yieldNow
+      yield* Deferred.succeed(streamGate, undefined)
+      yield* Fiber.join(resumed)
+      streamGate = undefined
+      streamStarted = undefined
+
+      expect(requests).toHaveLength(2)
+      expect(userTexts(requests[0]!)).toEqual(["Interrupt current work"])
+      expect(userTexts(requests[1]!)).toEqual(["Interrupt current work", "Run after interrupt"])
+    }),
+  )
+
+  it.effect("preserves durable steering input for a later resume after interruption", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const { db } = yield* Database.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Interrupt current work" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Steer after interrupt" }),
+      })
+      yield* session.interrupt(sessionID)
+      expect(yield* Fiber.await(run)).toMatchObject({ _tag: "Failure" })
+      expect(requests).toHaveLength(1)
+      expect(yield* SessionInput.hasPending(db, sessionID, "steer")).toBe(true)
+
+      const resumed = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      while (requests.length < 2) yield* Effect.yieldNow
+      yield* Deferred.succeed(streamGate, undefined)
+      yield* Fiber.join(resumed)
+      streamGate = undefined
+      streamStarted = undefined
+
+      expect(requests).toHaveLength(2)
+      expect(userTexts(requests[0]!)).toEqual(["Interrupt current work"])
+      expect(userTexts(requests[1]!)).toEqual(["Interrupt current work", "Steer after interrupt"])
+    }),
+  )
+
+  it.effect("promotes queued inputs one at a time in FIFO order", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Start working" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Queue first" }), delivery: "queue" })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Queue second" }), delivery: "queue" })
+      yield* Deferred.succeed(streamGate, undefined)
+      yield* Fiber.join(first)
+      streamGate = undefined
+      streamStarted = undefined
+
+      expect(requests).toHaveLength(3)
+      expect(userTexts(requests[0]!)).toEqual(["Start working"])
+      expect(userTexts(requests[1]!)).toEqual(["Start working", "Queue first"])
+      expect(userTexts(requests[2]!)).toEqual(["Start working", "Queue first", "Queue second"])
+    }),
+  )
+
+  it.effect("promotes queued input after steering continuation ends", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Start steering" }), resume: false })
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Queue for later" }),
+        delivery: "queue",
+        resume: false,
+      })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      expect(userTexts(requests[0]!)).toEqual(["Start steering"])
+      expect(userTexts(requests[1]!)).toEqual(["Start steering", "Queue for later"])
+    }),
+  )
+
+  it.effect("promotes steers before the next queued input", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Start working" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+      const firstGate = yield* Deferred.make<void>()
+      const secondGate = yield* Deferred.make<void>()
+      streamGate = firstGate
+
+      const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      while (requests.length < 1) yield* Effect.yieldNow
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Queue first" }), delivery: "queue" })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Queue second" }), delivery: "queue" })
+      streamGate = secondGate
+      yield* Deferred.succeed(firstGate, undefined)
+      while (requests.length < 2) yield* Effect.yieldNow
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Steer before next queued input" }) })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Also steer before next queued input" }) })
+      yield* Deferred.succeed(secondGate, undefined)
+      yield* Fiber.join(first)
+      streamGate = undefined
+
+      expect(requests).toHaveLength(4)
+      expect(userTexts(requests[0]!)).toEqual(["Start working"])
+      expect(userTexts(requests[1]!)).toEqual(["Start working", "Queue first"])
+      expect(userTexts(requests[2]!)).toEqual([
+        "Start working",
+        "Queue first",
+        "Steer before next queued input",
+        "Also steer before next queued input",
+      ])
+      expect(userTexts(requests[3]!)).toEqual([
+        "Start working",
+        "Queue first",
+        "Steer before next queued input",
+        "Also steer before next queued input",
+        "Queue second",
+      ])
+    }),
+  )
+
+  it.effect("coalesces multiple active steering prompts into one continuation turn", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Start working" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "First steer" }) })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Second steer" }) })
+      yield* Deferred.succeed(streamGate, undefined)
+      yield* Fiber.join(first)
+      streamGate = undefined
+      streamStarted = undefined
+      yield* Effect.yieldNow
+
+      expect(requests).toHaveLength(2)
+      expect(userTexts(requests[1]!)).toEqual(["Start working", "First steer", "Second steer"])
+      yield* (yield* SessionExecution.Service).wake(sessionID)
+      yield* Effect.yieldNow
+      expect(requests).toHaveLength(2)
+    }),
+  )
+
+  it.effect("runs steering input accepted while the active provider turn fails", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Start working" }), resume: false })
+
+      requests.length = 0
+      responses = undefined
+      response = []
+      streamFailure = providerUnavailable()
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Recover with this" }) })
+      yield* Deferred.succeed(streamGate, undefined)
+      expect(yield* Fiber.join(first).pipe(Effect.flip)).toBe(streamFailure)
+
+      streamFailure = undefined
+      streamGate = undefined
+      streamStarted = undefined
+      yield* Effect.yieldNow
+
+      expect(requests).toHaveLength(2)
+      expect(userTexts(requests[1]!)).toEqual(["Start working", "Recover with this"])
+    }),
+  )
+
+  it.effect("durably fails local tools left running by a prior process before continuing", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Recover interrupted tool" }), resume: false })
+      yield* SessionInput.promoteSteers((yield* Database.Service).db, events, sessionID, Number.MAX_SAFE_INTEGER)
+      const assistantMessageID = SessionMessage.ID.create()
+      yield* events.publish(SessionEvent.Step.Started, {
+        sessionID,
+        assistantMessageID,
+        timestamp: yield* DateTime.now,
+        agent: "build",
+        model: { id: ModelV2.ID.make("fake-model"), providerID: ProviderV2.ID.make("fake") },
+      })
+      yield* events.publish(SessionEvent.Tool.Input.Started, {
+        sessionID,
+        timestamp: yield* DateTime.now,
+        assistantMessageID,
+        callID: "call-interrupted",
+        name: "echo",
+      })
+      yield* events.publish(SessionEvent.Tool.Input.Ended, {
+        sessionID,
+        timestamp: yield* DateTime.now,
+        assistantMessageID,
+        callID: "call-interrupted",
+        text: '{"text":"stale"}',
+      })
+      yield* events.publish(SessionEvent.Tool.Called, {
+        sessionID,
+        timestamp: yield* DateTime.now,
+        assistantMessageID,
+        callID: "call-interrupted",
+        tool: "echo",
+        input: { text: "stale" },
+        provider: { executed: false },
+      })
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(requests[0]?.messages.map((message) => message.role)).toEqual(["user", "assistant", "tool"])
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Recover interrupted tool" },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "call-interrupted",
+              state: { status: "error", error: { type: "unknown", message: "Tool execution interrupted" } },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("durably fails hosted tools left running by a prior process before continuing inline", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Recover interrupted hosted tool" }),
+        resume: false,
+      })
+      yield* SessionInput.promoteSteers((yield* Database.Service).db, events, sessionID, Number.MAX_SAFE_INTEGER)
+      const assistantMessageID = SessionMessage.ID.create()
+      yield* events.publish(SessionEvent.Step.Started, {
+        sessionID,
+        assistantMessageID,
+        timestamp: yield* DateTime.now,
+        agent: "build",
+        model: { id: ModelV2.ID.make("fake-model"), providerID: ProviderV2.ID.make("fake") },
+      })
+      yield* events.publish(SessionEvent.Tool.Input.Started, {
+        sessionID,
+        timestamp: yield* DateTime.now,
+        assistantMessageID,
+        callID: "call-hosted-interrupted",
+        name: "web_search",
+      })
+      yield* events.publish(SessionEvent.Tool.Input.Ended, {
+        sessionID,
+        timestamp: yield* DateTime.now,
+        assistantMessageID,
+        callID: "call-hosted-interrupted",
+        text: '{"query":"stale"}',
+      })
+      yield* events.publish(SessionEvent.Tool.Called, {
+        sessionID,
+        timestamp: yield* DateTime.now,
+        assistantMessageID,
+        callID: "call-hosted-interrupted",
+        tool: "web_search",
+        input: { query: "stale" },
+        provider: { executed: true, metadata: { openai: { itemId: "call-hosted-interrupted" } } },
+      })
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(requests[0]?.messages.map((message) => message.role)).toEqual(["user", "assistant"])
+      expect(requests[0]?.messages[1]?.content).toMatchObject([
+        {
+          type: "tool-call",
+          id: "call-hosted-interrupted",
+          providerExecuted: true,
+          providerMetadata: { openai: { itemId: "call-hosted-interrupted" } },
+        },
+        { type: "tool-result", id: "call-hosted-interrupted", providerExecuted: true, result: { type: "error" } },
+      ])
+    }),
+  )
+
+  it.effect("durably fails pending tool input left by a prior process before continuing", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Recover interrupted tool input" }),
+        resume: false,
+      })
+      yield* SessionInput.promoteSteers((yield* Database.Service).db, events, sessionID, Number.MAX_SAFE_INTEGER)
+      const assistantMessageID = SessionMessage.ID.create()
+      yield* events.publish(SessionEvent.Step.Started, {
+        sessionID,
+        assistantMessageID,
+        timestamp: yield* DateTime.now,
+        agent: "build",
+        model: { id: ModelV2.ID.make("fake-model"), providerID: ProviderV2.ID.make("fake") },
+      })
+      yield* events.publish(SessionEvent.Tool.Input.Started, {
+        sessionID,
+        timestamp: yield* DateTime.now,
+        assistantMessageID,
+        callID: "call-pending-interrupted",
+        name: "echo",
+      })
+      requests.length = 0
+      response = []
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(requests[0]?.messages.map((message) => message.role)).toEqual(["user", "assistant", "tool"])
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Recover interrupted tool input" },
+        { type: "assistant", content: [{ type: "tool", id: "call-pending-interrupted", state: { status: "error" } }] },
+      ])
+    }),
+  )
+
+  it.effect("promotes the first queued input when woken while idle", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Wait in queue" }),
+        delivery: "queue",
+        resume: false,
+      })
+
+      requests.length = 0
+      yield* (yield* SessionExecution.Service).wake(sessionID)
+      yield* Effect.yieldNow
+
+      expect(requests).toHaveLength(1)
+      expect(userTexts(requests[0]!)).toEqual(["Wait in queue"])
+    }),
+  )
+
+  it.effect("retries inbox input after prompt projection rolls back", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      const defect = new Error("fail after prompt promotion")
+      let fail = true
+      yield* events.project(SessionEvent.Prompted, () => (fail ? Effect.die(defect) : Effect.void))
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Recover promoted input" }), resume: false })
+
+      expect(yield* session.resume(sessionID).pipe(Effect.catchDefect(Effect.succeed))).toBe(defect)
+      fail = false
+      requests.length = 0
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+        LLMEvent.finish({ reason: "stop" }),
+      ]
+
+      yield* (yield* SessionExecution.Service).wake(sessionID)
+      while (requests.length === 0) yield* Effect.yieldNow
+
+      expect(userTexts(requests[0]!)).toEqual(["Recover promoted input"])
+    }),
+  )
+
+  it.effect("does not strand a committed promotion when a post-commit listener defects", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const events = yield* EventV2.Service
+      yield* events.listen((event) =>
+        event.type === SessionEvent.Prompted.type ? Effect.die("fail after prompt promotion commits") : Effect.void,
+      )
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Run committed promotion" }),
+        resume: false,
+      })
+
+      requests.length = 0
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(userTexts(requests[0]!)).toEqual(["Run committed promotion"])
+    }),
+  )
+
+  it.effect("runs different sessions concurrently", () =>
+    Effect.gen(function* () {
+      yield* setup
+      yield* insertSession(otherSessionID)
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Run first" }), resume: false })
+      yield* session.prompt({ sessionID: otherSessionID, prompt: Prompt.make({ text: "Run second" }), resume: false })
+
+      requests.length = 0
+      responses = undefined
+      response = []
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      const second = yield* session.resume(otherSessionID).pipe(Effect.forkChild)
+      yield* Effect.yieldNow
+
+      expect(requests).toHaveLength(2)
+      expect(requests.map((request) => request.providerOptions?.openai?.promptCacheKey)).toEqual([
+        sessionID,
+        otherSessionID,
+      ])
+      yield* Deferred.succeed(streamGate, undefined)
+      yield* Fiber.join(first)
+      yield* Fiber.join(second)
+      streamGate = undefined
+      streamStarted = undefined
+    }),
+  )
+
+  it.effect("bounds 64-character session prompt cache keys", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const longSessionID = SessionV2.ID.make(`ses_${"a".repeat(64)}`)
+      const otherLongSessionID = SessionV2.ID.make(`ses_${"b".repeat(64)}`)
+      yield* insertSession(longSessionID)
+      yield* insertSession(otherLongSessionID)
+      const session = yield* SessionV2.Service
+      yield* session.prompt({
+        sessionID: longSessionID,
+        prompt: Prompt.make({ text: "Run long session" }),
+        resume: false,
+      })
+      yield* session.prompt({
+        sessionID: otherLongSessionID,
+        prompt: Prompt.make({ text: "Run other long session" }),
+        resume: false,
+      })
+
+      requests.length = 0
+      yield* session.resume(longSessionID)
+      yield* session.resume(otherLongSessionID)
+
+      const keys = requests.map((request) => request.providerOptions?.openai?.promptCacheKey)
+      expect(keys).toEqual([longSessionID.slice(4), otherLongSessionID.slice(4)])
+      expect(keys.every((key) => typeof key === "string" && key.length === 64)).toBe(true)
+      expect(keys[0]).not.toBe(keys[1])
+    }),
+  )
+
+  it.effect("fans out one failed run and allows a later retry", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Retry after failure" }), resume: false })
+
+      requests.length = 0
+      responses = undefined
+      response = []
+      streamFailure = providerUnavailable()
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      const second = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Effect.yieldNow
+
+      expect(requests).toHaveLength(1)
+      yield* Deferred.succeed(streamGate, undefined)
+      const [firstExit, secondExit] = yield* Effect.all([Fiber.await(first), Fiber.await(second)])
+      expect(secondExit).toEqual(firstExit)
+
+      streamFailure = undefined
+      streamGate = undefined
+      streamStarted = undefined
+      yield* session.resume(sessionID)
+      expect(requests).toHaveLength(2)
+    }),
+  )
+
+  it.effect("durably settles local tool failures before continuing", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Call missing" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-missing", name: "missing", input: {} }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.textStart({ id: "text-after-error" }),
+          LLMEvent.textDelta({ id: "text-after-error", text: "Recovered" }),
+          LLMEvent.textEnd({ id: "text-after-error" }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+      streamGate = undefined
+      streamStarted = undefined
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Call missing" },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "call-missing",
+              state: { status: "error", error: { message: "Unknown tool: missing" } },
+            },
+          ],
+        },
+        { type: "assistant", finish: "stop", content: [{ type: "text", id: "text-after-error", text: "Recovered" }] },
+      ])
+    }),
+  )
+
+  it.effect("returns unexpected local tool defects to the model and continues", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Call defect" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-defect", name: "defect", input: {} }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.textStart({ id: "text-after-defect" }),
+          LLMEvent.textDelta({ id: "text-after-defect", text: "Recovered" }),
+          LLMEvent.textEnd({ id: "text-after-defect" }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      expect(requests[1]?.messages.map((message) => message.role)).toEqual(["user", "assistant", "tool"])
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Call defect" },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "call-defect",
+              state: {
+                status: "error",
+                error: { type: "unknown", message: "Tool execution failed: unexpected tool defect" },
+              },
+            },
+          ],
+        },
+        { type: "assistant", finish: "stop", content: [{ type: "text", text: "Recovered" }] },
+      ])
+    }),
+  )
+
+  it.effect("returns policy-blocked tools to the model and continues", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const registry = yield* ToolRegistry.Service
+      yield* registry.register({
+        blocked: Tool.make({
+          description: "Fail because policy blocked execution",
+          input: Schema.Struct({}),
+          output: Schema.Struct({}),
+          execute: () =>
+            Effect.fail(new PermissionV2.BlockedError({ rules: [] })).pipe(
+              Effect.mapError(() => new Tool.Failure({ message: "Permission blocked" })),
+            ),
+        }),
+      })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Call blocked" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-blocked", name: "blocked", input: {} }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Call blocked" },
+        {
+          type: "assistant",
+          content: [
+            { type: "tool", id: "call-blocked", state: { status: "error", error: { message: "Permission blocked" } } },
+          ],
+        },
+        { type: "assistant", finish: "stop" },
+      ])
+    }),
+  )
+
+  it.effect("interrupts runner continuation when permission approval is declined", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const registry = yield* ToolRegistry.Service
+      yield* registry.register({
+        declined: Tool.make({
+          description: "Fail because the user declined approval",
+          input: Schema.Struct({}),
+          output: Schema.Struct({}),
+          execute: () => Effect.die(new PermissionV2.DeclinedError()),
+        }),
+      })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Call declined" }), resume: false })
+
+      requests.length = 0
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolCall({ id: "call-declined", name: "declined", input: {} }),
+        LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+        LLMEvent.finish({ reason: "tool-calls" }),
+      ]
+
+      const exit = yield* session.resume(sessionID).pipe(Effect.exit)
+
+      expect(exit._tag).toBe("Failure")
+      if (exit._tag === "Failure") expect(Cause.hasInterruptsOnly(exit.cause)).toBe(true)
+      expect(requests).toHaveLength(1)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Call declined" },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "call-declined",
+              state: { status: "error", error: { message: "Tool execution interrupted" } },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("returns permission corrections to the model and continues", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const registry = yield* ToolRegistry.Service
+      yield* registry.register({
+        corrected: Tool.make({
+          description: "Fail with user correction feedback",
+          input: Schema.Struct({}),
+          output: Schema.Struct({}),
+          execute: () =>
+            Effect.fail(new PermissionV2.CorrectedError({ feedback: "Use another tool" })).pipe(
+              Effect.mapError(() => new Tool.Failure({ message: "Use another tool" })),
+            ),
+        }),
+      })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Call corrected" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-corrected", name: "corrected", input: {} }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Call corrected" },
+        {
+          type: "assistant",
+          content: [
+            { type: "tool", id: "call-corrected", state: { status: "error", error: { message: "Use another tool" } } },
+          ],
+        },
+        { type: "assistant", finish: "stop" },
+      ])
+    }),
+  )
+
+  it.effect("interrupts runner continuation when a question is dismissed", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      const registry = yield* ToolRegistry.Service
+      const questions = yield* QuestionV2.Service
+      yield* registry.register({
+        question: Tool.make({
+          description: "Ask the user",
+          input: Schema.Struct({}),
+          output: Schema.Struct({}),
+          execute: (_, context) =>
+            questions.ask({ sessionID: context.sessionID, questions: [] }).pipe(Effect.as({}), Effect.orDie),
+        }),
+      })
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Ask then stop" }), resume: false })
+
+      requests.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-question", name: "question", input: {} }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [],
+      ]
+
+      const run = yield* session.resume(sessionID).pipe(Effect.exit, Effect.forkChild)
+      let pending = yield* questions.list()
+      while (pending.length === 0) {
+        yield* Effect.yieldNow
+        pending = yield* questions.list()
+      }
+      yield* questions.reject(pending[0]!.id)
+      const exit = yield* Fiber.join(run)
+
+      expect(exit._tag).toBe("Failure")
+      if (exit._tag === "Failure") expect(Cause.hasInterruptsOnly(exit.cause)).toBe(true)
+      expect(requests).toHaveLength(1)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Ask then stop" },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "call-question",
+              state: { status: "error", error: { type: "unknown", message: "Tool execution interrupted" } },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("awaits started local tools before surfacing provider stream failure", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Settle before failing" }), resume: false })
+      const failure = providerUnavailable()
+      toolExecutionGate = yield* Deferred.make<void>()
+      responseStream = Stream.concat(
+        Stream.fromIterable([
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-before-failure", name: "echo", input: { text: "settle" } }),
+        ]),
+        Stream.fail(failure),
+      )
+
+      const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      while (executions.length === 0) yield* Effect.yieldNow
+      yield* Effect.yieldNow
+      yield* Deferred.succeed(toolExecutionGate, undefined)
+      expect(yield* Fiber.join(run).pipe(Effect.flip)).toBe(failure)
+      toolExecutionGate = undefined
+
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Settle before failing" },
+        {
+          type: "assistant",
+          content: [
+            { type: "tool", id: "call-before-failure", state: { status: "completed", structured: { text: "settle" } } },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("durably fails blocked local tools when a provider turn is interrupted", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Interrupt blocked tool" }), resume: false })
+      executions.length = 0
+      toolExecutionGate = yield* Deferred.make<void>()
+      responseStream = Stream.concat(
+        Stream.fromIterable([
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-before-interrupt", name: "echo", input: { text: "blocked" } }),
+        ]),
+        Stream.never,
+      )
+
+      const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      while (executions.length === 0) yield* Effect.yieldNow
+      yield* session.interrupt(sessionID)
+      toolExecutionGate = undefined
+
+      expect(yield* Fiber.await(run)).toMatchObject({ _tag: "Failure" })
+      yield* session.interrupt(sessionID)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Interrupt blocked tool" },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "call-before-interrupt",
+              state: { status: "error", error: { type: "unknown", message: "Tool execution interrupted" } },
+            },
+          ],
+        },
+      ])
+
+      yield* replaySessionProjection(sessionID)
+
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Interrupt blocked tool" },
+        { type: "assistant", content: [{ type: "tool", id: "call-before-interrupt", state: { status: "error" } }] },
+      ])
+      requests.length = 0
+      responseStream = undefined
+      response = []
+      yield* session.resume(sessionID)
+      expect(requests[0]?.messages.map((message) => message.role)).toEqual(["user", "assistant", "tool"])
+    }),
+  )
+
+  it.effect("interrupts a blocked provider turn without local tool execution", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Interrupt provider" }), resume: false })
+      requests.length = 0
+      response = []
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      yield* session.interrupt(sessionID)
+      const exit = yield* Fiber.await(run)
+      streamGate = undefined
+      streamStarted = undefined
+
+      expect(Exit.isFailure(exit) && Cause.hasInterruptsOnly(exit.cause)).toBeTrue()
+      expect(requests).toHaveLength(1)
+      yield* session.interrupt(sessionID)
+    }),
+  )
+
+  it.effect("durably fails blocked local tools when interrupted while awaiting settlement", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Interrupt tool settlement" }), resume: false })
+      executions.length = 0
+      toolExecutionGate = yield* Deferred.make<void>()
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolCall({ id: "call-await-interrupt", name: "echo", input: { text: "blocked" } }),
+        LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+        LLMEvent.finish({ reason: "tool-calls" }),
+      ]
+
+      const runner = yield* SessionRunner.Service
+      const run = yield* runner.run({ sessionID, force: true }).pipe(Effect.forkChild)
+      while (executions.length === 0) yield* Effect.yieldNow
+      yield* Fiber.interrupt(run)
+      toolExecutionGate = undefined
+
+      expect(yield* Fiber.await(run)).toMatchObject({ _tag: "Failure" })
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Interrupt tool settlement" },
+        {
+          type: "assistant",
+          content: [
+            {
+              type: "tool",
+              id: "call-await-interrupt",
+              state: { status: "error", error: { type: "unknown", message: "Tool execution interrupted" } },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("forces a text response on an agent's configured final step", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const agents = yield* AgentV2.Service
+      yield* agents.transform((editor) =>
+        editor.update(AgentV2.ID.make("build"), (agent) => {
+          agent.steps = 2
+        }),
+      )
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Finish at the limit" }), resume: false })
+
+      requests.length = 0
+      executions.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-terminal", name: "echo", input: { text: "done" } }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-forbidden", name: "echo", input: { text: "forbidden" } }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      expect(requests[0]?.toolChoice).toBeUndefined()
+      expect(requests[1]?.toolChoice).toMatchObject({ type: "none" })
+      expect(requests[1]?.tools).toEqual([])
+      expect(requests[1]?.messages.at(-1)).toMatchObject({
+        role: "assistant",
+        content: [{ type: "text", text: expect.stringContaining("MAXIMUM STEPS REACHED") }],
+      })
+      expect(executions).toEqual(["done"])
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Finish at the limit" },
+        { type: "assistant", content: [{ type: "tool", id: "call-terminal", state: { status: "completed" } }] },
+        { type: "assistant", content: [{ type: "tool", id: "call-forbidden", state: { status: "error" } }] },
+      ])
+    }),
+  )
+
+  it.effect("resets the configured step allowance when steering input promotes", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const agents = yield* AgentV2.Service
+      yield* agents.transform((editor) =>
+        editor.update(AgentV2.ID.make("build"), (agent) => {
+          agent.steps = 2
+        }),
+      )
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Start work" }), resume: false })
+
+      requests.length = 0
+      executions.length = 0
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-before-steer", name: "echo", input: { text: "before" } }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-after-steer", name: "echo", input: { text: "after" } }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+          LLMEvent.finish({ reason: "stop" }),
+        ],
+      ]
+      streamGate = yield* Deferred.make<void>()
+      streamStarted = yield* Deferred.make<void>()
+
+      const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(streamStarted)
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Change direction" }) })
+      yield* Deferred.succeed(streamGate, undefined)
+      yield* Fiber.join(run)
+      streamGate = undefined
+      streamStarted = undefined
+
+      expect(requests).toHaveLength(3)
+      expect(requests[1]?.toolChoice).toBeUndefined()
+      expect(requests[1]?.tools).not.toEqual([])
+      expect(requests[2]?.toolChoice).toMatchObject({ type: "none" })
+      expect(executions).toEqual(["before", "after"])
+    }),
+  )
+
+  it.effect("projects provider errors as terminal assistant step failures", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Fail durably" }), resume: false })
+
+      requests.length = 0
+      responses = undefined
+      streamGate = undefined
+      streamStarted = undefined
+      response = [LLMEvent.stepStart({ index: 0 }), LLMEvent.providerError({ message: "Provider unavailable" })]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Fail durably" },
+        { type: "assistant", finish: "error", error: { type: "unknown", message: "Provider unavailable" } },
+      ])
+    }),
+  )
+
+  it.effect("projects provider errors emitted before assistant step start", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Fail before step" }), resume: false })
+
+      requests.length = 0
+      response = [LLMEvent.providerError({ message: "Provider unavailable" })]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Fail before step" },
+        { type: "assistant", finish: "error", error: { type: "unknown", message: "Provider unavailable" } },
+      ])
+    }),
+  )
+
+  it.effect("does not recover context overflow after durable assistant output", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Fail after output" }), resume: false })
+
+      requests.length = 0
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.textStart({ id: "text-partial" }),
+        LLMEvent.textDelta({ id: "text-partial", text: "Partial" }),
+        LLMEvent.textEnd({ id: "text-partial" }),
+        LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" }),
+      ]
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Fail after output" },
+        {
+          type: "assistant",
+          finish: "error",
+          error: { message: "prompt too long" },
+          content: [{ type: "text", text: "Partial" }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("projects raw provider stream failures as terminal assistant step failures", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Fail raw stream durably" }), resume: false })
+      const failure = providerUnavailable()
+      responseStream = Stream.fail(failure)
+
+      expect(yield* session.resume(sessionID).pipe(Effect.flip)).toBe(failure)
+      yield* replaySessionProjection(sessionID)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Fail raw stream durably" },
+        { type: "assistant", finish: "error", error: { type: "unknown", message: "Provider unavailable" } },
+      ])
+    }),
+  )
+
+  it.effect("does not continue automatically after a provider error follows a local tool call", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Do not continue failed provider" }),
+        resume: false,
+      })
+
+      requests.length = 0
+      const executionCount = executions.length
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolCall({ id: "call-before-provider-error", name: "echo", input: { text: "settled" } }),
+        LLMEvent.providerError({ message: "Provider unavailable" }),
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(executions.slice(executionCount)).toEqual(["settled"])
+    }),
+  )
+
+  it.effect("durably fails a hosted tool when its provider errors before returning a result", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Fail hosted tool durably" }), resume: false })
+
+      requests.length = 0
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolCall({
+          id: "call-hosted-provider-error",
+          name: "web_search",
+          input: { query: "effect" },
+          providerExecuted: true,
+        }),
+        LLMEvent.providerError({ message: "Provider unavailable" }),
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Fail hosted tool durably" },
+        {
+          type: "assistant",
+          content: [{ type: "tool", id: "call-hosted-provider-error", state: { status: "error" } }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("durably fails a hosted tool left unresolved at normal provider EOF", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Fail hosted tool at EOF" }), resume: false })
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolCall({
+          id: "call-hosted-eof",
+          name: "web_search",
+          input: { query: "effect" },
+          providerExecuted: true,
+        }),
+      ]
+
+      yield* session.resume(sessionID)
+      yield* replaySessionProjection(sessionID)
+
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Fail hosted tool at EOF" },
+        { type: "assistant", content: [{ type: "tool", id: "call-hosted-eof", state: { status: "error" } }] },
+      ])
+    }),
+  )
+
+  it.effect("durably fails a hosted tool left unresolved by a raw provider stream failure", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({
+        sessionID,
+        prompt: Prompt.make({ text: "Fail hosted tool on raw failure" }),
+        resume: false,
+      })
+      const failure = providerUnavailable()
+      responseStream = Stream.concat(
+        Stream.fromIterable([
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({
+            id: "call-hosted-raw-failure",
+            name: "web_search",
+            input: { query: "effect" },
+            providerExecuted: true,
+          }),
+        ]),
+        Stream.fail(failure),
+      )
+
+      expect(yield* session.resume(sessionID).pipe(Effect.flip)).toBe(failure)
+      yield* replaySessionProjection(sessionID)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Fail hosted tool on raw failure" },
+        {
+          type: "assistant",
+          finish: "error",
+          error: { type: "unknown", message: "Provider unavailable" },
+          content: [{ type: "tool", id: "call-hosted-raw-failure", state: { status: "error" } }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("keeps interleaved assistant text blocks separate", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Two blocks" }), resume: false })
+
+      responses = undefined
+      streamGate = undefined
+      streamStarted = undefined
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.textStart({ id: "text-1" }),
+        LLMEvent.textStart({ id: "text-2" }),
+        LLMEvent.textDelta({ id: "text-1", text: "First" }),
+        LLMEvent.textDelta({ id: "text-2", text: "Second" }),
+        LLMEvent.textEnd({ id: "text-1" }),
+        LLMEvent.textEnd({ id: "text-2" }),
+        LLMEvent.stepFinish({ index: 0, reason: "stop" }),
+        LLMEvent.finish({ reason: "stop" }),
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Two blocks" },
+        {
+          type: "assistant",
+          content: [
+            { type: "text", id: "text-1", text: "First" },
+            { type: "text", id: "text-2", text: "Second" },
+          ],
+        },
+      ])
+    }),
+  )
+
+  for (const kind of fragmentKinds) {
+    it.effect(`broadcasts provider ${kind} deltas without storing projection rewrites`, () =>
+      verifyEphemeralDeltas(kind),
+    )
+
+    it.effect(`durably closes partial ${kind} when the provider stream fails`, () => verifyPartialFlushOnFailure(kind))
+
+    it.effect(`durably closes partial ${kind} when the provider stream is interrupted`, () =>
+      verifyPartialFlushOnInterruption(kind),
+    )
+  }
+
+  it.effect("rejects duplicate streamed text starts", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      responses = undefined
+      streamGate = undefined
+      streamStarted = undefined
+      response = [LLMEvent.textStart({ id: "text-1" }), LLMEvent.textStart({ id: "text-1" })]
+
+      expect(yield* session.resume(sessionID).pipe(Effect.catchDefect(Effect.succeed))).toBe(
+        "Duplicate text start: text-1",
+      )
+    }),
+  )
+
+  it.effect("transitions streamed raw tool input to parsed called input", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "Call provider tool" }), resume: false })
+
+      responses = undefined
+      streamGate = undefined
+      streamStarted = undefined
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolInputStart({ id: "call-parsed", name: "web_search" }),
+        LLMEvent.toolInputDelta({ id: "call-parsed", name: "web_search", text: '{"query":"hello"}' }),
+        LLMEvent.toolInputEnd({ id: "call-parsed", name: "web_search" }),
+        LLMEvent.toolCall({ id: "call-parsed", name: "web_search", input: { query: "hello" }, providerExecuted: true }),
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Call provider tool" },
+        {
+          type: "assistant",
+          content: [{ type: "tool", id: "call-parsed", state: { status: "error", input: { query: "hello" } } }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("rejects malformed streamed tool input ordering", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      responses = undefined
+      streamGate = undefined
+      streamStarted = undefined
+      response = [LLMEvent.toolInputDelta({ id: "call-1", name: "read", text: "{}" })]
+
+      expect(yield* session.resume(sessionID).pipe(Effect.catchDefect(Effect.succeed))).toBe(
+        "Tool input delta before start: call-1",
+      )
+    }),
+  )
+})