refactor(ai): restore inline compaction beta handling

refactor(ai): centralize anthropic beta selection
refactor(ai): separate conversation and generation lowering
2026-09-01 06:26:24 +00:00 · 2026-08-31 20:01:57 +05:30 · 2026-08-31 19:52:15 +05:30 · 2026-08-31 19:50:39 +05:30 · 2026-08-31 19:30:04 +05:30 · 2026-08-31 19:29:23 +05:30
31 changed files with 2030 additions and 100 deletions
@@ -249,6 +249,107 @@ The published legacy `Service`, `layer`, `clientLayer`, and module-level control
 over the same implementation, including the legacy live `requests` array. New tests should use `Test` and
 `testLayer`.

+## Provider compaction
+
+Compaction is opt-in. The package supports automatic compaction in OpenAI/Azure Responses and Anthropic Messages (including Claude on Vertex and Bedrock Messages), and explicit compaction calls in OpenAI/Azure/xAI Responses. Model and deployment support still depends on the provider.
+
+This is different from prompt caching, server-side history storage, or truncation. Compaction returns provider-owned context that must be replayed to continue the conversation.
+
+### Automatic compaction
+
+Inside an `Effect.gen`, enable OpenAI compaction with typed provider options:
+
+```ts
+import { LLM, LLMClient, LLMRequest, Message } from "@opencode-ai/ai"
+import { OpenAI } from "@opencode-ai/ai/providers"
+
+const request = LLM.request({
+  model: OpenAI.configure({ apiKey }).responses("gpt-5.3-codex"),
+  messages,
+  providerOptions: {
+    contextManagement: [{ type: "compaction", compactThreshold: 200_000 }],
+  },
+})
+const response = yield * LLMClient.generate(request)
+const next = LLMRequest.update(request, {
+  messages: [...request.messages, response.message, Message.user("Continue")],
+})
+```
+
+`store: false` remains the default. Keep the entire `response.message`, not just `response.text`. Compaction events become ordered `CompactionPart`s alongside text and reasoning. The conversation contains everything needed to continue; there is no separate replay object or hidden provider transcript.
+
+A compaction part has `provider` and exactly one representation: `encrypted` for Responses, or `text` for Anthropic. Responses also preserves the optional checkpoint `id`. These fields survive message serialization without becoming visible assistant text. Sending a checkpoint to another provider or an incompatible API fails rather than silently losing context.
+
+```ts
+import { CompactionPart, ProviderID } from "@opencode-ai/ai"
+
+CompactionPart.make({ provider: ProviderID.make("openai"), id: "cmp_123", encrypted: "..." })
+CompactionPart.make({ provider: ProviderID.make("anthropic"), text: "Summary of the conversation..." })
+```
+
+For Anthropic, use:
+
+```ts
+providerOptions: {
+  contextManagement: {
+    edits: [{
+      type: "compact_20260112",
+      trigger: { type: "input_tokens", value: 150_000 },
+      pauseAfterCompaction: true,
+      instructions: "Summarize the task and decisions. Do not call tools while summarizing.",
+    }],
+  },
+}
+```
+
+- The trigger is optional (provider default: 150,000 tokens), with a minimum of 50,000.
+- Custom instructions replace Anthropic's default summarization instructions.
+- The route adds `compact-2026-01-12` to existing beta headers, including when replaying a checkpoint without enabling new compactions.
+- A pause is exposed as `response.finishReason.raw === "compaction"`. The caller explicitly issues the next request; the package never automatically resumes.
+- Anthropic can return a compaction block with `content: null` when summarization fails. This becomes a compaction part with `text: null`, which is **not** a successful replacement for prior history. The package never prunes history automatically.
+- `Usage` totals include all reported Anthropic `usage.iterations`, including compaction. `contextTokens` separately reports the final message iteration's inclusive input size, when available. A compaction-only pause does not report a post-compaction context size. Raw iteration usage remains in `providerMetadata`.
+
+Bedrock's Converse API does not support this feature. Select the native Claude Messages route explicitly; the default `.model(...)` remains Converse:
+
+```ts
+import { AmazonBedrock } from "@opencode-ai/ai/providers"
+
+const model = AmazonBedrock.configure({ region: "us-east-1", credentials }).messages("us.anthropic.claude-opus-4-6-v1")
+```
+
+The corresponding package entrypoint is `@opencode-ai/ai/providers/amazon-bedrock/messages`. It uses InvokeModelWithResponseStream, AWS event-stream framing, bearer or SigV4 auth, and `anthropic_beta` in the request body.
+
+### Explicit compaction
+
+`LLMClient.compact(request)` performs exactly one HTTP call to `/responses/compact`, using the selected route's endpoint, credentials, query, and HTTP middleware. It returns a `CompactionResponse` containing replacement `messages` and usage, not a normal generation response.
+
+```ts
+const compacted = yield * LLMClient.compact(request)
+const next = LLMRequest.update(request, {
+  messages: [...compacted.messages, Message.user("Continue")],
+})
+const response = yield * LLMClient.generate(next)
+```
+
+Replace the prior window with `compacted.messages`. Do not append it to the original transcript or extract only the encrypted item: the provider may retain additional messages in its output. Retained user and assistant messages remain ordinary messages with typed text, media, or reasoning parts, in their original order. Provider-specific message IDs, status, and phase use `providerMetadata`, not a raw output array hidden in an assistant message. Unsupported returned item types fail explicitly. Generation-only body overlays such as `stream` and `store` are not sent to the compact endpoint.
+
+The input must still fit the model's context window. Explicit compaction is not an overflow-recovery operation. xAI supports this explicit path, not the automatic OpenAI option. Unsupported routes, including Bedrock Mantle, do not inherit an explicit compact endpoint simply because they use a Responses protocol.
+
+### Ownership and verification
+
+The AI package transports options and typed conversation parts. It does not schedule compaction, persist Session checkpoints, select history, switch providers, or replace Core's existing local compaction policy. Native compaction is not enabled for OpenCode Sessions by this feature; Session integration must persist these parts before enabling it. The AI SDK bridge rejects native compaction parts rather than dropping them. Provider-executed tool APIs and persistence changes are a separate follow-up.
+
+Tests cover serialized round trips, real local HTTP plus a tool loop, AWS binary frames and signing, provider errors, malformed blocks, and usage accounting. Live provider tests are gated by `RECORD=true` and the relevant API keys:
+
+```sh
+# Run from packages/ai. Only records the selected new cassette group.
+RECORD=true RECORDED_PREFIX=openai-compaction bun test test/provider/compaction.recorded.test.ts
+RECORD=true RECORDED_PREFIX=xai-compaction bun test test/provider/compaction.recorded.test.ts
+RECORD=true RECORDED_PREFIX=anthropic-compaction bun test test/provider/compaction.recorded.test.ts
+```
+
+Provider references: [OpenAI](https://developers.openai.com/api/docs/guides/compaction), [Azure](https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/responses#server-side-compaction), [Anthropic](https://platform.claude.com/docs/en/build-with-claude/compaction), [Bedrock](https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-compaction.html), [xAI](https://docs.x.ai/developers/advanced-api-usage/context-compaction).
+
 ## Caching

 Prompt caching is **on by default**. Every `LLMRequest` resolves to `cache: "auto"` unless the caller opts out with `cache: "none"`. Each protocol translates `CacheHint`s to its wire format (`cache_control` on Anthropic, `cachePoint` on Bedrock; OpenAI and Gemini do implicit caching server-side and don't need inline markers — auto is a no-op there).
@@ -40,6 +40,7 @@ const RESPECTS_INLINE_HINTS = new Set([
  "anthropic-messages",
  "google-vertex-messages",
  "bedrock-converse",
+  "bedrock-messages",
  "openrouter",
 ])

@@ -6,8 +6,12 @@ import { Auth } from "../route/auth.js"
 import { Endpoint } from "../route/endpoint.js"
 import { Framing } from "../route/framing.js"
 import { Protocol } from "../route/protocol.js"
+import { Headers } from "effect/unstable/http"
+import { HttpTransport } from "../route/transport/index.js"
 import {
  AIError,
+  HttpOptions,
+  LLMRequest,
  LLMEvent,
  mergeJsonRecords,
  Usage,
@@ -15,7 +19,6 @@ import {
  type FinishReasonDetails,
  type FinishReason,
  type JsonSchema,
-  type LLMRequest,
  type MediaPart,
  type ProviderMetadata,
  type ToolCallPart,
@@ -61,6 +64,7 @@ export type ThinkingInput =
    ))

 export interface OptionsInput {
+  readonly contextManagement?: ContextManagement
  readonly [key: string]: unknown
  readonly thinking?: ThinkingInput
  readonly effort?: string
@@ -89,6 +93,23 @@ export interface OptionsInput {

 export type ProviderOptionsInput = OptionsInput

+export const ContextManagement = Schema.Struct({
+  edits: Schema.Array(
+    Schema.Struct({
+      type: Schema.Literal("compact_20260112"),
+      trigger: Schema.optional(
+        Schema.Struct({
+          type: Schema.Literal("input_tokens"),
+          value: Schema.Int.check(Schema.isGreaterThanOrEqualTo(50000)),
+        }),
+      ),
+      pauseAfterCompaction: Schema.optional(Schema.Boolean),
+      instructions: Schema.optional(Schema.String),
+    }),
+  ),
+})
+export type ContextManagement = typeof ContextManagement.Type
+
 // =============================================================================
 // Request Body Schema
 // =============================================================================
@@ -236,7 +257,12 @@ const AnthropicUserBlock = Schema.Union([
  AnthropicToolResultBlock,
 ])
 type AnthropicUserBlock = Schema.Schema.Type<typeof AnthropicUserBlock>
+const AnthropicCompactionBlock = Schema.Struct({
+  type: Schema.Literal("compaction"),
+  content: Schema.NullOr(Schema.String),
+})
 const AnthropicAssistantBlock = Schema.Union([
+  AnthropicCompactionBlock,
  AnthropicTextBlock,
  AnthropicThinkingBlock,
  AnthropicRedactedThinkingBlock,
@@ -312,6 +338,18 @@ const AnthropicContainer = Schema.Union([
 ])

 const AnthropicBodyFields = {
+  context_management: Schema.optional(
+    Schema.Struct({
+      edits: Schema.Array(
+        Schema.Struct({
+          type: Schema.Literal("compact_20260112"),
+          trigger: ContextManagement.fields.edits.value.fields.trigger,
+          pause_after_compaction: Schema.optional(Schema.Boolean),
+          instructions: Schema.optional(Schema.String),
+        }),
+      ),
+    }),
+  ),
  model: Schema.String,
  system: optionalArray(AnthropicTextBlock),
  messages: Schema.Array(AnthropicMessage),
@@ -335,7 +373,7 @@ const AnthropicBodyFields = {
 export const AnthropicMessagesBody = Schema.Struct(AnthropicBodyFields)
 export type AnthropicMessagesBody = Schema.Schema.Type<typeof AnthropicMessagesBody>

-const AnthropicUsage = Schema.StructWithRest(
+const AnthropicIterationUsage = Schema.StructWithRest(
  Schema.Struct({
    input_tokens: optionalNull(Schema.Number),
    output_tokens: Schema.optional(Schema.Number),
@@ -354,6 +392,13 @@ const AnthropicUsage = Schema.StructWithRest(
  }),
  [Schema.Record(Schema.String, Schema.Unknown)],
 )
+const AnthropicUsage = Schema.StructWithRest(
+  Schema.Struct({
+    ...AnthropicIterationUsage.schema.fields,
+    iterations: Schema.optional(Schema.Array(AnthropicIterationUsage)),
+  }),
+  [JsonObject],
+)
 type AnthropicUsage = Schema.Schema.Type<typeof AnthropicUsage>

 const AnthropicStreamBlock = Schema.Struct({
@@ -377,6 +422,7 @@ type AnthropicStreamBlock = Schema.Schema.Type<typeof AnthropicStreamBlock>
 const decodeAnthropicStreamBlock = Schema.decodeUnknownOption(AnthropicStreamBlock)

 const AnthropicStreamDelta = Schema.Struct({
+  content: optionalNull(Schema.String),
  type: Schema.optional(Schema.String),
  text: Schema.optional(Schema.String),
  thinking: Schema.optional(Schema.String),
@@ -406,6 +452,8 @@ const AnthropicEvent = Schema.Struct({
 type AnthropicEvent = Schema.Schema.Type<typeof AnthropicEvent>

 interface ParserState {
+  readonly provider: LLMRequest["model"]["provider"]
+  readonly compactions: Readonly<Record<number, string | null>>
  readonly providerMetadataKey: string
  readonly tools: ToolStream.State<number>
  readonly reasoningSignatures: Readonly<Record<number, string>>
@@ -848,6 +896,12 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
    if (message.role === "assistant") {
      const content: AnthropicAssistantBlock[] = []
      for (const part of message.content) {
+        if (part.type === "compaction") {
+          if (part.provider !== request.model.provider || part.text === undefined)
+            return yield* invalid("Compaction state must be replayed to its originating provider and API")
+          content.push({ type: "compaction", content: part.text })
+          continue
+        }
        if (part.type === "text") {
          if (part.text.trim().length === 0) continue
          content.push({ type: "text", text: part.text, cache_control: cacheControl(breakpoints, part.cache) })
@@ -1003,6 +1057,9 @@ const resolveThinking = Effect.fn("AnthropicMessages.resolveThinking")(function*
 })

 const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (request: LLMRequest) {
+  const management = yield* ProviderShared.validateWith(
+    Schema.decodeUnknownEffect(Schema.UndefinedOr(ContextManagement)),
+  )(request.providerOptions?.contextManagement)
  const generation = request.generation
  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
  // Allocate the 4-breakpoint budget in invalidation order: tools → system →
@@ -1037,7 +1094,7 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
    )
  }
  const options = yield* resolveOptions(request)
-  return {
+  const body = {
    model: request.model.id,
    system,
    messages,
@@ -1058,6 +1115,18 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
    metadata: options.metadata,
    service_tier: options.service_tier,
  }
+  if (!management) return body
+  return {
+    ...body,
+    context_management: {
+      edits: management.edits.map((edit) => ({
+        type: edit.type,
+        trigger: edit.trigger,
+        pause_after_compaction: edit.pauseAfterCompaction,
+        instructions: edit.instructions,
+      })),
+    },
+  }
 })

 // =============================================================================
@@ -1079,18 +1148,31 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
 // expose that subset through `output_tokens_details.thinking_tokens`.
 const mapUsage = (usage: AnthropicUsage | undefined, providerMetadataKey: string): Usage | undefined => {
  if (!usage) return undefined
-  const nonCached = usage.input_tokens ?? undefined
-  const cacheRead = usage.cache_read_input_tokens ?? undefined
-  const cacheWrite = usage.cache_creation_input_tokens ?? undefined
+  const iterations = usage.iterations?.length ? usage.iterations : [usage]
+  const last = usage.iterations?.at(-1)
+  const nonCached = ProviderShared.sumTokens(...iterations.map((item) => item.input_tokens ?? undefined))
+  const cacheRead = ProviderShared.sumTokens(...iterations.map((item) => item.cache_read_input_tokens ?? undefined))
+  const cacheWrite = ProviderShared.sumTokens(
+    ...iterations.map((item) => item.cache_creation_input_tokens ?? undefined),
+  )
  const inputTokens = ProviderShared.sumTokens(nonCached, cacheRead, cacheWrite)
+  const outputTokens = ProviderShared.sumTokens(...iterations.map((item) => item.output_tokens))
  return new Usage({
    inputTokens,
-    outputTokens: usage.output_tokens,
+    outputTokens,
+    contextTokens:
+      last?.type === "message"
+        ? ProviderShared.sumTokens(
+            last.input_tokens ?? undefined,
+            last.cache_read_input_tokens ?? undefined,
+            last.cache_creation_input_tokens ?? undefined,
+          )
+        : undefined,
    nonCachedInputTokens: nonCached,
    cacheReadInputTokens: cacheRead,
    cacheWriteInputTokens: cacheWrite,
-    reasoningTokens: usage.output_tokens_details?.thinking_tokens,
-    totalTokens: ProviderShared.totalTokens(inputTokens, usage.output_tokens, undefined),
+    reasoningTokens: ProviderShared.sumTokens(...iterations.map((item) => item.output_tokens_details?.thinking_tokens)),
+    totalTokens: ProviderShared.totalTokens(inputTokens, outputTokens, undefined),
    providerMetadata: { [providerMetadataKey]: usage },
  })
 }
@@ -1112,6 +1194,7 @@ const mergeUsage = (left: Usage | undefined, right: Usage | undefined, providerM
  return new Usage({
    inputTokens,
    outputTokens,
+    contextTokens: right.contextTokens ?? left.contextTokens,
    nonCachedInputTokens,
    cacheReadInputTokens,
    cacheWriteInputTokens,
@@ -1170,7 +1253,6 @@ const onContentBlockStart = (
  event: AnthropicEvent & { readonly content_block: AnthropicStreamBlock },
 ): StepResult => {
  const block = event.content_block
-  if (!block) return [state, NO_EVENTS]

  if (block.type === "tool_use" || block.type === "server_tool_use") {
    if (event.index === undefined || !block.id) return [state, NO_EVENTS]
@@ -1265,7 +1347,16 @@ const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(f
 ) {
  const delta = event.delta

-  if (delta?.type === "text_delta" && delta.text) {
+  if (delta.type === "compaction_delta") {
+    if (event.index === undefined || !(event.index in state.compactions) || delta.content === undefined)
+      return yield* ProviderShared.eventError(ADAPTER, "Compaction delta is missing its block or content")
+    return [
+      { ...state, compactions: { ...state.compactions, [event.index]: delta.content } },
+      NO_EVENTS,
+    ] satisfies StepResult
+  }
+
+  if (delta.type === "text_delta" && delta.text) {
    if (!state.lifecycle.text.has(`text-${event.index ?? 0}`)) return [state, NO_EVENTS] satisfies StepResult
    const events: LLMEvent[] = []
    return [
@@ -1274,7 +1365,7 @@ const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(f
    ] satisfies StepResult
  }

-  if (delta?.type === "thinking_delta" && delta.thinking) {
+  if (delta.type === "thinking_delta" && delta.thinking) {
    if (!state.lifecycle.reasoning.has(`reasoning-${event.index ?? 0}`)) return [state, NO_EVENTS] satisfies StepResult
    const events: LLMEvent[] = []
    return [
@@ -1286,7 +1377,7 @@ const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(f
    ] satisfies StepResult
  }

-  if (delta?.type === "signature_delta" && delta.signature) {
+  if (delta.type === "signature_delta" && delta.signature) {
    const index = event.index ?? 0
    if (!state.lifecycle.reasoning.has(`reasoning-${index}`)) return [state, NO_EVENTS] satisfies StepResult
    return [
@@ -1298,7 +1389,7 @@ const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(f
    ] satisfies StepResult
  }

-  if (delta?.type === "input_json_delta" && event.index !== undefined) {
+  if (delta.type === "input_json_delta" && event.index !== undefined) {
    if (!delta.partial_json) return [state, NO_EVENTS] satisfies StepResult
    if (!state.tools[event.index]) return [state, NO_EVENTS] satisfies StepResult
    const result = ToolStream.appendExisting(
@@ -1323,6 +1414,18 @@ const onContentBlockStop = Effect.fn("AnthropicMessages.onContentBlockStop")(fun
  event: AnthropicEvent,
 ) {
  if (event.index === undefined) return [state, NO_EVENTS] satisfies StepResult
+  if (event.index in state.compactions) {
+    const { [event.index]: content, ...compactions } = state.compactions
+    const events: LLMEvent[] = []
+    const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
+    events.push(
+      LLMEvent.compaction({
+        provider: state.provider,
+        text: content,
+      }),
+    )
+    return [{ ...state, compactions, lifecycle }, events] satisfies StepResult
+  }
  const result = yield* ToolStream.finish(ADAPTER, state.tools, event.index)
  const events: LLMEvent[] = []
  const resultEvents = result.events ?? []
@@ -1374,6 +1477,8 @@ const onMessageDelta = (
 }

 const onMessageStop = Effect.fn("AnthropicMessages.onMessageStop")(function* (state: ParserState) {
+  if (Object.keys(state.compactions).length)
+    return yield* ProviderShared.eventError(ADAPTER, "Response ended with an incomplete compaction block")
  const result = yield* ToolStream.finishAll(ADAPTER, state.tools)
  const events: LLMEvent[] = []
  const lifecycle = result.events.length ? Lifecycle.stepStart(state.lifecycle, events) : state.lifecycle
@@ -1418,16 +1523,21 @@ const onError = (event: AnthropicEvent) => {
  )
 }

-const isKnownStreamBlockType = (type: string) =>
-  type === "text" ||
-  type === "thinking" ||
-  type === "redacted_thinking" ||
-  type === "tool_use" ||
-  type === "server_tool_use" ||
-  isServerToolResultType(type)
-
-const isKnownStreamDeltaType = (type: string) =>
-  type === "text_delta" || type === "thinking_delta" || type === "signature_delta" || type === "input_json_delta"
+const STREAM_BLOCK_TYPES = new Set([
+  "compaction",
+  "text",
+  "thinking",
+  "redacted_thinking",
+  "tool_use",
+  "server_tool_use",
+])
+const STREAM_DELTA_TYPES = new Set([
+  "compaction_delta",
+  "text_delta",
+  "thinking_delta",
+  "signature_delta",
+  "input_json_delta",
+])

 const invalidStreamEvent = (event: AnthropicEvent) =>
  Effect.fail(
@@ -1456,7 +1566,16 @@ const step = (state: ParserState, event: AnthropicEvent) => {
  if (event.type === "content_block_start") {
    if (!ProviderShared.isRecord(event.content_block) || typeof event.content_block.type !== "string")
      return invalidStreamEvent(event)
-    if (!isKnownStreamBlockType(event.content_block.type)) return Effect.succeed<StepResult>([state, NO_EVENTS])
+    if (event.content_block.type === "compaction") {
+      const decoded = Schema.decodeUnknownOption(AnthropicCompactionBlock)(event.content_block)
+      if (event.index === undefined || Option.isNone(decoded)) return invalidStreamEvent(event)
+      return Effect.succeed<StepResult>([
+        { ...state, compactions: { ...state.compactions, [event.index]: decoded.value.content } },
+        NO_EVENTS,
+      ])
+    }
+    if (!STREAM_BLOCK_TYPES.has(event.content_block.type) && !isServerToolResultType(event.content_block.type))
+      return Effect.succeed<StepResult>([state, NO_EVENTS])
    const decoded = decodeAnthropicStreamBlock(event.content_block)
    if (Option.isNone(decoded)) return invalidStreamEvent(event)
    const block = decoded.value
@@ -1470,7 +1589,7 @@ const step = (state: ParserState, event: AnthropicEvent) => {
  }
  if (event.type === "content_block_delta") {
    if (!ProviderShared.isRecord(event.delta)) return invalidStreamEvent(event)
-    if (typeof event.delta.type === "string" && !isKnownStreamDeltaType(event.delta.type))
+    if (typeof event.delta.type === "string" && !STREAM_DELTA_TYPES.has(event.delta.type))
      return Effect.succeed<StepResult>([state, NO_EVENTS])
    const decoded = decodeAnthropicStreamDelta(event.delta)
    if (Option.isNone(decoded)) return invalidStreamEvent(event)
@@ -1504,6 +1623,8 @@ export const protocol = Protocol.make({
  stream: {
    event: Protocol.jsonEvent(AnthropicEvent),
    initial: (request) => ({
+      provider: request.model.provider,
+      compactions: {},
      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      tools: ToolStream.empty<number>(),
      reasoningSignatures: {},
@@ -1513,6 +1634,37 @@ export const protocol = Protocol.make({
  },
 })

+export const transport = <Body extends Pick<AnthropicMessagesBody, "messages" | "context_management">>() => {
+  const http = HttpTransport.httpJson<Body, string>({ framing })
+  return {
+    ...http,
+    prepare: (input: Parameters<typeof http.prepare>[0]) => {
+      if (
+        !input.body.context_management?.edits.length &&
+        !input.body.messages.some((message) => message.content.some((block) => block.type === "compaction"))
+      )
+        return http.prepare(input)
+      const headers = Headers.fromInput(input.request.http?.headers)
+      const betas = new Set(
+        (headers["anthropic-beta"] ?? "")
+          .split(",")
+          .map((item) => item.trim())
+          .filter(Boolean),
+      )
+      betas.add("compact-2026-01-12")
+      return http.prepare({
+        ...input,
+        request: LLMRequest.update(input.request, {
+          http: new HttpOptions({
+            ...input.request.http,
+            headers: { ...headers, "anthropic-beta": [...betas].join(",") },
+          }),
+        }),
+      })
+    },
+  }
+}
+
 export const route = Route.make({
  id: ADAPTER,
  provider: "anthropic",
@@ -1522,7 +1674,7 @@ export const route = Route.make({
    baseURL: DEFAULT_BASE_URL,
  }),
  auth: Auth.none,
-  framing,
+  transport: transport<AnthropicMessagesBody>(),
  headers: () => ({ "anthropic-version": "2023-06-01" }),
 })

@@ -0,0 +1,96 @@
+import { Effect, Encoding, Schema, Struct } from "effect"
+import { Headers } from "effect/unstable/http"
+import { AIError } from "../schema/index.js"
+import { Route } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
+import { Protocol } from "../route/protocol.js"
+import { classifyProviderFailure } from "../provider-error.js"
+import { AnthropicMessages } from "./anthropic-messages.js"
+import { BedrockEventStream } from "./bedrock-event-stream.js"
+import { BedrockAuth } from "./utils/bedrock-auth.js"
+import { JsonObject, ProviderShared } from "./shared.js"
+
+const ID = "bedrock-messages"
+const VERSION = "bedrock-2023-05-31"
+const Body = Schema.Struct({
+  ...Struct.omit(AnthropicMessages.AnthropicMessagesBody.fields, ["model", "stream"]),
+  anthropic_version: Schema.Literal(VERSION),
+  anthropic_beta: Schema.optional(Schema.Array(Schema.String)),
+})
+const Event = Schema.Struct({
+  chunk: Schema.optional(Schema.Struct({ bytes: Schema.String })),
+  exception: Schema.optional(
+    Schema.Struct({
+      type: Schema.String,
+      details: Schema.StructWithRest(
+        Schema.Struct({ message: Schema.optional(Schema.String), originalMessage: Schema.optional(Schema.String) }),
+        [JsonObject],
+      ),
+    }),
+  ),
+})
+
+export const protocol = Protocol.make({
+  id: ID,
+  body: {
+    schema: Body,
+    from: Effect.fn("BedrockMessages.fromRequest")(function* (request) {
+      const body = yield* AnthropicMessages.protocol.body.from(request)
+      const headers = Headers.fromInput(request.http?.headers)
+      const betas = new Set(
+        (headers["anthropic-beta"] ?? "")
+          .split(",")
+          .map((value) => value.trim())
+          .filter(Boolean),
+      )
+      if (
+        body.context_management?.edits.length ||
+        body.messages.some((message) => message.content.some((block) => block.type === "compaction"))
+      )
+        betas.add("compact-2026-01-12")
+      return {
+        ...Struct.omit(body, ["model", "stream"]),
+        anthropic_version: VERSION,
+        anthropic_beta: betas.size ? [...betas] : undefined,
+      } satisfies typeof Body.Type
+    }),
+  },
+  stream: {
+    event: Event,
+    initial: AnthropicMessages.protocol.stream.initial,
+    step: Effect.fn("BedrockMessages.step")(function* (state, event) {
+      if (event.exception)
+        return yield* new AIError({
+          reason: classifyProviderFailure({
+            message: event.exception.details.message ?? event.exception.details.originalMessage ?? event.exception.type,
+            rawBody: ProviderShared.encodeJson(event),
+          }),
+        })
+      if (!event.chunk) return yield* ProviderShared.eventError(ID, "Bedrock Messages event is missing its chunk")
+      const text = yield* Effect.fromResult(Encoding.decodeBase64String(event.chunk.bytes)).pipe(
+        Effect.mapError((cause) =>
+          ProviderShared.eventError(ID, "Invalid Bedrock Messages chunk encoding", undefined, cause),
+        ),
+      )
+      const decoded = yield* Schema.decodeUnknownEffect(AnthropicMessages.protocol.stream.event)(text).pipe(
+        Effect.mapError((cause) => ProviderShared.eventError(ID, "Invalid Bedrock Messages event", undefined, cause)),
+      )
+      return yield* AnthropicMessages.protocol.stream.step(state, decoded)
+    }),
+  },
+})
+
+export const route = Route.make({
+  id: ID,
+  provider: "amazon-bedrock",
+  providerMetadataKey: "anthropic",
+  protocol,
+  endpoint: Endpoint.path(
+    ({ request }) => `/model/${encodeURIComponent(request.model.id)}/invoke-with-response-stream`,
+    { baseURL: "https://bedrock-runtime.us-east-1.amazonaws.com" },
+  ),
+  auth: BedrockAuth.auth,
+  framing: BedrockEventStream.framing(ID),
+})
+
+export * as BedrockMessages from "./bedrock-messages.js"
@@ -1,5 +1,6 @@
 export * as AnthropicMessages from "./anthropic-messages.js"
 export * as BedrockConverse from "./bedrock-converse.js"
+export { BedrockMessages } from "./bedrock-messages.js"
 export * as Gemini from "./gemini.js"
 export * as MistralChat from "./mistral-chat.js"
 export * as OpenAIChat from "./openai-chat.js"
@@ -32,13 +32,14 @@ export const PATH = "/responses"
 // =============================================================================
 // Request Body Schema
 // =============================================================================
-const OpenResponsesInputText = Schema.Struct({
+export const OpenResponsesInputText = Schema.Struct({
  type: Schema.tag("input_text"),
  text: Schema.String,
 })
-const OpenResponsesInputImage = Schema.Struct({
+export const OpenResponsesInputImage = Schema.Struct({
  type: Schema.tag("input_image"),
  image_url: Schema.String,
+  detail: Schema.optional(Schema.String),
 })
 const OpenResponsesInputFile = Schema.Struct({
  type: Schema.tag("input_file"),
@@ -54,7 +55,7 @@ const MediaInput = Schema.Union([OpenResponsesInputImage, OpenResponsesInputFile
 export type MediaInput = Schema.Schema.Type<typeof MediaInput>
 const OpenResponsesInputContent = Schema.Union([OpenResponsesInputText, MediaInput])

-const OpenResponsesOutputText = Schema.Struct({
+export const OpenResponsesOutputText = Schema.Struct({
  type: Schema.tag("output_text"),
  text: Schema.String,
 })
@@ -62,6 +63,13 @@ const OpenResponsesOutputText = Schema.Struct({
 export const MessagePhase = Schema.NullOr(Schema.Literals(["commentary", "final_answer"]))
 type MessagePhase = Schema.Schema.Type<typeof MessagePhase>

+export const MessageMetadata = Schema.Struct({
+  itemId: Schema.optional(Schema.String),
+  type: Schema.optional(Schema.Literal("message")),
+  status: Schema.optional(Schema.String),
+  phase: Schema.optional(MessagePhase),
+})
+
 const messagePhase = (value: unknown): MessagePhase | undefined => {
  if (value === null || value === "commentary" || value === "final_answer") return value
  return undefined
@@ -72,7 +80,7 @@ const OpenResponsesReasoningSummaryText = Schema.Struct({
  text: Schema.String,
 })

-const OpenResponsesReasoningItem = Schema.Struct({
+export const OpenResponsesReasoningItem = Schema.Struct({
  type: Schema.tag("reasoning"),
  id: Schema.optionalKey(Schema.String),
  summary: Schema.Array(OpenResponsesReasoningSummaryText),
@@ -149,16 +157,30 @@ const OpenResponsesFunctionCallOutput = Schema.Union([
  Schema.Array(OpenResponsesFunctionCallOutputContent),
 ])

+export const CompactionItem = Schema.Struct({
+  type: Schema.Literal("compaction"),
+  id: optionalNull(Schema.String),
+  encrypted_content: Schema.String,
+})
+
 export const InputItem = Schema.Union([
+  CompactionItem,
  Schema.Struct({ role: Schema.tag("system"), content: Schema.String }),
  Schema.Struct({ role: Schema.tag("developer"), content: Schema.String }),
-  Schema.Struct({ role: Schema.tag("user"), content: Schema.Array(OpenResponsesInputContent) }),
+  Schema.Struct({
+    role: Schema.tag("user"),
+    content: Schema.Array(OpenResponsesInputContent),
+    type: Schema.optional(Schema.Literal("message")),
+    id: Schema.optional(Schema.String),
+    status: Schema.optional(Schema.String),
+  }),
  Schema.Struct({
    type: Schema.tag("message"),
    id: Schema.optionalKey(Schema.String),
    role: Schema.tag("assistant"),
    content: Schema.Array(OpenResponsesOutputText),
    phase: Schema.optionalKey(MessagePhase),
+    status: Schema.optional(Schema.String),
  }),
  OpenResponsesReasoningItem,
  Schema.Struct({
@@ -267,7 +289,7 @@ const OpenResponsesBody = Schema.Struct({
 })
 export type OpenResponsesBody = Schema.Schema.Type<typeof OpenResponsesBody>

-const OpenResponsesUsage = Schema.Struct({
+export const OpenResponsesUsage = Schema.Struct({
  input_tokens: Schema.optional(Schema.Number),
  input_tokens_details: optionalNull(
    Schema.Struct({
@@ -387,6 +409,8 @@ export interface Extension {
 const BASE: Extension = { id: ADAPTER, name: NAME }

 export interface ParserState {
+  readonly provider: LLMRequest["model"]["provider"]
+  readonly completedCompactions: ReadonlySet<string>
  readonly id: string
  readonly name: string
  readonly providerMetadataKey: string
@@ -501,7 +525,13 @@ const lowerMedia = Effect.fn("OpenResponses.lowerMedia")(function* (
      ...(url ? { file_url: url } : { file_data: media.dataUrl }),
    }
  }
-  return { type: "input_image" as const, image_url: url ?? media.dataUrl }
+  return {
+    type: "input_image" as const,
+    image_url: url ?? media.dataUrl,
+    detail: yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(OpenResponsesInputImage.fields.detail))(
+      part.providerMetadata?.[request.model.route.providerMetadataKey ?? "openresponses"]?.detail,
+    ),
+  }
 })

 const lowerUserContent = Effect.fnUntraced(function* (
@@ -568,6 +598,9 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
  const providerMetadataKey = request.model.route.providerMetadataKey ?? "openresponses"

  for (const message of request.messages) {
+    const metadata = yield* ProviderShared.validateWith(
+      Schema.decodeUnknownEffect(Schema.UndefinedOr(MessageMetadata)),
+    )(message.providerMetadata?.[providerMetadataKey])
    if (message.role === "system") {
      input.push({
        role: "developer",
@@ -578,7 +611,8 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques

    if (message.role === "user") {
      const content = yield* Effect.forEach(message.content, (part) => lowerUserContent(part, request, extension))
-      if (content.length > 0) input.push({ role: "user", content })
+      if (content.length > 0)
+        input.push({ role: "user", content, type: metadata?.type, id: metadata?.itemId, status: metadata?.status })
      continue
    }

@@ -591,9 +625,10 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
        const groups = content.reduce<
          Array<{ id: string | undefined; phase: MessagePhase | null | undefined; parts: TextPart[] }>
        >((groups, part) => {
-          const metadata = part.providerMetadata?.[providerMetadataKey]
-          const id = itemID(part.providerMetadata, providerMetadataKey)
-          const phase = ProviderShared.isRecord(metadata) ? messagePhase(metadata.phase) : undefined
+          const partMetadata = part.providerMetadata?.[providerMetadataKey]
+          const id = itemID(part.providerMetadata, providerMetadataKey) ?? metadata?.itemId
+          const partPhase = messagePhase(partMetadata?.phase)
+          const phase = partPhase === undefined ? metadata?.phase : partPhase
          const group = groups.at(-1)
          if (group && group.id === id && group.phase === phase) group.parts.push(part)
          else groups.push({ id, phase, parts: [part] })
@@ -604,6 +639,7 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
            type: "message" as const,
            ...(group.id === undefined ? {} : { id: group.id }),
            role: "assistant" as const,
+            status: metadata?.status,
            content: group.parts.map((part) => ({ type: "output_text" as const, text: part.text })),
            ...(group.phase === undefined ? {} : { phase: group.phase }),
          })),
@@ -611,6 +647,15 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
        content.splice(0, content.length)
      }
      for (const part of message.content) {
+        if (part.type === "compaction") {
+          flushText()
+          if (part.provider !== request.model.provider || part.encrypted === undefined)
+            return yield* ProviderShared.invalidRequest(
+              "Compaction state must be replayed to its originating provider and API",
+            )
+          input.push({ type: "compaction", id: part.id, encrypted_content: part.encrypted })
+          continue
+        }
        if (part.type === "text") {
          content.push(part)
          continue
@@ -689,13 +734,30 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
  return input
 })

-const lowerOptions = (request: LLMRequest) => {
-  const options = OpenResponsesOptions.resolve(request)
+export const lowerConversation = Effect.fn("OpenResponses.lowerConversation")(function* (
+  request: LLMRequest,
+  extension: Extension,
+) {
  const instructions = ProviderShared.joinText(request.system)
+  return {
+    model: request.model.id,
+    input: yield* lowerMessages(request, extension),
+    ...(instructions ? { instructions } : {}),
+  }
+})
+
+export const lowerGeneration = (request: LLMRequest) => {
+  const options = OpenResponsesOptions.resolve(request)
+  const generation = request.generation
  const cacheKey = ProviderShared.promptCacheKey(request)
  const parallelToolCalls = resolveParallelToolCalls(request)
  return {
-    ...(instructions ? { instructions } : {}),
+    stream: true as const,
+    max_output_tokens: generation?.maxTokens,
+    temperature: generation?.temperature,
+    top_p: generation?.topP,
+    presence_penalty: generation?.presencePenalty,
+    frequency_penalty: generation?.frequencyPenalty,
    ...(options.store !== undefined ? { store: options.store } : {}),
    ...(options.metadata ? { metadata: options.metadata } : {}),
    ...(options.safetyIdentifier ? { safety_identifier: options.safetyIdentifier } : {}),
@@ -723,7 +785,7 @@ export const resolveParallelToolCalls = (request: LLMRequest) => {
  return disabled === undefined ? undefined : !disabled
 }

-const allowedToolChoice = (request: LLMRequest) => {
+export const allowedToolChoice = (request: LLMRequest) => {
  const allowed = OpenResponsesOptions.resolve(request).allowedTools
  if (!allowed) return undefined
  return {
@@ -737,11 +799,10 @@ export const fromRequestWithExtension = Effect.fn("OpenResponses.fromRequestWith
  request: LLMRequest,
  extension: Extension,
 ) {
-  const generation = request.generation
  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
  return {
-    model: request.model.id,
-    input: yield* lowerMessages(request, extension),
+    ...(yield* lowerConversation(request, extension)),
+    ...lowerGeneration(request),
    tools:
      request.tools.length === 0
        ? undefined
@@ -755,13 +816,6 @@ export const fromRequestWithExtension = Effect.fn("OpenResponses.fromRequestWith
    tool_choice:
      allowedToolChoice(request) ??
      (request.toolChoice ? yield* lowerToolChoice(extension.name, request.toolChoice) : undefined),
-    stream: true as const,
-    max_output_tokens: generation?.maxTokens,
-    temperature: generation?.temperature,
-    top_p: generation?.topP,
-    presence_penalty: generation?.presencePenalty,
-    frequency_penalty: generation?.frequencyPenalty,
-    ...lowerOptions(request),
  }
 })

@@ -778,7 +832,7 @@ export const fromRequest = Effect.fn("OpenResponses.fromRequest")(function* (req
 // cached-read and cache-write subsets, and `output_tokens` (inclusive total)
 // with a `reasoning_tokens` subset. Pass the totals through and derive the
 // non-cached breakdown.
-const mapUsage = (usage: OpenResponsesUsage | null | undefined, providerMetadataKey: string) => {
+export const mapUsage = (usage: OpenResponsesUsage | null | undefined, providerMetadataKey: string) => {
  if (!usage) return undefined
  const cached = usage.input_tokens_details?.cached_tokens
  const cacheWrite = usage.input_tokens_details?.cache_write_tokens
@@ -1084,6 +1138,25 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
 ) {
  if (!item) return [state, NO_EVENTS] satisfies StepResult

+  if (item.type === "compaction") {
+    if (!item.id || typeof item.encrypted_content !== "string")
+      return yield* ProviderShared.eventError(state.id, "Compaction output is missing its id or encrypted content")
+    if (state.completedCompactions.has(item.id)) return [state, NO_EVENTS] satisfies StepResult
+    const events: LLMEvent[] = []
+    const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
+    events.push(
+      LLMEvent.compaction({
+        provider: state.provider,
+        id: item.id,
+        encrypted: item.encrypted_content,
+      }),
+    )
+    return [
+      { ...state, lifecycle, completedCompactions: new Set([...state.completedCompactions, item.id]) },
+      events,
+    ] satisfies StepResult
+  }
+
  if (item.type === "message" && item.id !== undefined) {
    const message = state.message?.id === item.id ? state.message : undefined
    const itemPhase = messagePhase(item.phase)
@@ -1243,26 +1316,34 @@ const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (
  const events: LLMEvent[] = []
  if (event.type === "response.completed") {
    for (const item of event.response?.output ?? []) {
-      const id = item.id ?? (item.type === "function_call" ? item.call_id : undefined)
-      if (id === undefined) continue
-      if (item.type !== "function_call" || !current.tools[id]) continue
+      if (item.type !== "compaction" && item.type !== "function_call") continue
+      if (item.type === "compaction") {
+        // Terminal recovery cannot insert a checkpoint before already-emitted content.
+        if (state.lifecycle.stepStarted && !state.completedCompactions.has(item.id ?? ""))
+          return yield* ProviderShared.eventError(
+            state.id,
+            "Cannot recover a compaction checkpoint after output has been emitted",
+          )
+      }
+      if (item.type === "function_call" && !current.tools[item.id ?? item.call_id ?? ""]) continue
      const [next, emitted] = yield* onOutputItemDone(current, item)
      current = next
      events.push(...emitted)
    }
+    // Some compatible providers omit output_item.done even after completing the response.
+    const pending = yield* ToolStream.finishAll(current.id, current.tools)
+    current = {
+      ...current,
+      tools: pending.tools,
+      hasFunctionCall:
+        current.hasFunctionCall ||
+        pending.events.some((event) => LLMEvent.is.toolCall(event) || LLMEvent.is.toolInputError(event)),
+    }
+    events.push(...pending.events)
  }
-  // Some compatible providers omit output_item.done even after completing the response.
-  const pending =
-    event.type === "response.completed"
-      ? yield* ToolStream.finishAll(current.id, current.tools)
-      : { tools: current.tools, events: NO_EVENTS }
-  events.push(...pending.events)
-  const hasFunctionCall =
-    pending.events.some((event) => LLMEvent.is.toolCall(event) || LLMEvent.is.toolInputError(event)) ||
-    current.hasFunctionCall
  const lifecycle = Lifecycle.finish(current.lifecycle, events, {
    reason: {
-      normalized: mapFinishReason(event, hasFunctionCall),
+      normalized: mapFinishReason(event, current.hasFunctionCall),
      raw: event.response?.incomplete_details?.reason,
    },
    usage: mapUsage(event.response?.usage, current.providerMetadataKey),
@@ -1274,7 +1355,7 @@ const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (
          })
        : undefined,
  })
-  return [{ ...current, lifecycle, hasFunctionCall, tools: pending.tools }, events] satisfies StepResult
+  return [{ ...current, lifecycle }, events] satisfies StepResult
 })

 // Build the prettiest summary available from whatever the provider supplied.
@@ -1409,6 +1490,8 @@ export const step = (state: ParserState, input: Event) => {
 * implementations compose this baseline with their own tools and event variants.
 */
 export const initial = (request: LLMRequest, extension: Extension = BASE): ParserState => ({
+  provider: request.model.provider,
+  completedCompactions: new Set<string>(),
  id: extension.id,
  name: extension.name,
  providerMetadataKey: request.model.route.providerMetadataKey ?? "openresponses",
@@ -5,13 +5,14 @@ import { Auth } from "../route/auth.js"
 import { Endpoint } from "../route/endpoint.js"
 import { Protocol } from "../route/protocol.js"
 import { HttpTransport } from "../route/transport/index.js"
-import { LLMRequest, type JsonSchema, type ToolDefinition } from "../schema/index.js"
+import type { LLMRequest, JsonSchema, ToolDefinition } from "../schema/index.js"
 import { OpenResponses } from "./open-responses.js"
 import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
 import { OpenAIImage } from "./utils/openai-image.js"
 import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"
 import { ToolSchemaProjection } from "./utils/tool-schema.js"
 import { OpenResponsesChannel } from "./open-responses-channel.js"
+import { ResponsesCompaction } from "./utils/responses-compaction.js"

 const ADAPTER = "openai-responses"
 const NAME = "OpenAI Responses"
@@ -20,6 +21,14 @@ const WEBSOCKET_ROTATE_AFTER_MS = 55 * 60 * 1000
 export const DEFAULT_BASE_URL = "https://api.openai.com/v1"
 export const PATH = OpenResponses.PATH

+export const ContextManagement = Schema.Array(
+  Schema.Struct({
+    type: Schema.Literal("compaction"),
+    compactThreshold: Schema.optional(Schema.Int.check(Schema.isGreaterThan(0))),
+  }),
+)
+export type ContextManagement = typeof ContextManagement.Type
+
 const OpenAIResponsesImageGenerationTool = Schema.Struct({
  type: Schema.tag("image_generation"),
  action: Schema.optional(Schema.Literals(["auto", "generate", "edit"])),
@@ -78,6 +87,14 @@ const OpenAIResponsesCoreFields = {
  input: Schema.Array(Schema.Union([OpenResponses.InputItem, OpenAIResponsesHostedToolItem])),
  tools: optionalArray(OpenAIResponsesTools),
  tool_choice: Schema.optional(OpenAIResponsesToolChoice),
+  context_management: Schema.optional(
+    Schema.Array(
+      Schema.Struct({
+        type: Schema.Literal("compaction"),
+        compact_threshold: Schema.optional(Schema.Int.check(Schema.isGreaterThan(0))),
+      }),
+    ),
+  ),
 }

 const OpenAIResponsesBody = Schema.Struct({
@@ -125,15 +142,14 @@ const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>, tool
 const decodeBody = ProviderShared.validateWith(Schema.decodeUnknownEffect(OpenAIResponsesBody))

 const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request: LLMRequest) {
-  const body = yield* OpenResponses.fromRequestWithExtension(
-    LLMRequest.update(request, { tools: [], toolChoice: undefined }),
-    extension,
-  )
+  const management = yield* ProviderShared.validateWith(
+    Schema.decodeUnknownEffect(Schema.UndefinedOr(ContextManagement)),
+  )(request.providerOptions?.contextManagement)
  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
-  const parallelToolCalls = OpenResponses.resolveParallelToolCalls(request)
  return yield* decodeBody({
-    ...body,
-    ...(parallelToolCalls === undefined ? {} : { parallel_tool_calls: parallelToolCalls }),
+    ...(yield* OpenResponses.lowerConversation(request, extension)),
+    ...OpenResponses.lowerGeneration(request),
+    context_management: management?.map((edit) => ({ type: edit.type, compact_threshold: edit.compactThreshold })),
    tools:
      request.tools.length === 0
        ? undefined
@@ -141,7 +157,8 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
            lowerTool(tool, ToolSchemaProjection.modelCompatibility(tool.inputSchema, toolSchemaCompatibility)),
          ),
    tool_choice:
-      body.tool_choice ?? (request.toolChoice ? yield* lowerToolChoice(request.toolChoice, request.tools) : undefined),
+      OpenResponses.allowedToolChoice(request) ??
+      (request.toolChoice ? yield* lowerToolChoice(request.toolChoice, request.tools) : undefined),
  })
 })

@@ -223,6 +240,7 @@ export const transport = channelTransport({
 })

 export const route = Route.make({
+  compact: ResponsesCompaction.make(extension),
  id: ADAPTER,
  provider: "openai",
  providerMetadataKey: "openai",
@@ -0,0 +1,152 @@
+import { Effect, Schema, Stream } from "effect"
+import {
+  AIError,
+  InvalidProviderOutputError,
+  CompactionPart,
+  CompactionResponse,
+  HttpOptions,
+  LLMRequest,
+  Message,
+  type ContentPart,
+  mergeJsonRecords,
+} from "../../schema/index.js"
+import type { CompactOperation } from "../../route/client.js"
+import { Endpoint } from "../../route/endpoint.js"
+import { RequestExecutor } from "../../route/executor.js"
+import { HttpTransport } from "../../route/transport/index.js"
+import { OpenResponses } from "../open-responses.js"
+import { JsonObject, ProviderShared } from "../shared.js"
+
+const Body = Schema.Struct({
+  model: Schema.String,
+  input: Schema.Array(Schema.Unknown),
+  instructions: Schema.optional(Schema.String),
+  previous_response_id: Schema.optional(Schema.String),
+})
+
+const Text = Schema.Union([OpenResponses.OpenResponsesInputText, OpenResponses.OpenResponsesOutputText])
+const File = Schema.Union([
+  Schema.Struct({ type: Schema.Literal("input_file"), filename: Schema.String, file_url: Schema.String }),
+  Schema.Struct({ type: Schema.Literal("input_file"), filename: Schema.String, file_data: Schema.String }),
+])
+const MessageFields = {
+  type: Schema.Literal("message"),
+  id: Schema.optional(Schema.String),
+  status: Schema.optional(Schema.String),
+  phase: Schema.optional(OpenResponses.MessagePhase),
+}
+const Response = Schema.Struct({
+  object: Schema.Literal("response.compaction"),
+  output: Schema.Array(
+    Schema.Union([
+      OpenResponses.CompactionItem,
+      OpenResponses.OpenResponsesReasoningItem,
+      Schema.Struct({
+        ...MessageFields,
+        role: Schema.Literal("user"),
+        content: Schema.Array(Schema.Union([Text, OpenResponses.OpenResponsesInputImage, File])).check(
+          Schema.isMinLength(1),
+        ),
+      }),
+      Schema.Struct({
+        ...MessageFields,
+        role: Schema.Literal("assistant"),
+        content: Schema.Array(Text).check(Schema.isMinLength(1)),
+      }),
+    ]),
+  ),
+  usage: Schema.optional(Schema.StructWithRest(OpenResponses.OpenResponsesUsage, [JsonObject])),
+})
+
+export const make = (extension: OpenResponses.Extension): CompactOperation =>
+  Effect.fn("ResponsesCompaction.execute")(function* (request, executor, options) {
+    const route = request.model.route
+    const native = yield* OpenResponses.lowerConversation(request, extension)
+    const body = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Body))(
+      mergeJsonRecords(native, request.http?.body),
+    )
+    const url = Endpoint.render(route.endpoint, { request, body: native })
+    url.pathname = `${url.pathname.replace(/\/$/, "")}/compact`
+    const parts = yield* HttpTransport.jsonRequestParts({
+      request: LLMRequest.update(request, {
+        http: request.http === undefined ? undefined : new HttpOptions({ ...request.http, body: undefined }),
+      }),
+      body,
+      endpoint: Endpoint.path(url.toString()),
+      auth: route.auth,
+      encodeBody: Schema.encodeSync(Schema.fromJsonString(Body)),
+    })
+    const response = yield* executor.execute(
+      ProviderShared.jsonPost({ url: parts.url, body: parts.bodyText, headers: parts.headers }),
+      options?.http,
+    )
+    const text = yield* RequestExecutor.responseStream(response).pipe(
+      Stream.decodeText(),
+      Stream.runFold(
+        () => "",
+        (text, chunk) => text + chunk,
+      ),
+    )
+    const invalid = (message: string, cause?: unknown) =>
+      new AIError({
+        reason: new InvalidProviderOutputError({
+          route: route.id,
+          message,
+          body: text,
+          cause,
+          http: RequestExecutor.responseHttp(response),
+        }),
+      })
+    const result = yield* Schema.decodeUnknownEffect(Schema.fromJsonString(Response))(text).pipe(
+      Effect.mapError((cause) => invalid("Invalid compaction response", cause)),
+    )
+    if (!result.output.some((item) => item.type === "compaction"))
+      return yield* invalid("Compaction response did not contain a checkpoint")
+    return new CompactionResponse({
+      messages: result.output.map((item) => toMessage(item, request.model)),
+      usage: OpenResponses.mapUsage(result.usage, route.providerMetadataKey ?? String(request.model.provider)),
+    })
+  })
+
+function toMessage(item: (typeof Response.Type.output)[number], model: LLMRequest["model"]): Message {
+  if (item.type === "compaction")
+    return Message.assistant(
+      CompactionPart.make({ provider: model.provider, id: item.id ?? undefined, encrypted: item.encrypted_content }),
+    )
+
+  const key = model.route.providerMetadataKey ?? String(model.provider)
+  if (item.type === "reasoning") {
+    const summary = item.summary.length ? item.summary : [{ text: "" }]
+    return Message.assistant(
+      summary.map((part) => ({
+        type: "reasoning" as const,
+        text: part.text,
+        providerMetadata: { [key]: { itemId: item.id, reasoningEncryptedContent: item.encrypted_content } },
+      })),
+    )
+  }
+
+  return Message.make({
+    role: item.role,
+    providerMetadata: { [key]: { itemId: item.id, type: item.type, status: item.status, phase: item.phase } },
+    content: item.content.map((part): ContentPart => {
+      if (part.type === "input_text" || part.type === "output_text") return { type: "text", text: part.text }
+      if (part.type === "input_image")
+        return {
+          type: "media",
+          data: part.image_url,
+          mediaType: /^data:([^;,]+)/.exec(part.image_url)?.[1] ?? "image/*",
+          providerMetadata: part.detail === undefined ? undefined : { [key]: { detail: part.detail } },
+        }
+      const data = "file_url" in part ? part.file_url : part.file_data
+      return {
+        type: "media",
+        data,
+        filename: part.filename,
+        mediaType: /^data:([^;,]+)/.exec(data)?.[1] ?? "application/octet-stream",
+      }
+    }),
+  })
+}
+
+export * as ResponsesCompaction from "./responses-compaction.js"
@@ -4,6 +4,7 @@ import type { LLMRequest } from "../schema/index.js"
 import { OpenResponses } from "./open-responses.js"
 import { JsonObject, optionalNull, ProviderShared } from "./shared.js"
 import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"
+import { ResponsesCompaction } from "./utils/responses-compaction.js"

 const ADAPTER = "xai-responses"
 const NAME = "xAI Responses"
@@ -44,6 +45,10 @@ const extension = {

 const decodeBody = ProviderShared.validateWith(Schema.decodeUnknownEffect(XAIResponsesBody))
 const fromRequest = Effect.fn("XAIResponses.fromRequest")(function* (request: LLMRequest) {
+  if (request.providerOptions?.contextManagement !== undefined)
+    return yield* ProviderShared.invalidRequest(
+      "xAI requires explicit compaction through LLMClient.compact; automatic context management is not supported",
+    )
  return yield* decodeBody(yield* OpenResponses.fromRequestWithExtension(request, extension))
 })

@@ -84,4 +89,6 @@ export const protocol = Protocol.make({
  },
 })

+export const compact = ResponsesCompaction.make(extension)
+
 export * as XAIResponses from "./xai-responses.js"
@@ -1,9 +1,11 @@
-import type { RouteDefaultsInput } from "../route/client.js"
+import type { Route, RouteDefaultsInput } from "../route/client.js"
 import { Auth } from "../route/auth.js"
 import type { ProviderPackage } from "../provider-package.js"
 import { ProviderID, type ModelID } from "../schema/index.js"
 import * as BedrockConverse from "../protocols/bedrock-converse.js"
 import type { BedrockCredentials } from "../protocols/bedrock-converse.js"
+import { BedrockMessages } from "../protocols/bedrock-messages.js"
+import type { AnthropicMessages } from "../protocols/anthropic-messages.js"

 export const id = ProviderID.make("amazon-bedrock")

@@ -25,38 +27,40 @@ export interface Settings extends ProviderPackage.Settings {
  readonly region?: string
  readonly topP?: number
 }
-export const routes = [BedrockConverse.route]
+export const routes = [BedrockConverse.route, BedrockMessages.route]

 const bedrockBaseURL = (region: string) => `https://bedrock-runtime.${region}.amazonaws.com`

-const configuredRoute = (input: Config) => {
+const configuredRoute = <Body, Prepared>(route: Route<Body, Prepared>, input: Config) => {
  const { apiKey, credentials, region, baseURL, ...rest } = input
  const resolvedRegion = region ?? credentials?.region ?? "us-east-1"
-  return BedrockConverse.route.with({
+  return route.with({
    ...rest,
    provider: id,
-    providerMetadataKey: "bedrock",
+    providerMetadataKey: route.providerMetadataKey,
    endpoint: { baseURL: baseURL ?? bedrockBaseURL(resolvedRegion) },
    auth: apiKey === undefined ? BedrockConverse.sigV4Auth(credentials) : Auth.bearer(apiKey),
  })
 }

 export const configure = (input: Config = {}) => {
-  const route = configuredRoute(input)
+  const route = configuredRoute(BedrockConverse.route, input)
+  const messages = configuredRoute(BedrockMessages.route, input)
  return {
    id,
    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    messages: (modelID: string | ModelID) => messages.model<AnthropicMessages.ProviderOptionsInput>({ id: modelID }),
    configure,
  }
 }

 export const provider = configure()
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+const config = (settings: Settings): Config => {
  if (settings.auth === "bearer" && settings.apiKey === undefined)
    throw new Error("Amazon Bedrock bearer auth requires apiKey")
  if (settings.auth === "sigv4" && settings.apiKey !== undefined)
    throw new Error("Amazon Bedrock SigV4 auth does not accept apiKey")
-  return configure({
+  return {
    apiKey: settings.auth === "sigv4" ? undefined : settings.apiKey,
    baseURL: settings.baseURL,
    credentials: settings.credentials,
@@ -64,5 +68,13 @@ export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, se
    headers: settings.headers === undefined ? undefined : { ...settings.headers },
    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
    region: settings.region,
-  }).model(modelID)
+  }
 }
+
+export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
+  configure(config(settings)).model(modelID)
+export const messagesModel: ProviderPackage.Definition<
+  Settings & { readonly providerOptions?: AnthropicMessages.ProviderOptionsInput },
+  AnthropicMessages.ProviderOptionsInput
+>["model"] = (modelID, settings) =>
+  configure({ ...config(settings), providerOptions: settings.providerOptions }).messages(modelID)
@@ -0,0 +1 @@
+export { messagesModel as model } from "../amazon-bedrock.js"
@@ -57,7 +57,9 @@ const route = Route.make({
  }),
  endpoint: Endpoint.path(({ request }) => `/${request.model.id}:streamRawPredict`),
  auth: Auth.none,
-  framing: AnthropicMessages.framing,
+  transport: AnthropicMessages.transport<
+    Omit<AnthropicMessages.AnthropicMessagesBody, "model"> & { readonly anthropic_version: typeof VERSION }
+  >(),
  headers: () => ({ "anthropic-version": HEADER_VERSION }),
 })

@@ -1,10 +1,12 @@
 import { mergeProviderOptions, type ProviderOptions } from "../schema/index.js"
 import type { OpenAIServiceTier } from "../protocols/utils/openai-options.js"
 import type { Options } from "../protocols/utils/open-responses-options.js"
+import type { ContextManagement } from "../protocols/openai-responses.js"

 export type { OpenAIResponseIncludable, OpenAIServiceTier } from "../protocols/utils/openai-options.js"

 export type OpenAIOptionsInput = Omit<Options, "serviceTier"> & {
+  readonly contextManagement?: ContextManagement
  readonly serviceTier?: OpenAIServiceTier
  readonly [key: string]: unknown
 }
@@ -13,7 +13,7 @@ import type { ProviderPackage } from "../provider-package.js"

 export const id = ProviderID.make("xai")

-export type XAIProviderOptionsInput = OpenAIOptionsInput
+export type XAIProviderOptionsInput = OpenAIOptionsInput & { readonly contextManagement?: never }

 export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
  ProviderAuthOption<"optional"> & {
@@ -32,6 +32,7 @@ export type { XAIImageOptions } from "../protocols/xai-images.js"
 const RESPONSES_WEBSOCKET_ROTATE_AFTER_MS = 24 * 60 * 1000

 const responsesRoute = Route.make({
+  compact: XAIResponses.compact,
  id: "openai-responses",
  provider: id,
  providerMetadataKey: "xai",
@@ -13,6 +13,7 @@ import * as ProviderShared from "../protocols/shared.js"
 import type { ProtocolID, ProviderOptions } from "../schema/index.js"
 import {
  AIError,
+  CompactionResponse,
  AIErrorReason,
  GenerationOptions,
  HttpOptions,
@@ -35,6 +36,7 @@ export interface RouteBody<Body> {
 }

 export interface Route<Body, Prepared = unknown> {
+  readonly compact?: CompactOperation
  readonly id: string
  readonly provider?: ProviderID
  /** ProviderMetadata namespace emitted and consumed by this route. */
@@ -42,6 +44,8 @@ export interface Route<Body, Prepared = unknown> {
  readonly protocol: ProtocolID
  readonly endpoint: Endpoint.Definition<Body>
  readonly auth: Auth.Definition
+  /** Deployment headers resolved once for every operation, before transport authentication. */
+  readonly headers?: (input: { readonly request: LLMRequest }) => Record<string, string>
  readonly transport: Transport<Body, Prepared, unknown>
  readonly defaults: RouteDefaults
  readonly body: RouteBody<Body>
@@ -150,6 +154,10 @@ export const httpOptions = (input: HttpOptionsInput | undefined) => {
 }

 export interface Interface {
+  readonly compact: (
+    request: LLMRequest,
+    options?: Pick<StreamOptions, "http">,
+  ) => Effect.Effect<CompactionResponse, AIError>
  readonly stream: StreamMethod
  readonly generate: GenerateMethod
 }
@@ -167,6 +175,12 @@ export interface GenerateMethod {
  (request: LLMRequest, options?: StreamOptions): Effect.Effect<LLMResponse, AIError>
 }

+export type CompactOperation = (
+  request: LLMRequest,
+  executor: RequestExecutor.Interface,
+  options?: Pick<StreamOptions, "http">,
+) => Effect.Effect<CompactionResponse, AIError>
+
 export class Service extends Context.Service<Service, Interface>()("@opencode/LLMClient") {}

 const resolveRequestOptions = (request: LLMRequest) => {
@@ -187,6 +201,7 @@ const resolveRequestOptions = (request: LLMRequest) => {
 }

 export interface MakeInput<Body, Frame, Event, State> {
+  readonly compact?: CompactOperation
  /** Route id used in diagnostics and prepared request metadata. */
  readonly id: string
  /** Provider identity for route-owned model construction. */
@@ -208,6 +223,7 @@ export interface MakeInput<Body, Frame, Event, State> {
 }

 export interface MakeTransportInput<Body, Prepared, Frame, Event, State> {
+  readonly compact?: CompactOperation
  /** Route id used in diagnostics and prepared request metadata. */
  readonly id: string
  /** Provider identity for route-owned model construction. */
@@ -283,12 +299,14 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(

  const build = (routeInput: BuiltRouteInput): Route<Body, Prepared> => {
    const route: Route<Body, Prepared> = {
+      compact: routeInput.compact,
      id: routeInput.id,
      provider: routeInput.provider === undefined ? undefined : ProviderID.make(routeInput.provider),
      providerMetadataKey: routeInput.providerMetadataKey,
      protocol: protocol.id,
      endpoint: routeInput.endpoint,
      auth: routeInput.auth ?? Auth.none,
+      headers: routeInput.headers,
      transport: routeInput.transport,
      defaults: routeInput.defaults ?? {},
      body: protocol.body,
@@ -318,7 +336,6 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
          endpoint: routeInput.endpoint,
          auth: routeInput.auth ?? Auth.none,
          encodeBody,
-          headers: routeInput.headers,
          middleware: options?.http,
          webSocket: options?.webSocket,
        }),
@@ -435,6 +452,7 @@ export function make<Body, Prepared, Frame, Event, State>(
  if ("transport" in input) return makeFromTransport(input)
  const protocol = input.protocol
  return makeFromTransport({
+    compact: input.compact,
    id: input.id,
    provider: input.provider,
    providerMetadataKey: input.providerMetadataKey,
@@ -447,11 +465,19 @@ export function make<Body, Prepared, Frame, Event, State>(
  })
 }

-const compile = Effect.fn("LLM.compile")(function* (request: LLMRequest, options?: StreamOptions) {
+const prepareRequest = (request: LLMRequest) => {
  const original = applyCachePolicy(resolveRequestOptions(request))
  const sanitized = LLMRequest.update(original, sanitizeSurrogates({ ...LLMRequest.input(original), model: undefined }))
  const tools = [...new Map(sanitized.tools.map((tool) => [tool.name, tool])).values()]
  const resolved = tools.length === sanitized.tools.length ? sanitized : LLMRequest.update(sanitized, { tools })
+  const headers = resolved.model.route.headers?.({ request: resolved })
+  return headers === undefined
+    ? resolved
+    : LLMRequest.update(resolved, { http: mergeHttpOptions(new HttpOptions({ headers }), resolved.http) })
+}
+
+const compile = Effect.fn("LLM.compile")(function* (request: LLMRequest, options?: StreamOptions) {
+  const resolved = prepareRequest(request)
  const route = resolved.model.route

  const body = yield* route.body
@@ -510,6 +536,15 @@ export function generate(request: LLMRequest, options?: StreamOptions): Effect.E
  })
 }

+export const compact = (
+  request: LLMRequest,
+  options?: Pick<StreamOptions, "http">,
+): Effect.Effect<CompactionResponse, AIError, Service> =>
+  Effect.gen(function* () {
+    const client = yield* Service
+    return yield* client.compact(request, options)
+  })
+
 export const streamRequest = (request: LLMRequest, options?: StreamOptions) =>
  Stream.unwrap(
    Effect.gen(function* () {
@@ -520,16 +555,28 @@ export const streamRequest = (request: LLMRequest, options?: StreamOptions) =>
 export const layer: Layer.Layer<Service, never, RequestExecutor.Service> = Layer.effect(
  Service,
  Effect.gen(function* () {
-    const stream = streamRequestWith({
-      http: yield* RequestExecutor.Service,
+    const executor = yield* RequestExecutor.Service
+    const stream = streamRequestWith({ http: executor })
+    return Service.of({
+      stream,
+      generate: generateWith(stream),
+      compact: (request, options) =>
+        Effect.suspend(() => {
+          const operation = request.model.route.compact
+          if (!operation)
+            return ProviderShared.invalidRequest(
+              `${request.model.provider}/${request.model.route.id} does not support explicit compaction`,
+            )
+          return operation(prepareRequest(request), executor, options)
+        }),
    })
-    return Service.of({ stream, generate: generateWith(stream) })
  }),
 )

 export const Route = { make } as const

 export const LLMClient = {
+  compact,
  Service,
  layer,
  stream,
@@ -3,6 +3,7 @@ import { LLM } from "@opencode-ai/schema/llm"
 import { ContentBlockID, ToolCallID } from "./ids.js"
 import {
  Message,
+  CompactionPart,
  ProviderMetadata,
  ToolCallPart,
  ToolOutput,
@@ -62,6 +63,8 @@ export { ProviderMetadata } from "./messages.js"
 * Matches the same escape-hatch field on `LLMEvent`.
 */
 export class Usage extends Schema.Class<Usage>("AI.Usage")({
+  /** Effective input size of the final message iteration, when reported; not billed totals. */
+  contextTokens: Schema.optional(Schema.Number),
  inputTokens: Schema.optional(Schema.Number),
  outputTokens: Schema.optional(Schema.Number),
  nonCachedInputTokens: Schema.optional(Schema.Number),
@@ -72,7 +75,7 @@ export class Usage extends Schema.Class<Usage>("AI.Usage")({
  providerMetadata: Schema.optional(ProviderMetadata),
 }) {
  /**
-   * Visible output tokens — `outputTokens` minus `reasoningTokens`, clamped
+   * Non-reasoning output tokens (including compaction summaries) — `outputTokens` minus `reasoningTokens`, clamped
   * to zero. The one place subtraction happens in this contract; the clamp
   * means a provider reporting `reasoningTokens > outputTokens` produces a
   * harmless zero rather than a negative that crashes downstream schemas.
@@ -88,6 +91,12 @@ export class Usage extends Schema.Class<Usage>("AI.Usage")({

 export type UsageInput = Usage | ConstructorParameters<typeof Usage>[0]

+/** A replacement context window. Replace prior history with these messages. */
+export class CompactionResponse extends Schema.Class<CompactionResponse>("LLM.CompactionResponse")({
+  messages: Schema.Array(Message),
+  usage: Schema.optional(Usage),
+}) {}
+
 export const StepStart = Schema.Struct({
  type: Schema.tag("step-start"),
  index: Schema.Number,
@@ -241,6 +250,7 @@ export const ProviderErrorEvent = Schema.Struct({
 export type ProviderErrorEvent = Schema.Schema.Type<typeof ProviderErrorEvent>

 const llmEventTagged = Schema.Union([
+  CompactionPart,
  StepStart,
  TextStart,
  TextDelta,
@@ -274,6 +284,7 @@ const toolCallID = (value: ToolCallID | string) => ToolCallID.make(value)
 * `events.filter(LLMEvent.guards["tool-call"])`.
 */
 export const LLMEvent = Object.assign(llmEventTagged, {
+  compaction: CompactionPart.make,
  stepStart: StepStart.make,
  textStart: (input: WithID<TextStart, ContentBlockID>) => TextStart.make({ ...input, id: contentBlockID(input.id) }),
  textDelta: (input: WithID<TextDelta, ContentBlockID>) => TextDelta.make({ ...input, id: contentBlockID(input.id) }),
@@ -311,6 +322,7 @@ export const LLMEvent = Object.assign(llmEventTagged, {
    }),
  providerError: ProviderErrorEvent.make,
  is: {
+    compaction: llmEventTagged.guards.compaction,
    stepStart: llmEventTagged.guards["step-start"],
    textStart: llmEventTagged.guards["text-start"],
    textDelta: llmEventTagged.guards["text-delta"],
@@ -333,10 +345,10 @@ export const LLMEvent = Object.assign(llmEventTagged, {
 export type LLMEvent = Schema.Schema.Type<typeof llmEventTagged>

 /** Joins deltas per fragment, letting an authoritative end value replace that fragment's accumulated deltas. */
-const joinFragments = <Delta extends { id: string; text: string }, End extends { id: string; text?: string }>(
+const joinFragments = (
  events: ReadonlyArray<LLMEvent>,
-  isDelta: (event: LLMEvent) => event is Extract<LLMEvent, Delta>,
-  isEnd: (event: LLMEvent) => event is Extract<LLMEvent, End>,
+  isDelta: (event: LLMEvent) => event is LLMEvent & { id: string; text: string },
+  isEnd: (event: LLMEvent) => event is LLMEvent & { id: string; text?: string },
 ) => {
  const order: string[] = []
  const parts = new Map<string, string>()
@@ -563,6 +575,8 @@ const reduceToolCall = (state: ResponseState, event: ToolCall): ResponseState =>
 const reduceResponseState = (state: ResponseState, event: LLMEvent): ResponseState => {
  const next = appendEvent(state, event)
  switch (event.type) {
+    case "compaction":
+      return appendContent(next, event)
    case "text-start":
      return ensureText(next, event.id, event.providerMetadata)
    case "text-delta":
@@ -9,6 +9,7 @@ import {
  LanguageModelSchema,
  ProviderOptions,
 } from "./options.js"
+import { ProviderID } from "./ids.js"

 export const MessageRole = Schema.Literals(["system", "user", "assistant", "tool"])
 export type MessageRole = Schema.Schema.Type<typeof MessageRole>
@@ -52,6 +53,7 @@ export const MediaPart = Schema.Struct({
  filename: Schema.optional(Schema.String),
  cache: Schema.optional(CacheHint),
  metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
+  providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Content.Media" })
 export type MediaPart = Schema.Schema.Type<typeof MediaPart>

@@ -185,9 +187,40 @@ export const ReasoningPart = Schema.Struct({
 }).annotate({ identifier: "LLM.Content.Reasoning" })
 export type ReasoningPart = Schema.Schema.Type<typeof ReasoningPart>

-export const ContentPart = Schema.Union([TextPart, MediaPart, ToolCallPart, ToolResultPart, ReasoningPart]).pipe(
-  Schema.toTaggedUnion("type"),
-)
+/** A provider-generated context checkpoint, distinct from visible assistant text. */
+type CompactionContent =
+  | { readonly encrypted: string; readonly text?: never }
+  | { readonly text: string | null; readonly encrypted?: never }
+
+const compactionPartSchema = Schema.Struct({
+  type: Schema.Literal("compaction"),
+  provider: ProviderID,
+  id: Schema.optional(Schema.String),
+  encrypted: Schema.optional(Schema.String),
+  /** Null means the provider failed to produce a summary; prior history must be retained. */
+  text: Schema.optional(Schema.NullOr(Schema.String)),
+})
+  .pipe(
+    Schema.refine(
+      (part): part is typeof part & CompactionContent => (part.encrypted !== undefined) !== (part.text !== undefined),
+      { message: "Compaction requires either encrypted content or a summary" },
+    ),
+  )
+  .annotate({ identifier: "LLM.Content.Compaction" })
+export type CompactionPart = typeof compactionPartSchema.Type
+export const CompactionPart = Object.assign(compactionPartSchema, {
+  make: (input: Omit<CompactionPart, "type" | "encrypted" | "text"> & CompactionContent): CompactionPart =>
+    Schema.decodeUnknownSync(compactionPartSchema)({ type: "compaction", ...input }),
+})
+
+export const ContentPart = Schema.Union([
+  TextPart,
+  MediaPart,
+  ToolCallPart,
+  ToolResultPart,
+  ReasoningPart,
+  CompactionPart,
+]).pipe(Schema.toTaggedUnion("type"))
 export type ContentPart = Schema.Schema.Type<typeof ContentPart>

 export class Message extends Schema.Class<Message>("LLM.Message")({
@@ -195,6 +228,7 @@ export class Message extends Schema.Class<Message>("LLM.Message")({
  role: MessageRole,
  content: Schema.Array(ContentPart),
  metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
+  providerMetadata: Schema.optional(ProviderMetadata),
  native: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
 }) {}

@@ -4,6 +4,7 @@ import { LLMClient } from "./route/client.js"
 import {
  LLMEvent,
  LLMResponse,
+  CompactionResponse,
  type FinishReasonDetails,
  type AIError,
  type LLMRequest,
@@ -133,6 +134,16 @@ const make = (options: LayerOptions) =>
        }
      })
    const test = Test.of({
+      compact: (request) =>
+        stream(request).pipe(
+          Stream.runFold(LLMResponse.empty, LLMResponse.reduce),
+          Effect.flatMap((state) => {
+            const response = LLMResponse.complete(state)
+            if (!response?.message.content.some((part) => part.type === "compaction"))
+              return Effect.die("TestLLM compaction response must contain a checkpoint and terminal finish event")
+            return Effect.succeed(new CompactionResponse({ messages: [response.message], usage: response.usage }))
+          }),
+        ),
      stream,
      generate: (request) =>
        stream(request).pipe(
@@ -56,6 +56,7 @@ function normalizeToolMessage(message: Message, pending: Map<string, ToolCallPar
    role: message.role,
    content,
    metadata: message.metadata,
+    providerMetadata: message.providerMetadata,
    native: message.native,
  })
 }
@@ -126,6 +126,38 @@ describe("applyCachePolicy", () => {
    }),
  )

+  for (const fixture of [
+    { name: "default", cache: undefined, control: { type: "ephemeral" } },
+    { name: "auto", cache: "auto", control: { type: "ephemeral" } },
+    {
+      name: "explicit one-hour",
+      cache: { tools: true, system: true, messages: { tail: 1 }, ttlSeconds: 3600 },
+      control: { type: "ephemeral", ttl: "1h" },
+    },
+    { name: "disabled", cache: "none", control: undefined },
+  ] as const) {
+    it.effect(`Bedrock Messages respects ${fixture.name} caching`, () =>
+      Effect.gen(function* () {
+        const prepared = yield* compileRequest(
+          LLM.request({
+            model: AmazonBedrock.configure({ apiKey: "test" }).messages("anthropic.claude-opus-4-6-v1"),
+            system: "Stable instructions",
+            tools: [
+              { name: "lookup", description: "Look up a value", inputSchema: { type: "object", properties: {} } },
+            ],
+            prompt: "hello",
+            cache: fixture.cache,
+          }),
+        )
+        expect(prepared.body).toMatchObject({
+          tools: [{ name: "lookup", cache_control: fixture.control }],
+          system: [{ type: "text", text: "Stable instructions", cache_control: fixture.control }],
+          messages: [{ role: "user", content: [{ type: "text", text: "hello", cache_control: fixture.control }] }],
+        })
+      }),
+    )
+  }
+
  it.effect("'auto' is a no-op on Gemini (out-of-band caching protocol)", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -0,0 +1,97 @@
+import { expect } from "bun:test"
+import { Effect } from "effect"
+import { FetchHttpClient } from "effect/unstable/http"
+import { LLM, LLMRequest, Message } from "../src/index.js"
+import { LLMClient } from "../src/route/client.js"
+import { OpenAI } from "../src/providers.js"
+import { testEffect } from "./lib/effect.js"
+import { runtimeLayer } from "./lib/http.js"
+import { sseEvents } from "./lib/sse.js"
+
+testEffect(runtimeLayer(FetchHttpClient.layer)).live("compaction and a tool loop work end to end over HTTP", () =>
+  Effect.gen(function* () {
+    const checkpoint = { type: "compaction", id: "cmp_local", encrypted_content: "opaque-local-state" }
+    const calls: string[] = []
+    const server = yield* Effect.acquireRelease(
+      Effect.sync(() =>
+        Bun.serve({
+          hostname: "127.0.0.1",
+          port: 0,
+          async fetch(request) {
+            const path = new URL(request.url).pathname
+            calls.push(path)
+            const body = await request.json()
+            expect(request.headers.get("authorization")).toBe("Bearer fixture")
+            if (path === "/v1/responses/compact") {
+              expect(body.stream).toBeUndefined()
+              return Response.json({
+                object: "response.compaction",
+                output: [checkpoint],
+                usage: { input_tokens: 100, output_tokens: 10, total_tokens: 110 },
+              })
+            }
+            expect(body.input[0]).toEqual(checkpoint)
+            expect(body.stream).toBe(true)
+            if (calls.length === 2)
+              return new Response(
+                sseEvents(
+                  {
+                    type: "response.output_item.done",
+                    item: { type: "function_call", id: "fc_1", call_id: "call_1", name: "lookup", arguments: "{}" },
+                  },
+                  { type: "response.completed", response: { id: "resp_1" } },
+                ),
+                { headers: { "content-type": "text/event-stream" } },
+              )
+            expect(body.input.at(-2)).toMatchObject({ type: "function_call", call_id: "call_1" })
+            expect(body.input.at(-1)).toEqual({ type: "function_call_output", call_id: "call_1", output: "42" })
+            const output = sseEvents(
+              { type: "response.output_item.added", item: { type: "message", id: "msg_1" } },
+              { type: "response.output_text.delta", item_id: "msg_1", delta: "The answer is 42." },
+              {
+                type: "response.output_item.done",
+                item: { type: "message", id: "msg_1", content: [{ type: "output_text", text: "The answer is 42." }] },
+              },
+              { type: "response.completed", response: { id: "resp_2" } },
+            )
+            return new Response(
+              new ReadableStream({
+                start(controller) {
+                  controller.enqueue(new TextEncoder().encode(output.slice(0, 37)))
+                  controller.enqueue(new TextEncoder().encode(output.slice(37)))
+                  controller.close()
+                },
+              }),
+              { headers: { "content-type": "text/event-stream" } },
+            )
+          },
+        }),
+      ),
+      (server) => Effect.sync(() => server.stop(true)),
+    )
+    const model = OpenAI.configure({ apiKey: "fixture", baseURL: `http://127.0.0.1:${server.port}/v1` }).responses(
+      "fixture",
+    )
+    const request = LLM.request({
+      model,
+      prompt: "original",
+      tools: [{ name: "lookup", description: "Lookup a number", inputSchema: { type: "object", properties: {} } }],
+    })
+    const compacted = yield* LLMClient.compact(request)
+    const messages = [...compacted.messages, Message.user("Look up the answer")]
+    const first = yield* LLMClient.generate(LLMRequest.update(request, { messages }))
+    expect(first.toolCalls).toHaveLength(1)
+    const call = first.toolCalls[0]!
+    const last = yield* LLMClient.generate(
+      LLMRequest.update(request, {
+        messages: [
+          ...messages,
+          first.message,
+          Message.tool({ id: call.id, name: call.name, result: "42", resultType: "text" }),
+        ],
+      }),
+    )
+    expect(last.text).toBe("The answer is 42.")
+    expect(calls).toEqual(["/v1/responses/compact", "/v1/responses", "/v1/responses"])
+  }),
+)
@@ -0,0 +1,49 @@
+import { expect, test } from "bun:test"
+import { Schema } from "effect"
+import { CompactionPart, LLMEvent, LLMResponse, Message, ProviderID } from "../src/schema/index.js"
+
+test("compaction survives event assembly and message serialization without becoming text", () => {
+  const part = CompactionPart.make({
+    provider: ProviderID.make("openai"),
+    id: "cmp_1",
+    encrypted: "opaque",
+  })
+  const response = LLMResponse.fromEvents([
+    LLMEvent.textStart({ id: "before" }),
+    LLMEvent.textDelta({ id: "before", text: "Before" }),
+    LLMEvent.textEnd({ id: "before" }),
+    part,
+    LLMEvent.textStart({ id: "after" }),
+    LLMEvent.textDelta({ id: "after", text: "After" }),
+    LLMEvent.textEnd({ id: "after" }),
+    LLMEvent.finish({ reason: { normalized: "stop" } }),
+  ])!
+  expect(response.message.content.map((part) => part.type)).toEqual(["text", "compaction", "text"])
+  expect(response.text).toBe("BeforeAfter")
+  expect(response.reasoning).toBe("")
+  expect(response.events.filter(LLMEvent.is.compaction)).toEqual([part])
+  const codec = Schema.fromJsonString(Message)
+  expect(Schema.decodeSync(codec)(Schema.encodeSync(codec)(response.message))).toEqual(response.message)
+})
+
+test("compaction requires exactly one typed representation", () => {
+  const provider = ProviderID.make("anthropic")
+  expect(CompactionPart.make({ provider, text: null })).toEqual({ type: "compaction", provider, text: null })
+  const decode = Schema.decodeUnknownSync(CompactionPart)
+  expect(() => decode({ type: "compaction", provider })).toThrow()
+  expect(() => decode({ type: "compaction", provider, text: "summary", encrypted: "opaque" })).toThrow()
+})
+
+test("tagged content and event guards accept both checkpoint representations", () => {
+  for (const part of [
+    CompactionPart.make({ provider: ProviderID.make("openai"), encrypted: "opaque" }),
+    CompactionPart.make({ provider: ProviderID.make("anthropic"), text: "summary" }),
+    CompactionPart.make({ provider: ProviderID.make("anthropic"), text: null }),
+  ]) {
+    expect(LLMEvent.is.compaction(part)).toBe(true)
+    expect(LLMEvent.guards.compaction(part)).toBe(true)
+    const codec = Schema.fromJsonString(Message)
+    const message = Message.assistant(part)
+    expect(Schema.decodeSync(codec)(Schema.encodeSync(codec)(message))).toEqual(message)
+  }
+})
@@ -0,0 +1,69 @@
+import { Effect } from "effect"
+import { CompactionPart, LLM, LLMClient, LLMEvent, Message, ProviderID } from "../../src/index.js"
+import { OpenAI, Anthropic, AmazonBedrock } from "../../src/providers.js"
+
+const openai = OpenAI.configure({
+  apiKey: "test",
+  providerOptions: { contextManagement: [{ type: "compaction", compactThreshold: 100000 }] },
+}).responses("gpt-5.3-codex")
+LLMClient.compact(LLM.request({ model: openai, prompt: "hello" }))
+
+const checkpoint = CompactionPart.make({ provider: ProviderID.make("openai"), id: "cmp_1", encrypted: "opaque" })
+const provider = ProviderID.make("anthropic")
+CompactionPart.make({ provider, text: "summary" })
+CompactionPart.make({ provider, text: null })
+// @ts-expect-error A checkpoint must have a representation.
+CompactionPart.make({ provider })
+// @ts-expect-error Encrypted and summary representations are mutually exclusive.
+CompactionPart.make({ provider, encrypted: "opaque", text: "summary" })
+// @ts-expect-error A failed summary cannot also carry encrypted content.
+LLMEvent.compaction({ provider, encrypted: "opaque", text: null })
+// @ts-expect-error The canonical message type also enforces the invariant.
+Message.assistant({ type: "compaction", provider })
+if (checkpoint.encrypted !== undefined) {
+  checkpoint.encrypted satisfies string
+  checkpoint.text satisfies undefined
+}
+if (checkpoint.text !== undefined) {
+  checkpoint.text satisfies string | null
+  checkpoint.encrypted satisfies undefined
+}
+checkpoint.encrypted
+// @ts-expect-error Compaction parts do not contain a generic provider payload.
+checkpoint.value
+LLMClient.compact(LLM.request({ model: openai, prompt: "hello" })).pipe(
+  Effect.map((result) => {
+    result.messages
+    // @ts-expect-error Compaction returns replacement history, not a synthetic assistant message.
+    result.message
+  }),
+)
+LLM.request({
+  model: openai,
+  providerOptions: {
+    // @ts-expect-error A token threshold is numeric.
+    contextManagement: [{ type: "compaction", compactThreshold: "100000" }],
+  },
+})
+for (const model of [
+  Anthropic.configure().model("claude-opus-4-6"),
+  AmazonBedrock.configure().messages("anthropic.claude-opus-4-6-v1"),
+]) {
+  LLM.request({
+    model,
+    providerOptions: {
+      contextManagement: {
+        edits: [
+          { type: "compact_20260112", pauseAfterCompaction: true, instructions: "Summarize without using tools" },
+        ],
+      },
+    },
+  })
+  LLM.request({
+    model,
+    providerOptions: {
+      // @ts-expect-error A pause setting is boolean.
+      contextManagement: { edits: [{ type: "compact_20260112", pauseAfterCompaction: "yes" }] },
+    },
+  })
+}
@@ -0,0 +1,170 @@
+import { expect } from "bun:test"
+import { Effect, Schema } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient } from "../../src/route/client.js"
+import { Anthropic, GoogleVertexMessages } from "../../src/providers/index.js"
+import { testEffect } from "../lib/effect.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
+
+for (const fixture of [
+  {
+    name: "empty iterations fall back to top-level usage",
+    usage: { input_tokens: 2, output_tokens: 3, cache_read_input_tokens: null, iterations: [] },
+    expected: { inputTokens: 2, outputTokens: 3, totalTokens: 5, contextTokens: undefined },
+  },
+  {
+    name: "compaction-only usage has no post-compaction context size",
+    usage: {
+      input_tokens: 0,
+      output_tokens: 0,
+      iterations: [{ type: "compaction", input_tokens: 7, cache_read_input_tokens: 3, output_tokens: 2 }],
+    },
+    expected: { inputTokens: 10, outputTokens: 2, totalTokens: 12, contextTokens: undefined },
+  },
+  {
+    name: "partially reported iterations preserve known totals",
+    usage: {
+      iterations: [
+        { type: "compaction", input_tokens: 7, cache_creation_input_tokens: 2 },
+        { type: "message", output_tokens: 3 },
+      ],
+    },
+    expected: { inputTokens: 9, outputTokens: 3, totalTokens: 12, contextTokens: undefined },
+  },
+  {
+    name: "missing counters remain unknown rather than zero",
+    usage: { iterations: [{ type: "message" }] },
+    expected: { inputTokens: undefined, outputTokens: undefined, totalTokens: undefined, contextTokens: undefined },
+  },
+]) {
+  testEffect(
+    fixedResponse(
+      sseEvents(
+        { type: "message_start", message: { usage: fixture.usage } },
+        { type: "message_delta", delta: { stop_reason: "end_turn" } },
+        { type: "message_stop" },
+      ),
+    ),
+  ).effect(fixture.name, () =>
+    Effect.gen(function* () {
+      const result = yield* LLMClient.generate(
+        LLM.request({
+          model: Anthropic.configure({ apiKey: "test" }).model("claude-opus-4-6"),
+          prompt: "hello",
+        }),
+      )
+      expect(result.usage).toMatchObject(fixture.expected)
+    }),
+  )
+}
+
+for (const model of [
+  Anthropic.configure({ apiKey: "test" }).model("claude-opus-4-6"),
+  GoogleVertexMessages.configure({ accessToken: "test", project: "test" }).model("claude-opus-4-6"),
+]) {
+  for (const summary of ["Summary of the conversation", null]) {
+    const block = { type: "compaction", content: summary }
+    testEffect(
+      dynamicResponse(({ request, text, respond }) =>
+        Effect.sync(() => {
+          const body = JSON.parse(text)
+          expect(request.headers["anthropic-beta"]).toBe("existing-beta,compact-2026-01-12")
+          if (body.messages.length === 1) {
+            expect(body.context_management.edits).toEqual([
+              {
+                type: "compact_20260112",
+                trigger: { type: "input_tokens", value: 50000 },
+                pause_after_compaction: true,
+              },
+            ])
+          }
+          if (body.messages.length > 1) {
+            expect(body.messages[1].content).toEqual([block])
+            expect(body.context_management).toBeUndefined()
+          }
+          return respond(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 50000, output_tokens: 0 } } },
+              { type: "content_block_start", index: 0, content_block: { type: "compaction", content: null } },
+              { type: "content_block_delta", index: 0, delta: { type: "compaction_delta", content: summary } },
+              { type: "content_block_stop", index: 0 },
+              {
+                type: "message_delta",
+                delta: { stop_reason: "compaction" },
+                usage: {
+                  input_tokens: 1000,
+                  output_tokens: 5,
+                  iterations: [
+                    { type: "compaction", input_tokens: 50000, output_tokens: 1000, cache_read_input_tokens: 10 },
+                    { type: "message", input_tokens: 1000, output_tokens: 5 },
+                  ],
+                },
+              },
+              { type: "message_stop" },
+            ),
+            { headers: { "content-type": "text/event-stream" } },
+          )
+        }),
+      ),
+    ).effect(
+      `${model.provider} replays ${summary === null ? "failed" : "successful"} compaction with billing and context usage`,
+      () =>
+        Effect.gen(function* () {
+          const request = LLM.request({
+            model,
+            prompt: "hello",
+            http: { headers: { "anthropic-beta": "existing-beta" } },
+            providerOptions: {
+              contextManagement: {
+                edits: [
+                  {
+                    type: "compact_20260112",
+                    trigger: { type: "input_tokens", value: 50000 },
+                    pauseAfterCompaction: true,
+                  },
+                ],
+              },
+            },
+          })
+          const first = yield* LLMClient.generate(request)
+          expect(first.finishReason.raw).toBe("compaction")
+          expect(first.message.content).toEqual([{ type: "compaction", provider: model.provider, text: summary }])
+          expect(first.text).toBe("")
+          expect(first.usage?.inputTokens).toBe(51010)
+          expect(first.usage?.outputTokens).toBe(1005)
+          expect(first.usage?.totalTokens).toBe(52015)
+          expect(first.usage?.contextTokens).toBe(1000)
+          const codec = Schema.fromJsonString(Message)
+          const message = Schema.decodeSync(codec)(Schema.encodeSync(codec)(first.message))
+          yield* LLMClient.generate(
+            LLMRequest.update(request, {
+              providerOptions: {},
+              messages: [...request.messages, message, Message.user("continue")],
+            }),
+          )
+        }),
+    )
+  }
+}
+
+for (const events of [
+  [{ type: "content_block_start", index: 0, content_block: { type: "compaction", content: 42 } }],
+  [{ type: "content_block_delta", index: 0, delta: { type: "compaction_delta", content: "no start" } }],
+  [
+    { type: "content_block_start", index: 0, content_block: { type: "compaction", content: null } },
+    { type: "message_stop" },
+  ],
+]) {
+  testEffect(fixedResponse(sseEvents(...events))).effect(
+    `rejects malformed compaction lifecycle: ${JSON.stringify(events)}`,
+    () =>
+      Effect.gen(function* () {
+        const error = yield* LLMClient.generate(
+          LLM.request({ model: Anthropic.configure({ apiKey: "test" }).model("claude-opus-4-6"), prompt: "hello" }),
+        ).pipe(Effect.flip)
+        expect(error.reason._tag).toBe("InvalidProviderOutput")
+        expect(error.reason.http?.status).toBe(200)
+      }),
+  )
+}
@@ -0,0 +1,89 @@
+import { EventStreamCodec } from "@smithy/eventstream-codec"
+import { fromUtf8, toUtf8 } from "@smithy/util-utf8"
+import { expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient } from "../../src/route/client.js"
+import { AmazonBedrock } from "../../src/providers/index.js"
+import { testEffect } from "../lib/effect.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
+
+const codec = new EventStreamCodec(toUtf8, fromUtf8)
+const frame = (event: object) =>
+  codec.encode({
+    headers: { ":message-type": { type: "string", value: "event" }, ":event-type": { type: "string", value: "chunk" } },
+    body: new TextEncoder().encode(JSON.stringify({ bytes: Buffer.from(JSON.stringify(event)).toString("base64") })),
+  })
+const response = Buffer.concat(
+  [
+    { type: "message_start", message: { usage: { input_tokens: 60000 } } },
+    { type: "content_block_start", index: 0, content_block: { type: "compaction", content: null } },
+    { type: "content_block_delta", index: 0, delta: { type: "compaction_delta", content: "Summary" } },
+    { type: "content_block_stop", index: 0 },
+    { type: "content_block_start", index: 1, content_block: { type: "text", text: "" } },
+    { type: "content_block_delta", index: 1, delta: { type: "text_delta", text: "Hello" } },
+    { type: "content_block_stop", index: 1 },
+    { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 10 } },
+    { type: "message_stop" },
+  ].map(frame),
+)
+
+for (const auth of [
+  { apiKey: "test" },
+  { credentials: { accessKeyId: "test", secretAccessKey: "test", region: "us-west-2" } },
+]) {
+  testEffect(
+    dynamicResponse(({ request, text, respond }) =>
+      Effect.sync(() => {
+        expect(request.url).toBe(
+          "https://bedrock-runtime.us-west-2.amazonaws.com/model/anthropic.claude-opus-4-6-v1%3A0/invoke-with-response-stream",
+        )
+        expect(request.headers.authorization).toStartWith(auth.apiKey ? "Bearer test" : "AWS4-HMAC-SHA256")
+        const body = JSON.parse(text)
+        expect(body.model).toBeUndefined()
+        expect(body.stream).toBeUndefined()
+        expect(body.anthropic_version).toBe("bedrock-2023-05-31")
+        expect(body.anthropic_beta).toEqual(["compact-2026-01-12"])
+        expect(body.context_management.edits).toEqual([{ type: "compact_20260112" }])
+        if (body.messages.length > 1)
+          expect(body.messages[1].content[0]).toEqual({ type: "compaction", content: "Summary" })
+        return respond(response, { headers: { "content-type": "application/vnd.amazon.eventstream" } })
+      }),
+    ),
+  ).effect(`Bedrock Messages compaction round trip with ${auth.apiKey ? "bearer" : "SigV4"} authentication`, () =>
+    Effect.gen(function* () {
+      const model = AmazonBedrock.configure({ ...auth, region: "us-west-2" }).messages("anthropic.claude-opus-4-6-v1:0")
+      const request = LLM.request({
+        model,
+        prompt: "hello",
+        providerOptions: { contextManagement: { edits: [{ type: "compact_20260112" }] } },
+      })
+      const first = yield* LLMClient.generate(request)
+      expect(first.text).toBe("Hello")
+      expect(first.message.content.map((part) => part.type)).toEqual(["compaction", "text"])
+      yield* LLMClient.generate(
+        LLMRequest.update(request, { messages: [...request.messages, first.message, Message.user("continue")] }),
+      )
+    }),
+  )
+}
+
+testEffect(
+  fixedResponse(
+    codec.encode({
+      headers: {
+        ":message-type": { type: "string", value: "exception" },
+        ":exception-type": { type: "string", value: "throttlingException" },
+      },
+      body: new TextEncoder().encode(JSON.stringify({ message: "Too many requests", trace: "keep-original" })),
+    }),
+  ),
+).effect("Bedrock Messages retains the original exception frame", () =>
+  Effect.gen(function* () {
+    const error = yield* LLMClient.generate(
+      LLM.request({ model: AmazonBedrock.configure({ apiKey: "test" }).messages("claude"), prompt: "hello" }),
+    ).pipe(Effect.flip)
+    expect(error.reason.body).toContain("keep-original")
+    expect(error.reason.http?.status).toBe(200)
+  }),
+)
@@ -0,0 +1,50 @@
+import { expect } from "bun:test"
+import { Effect, Stream } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient, WebSocketTransport } from "../../src/route.js"
+import { OpenAI } from "../../src/providers.js"
+import { testEffect } from "../lib/effect.js"
+import { fixedResponse } from "../lib/http.js"
+
+testEffect(fixedResponse("unexpected HTTP fallback")).effect(
+  "WebSocket responses preserve compaction options and replay state",
+  () =>
+    Effect.gen(function* () {
+      const checkpoint = { type: "compaction", id: "cmp_ws", encrypted_content: "opaque" }
+      const sent: unknown[] = []
+      const webSocket = WebSocketTransport.makeDirect({
+        open: () =>
+          Effect.succeed({
+            sendText: (message) =>
+              Effect.sync(() => {
+                const body = JSON.parse(message)
+                expect(body.context_management).toEqual([{ type: "compaction", compact_threshold: 100000 }])
+                expect(body.stream).toBeUndefined()
+                if (sent.length) expect(body.input[1]).toEqual(checkpoint)
+                sent.push(body)
+              }),
+            messages: Stream.fromIterable(
+              [
+                { type: "response.created", response: { id: "resp_ws" } },
+                { type: "response.output_item.done", item: checkpoint },
+                { type: "response.completed", response: { id: "resp_ws", output: [checkpoint] } },
+              ].map((event) => JSON.stringify(event)),
+            ),
+            close: Effect.void,
+          }),
+      })
+      const request = LLM.request({
+        model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"),
+        prompt: "hello",
+        providerOptions: { contextManagement: [{ type: "compaction", compactThreshold: 100000 }] },
+      })
+      const first = yield* LLMClient.generate(request, { webSocket })
+      expect(first.message.content).toHaveLength(1)
+      expect(first.message.content[0]?.type).toBe("compaction")
+      yield* LLMClient.generate(
+        LLMRequest.update(request, { messages: [...request.messages, first.message, Message.user("continue")] }),
+        { webSocket },
+      )
+      expect(sent).toHaveLength(2)
+    }),
+)
@@ -0,0 +1,91 @@
+import { expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient } from "../../src/route/client.js"
+import { OpenAI, XAI, Anthropic } from "../../src/providers.js"
+import { recordedTests } from "../recorded-test.js"
+
+const history = [
+  Message.user("Remember the project codename COPPER-ORBIT-42."),
+  Message.assistant(
+    "The project codename is COPPER-ORBIT-42. " + "We reviewed the implementation and tests. ".repeat(1000),
+  ),
+]
+
+for (const provider of [
+  {
+    id: "openai",
+    key: "OPENAI_API_KEY",
+    model: OpenAI.configure({ apiKey: process.env.OPENAI_API_KEY ?? "fixture" }).responses("gpt-5.3-codex"),
+  },
+  {
+    id: "xai",
+    key: "XAI_API_KEY",
+    model: XAI.configure({ apiKey: process.env.XAI_API_KEY ?? "fixture" }).responses("grok-4.6"),
+  },
+]) {
+  recordedTests({ prefix: `${provider.id}-compaction`, provider: provider.id, requires: [provider.key] }).effect(
+    "compacts and continues with the provider checkpoint",
+    () =>
+      Effect.gen(function* () {
+        const request = LLM.request({ model: provider.model, messages: history, generation: { maxTokens: 1024 } })
+        const compacted = yield* LLMClient.compact(request)
+        const result = yield* LLMClient.generate(
+          LLMRequest.update(request, {
+            messages: [
+              ...compacted.messages,
+              Message.user("What is the project codename? Reply only with the codename."),
+            ],
+          }),
+        )
+        expect(result.text).toContain("COPPER-ORBIT-42")
+      }),
+    120000,
+  )
+}
+
+recordedTests({
+  prefix: "anthropic-compaction",
+  provider: "anthropic",
+  requires: ["ANTHROPIC_API_KEY"],
+  options: { redact: { allowRequestHeaders: ["anthropic-version", "anthropic-beta"] } },
+}).effect(
+  "automatically compacts and continues after a pause",
+  () =>
+    Effect.gen(function* () {
+      const model = Anthropic.configure({ apiKey: process.env.ANTHROPIC_API_KEY ?? "fixture" }).model(
+        "claude-sonnet-4-6",
+      )
+      const request = LLM.request({
+        model,
+        messages: [
+          Message.user(
+            "Remember the project codename COPPER-ORBIT-42. " +
+              "The implementation and tests were reviewed. ".repeat(10000),
+          ),
+        ],
+        generation: { maxTokens: 4096 },
+        providerOptions: {
+          contextManagement: {
+            edits: [
+              { type: "compact_20260112", trigger: { type: "input_tokens", value: 50000 }, pauseAfterCompaction: true },
+            ],
+          },
+        },
+      })
+      const first = yield* LLMClient.generate(request)
+      expect(first.finishReason.raw).toBe("compaction")
+      expect(first.message.content.some((part) => part.type === "compaction")).toBe(true)
+      const result = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          messages: [
+            ...request.messages,
+            first.message,
+            Message.user("What is the project codename? Reply only with the codename."),
+          ],
+        }),
+      )
+      expect(result.text).toContain("COPPER-ORBIT-42")
+    }),
+  120000,
+)
@@ -0,0 +1,136 @@
+import { expect } from "bun:test"
+import { Effect, Schema } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient } from "../../src/route/client.js"
+import { OpenAI, Azure, XAI } from "../../src/providers/index.js"
+import { testEffect } from "../lib/effect.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
+
+const checkpoint = { type: "compaction", id: "cmp_1", encrypted_content: "opaque" }
+const response = sseEvents(
+  { type: "response.output_item.done", item: checkpoint },
+  {
+    type: "response.completed",
+    response: { id: "resp_1", output: [checkpoint], usage: { input_tokens: 10, output_tokens: 2, total_tokens: 12 } },
+  },
+)
+
+for (const model of [
+  OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"),
+  Azure.configure({ apiKey: "test", resourceName: "test" }).responses("deployment"),
+]) {
+  testEffect(
+    dynamicResponse(({ text, respond }) =>
+      Effect.sync(() => {
+        const body = JSON.parse(text)
+        expect(body.context_management).toEqual([{ type: "compaction", compact_threshold: 100000 }])
+        expect(body.store).toBe(false)
+        if (body.input.length > 1) expect(body.input[1]).toEqual(checkpoint)
+        return respond(response, { headers: { "content-type": "text/event-stream" } })
+      }),
+    ),
+  ).effect(`${model.provider} compaction survives generation, serialization, and a second request`, () =>
+    Effect.gen(function* () {
+      const request = LLM.request({
+        model,
+        prompt: "hello",
+        providerOptions: { contextManagement: [{ type: "compaction", compactThreshold: 100000 }] },
+      })
+      const first = yield* LLMClient.generate(request)
+      expect(first.message.content).toHaveLength(1)
+      expect(first.message.content[0]?.type).toBe("compaction")
+      expect(first.text).toBe("")
+      const codec = Schema.fromJsonString(Message)
+      const message = Schema.decodeSync(codec)(Schema.encodeSync(codec)(first.message))
+      yield* LLMClient.generate(
+        LLMRequest.update(request, { messages: [...request.messages, message, Message.user("continue")] }),
+      )
+      const rejected = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          model: XAI.configure({ apiKey: "test" }).responses("grok-4.6"),
+          providerOptions: {},
+          messages: [message],
+        }),
+      ).pipe(Effect.flip)
+      expect(rejected.reason._tag).toBe("InvalidRequest")
+    }),
+  )
+}
+
+testEffect(fixedResponse(sseEvents({ type: "response.completed", response: { output: [checkpoint] } }))).effect(
+  "recovers compaction from the terminal output when item completion is absent",
+  () =>
+    Effect.gen(function* () {
+      const result = yield* LLMClient.generate(
+        LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"), prompt: "hello" }),
+      )
+      expect(result.message.content).toHaveLength(1)
+      expect(result.message.content[0]?.type).toBe("compaction")
+    }),
+)
+
+testEffect(
+  fixedResponse(sseEvents({ type: "response.output_item.done", item: { type: "compaction", id: "cmp_bad" } })),
+).effect("rejects incomplete compaction payloads without publishing a checkpoint", () =>
+  Effect.gen(function* () {
+    const error = yield* LLMClient.generate(
+      LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"), prompt: "hello" }),
+    ).pipe(Effect.flip)
+    expect(error.reason._tag).toBe("InvalidProviderOutput")
+    expect(error.reason.body).toContain("cmp_bad")
+  }),
+)
+
+const textItem = {
+  type: "message",
+  id: "msg_after",
+  role: "assistant",
+  content: [{ type: "output_text", text: "After checkpoint" }],
+}
+
+for (const completed of [false, true]) {
+  testEffect(
+    fixedResponse(
+      sseEvents(
+        { type: "response.output_item.added", output_index: 0, item: { type: "compaction", id: checkpoint.id } },
+        ...(completed ? [{ type: "response.output_item.done", output_index: 0, item: checkpoint }] : []),
+        { type: "response.output_item.added", output_index: 1, item: textItem },
+        { type: "response.output_text.delta", output_index: 1, item_id: textItem.id, delta: "After checkpoint" },
+        { type: "response.output_item.done", output_index: 1, item: textItem },
+        { type: "response.completed", response: { id: "resp_1", output: [checkpoint, textItem] } },
+      ),
+    ),
+  ).effect(completed ? "keeps streamed checkpoints before later text" : "rejects order-unsafe terminal recovery", () =>
+    Effect.gen(function* () {
+      const request = LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("fixture"), prompt: "hello" })
+      if (completed) {
+        const response = yield* LLMClient.generate(request)
+        expect(response.message.content.map((part) => part.type)).toEqual(["compaction", "text"])
+        return
+      }
+      const error = yield* LLMClient.generate(request).pipe(Effect.flip)
+      expect(error.reason._tag).toBe("InvalidProviderOutput")
+      expect(error.message).toContain("Cannot recover a compaction checkpoint")
+      expect(error.reason.body).toContain("response.completed")
+      expect(error.reason.http?.status).toBe(200)
+    }),
+  )
+}
+
+testEffect(
+  fixedResponse(
+    sseEvents({
+      type: "response.completed",
+      response: { output: [{ type: "compaction", encrypted_content: "opaque" }] },
+    }),
+  ),
+).effect("rejects terminal checkpoints missing an id", () =>
+  Effect.gen(function* () {
+    const error = yield* LLMClient.generate(
+      LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("fixture"), prompt: "hello" }),
+    ).pipe(Effect.flip)
+    expect(error.reason._tag).toBe("InvalidProviderOutput")
+    expect(error.message).toContain("missing its id")
+  }),
+)
@@ -0,0 +1,377 @@
+import { expect } from "bun:test"
+import { Effect, Schema } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient, Route } from "../../src/route/client.js"
+import { Auth } from "../../src/route/auth.js"
+import { Endpoint } from "../../src/route/endpoint.js"
+import { OpenAIResponses } from "../../src/protocols/openai-responses.js"
+import { OpenAI, Azure, XAI, Anthropic, AmazonBedrockMantle } from "../../src/providers/index.js"
+import { testEffect } from "../lib/effect.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
+
+const checkpoint = { type: "compaction", id: "cmp_1", encrypted_content: "opaque" }
+const retained = {
+  type: "message",
+  role: "user",
+  id: "msg_1",
+  status: "completed",
+  content: [{ type: "input_text", text: "retained" }],
+}
+const output = [retained, checkpoint]
+
+testEffect(
+  dynamicResponse(({ request, text, respond }) =>
+    Effect.sync(() => {
+      expect(request.headers["x-deployment"]).toBe("fixture")
+      expect(request.headers["x-override"]).toBe("request")
+      expect(request.headers["x-default"]).toBe("configured")
+      expect(request.headers.authorization).toBe("Bearer test")
+      expect(new URL(request.url).searchParams.get("api-version")).toBe("fixture")
+      expect(new URL(request.url).searchParams.get("trace")).toBe("request")
+      if (new URL(request.url).pathname.endsWith("/compact")) {
+        expect(JSON.parse(text)).toEqual({
+          model: "overlaid",
+          input: [{ role: "user", content: [{ type: "input_text", text: "hello" }] }],
+          instructions: "request instructions",
+          previous_response_id: "resp_previous",
+        })
+        return respond(JSON.stringify({ object: "response.compaction", output }))
+      }
+      return respond(sseEvents({ type: "response.completed", response: { id: "resp_1" } }))
+    }),
+  ),
+).effect("generation and compaction share deployment headers, defaults, auth, query, and middleware", () =>
+  Effect.gen(function* () {
+    const headers: string[] = []
+    const middleware: string[] = []
+    const route = Route.make({
+      id: "compaction-headers",
+      provider: "openai",
+      protocol: OpenAIResponses.protocol,
+      compact: OpenAIResponses.route.compact,
+      transport: OpenAIResponses.httpTransport,
+      endpoint: Endpoint.path(({ body }) => `/${body.model}/responses`, {
+        baseURL: "https://example.com",
+        query: { "api-version": "fixture" },
+      }),
+      auth: Auth.bearer("test"),
+      headers: ({ request }) => {
+        expect(request.providerOptions?.store).toBe(false)
+        headers.push(String(request.model.id))
+        return { "x-deployment": "fixture", "x-override": "route" }
+      },
+      defaults: {
+        headers: { "x-default": "configured", "x-override": "configured" },
+        providerOptions: { store: false },
+        http: { body: { instructions: "default instructions" } },
+      },
+    })
+    const request = LLM.request({
+      model: route.model({ id: "fixture" }),
+      prompt: "hello",
+      system: "system instructions",
+      http: {
+        headers: { "x-override": "request" },
+        query: { trace: "request" },
+        body: {
+          model: "overlaid",
+          instructions: "request instructions",
+          previous_response_id: "resp_previous",
+          store: false,
+          stream: true,
+        },
+      },
+    })
+    const options: Parameters<typeof LLMClient.compact>[1] = {
+      http: (request, next) => {
+        middleware.push(new URL(request.url).pathname)
+        return next(request)
+      },
+    }
+    yield* LLMClient.generate(request, options)
+    yield* LLMClient.compact(request, options)
+    expect(headers).toEqual(["fixture", "fixture"])
+    expect(middleware).toEqual(["/fixture/responses", "/fixture/responses/compact"])
+  }),
+)
+
+for (const model of [
+  OpenAI.configure({ apiKey: "test" }).responses("fixture"),
+  Azure.configure({ apiKey: "test", resourceName: "test" }).responses("fixture"),
+  XAI.configure({ apiKey: "test" }).responses("fixture"),
+]) {
+  const item = {
+    type: model.provider === "xai" ? "x_search_call" : "computer_call",
+    id: "hosted_1",
+    status: "completed",
+  }
+  testEffect(
+    dynamicResponse(({ request, text, respond }) =>
+      Effect.sync(() => {
+        expect(new URL(request.url).pathname).toEndWith("/responses/compact")
+        expect(JSON.parse(text)).toEqual({ model: "fixture", input: [item], instructions: "Keep the context" })
+        return respond(JSON.stringify({ object: "response.compaction", output: [checkpoint] }))
+      }),
+    ),
+  ).effect(`${model.provider} compacts provider-specific history without lowering generation settings`, () =>
+    Effect.gen(function* () {
+      const request = LLM.request({
+        model,
+        system: "Keep the context",
+        messages: [
+          Message.assistant({
+            type: "tool-result",
+            id: item.id,
+            name: item.type,
+            result: { type: "json", value: item },
+            providerExecuted: true,
+            providerMetadata: { [model.route.providerMetadataKey ?? model.provider]: { itemId: item.id } },
+          }),
+        ],
+      })
+      for (const candidate of [
+        LLMRequest.update(request, {
+          tools: [
+            { name: "unsupported", description: "Generation only", inputSchema: {}, native: { unsupported: {} } },
+          ],
+        }),
+        LLMRequest.update(request, { providerOptions: { contextManagement: "invalid-generation-option" } }),
+      ]) {
+        const error = yield* LLMClient.generate(candidate).pipe(Effect.flip)
+        expect(error.reason._tag).toBe("InvalidRequest")
+        const response = yield* LLMClient.compact(candidate)
+        expect(response.messages[0]?.content[0]?.type).toBe("compaction")
+      }
+    }),
+  )
+}
+
+const retainedItems = [
+  retained,
+  {
+    type: "message",
+    id: "msg_assistant",
+    role: "assistant",
+    status: "completed",
+    phase: "commentary",
+    content: [
+      { type: "output_text", text: "First" },
+      { type: "output_text", text: "Second" },
+    ],
+  },
+  {
+    type: "reasoning",
+    id: "rs_1",
+    summary: [
+      { type: "summary_text", text: "Thinking" },
+      { type: "summary_text", text: "More thinking" },
+    ],
+    encrypted_content: "reasoning-state",
+  },
+  { type: "reasoning", id: "rs_2", summary: [], encrypted_content: "hidden-reasoning" },
+  {
+    type: "message",
+    id: "msg_media",
+    role: "user",
+    content: [
+      { type: "input_image", image_url: "https://example.com/image.png" },
+      { type: "input_file", filename: "report.pdf", file_data: "data:application/pdf;base64,cGRm" },
+      { type: "input_file", filename: "other.pdf", file_url: "https://example.com/report.pdf" },
+    ],
+  },
+  checkpoint,
+]
+
+testEffect(
+  dynamicResponse(({ request, text, respond }) =>
+    Effect.sync(() => {
+      if (new URL(request.url).pathname.endsWith("/compact"))
+        return respond(JSON.stringify({ object: "response.compaction", output: retainedItems }))
+      expect(JSON.parse(text).input).toEqual(retainedItems)
+      return respond(sseEvents({ type: "response.completed", response: { id: "resp_1" } }), {
+        headers: { "content-type": "text/event-stream" },
+      })
+    }),
+  ),
+).effect("retained messages, reasoning, and media are ordinary typed conversation parts", () =>
+  Effect.gen(function* () {
+    const request = LLM.request({
+      model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"),
+      prompt: "hello",
+    })
+    const compacted = yield* LLMClient.compact(request)
+    expect(compacted.messages.map((message) => message.role)).toEqual([
+      "user",
+      "assistant",
+      "assistant",
+      "assistant",
+      "user",
+      "assistant",
+    ])
+    expect(compacted.messages[1]?.content).toEqual([
+      { type: "text", text: "First" },
+      { type: "text", text: "Second" },
+    ])
+    expect(compacted.messages[2]?.content.map((part) => part.type)).toEqual(["reasoning", "reasoning"])
+    expect(compacted.messages[4]?.content.map((part) => part.type)).toEqual(["media", "media", "media"])
+    const codec = Schema.fromJsonString(Schema.Array(Message))
+    const messages = Schema.decodeSync(codec)(Schema.encodeSync(codec)(compacted.messages))
+    yield* LLMClient.generate(LLMRequest.update(request, { messages }))
+  }),
+)
+
+for (const item of [
+  { type: "unknown_provider_item", data: "do not hide in a compaction part" },
+  {
+    type: "message",
+    role: "user",
+    content: [{ type: "input_image", image_url: "https://example.com/image.png", detail: 42 }],
+  },
+  { type: "message", role: "user", content: [] },
+  {
+    type: "message",
+    role: "assistant",
+    content: [{ type: "input_image", image_url: "https://example.com/image.png" }],
+  },
+  { type: "message", role: "user", content: [{ type: "input_file", filename: "missing.pdf" }] },
+]) {
+  testEffect(fixedResponse(JSON.stringify({ object: "response.compaction", output: [item, checkpoint] }))).effect(
+    `rejects unsupported compact output: ${JSON.stringify(item)}`,
+    () =>
+      Effect.gen(function* () {
+        const error = yield* LLMClient.compact(
+          LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"), prompt: "hello" }),
+        ).pipe(Effect.flip)
+        expect(error.reason._tag).toBe("InvalidProviderOutput")
+        expect(error.reason.body).toContain(JSON.stringify(item))
+        expect(error.reason.http?.status).toBe(200)
+      }),
+  )
+}
+
+for (const model of [
+  OpenAI.configure({ apiKey: "test" }).responses("fixture"),
+  Azure.configure({ apiKey: "test", resourceName: "test" }).responses("fixture"),
+  XAI.configure({ apiKey: "test" }).responses("fixture"),
+]) {
+  const images = [undefined, "low", "high", "auto"].map((detail) => ({
+    type: "input_image",
+    image_url: "https://example.com/image.png",
+    ...(detail === undefined ? {} : { detail }),
+  }))
+  testEffect(
+    dynamicResponse(({ request, text, respond }) =>
+      Effect.sync(() => {
+        if (new URL(request.url).pathname.endsWith("/compact"))
+          return respond(
+            JSON.stringify({
+              object: "response.compaction",
+              output: [{ type: "message", role: "user", content: images }, checkpoint],
+            }),
+          )
+        expect(JSON.parse(text).input[0].content).toEqual(images)
+        return respond(sseEvents({ type: "response.completed", response: { id: "resp_1" } }))
+      }),
+    ),
+  ).effect(`${model.provider} preserves retained image detail through serialization and replay`, () =>
+    Effect.gen(function* () {
+      const request = LLM.request({ model, prompt: "hello" })
+      const compacted = yield* LLMClient.compact(request)
+      const codec = Schema.fromJsonString(Schema.Array(Message))
+      const messages = Schema.decodeSync(codec)(Schema.encodeSync(codec)(compacted.messages))
+      yield* LLMClient.generate(LLMRequest.update(request, { messages }))
+    }),
+  )
+}
+
+testEffect(fixedResponse("must not execute")).effect("xAI rejects automatic compaction options", () =>
+  Effect.gen(function* () {
+    const request = LLMRequest.update(
+      LLM.request({ model: XAI.configure({ apiKey: "test" }).responses("grok-4.6"), prompt: "hello" }),
+      { providerOptions: { contextManagement: [{ type: "compaction" }] } },
+    )
+    const error = yield* LLMClient.generate(request).pipe(Effect.flip)
+    expect(error.reason._tag).toBe("InvalidRequest")
+    expect(error.message).toContain("LLMClient.compact")
+  }),
+)
+
+for (const model of [
+  OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"),
+  Azure.configure({ apiKey: "test", resourceName: "test" }).responses("deployment"),
+  XAI.configure({ apiKey: "test" }).responses("grok-4.6"),
+]) {
+  testEffect(
+    dynamicResponse(({ request, text, respond }) =>
+      Effect.sync(() => {
+        const body = JSON.parse(text)
+        expect(request.method).toBe("POST")
+        expect(request.headers[model.provider === "azure" ? "api-key" : "authorization"]).toBe(
+          model.provider === "azure" ? "test" : "Bearer test",
+        )
+        if (new URL(request.url).pathname.endsWith("/responses/compact")) {
+          expect(body).toEqual({
+            model: model.id,
+            input: [{ role: "user", content: [{ type: "input_text", text: "original" }] }],
+            instructions: "system",
+          })
+          return respond(
+            JSON.stringify({
+              object: "response.compaction",
+              output,
+              usage: { input_tokens: 1000, output_tokens: 10, total_tokens: 1010 },
+            }),
+            { headers: { "content-type": "application/json" } },
+          )
+        }
+        expect(new URL(request.url).pathname.endsWith("/responses")).toBe(true)
+        expect(body.input).toEqual([...output, { role: "user", content: [{ type: "input_text", text: "continue" }] }])
+        return respond(sseEvents({ type: "response.completed", response: { id: "resp_1", output: [] } }), {
+          headers: { "content-type": "text/event-stream" },
+        })
+      }),
+    ),
+  ).effect(`${model.provider} explicitly compacts and replays the entire canonical window`, () =>
+    Effect.gen(function* () {
+      const request = LLM.request({ model, prompt: "original", system: "system", http: { body: { store: false } } })
+      const compacted = yield* LLMClient.compact(request)
+      expect(compacted.usage?.totalTokens).toBe(1010)
+      expect(compacted.messages.map((message) => message.role)).toEqual(["user", "assistant"])
+      expect(compacted.messages[0]?.content).toEqual([{ type: "text", text: "retained" }])
+      expect(compacted.messages[1]?.content).toEqual([
+        { type: "compaction", provider: model.provider, id: "cmp_1", encrypted: "opaque" },
+      ])
+      const codec = Schema.fromJsonString(Schema.Array(Message))
+      const messages = Schema.decodeSync(codec)(Schema.encodeSync(codec)(compacted.messages))
+      yield* LLMClient.generate(LLMRequest.update(request, { messages: [...messages, Message.user("continue")] }))
+    }),
+  )
+}
+
+for (const model of [
+  Anthropic.configure({ apiKey: "test" }).model("claude-opus-4-6"),
+  AmazonBedrockMantle.configure({ apiKey: "test" }).responses("model"),
+]) {
+  testEffect(fixedResponse("must not execute")).effect(
+    `${model.route.id} does not inherit an unsupported compact endpoint`,
+    () =>
+      Effect.gen(function* () {
+        const error = yield* LLMClient.compact(LLM.request({ model, prompt: "hello" })).pipe(Effect.flip)
+        expect(error.reason._tag).toBe("InvalidRequest")
+      }),
+  )
+}
+
+testEffect(
+  fixedResponse(JSON.stringify({ object: "response.compaction", output: [retained], debug: "original payload" })),
+).effect("invalid explicit compaction preserves the original response and HTTP context", () =>
+  Effect.gen(function* () {
+    const error = yield* LLMClient.compact(
+      LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"), prompt: "hello" }),
+    ).pipe(Effect.flip)
+    expect(error.reason._tag).toBe("InvalidProviderOutput")
+    expect(error.reason.body).toContain("original payload")
+    expect(error.reason.http?.status).toBe(200)
+  }),
+)
@@ -327,7 +327,12 @@ function modelFromLanguage(info: Info, language: LanguageModelV3) {
    },
    body: {
      schema: Schema.Unknown,
-      from: (request) => Effect.succeed(callOptions(request, packageName, info.modelID ?? info.id, optionKey)),
+      from: (request) =>
+        Effect.try({
+          try: () => callOptions(request, packageName, info.modelID ?? info.id, optionKey),
+          catch: (cause) =>
+            cause instanceof AIError ? cause : ProviderShared.invalidRequest("Invalid AI SDK request", cause),
+        }),
    },
    with: () => route,
    model: (input) =>
@@ -512,6 +517,8 @@ function userPart(part: ContentPart): UserContent {

 function assistantPart(part: ContentPart): AssistantContent {
  switch (part.type) {
+    case "compaction":
+      throw ProviderShared.invalidRequest("AI SDK routes cannot replay native provider compaction state")
    case "text":
      return [{ type: "text", text: part.text, providerOptions: metadataProviderOptions(part.providerMetadata) }]
    case "media":
@@ -10,6 +10,8 @@ import { Provider } from "@opencode-ai/core/provider"
 import {
  LLM,
  AIError,
+  CompactionPart,
+  ProviderID,
  HttpContext,
  LLMEvent,
  Message,
@@ -68,6 +70,31 @@ const client = LLMClient.layer.pipe(
  ),
 )

+it.effect("rejects native provider compaction rather than silently dropping replay state", () =>
+  Effect.gen(function* () {
+    const aisdk = yield* AISDK.Service
+    yield* aisdk.hook.sdk((event) => {
+      event.sdk = { languageModel: () => streamModel([]) }
+    })
+    const resolved = yield* aisdk.model(model("@ai-sdk/openai"))
+    const error = yield* compileRequest(
+      LLM.request({
+        model: resolved,
+        messages: [
+          Message.assistant(
+            CompactionPart.make({
+              provider: ProviderID.make("test-provider"),
+              encrypted: "opaque",
+            }),
+          ),
+        ],
+      }),
+    ).pipe(Effect.flip)
+    expect(error.reason._tag).toBe("InvalidRequest")
+    expect(error.message).toContain("cannot replay")
+  }),
+)
+
 it.effect("keys language models by package and flattened overlays", () =>
  Effect.gen(function* () {
    const aisdk = yield* AISDK.Service
Author	SHA1	Message	Date
Shoubhit Dash	a69f0ef2f8	refactor(ai): restore inline compaction beta handling	2026-08-31 20:01:57 +05:30
Shoubhit Dash	0934b9e01d	refactor(ai): centralize anthropic beta selection	2026-08-31 19:52:15 +05:30
Shoubhit Dash	716393cdfa	refactor(ai): separate conversation and generation lowering	2026-08-31 19:50:39 +05:30
Shoubhit Dash	48aedea579	refactor(ai): simplify response fragment guards	2026-08-31 19:30:04 +05:30
Shoubhit Dash	96c0f0d8c3	fix(ai): enforce exclusive checkpoint representations	2026-08-31 19:29:23 +05:30
Shoubhit Dash	4b7d49c9a4	fix(ai): preserve compacted image detail on replay	2026-08-31 19:27:34 +05:30
Shoubhit Dash	0e1081b0a5	fix(ai): apply cache policy to bedrock messages	2026-08-31 19:26:21 +05:30
Shoubhit Dash	ce96bbe8a7	refactor(ai): share request preparation across operations	2026-08-31 19:25:28 +05:30
Shoubhit Dash	027f0e7cc9	fix(ai): reject order-unsafe checkpoint recovery	2026-08-31 19:23:58 +05:30
Shoubhit Dash	0a442fa2f8	chore: merge v2 into provider-compaction	2026-08-31 19:08:10 +05:30
Shoubhit Dash	59d1ab783c	refactor(ai): make compaction conversion synchronous	2026-08-31 17:29:38 +05:30
Shoubhit Dash	6b4426bbd3	docs(ai): describe typed compaction history	2026-08-31 17:19:57 +05:30
Shoubhit Dash	1b9f762ea5	refactor(ai): model compaction as typed conversation parts	2026-08-31 17:19:45 +05:30
Shoubhit Dash	c8b4963d98	refactor(ai): simplify compaction control flow	2026-08-31 16:24:01 +05:30
Shoubhit Dash	14a90331bc	docs(ai): explain provider compaction and replay ownership	2026-08-31 16:11:38 +05:30
Shoubhit Dash	cd2880075a	test(ai): cover compaction across http and websocket flows	2026-08-31 16:11:17 +05:30
Shoubhit Dash	af22c16249	fix(ai): validate compaction boundaries and incomplete blocks	2026-08-31 16:09:00 +05:30
Shoubhit Dash	bc4db825a4	fix(core): reject unsupported compaction replay in ai sdk routes	2026-08-31 16:08:32 +05:30
Shoubhit Dash	4f0ba4f06e	feat(ai): add bedrock messages route for claude compaction	2026-08-31 16:02:30 +05:30
Shoubhit Dash	8dd92e2d95	feat(ai): add explicit responses compaction calls	2026-08-31 15:58:29 +05:30
Shoubhit Dash	cb5a6c7db4	feat(ai): support anthropic compaction and iteration usage	2026-08-31 15:53:57 +05:30
Shoubhit Dash	eaa3dfe04e	feat(ai): support automatic responses compaction	2026-08-31 15:51:01 +05:30
Shoubhit Dash	5fd800d169	feat(ai): preserve provider compaction in messages and events	2026-08-31 15:48:42 +05:30
				`@@ -0,0 +1 @@`
				`export { messagesModel as model } from "../amazon-bedrock.js"`