@bitkyc08/opencodex 2.8.2-preview.20260731 → 2.9.0

package/src/adapters/mimo-free.ts

@@ -5,7 +5,7 @@ import { getConfigDir } from "../config";
 import { recordOwnedConfigPath } from "../lib/config-ownership";
 import type { OcxProviderConfig, OcxParsedRequest } from "../types";
 import { createOpenAIChatAdapter } from "./openai-chat";
-import type { ProviderAdapter, AdapterRequest } from "./base";
+import type { ProviderAdapter, AdapterRequest, IncomingMeta } from "./base";
 
 const BOOTSTRAP_URL = "https://api.xiaomimimo.com/api/free-ai/bootstrap";
 export const MIMO_CHAT_URL = "https://api.xiaomimimo.com/api/free-ai/openai/chat";
@@ -27,6 +27,8 @@ const USER_AGENTS = [
 const JWT_FALLBACK_TTL_MS = 3_000_000; // 50 min
 const JWT_EXPIRY_BUFFER_MS = 300_000;  // 5 min early refresh
 const BOOTSTRAP_TIMEOUT_MS = 15_000;
+const MIMO_BOOTSTRAP_MAX_BYTES = 128 * 1024;
+const MIMO_JWT_MAX_BYTES = 64 * 1024;
 
 // In-process JWT cache -- survives across requests, reset on restart.
 let cachedJwt: string | null = null;
@@ -107,8 +109,38 @@ async function fetchJwt(signal?: AbortSignal): Promise<string> {
     try { await response.body?.cancel(); } catch { /* already consumed */ }
     throw new Error(`MiMo bootstrap failed: ${response.status}`);
   }
-  const data = await response.json() as { jwt?: string };
+  const announced = Number(response.headers.get("content-length") ?? 0);
+  if (Number.isFinite(announced) && announced > MIMO_BOOTSTRAP_MAX_BYTES) {
+    await response.body?.cancel().catch(() => {});
+    throw new Error("MiMo bootstrap response too large");
+  }
+  const reader = response.body?.getReader();
+  const chunks: Uint8Array[] = [];
+  let bytes = 0;
+  if (reader) {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+      if (!value) continue;
+      bytes += value.byteLength;
+      if (bytes > MIMO_BOOTSTRAP_MAX_BYTES) {
+        await reader.cancel().catch(() => {});
+        throw new Error("MiMo bootstrap response too large");
+      }
+      chunks.push(value);
+    }
+  }
+  const body = new Uint8Array(bytes);
+  let offset = 0;
+  for (const chunk of chunks) {
+    body.set(chunk, offset);
+    offset += chunk.byteLength;
+  }
+  const data = JSON.parse(new TextDecoder("utf-8", { fatal: true }).decode(body)) as { jwt?: string };
   if (!data.jwt) throw new Error("MiMo bootstrap returned no JWT");
+  if (new TextEncoder().encode(data.jwt).byteLength > MIMO_JWT_MAX_BYTES) {
+    throw new Error("MiMo bootstrap response too large");
+  }
   return data.jwt;
 }
 
@@ -170,12 +202,12 @@ export function createMimoFreeAdapter(provider: OcxProviderConfig): ProviderAdap
     ...base,
     name: "mimo-free",
 
-    async buildRequest(parsed: OcxParsedRequest): Promise<AdapterRequest> {
+    async buildRequest(parsed: OcxParsedRequest, incoming: IncomingMeta): Promise<AdapterRequest> {
       const jwt = await getMimoJwt();
 
       // Let the base adapter build the wire body (handles reasoning, tools, etc.)
       // but override the URL and headers after.
-      const baseReq = base.buildRequest(parsed) as AdapterRequest;
+      const baseReq = base.buildRequest(parsed, incoming) as AdapterRequest;
       const baseBody = JSON.parse(baseReq.body as string) as unknown;
       const markedBody = injectMimoSystemMarker(baseBody);
 

package/src/adapters/openai-chat.ts

@@ -10,6 +10,12 @@ import { contentPartsToText } from "./image";
 import { neutralizeIdentity } from "./identity";
 import { buildNonOpenAIToolCatalogNudgeForTools, shouldInjectNonOpenAIToolCatalogNudge } from "./tool-catalog-nudge";
 import { openRouterProviderPayload, resolveOpenRouterRouting } from "../providers/openrouter-routing";
+import {
+  isTranslatorBudgetExceededError,
+  retainTranslatedEventBatch,
+  TRANSLATOR_MAX_SSE_EVENT_BYTES,
+  type TranslatorBudget,
+} from "../lib/translator-budget";
 
 // Providers may opt into stripping one trailing "[...]" group from the wire model id.
 // Z.AI needs this because its OpenAI path rejects glm-5.2[1m] with 400 code 1211;
@@ -676,7 +682,7 @@ export function createOpenAIChatAdapter(provider: OcxProviderConfig): ProviderAd
       };
     },
 
-    async *parseStream(response: Response): AsyncGenerator<AdapterEvent> {
+    async *parseStream(response: Response, budget: TranslatorBudget): AsyncGenerator<AdapterEvent> {
       if (!response.body) {
         yield { type: "error", message: "No response body" };
         return;
@@ -684,7 +690,9 @@ export function createOpenAIChatAdapter(provider: OcxProviderConfig): ProviderAd
 
       const reader = response.body.getReader();
       const decoder = new TextDecoder();
+      const budgetEncoder = new TextEncoder();
       let buffer = "";
+      let bufferBytes = 0;
       // Streamed tool calls are BUFFERED until a terminal signal, then flushed as atomic
       // start/delta/end sequences. The bridge treats text/reasoning deltas as barriers that
       // close an open tool-call item (bridge.ts closeCurrentToolCall on text_delta), so
@@ -693,15 +701,18 @@ export function createOpenAIChatAdapter(provider: OcxProviderConfig): ProviderAd
       // chunks, whole-chunk calls) cannot be represented live without overlapping sequences.
       // Keyed by `index` (OpenAI wire standard), falling back to `id`, falling back to the
       // last-seen call for providers that omit both on continuation chunks.
-      interface PendingToolCall { key: string; id: string; name: string; args: string }
+      interface PendingToolCall { key: string; id: string; name: string; args: string; argsBytes: number }
       const pendingToolCalls: PendingToolCall[] = [];
       let toolCallSeq = 0;
       const flushToolCalls = function* (): Generator<AdapterEvent> {
+        // Do not treat flushed tool calls as user-facing output for the finish-less EOF
+        // fallback — incomplete tool args must stay on the truncation path.
         for (const call of pendingToolCalls) {
           if (!call.id) call.id = `call_${++toolCallSeq}`;
           yield { type: "tool_call_start", id: call.id, name: call.name };
           if (call.args.length > 0) yield { type: "tool_call_delta", arguments: call.args };
           yield { type: "tool_call_end" };
+          budget.closeCall(call.key);
         }
         pendingToolCalls.length = 0;
       };
@@ -711,6 +722,9 @@ export function createOpenAIChatAdapter(provider: OcxProviderConfig): ProviderAd
       // explicit `[DONE]` sentinel OR a chunk carrying a non-null `finish_reason` (some
       // OpenAI-compatible providers omit `[DONE]` but do send finish_reason).
       let finishReason: string | undefined;
+      // Only answer text enables the finish-less EOF fallback. Reasoning-only streams can be
+      // suppressed by hideThinkingSummary and must not complete as empty successful turns.
+      let sawUserFacingOutput = false;
 
       // Single per-line handler shared by the streaming loop and the EOF residual-frame flush, so
       // a final frame is parsed identically wherever it lands (no duplicated, drift-prone parsing).
@@ -780,6 +794,7 @@ export function createOpenAIChatAdapter(provider: OcxProviderConfig): ProviderAd
             yield { type: "reasoning_raw_delta", text: delta.reasoning_content };
           }
           if (typeof delta.content === "string" && delta.content.length > 0) {
+            sawUserFacingOutput = true;
             yield { type: "text_delta", text: delta.content };
           }
 
@@ -797,12 +812,27 @@ export function createOpenAIChatAdapter(provider: OcxProviderConfig): ProviderAd
               // splitting into two calls that share one call_id downstream.
               if (!call && tc.id) call = pendingToolCalls.find(c => c.id === tc.id);
               if (!call) {
-                call = { key: key ?? `seq:${pendingToolCalls.length}`, id: "", name: "", args: "" };
+                call = { key: key ?? `seq:${pendingToolCalls.length}`, id: "", name: "", args: "", argsBytes: 0 };
                 pendingToolCalls.push(call);
+                budget.openCall(call.key);
               }
               if (tc.id && !call.id) call.id = tc.id;
               if (tc.function?.name && !call.name) call.name = tc.function.name;
-              if (tc.function?.arguments) call.args += tc.function.arguments;
+              if (tc.function?.arguments) {
+                const previousBytes = call.argsBytes;
+                const nextBytes = previousBytes + budgetEncoder.encode(tc.function.arguments).byteLength;
+                const scope = { kind: "tool_args" as const, callId: call.key };
+                const reservation = budget.reserveTransient(nextBytes, scope);
+                try {
+                  call.args += tc.function.arguments;
+                  reservation.commitRetained();
+                  budget.releaseRetained(previousBytes, scope);
+                  call.argsBytes = nextBytes;
+                } catch (error) {
+                  reservation.release();
+                  throw error;
+                }
+              }
             }
           }
         }
@@ -819,10 +849,31 @@ export function createOpenAIChatAdapter(provider: OcxProviderConfig): ProviderAd
         while (true) {
           const { done, value } = await reader.read();
           if (done) break;
-          buffer += decoder.decode(value, { stream: true });
+          const decoded = decoder.decode(value, { stream: true });
+          const nextBufferBytes = bufferBytes + budgetEncoder.encode(decoded).byteLength;
+          if (nextBufferBytes > TRANSLATOR_MAX_SSE_EVENT_BYTES) {
+            throw new Error(`translation SSE event exceeded ${TRANSLATOR_MAX_SSE_EVENT_BYTES} bytes`, {
+              cause: { code: "translation_buffer_limit" },
+            });
+          }
+          const appendReservation = budget.reserveTransient(nextBufferBytes, { kind: "live_transient" });
+          try {
+            buffer += decoded;
+            appendReservation.commitRetained();
+            budget.releaseRetained(bufferBytes, { kind: "live_transient" });
+          } catch (error) {
+            appendReservation.release();
+            throw error;
+          }
+          bufferBytes = nextBufferBytes;
 
           const lines = buffer.split("\n");
           buffer = lines.pop() ?? "";
+          const residualBytes = budgetEncoder.encode(buffer).byteLength;
+          const residualReservation = budget.reserveTransient(residualBytes, { kind: "live_transient" });
+          residualReservation.commitRetained();
+          budget.releaseRetained(bufferBytes, { kind: "live_transient" });
+          bufferBytes = residualBytes;
 
           for (const line of lines) {
             if ((yield* handleDataLine(line)) === "terminate") return;
@@ -837,10 +888,8 @@ export function createOpenAIChatAdapter(provider: OcxProviderConfig): ProviderAd
         if (buffer.length > 0) {
           if ((yield* handleDataLine(buffer)) === "terminate") return;
         }
-        // Reader EOF. A graceful close shows at least one terminal signal: `[DONE]` (returns above),
-        // a non-null finish_reason (sawFinish), or a trailing usage chunk (providers emit usage only
-        // at end-of-generation). If NONE of those were seen, the stream was cut mid-flight — fail
-        // closed so the bridge emits a classified response.failed rather than a silent truncation.
+        // Reader EOF. Prefer failing closed before flushing pending tool calls so the bridge
+        // never sees a fabricated tool_call_end on a truncated mid-assembly stream.
         //
         // Checked BEFORE flushToolCalls(), because that helper emits tool_call_end and there is no
         // taking it back: a half-assembled argument string would reach the client as a completed
@@ -856,29 +905,51 @@ export function createOpenAIChatAdapter(provider: OcxProviderConfig): ProviderAd
           yield { type: "error", message: "upstream stream ended mid tool call without a terminal signal — possible truncation" };
           return;
         }
-        yield* flushToolCalls();
-        if (!sawFinish && pendingUsage === undefined) {
+        // Finish-less EOF is only safe when answer text was emitted. Reasoning-only / usage-only
+        // truncations must stay on the error path (hideThinkingSummary can suppress reasoning).
+        // Trailing usage alone is not a terminal signal for this adapter (#735 / restore #773).
+        if (!sawFinish && !sawUserFacingOutput) {
           debugProviderDiagnostic("openai-chat", "stream-truncated", {
             finishReason: finishReason ?? null,
-            hadUsage: false,
+            hadUsage: pendingUsage !== undefined,
           });
           yield { type: "error", message: "upstream stream ended without a terminal signal ([DONE] or finish_reason) — possible truncation" };
           return;
         }
-        // Graceful close that omitted [DONE] but delivered finish_reason and/or final usage.
+        yield* flushToolCalls();
+        // Graceful close that omitted [DONE] but delivered finish_reason and/or answer text.
         const stopReason = finishReason === "length"
           ? "max_tokens"
           : finishReason === "content_filter"
             ? "content_filter"
             : undefined;
         yield { type: "done", usage: pendingUsage, ...(stopReason ? { stopReason } : {}) };
+      } catch (error) {
+        if (isTranslatorBudgetExceededError(error)
+          || (error instanceof Error && (error.cause as { code?: unknown } | undefined)?.code === "translation_buffer_limit")) {
+          yield {
+            type: "error",
+            status: 502,
+            errorType: "upstream_error",
+            code: "translation_buffer_limit",
+            message: "upstream translation buffer exceeded the safe limit",
+          };
+          try { await reader.cancel(error); } catch { /* already closed */ }
+          return;
+        }
+        throw error;
       } finally {
+        budget.releaseRetained(bufferBytes, { kind: "live_transient" });
+        for (const call of pendingToolCalls) budget.closeCall(call.key);
         reader.releaseLock();
       }
     },
 
-    async parseResponse(response: Response): Promise<AdapterEvent[]> {
+    async parseResponse(response: Response, budget: TranslatorBudget): Promise<AdapterEvent[]> {
       const json = await response.json() as Record<string, unknown>;
+      const responseBytes = new TextEncoder().encode(JSON.stringify(json)).byteLength;
+      budget.chargeRetained(responseBytes, { kind: "retained_collectors" });
+      try {
       if (json.error) {
         const upstreamError = json.error as { message?: unknown; code?: unknown; type?: unknown; status?: unknown };
         const message = typeof upstreamError.message === "string" ? upstreamError.message : "upstream error";
@@ -924,7 +995,11 @@ export function createOpenAIChatAdapter(provider: OcxProviderConfig): ProviderAd
         type: "done",
         usage: usageFromOpenAIChat(usage),
       });
+      retainTranslatedEventBatch(events, budget);
       return events;
+      } finally {
+        budget.releaseRetained(responseBytes, { kind: "retained_collectors" });
+      }
     },
   };
 }

package/src/adapters/openai-responses.ts

@@ -8,6 +8,7 @@ import { decodeServerSentEvents } from "../lib/sse-decoder";
 import { isCanonicalOpenAiForwardProvider } from "../providers/openai-tiers";
 import { OCX_REASONING_PREFIX } from "../responses/reasoning-envelope";
 import { modelRecordValue } from "../reasoning-effort";
+import type { TranslatorBudget } from "../lib/translator-budget";
 
 // Headers relayed verbatim from the caller in OAuth-passthrough ("forward") mode.
 // Exported so the web-search sidecar reuses the exact same forwarded-auth set for its ChatGPT call.
@@ -354,6 +355,47 @@ function isPlainObject(v: unknown): v is Record<string, unknown> {
   return !!v && typeof v === "object" && !Array.isArray(v);
 }
 
+function normalizeFunctionToolSchema(tool: unknown): unknown {
+  if (!isPlainObject(tool) || tool.type !== "function") return tool;
+  if (isPlainObject(tool.parameters) && tool.parameters.type === "object") return tool;
+  return {
+    ...tool,
+    parameters: { ...(isPlainObject(tool.parameters) ? tool.parameters : {}), type: "object" },
+  };
+}
+
+function normalizeToolSchemas(body: unknown): unknown {
+  if (!isPlainObject(body)) return body;
+
+  const normalizeTools = (tools: unknown[]): unknown[] => {
+    let changed = false;
+    const normalized = tools.map((tool) => {
+      const fixed = normalizeFunctionToolSchema(tool);
+      if (fixed !== tool) changed = true;
+      return fixed;
+    });
+    return changed ? normalized : tools;
+  };
+
+  let normalizedBody = body;
+  if (Array.isArray(body.tools)) {
+    const tools = normalizeTools(body.tools);
+    if (tools !== body.tools) normalizedBody = { ...normalizedBody, tools };
+  }
+  if (Array.isArray(normalizedBody.input)) {
+    let inputChanged = false;
+    const input = normalizedBody.input.map((item) => {
+      if (!isPlainObject(item) || item.type !== "additional_tools" || !Array.isArray(item.tools)) return item;
+      const tools = normalizeTools(item.tools);
+      if (tools === item.tools) return item;
+      inputChanged = true;
+      return { ...item, tools };
+    });
+    if (inputChanged) normalizedBody = { ...normalizedBody, input };
+  }
+  return normalizedBody;
+}
+
 const MAX_RESPONSES_CALL_ID_LENGTH = 64;
 const REPAIRED_CALL_ID_PREFIX = "call_ocx_";
 const REPAIRED_CALL_ID_DIGEST_LENGTH = MAX_RESPONSES_CALL_ID_LENGTH - REPAIRED_CALL_ID_PREFIX.length;
@@ -484,6 +526,40 @@ function stripPreviousResponseId(body: unknown, strip: boolean): unknown {
   return rest;
 }
 
+/**
+ * Drop request parameters a stateless Responses upstream cannot implement, and pin
+ * `store` false.
+ *
+ * `previous_response_id` is listed here as well as in `stripPreviousResponseId`
+ * because that helper's strip is conditional on replay expansion, and it keeps the
+ * field for API-key providers on the premise that the platform offers real
+ * server-side storage. DeepSeek documents the opposite: "the API is stateless:
+ * responses and conversations are not stored on the server", so the field can never
+ * be honoured regardless of expansion state.
+ *
+ * `prompt` is a reference to a server-stored prompt template — the most stateful
+ * field in the accepted schema.
+ *
+ * `service_tier` is deliberately NOT dropped: the server writes it for fast mode
+ * (`responses/core.ts`), and silently deleting a configured knob inside an adapter is
+ * worse than forwarding a parameter the upstream ignores.
+ *
+ * MUST run before the composed sanitize chain below: `stripItemIdsWhenUnstored` keys
+ * off `store === false`, and a stateless upstream cannot resolve a stored item id.
+ * Returns a copy, so `parsed._rawBody` keeps the client's original `store` value and
+ * the local replay cache still records the turn.
+ */
+function stripStatefulResponsesParams(body: unknown): unknown {
+  if (!isPlainObject(body)) return body;
+  const drop = ["previous_response_id", "conversation", "background", "metadata", "prompt"] as const;
+  const present = drop.some(key => Object.prototype.hasOwnProperty.call(body, key));
+  if (!present && body.store === false) return body;
+  const next: Record<string, unknown> = { ...body };
+  for (const key of drop) delete next[key];
+  next.store = false;
+  return next;
+}
+
 /**
  * Remove top-level parameters the ChatGPT backend (`authMode: "forward"`) rejects
  * with `{"detail":"Unsupported parameter: …"}` (strict allowlist). Codex CLI never
@@ -881,7 +957,8 @@ export function createResponsesPassthroughAdapter(provider: OcxProviderConfig):
     name: "openai-responses",
     passthrough: true as const,
 
-    buildRequest(parsed: OcxParsedRequest, incoming?: IncomingMeta) {
+    buildRequest(parsed: OcxParsedRequest, incoming: IncomingMeta) {
+      const translatorBudget = incoming.translatorBudget;
       const headers: Record<string, string> = { "Content-Type": "application/json" };
       let url: string;
 
@@ -923,8 +1000,17 @@ export function createResponsesPassthroughAdapter(provider: OcxProviderConfig):
         parsed._rawBody,
         forward || parsed._previousResponseInputExpanded === true,
       );
-      if (forward) {
+      const stateless = provider.statelessResponses === true;
+      if (stateless) outBody = stripStatefulResponsesParams(outBody);
+      // A replay miss can leave a function_call_output whose paired function_call sat
+      // in the prefix that was never expanded. A stateless upstream cannot resolve the
+      // pair from its own storage either, so it needs the same repair the forward
+      // backend gets — dropping previous_response_id is not much use if the body that
+      // reaches the wire is unparseable.
+      if (forward || stateless) {
         outBody = repairOrphanedInputItems(outBody, unexpandedMiss);
+      }
+      if (forward) {
         outBody = stripUnsupportedForwardParams(outBody);
       }
       else outBody = normalizeImageGenClientTools(outBody);
@@ -938,41 +1024,62 @@ export function createResponsesPassthroughAdapter(provider: OcxProviderConfig):
       if (parsed._compactionRequest === true && !isCanonicalOpenAiForwardProvider(provider)) {
         outBody = buildRoutedCompactionBody(outBody);
       }
-      const sanitizedBody = stripSparkCompatibility(stripUnsupportedReasoningParams(stripItemIdsWhenUnstored(stripInvalidItemIds(stripUnsupportedHostedTools(sanitizeReasoningInputContent(scrubOcxCompactionItems(outBody)))))));
+      const sanitizedBody = normalizeToolSchemas(stripSparkCompatibility(stripUnsupportedReasoningParams(stripItemIdsWhenUnstored(stripInvalidItemIds(stripUnsupportedHostedTools(sanitizeReasoningInputContent(scrubOcxCompactionItems(outBody))))))));
+      const body = JSON.stringify(stripDisabledReasoningSummaries(
+        normalizeConfiguredReasoningSummaryDelivery(sanitizedBody, provider, parsed.modelId),
+        provider,
+        parsed.modelId,
+      ));
+      const releaseBodyObservation = translatorBudget.observeExternallyCapped(
+        "passthrough_serialization",
+        new TextEncoder().encode(body).byteLength,
+      );
       return {
         url,
         method: "POST",
         headers,
-        body: JSON.stringify(stripDisabledReasoningSummaries(
-          normalizeConfiguredReasoningSummaryDelivery(sanitizedBody, provider, parsed.modelId),
-          provider,
-          parsed.modelId,
-        )),
+        body,
+        releaseBodyObservation,
       };
     },
 
     // The passthrough normally relays the upstream stream verbatim and never parses.
     // The exception is a routed compaction turn: the server drives this adapter like
     // an ordinary one so the bridge can build the single compaction item (#422).
-    async *parseStream(response: Response): AsyncGenerator<AdapterEvent> {
+    async *parseStream(response: Response, budget: TranslatorBudget): AsyncGenerator<AdapterEvent> {
       if (!response.body) {
         yield { type: "error", message: "passthrough adapter received no response body" };
         return;
       }
+      const budgetEncoder = new TextEncoder();
       let deltas = "";
       let doneText = "";
       let snapshot = "";
       let usage: OcxUsage | undefined;
-      for await (const event of decodeServerSentEvents(response.body)) {
+      for await (const event of decodeServerSentEvents(response.body, { translatorBudget: budget })) {
         let payload: unknown;
         try { payload = JSON.parse(event.data); } catch { continue; }
         if (!isPlainObject(payload)) continue;
         switch (payload.type) {
           case "response.output_text.delta":
-            if (typeof payload.delta === "string") deltas += payload.delta;
+            if (typeof payload.delta === "string") {
+              const next = deltas + payload.delta;
+              const previousBytes = budgetEncoder.encode(deltas).byteLength;
+              const reservation = budget.reserveTransient(budgetEncoder.encode(next).byteLength, { kind: "retained_collectors" });
+              deltas = next;
+              reservation.commitRetained();
+              budget.releaseRetained(previousBytes, { kind: "retained_collectors" });
+            }
             break;
           case "response.output_text.done":
-            if (typeof payload.text === "string") doneText += payload.text;
+            if (typeof payload.text === "string") {
+              const next = doneText + payload.text;
+              const previousBytes = budgetEncoder.encode(doneText).byteLength;
+              const reservation = budget.reserveTransient(budgetEncoder.encode(next).byteLength, { kind: "retained_collectors" });
+              doneText = next;
+              reservation.commitRetained();
+              budget.releaseRetained(previousBytes, { kind: "retained_collectors" });
+            }
             break;
           case "response.failed":
           case "error":
@@ -982,7 +1089,14 @@ export function createResponsesPassthroughAdapter(provider: OcxProviderConfig):
             yield { type: "incomplete", reason: responsesErrorMessage(payload.response ?? payload) };
             return;
           case "response.completed":
-            snapshot = responsesPayloadText(payload.response);
+            {
+              const next = responsesPayloadText(payload.response);
+              const previousBytes = budgetEncoder.encode(snapshot).byteLength;
+              const reservation = budget.reserveTransient(budgetEncoder.encode(next).byteLength, { kind: "retained_collectors" });
+              snapshot = next;
+              reservation.commitRetained();
+              budget.releaseRetained(previousBytes, { kind: "retained_collectors" });
+            }
             usage = usageFromResponsesPayload(payload.response);
             break;
         }
@@ -991,14 +1105,16 @@ export function createResponsesPassthroughAdapter(provider: OcxProviderConfig):
       // completed snapshot so text is never double-counted.
       const text = snapshot || doneText || deltas;
       if (text) yield { type: "text_delta", text };
+      budget.releaseRetained(budgetEncoder.encode(deltas).byteLength + budgetEncoder.encode(doneText).byteLength + budgetEncoder.encode(snapshot).byteLength, { kind: "retained_collectors" });
       yield { type: "done", ...(usage ? { usage } : {}) };
     },
 
-    async parseResponse(response: Response): Promise<AdapterEvent[]> {
+    async parseResponse(response: Response, budget: TranslatorBudget): Promise<AdapterEvent[]> {
       let payload: unknown;
       try { payload = await response.json(); } catch {
         return [{ type: "error", message: "malformed upstream compaction response" }];
       }
+      budget.chargeRetained(new TextEncoder().encode(JSON.stringify(payload)).byteLength, { kind: "retained_collectors" });
       if (!isPlainObject(payload)) {
         return [{ type: "error", message: "malformed upstream compaction response" }];
       }

package/src/adapters/run-turn-queue.ts

@@ -2,6 +2,8 @@ import type { AdapterEvent } from "../types";
 
 type QueueReader = (result: IteratorResult<AdapterEvent>) => void;
 
+export const PREFLIGHT_HEARTBEAT_RETAIN_LIMIT = 16;
+
 export interface AdapterEventQueue {
   push(event: AdapterEvent): void;
   close(): void;
@@ -39,8 +41,12 @@ export async function preflightAdapterEvents(
   while (true) {
     const next = await iterator.next();
     if (next.done) return { stream: replay(buffered, iterator), empty: true };
+    if (next.value.type === "heartbeat") {
+      buffered.push(next.value);
+      if (buffered.length > PREFLIGHT_HEARTBEAT_RETAIN_LIMIT) buffered.shift();
+      continue;
+    }
     buffered.push(next.value);
-    if (next.value.type === "heartbeat") continue;
     if (next.value.type === "error") {
       await iterator.return?.();
       return { stream: replay(buffered, iterator), error: next.value, empty: false };