agent-accelerator 0.5.5 → 0.5.6

package/README.md

@@ -713,9 +713,12 @@ Thinking levels map to `reasoning.effort` verbatim, the same as OpenAI (`dynamic
 It also:
 
 * passes `service_tier` (`flex` | `priority`);
-* enables the `file-parser` plugin when file parts are present (remote file URLs are additionally surfaced in text);
+* sends audio with its real format (`wav`, `mp3`, `aiff`, `aac`, `ogg`, `flac`, `m4a`) and QuickTime video as `video/mov`;
+* sends file parts as `file.file_data` (a public URL or a base64 data URL) and never adds a plugin: OpenRouter reads the file natively or with its own default parser;
 * maps `tool_choice` (`auto` omitted; `none`, `required`, and function pins supported);
-* reports `cached_tokens` / `cache_write_tokens` / reasoning tokens, preferring provider-reported `cost` when present.
+* treats an error chunk inside a stream (HTTP 200) like an HTTP error: its status comes from the chunk's `code`, or from its `error_type` (`rate_limit_exceeded` 429, `provider_overloaded` 503, `provider_unavailable` 502, `timeout` 504, `server` 500), so `runRetries` retries the retryable ones;
+* reports `cached_tokens` / `cache_write_tokens` / reasoning tokens, preferring provider-reported `cost` when present (a reported `0`, a free model or a cache hit, stays a known `$0`);
+* accepts `anthropic/...` and `~anthropic/...` model names for the `5m` / `1h` cache setting, and sends at most 4 `stop` sequences (the Chat Completions maximum; extra ones are dropped with one `provider_warning`).
 
 `serviceTier` values other than `flex` / `priority` are omitted (standard routing).
 
@@ -1300,8 +1303,11 @@ documents. The supported lists are the `GOOGLE_*_MIME` constants at the top of
 `resolution` (`MediaResolution`: `"low" | "medium" | "high" | "ultra_high"`)
 sets how much detail Google spends on one image or video, and so how many
 input tokens it costs (image: 280 / 560 / 1120 / 2240; video per frame:
-`low` and `medium` 70, `high` 280). Omit it for Google's default. Other
-providers ignore it with one `provider_warning`.
+`low` and `medium` 70, `high` 280). Omit it for Google's default. On
+OpenRouter an image's `resolution` is sent as `image_url.detail` (`low` → `low`,
+`medium` → `auto`, `high` → `high`, `ultra_high` → `original`, which OpenRouter
+downgrades to `high` where unsupported); a video's `resolution` is ignored. Other
+providers (and OpenRouter videos) ignore it with one `provider_warning`.
 
 Before any network call, the executor checks `image` / `audio` / `video` /
 `file`-as-PDF parts against the catalog's `modalities.input` for that

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-accelerator",
-  "version": "0.5.5",
+  "version": "0.5.6",
   "description": "High-performance SDK for AI agents with a unified API across AI providers. Optimized for high cache hit rates, observability, and seamless agent orchestration. Built to power production-grade agentic systems and harnesses for teams at any scale.",
   "license": "MIT",
   "author": "Sashvat Bharat",

package/src/agent/loop.ts

@@ -418,7 +418,7 @@ async function generateWithTimeout<T>(task: () => Promise<T>, timeoutMs: number)
 /** Prices one usage snapshot with the one cost model (`priceUsage`): provider-reported total first, else catalog pricing scaled by `serviceTier`; undefined when unknown. In: loop turns and `accumulateUsage`, out: `usage.cost`. */
 export function computeCostFromPricing(usage: TokenUsage, spec?: ModelSpec, serviceTier?: string): TokenUsage["cost"] {
   try {
-    if (usage.cost && usage.cost.totalCost !== undefined && usage.cost.totalCost > 0) {
+    if (usage.cost && typeof usage.cost.totalCost === "number" && usage.cost.totalCost >= 0) {
       return usage.cost;
     }
     if (!spec) return usage.cost;

package/src/providers/google.ts

@@ -58,6 +58,8 @@ import {
   retryOptionsFrom,
   headersToRecord,
   parseRetryAfterMs,
+  linkStreamAbort,
+  withAbortTimer,
 } from "./shared.ts";
 import { telemetryBus } from "../telemetry.ts";
 import { createGenericModelSpec } from "../models/catalog.ts";
@@ -1097,23 +1099,7 @@ export class GoogleInteractionsProvider implements Provider {
     const startTime = Date.now();
     const clean = cleanModelId(model);
 
-    const linked = new AbortController();
-    const forwardUserAbort = () => {
-      try {
-        linked.abort((options?.signal as { reason?: unknown })?.reason);
-      } catch {
-        try {
-          linked.abort();
-        } catch {}
-      }
-    };
-    if (options?.signal?.aborted) forwardUserAbort();
-    else options?.signal?.addEventListener("abort", forwardUserAbort, { once: true });
-    const removeStreamCancel = eventStream.onCancel(() => {
-      try {
-        linked.abort();
-      } catch {}
-    });
+    const { linked, release } = linkStreamAbort(eventStream, options?.signal);
 
     (async () => {
       // Hoisted for the catch-block self-heal (consts inside try are not
@@ -1159,21 +1145,7 @@ export class GoogleInteractionsProvider implements Provider {
         // (429/5xx/network) are retried before the first byte; a live stream is never retried.
         const open = (target: string, method: "POST" | "GET"): Promise<Response> =>
           openStreamResponse({
-            fetchOnce: async () => {
-              const timer =
-                streamTimeoutMs && streamTimeoutMs > 0
-                  ? setTimeout(() => {
-                      try {
-                        linked.abort();
-                      } catch {}
-                    }, streamTimeoutMs)
-                  : undefined;
-              try {
-                return await (options?.fetch ?? fetch)(target, this.requestInit(method, method === "POST" ? body : undefined, options, apiKey, linked.signal));
-              } finally {
-                if (timer) clearTimeout(timer);
-              }
-            },
+            fetchOnce: () => withAbortTimer(linked, streamTimeoutMs, () => (options?.fetch ?? fetch)(target, this.requestInit(method, method === "POST" ? body : undefined, options, apiKey, linked.signal))),
             onHttpError: (r, text) => this.throwIfError(r.status, r.statusText, target, text, clean, headersToRecord(r.headers)),
             noBodyError: () => toConciseProviderError(new Error("Google streaming response had no body"), "google", clean),
             retry: retryOptionsFrom(options),
@@ -1210,7 +1182,7 @@ export class GoogleInteractionsProvider implements Provider {
           () => {
             aborted = true;
             try {
-              void reader?.cancel();
+              void reader?.cancel().catch(() => {}); // an aborted body is already errored: its cancel() rejects, and nobody else listens
             } catch {}
           },
           { once: true }
@@ -1465,12 +1437,7 @@ export class GoogleInteractionsProvider implements Provider {
           eventStream.fail(raw);
         }
       } finally {
-        try {
-          options?.signal?.removeEventListener("abort", forwardUserAbort);
-        } catch {}
-        try {
-          removeStreamCancel();
-        } catch {}
+        release();
       }
     })();
 

package/src/providers/openai-compat.ts

@@ -52,7 +52,6 @@ import type { ProviderContext, ContentPart } from "../types/message.ts";
 import type { StandardToolDeclaration, ToolCallRecord } from "../types/tool.ts";
 import type { TokenUsage, ThinkingConfig } from "../types/core.ts";
 import { AssistantMessageEventStream } from "../streaming/event-stream.ts";
-import { SSEParser } from "../streaming/sse-parser.ts";
 import { AgentResponse } from "../types/response.ts";
 import { getApiKey, getEnv } from "../utils/env.ts";
 import { buildSessionHeaders, setHeaderCaseInsensitive, shouldEchoSignature } from "../utils/headers.ts";
@@ -68,7 +67,14 @@ import {
   redactedHeaders,
   readErrorPayload,
   incompleteStreamError,
-  hasTruncatedToolArguments,
+  hasUnfinishedToolCall,
+  linkStreamAbort,
+  withAbortTimer,
+  readSseData,
+  incompleteStreamCheck,
+  failStream,
+  toolArgumentsJson,
+  toolTurnFinishReason,
   combinedSignal,
   timeoutFor,
   hasAuthHeader,
@@ -450,7 +456,7 @@ async function fullHistoryMessages(context: ProviderContext): Promise<ChatMessag
           } = {
             id: clampToolCallId(part.callId || part.id),
             type: "function",
-            function: { name: part.name, arguments: JSON.stringify(part.arguments || {}) },
+            function: { name: part.name, arguments: toolArgumentsJson(part.arguments) },
           };
           // Echo provider-issued thought signatures verbatim so endpoints
           // that minted them (Gemini behind a compat proxy) keep working.
@@ -741,7 +747,7 @@ export class OpenAICompatibleChatProvider implements Provider {
     const compatTools = toCompatTools(tools);
     if (compatTools) body.tools = compatTools;
     // Standard Chat Completions generation fields, only when the caller set them.
-    Object.assign(body, mapGenerationParamsToChat(options));
+    Object.assign(body, mapGenerationParamsToChat(options, this.prefix, `${this.prefix}/${modelId}`));
     const toolChoice = mapToolChoiceToOpenAI(
       options?.toolChoice as "auto" | "none" | "required" | { type: "function"; function: { name: string } } | undefined
     );
@@ -949,23 +955,7 @@ export class OpenAICompatibleChatProvider implements Provider {
     const prefix = this.prefix;
     const clean = cleanModelId(this.prefix, model);
 
-    const linked = new AbortController();
-    const forwardUserAbort = () => {
-      try {
-        linked.abort((options?.signal as { reason?: unknown })?.reason);
-      } catch {
-        try {
-          linked.abort();
-        } catch {}
-      }
-    };
-    if (options?.signal?.aborted) forwardUserAbort();
-    else options?.signal?.addEventListener("abort", forwardUserAbort, { once: true });
-    const removeStreamCancel = eventStream.onCancel(() => {
-      try {
-        linked.abort();
-      } catch {}
-    });
+    const { linked, release } = linkStreamAbort(eventStream, options?.signal);
 
     (async () => {
       try {
@@ -994,21 +984,7 @@ export class OpenAICompatibleChatProvider implements Provider {
         // Transient failures (429/5xx/network) are retried here, before the
         // first byte; once the stream is live it is never retried.
         const res = await openStreamResponse({
-          fetchOnce: async () => {
-            const timer =
-              streamTimeoutMs && streamTimeoutMs > 0
-                ? setTimeout(() => {
-                    try {
-                      linked.abort();
-                    } catch {}
-                  }, streamTimeoutMs)
-                : undefined;
-            try {
-              return await (options?.fetch ?? fetch)(url, this.requestInit(body, options, apiKey, sessionId, linked.signal));
-            } finally {
-              if (timer) clearTimeout(timer);
-            }
-          },
+          fetchOnce: () => withAbortTimer(linked, streamTimeoutMs, () => (options?.fetch ?? fetch)(url, this.requestInit(body, options, apiKey, sessionId, linked.signal))),
           onHttpError: (r, text) => this.throwIfError(r.status, url, text, clean, headersToRecord(r.headers)),
           noBodyError: () => toConciseProviderError(new Error("OpenAI-compatible streaming response had no body"), prefix, clean),
           retry: retryOptionsFrom(options),
@@ -1022,9 +998,6 @@ export class OpenAICompatibleChatProvider implements Provider {
         const responseMeta: Record<string, unknown> & { status: number; statusText: string; headers: Record<string, string> } = { status: res.status, statusText: res.statusText, headers: responseHeaders };
         eventStream.push({ type: "start", raw: { request: rawRequest } } as never);
 
-        const parser = new SSEParser();
-        const reader = res.body!.getReader(); // openStreamResponse guarantees a body
-        const decoder = new TextDecoder();
         let text = "";
         let thinking = "";
         // Tool calls keyed by choice index: { id, name, startArgs, deltaArgs }.
@@ -1040,18 +1013,6 @@ export class OpenAICompatibleChatProvider implements Provider {
         let framesSeen = 0;
         let responseId: string | undefined;
         let completedBody: unknown = undefined;
-        let aborted = false;
-        linked.signal.addEventListener(
-          "abort",
-          () => {
-            aborted = true;
-            try {
-              void reader.cancel();
-            } catch {}
-          },
-          { once: true }
-        );
-
         const handleMessage = (data: string): void => {
           if (!data) return;
           if (data === "[DONE]") {
@@ -1176,26 +1137,7 @@ export class OpenAICompatibleChatProvider implements Provider {
           }
         };
 
-        while (true) {
-          if (linked.signal.aborted || eventStream.isCancelled()) {
-            try {
-              await reader.cancel();
-            } catch {}
-            throw Object.assign(new Error("Stream aborted"), { name: "AbortError" });
-          }
-          const { done, value } = await reader.read();
-          if (done) break;
-          const chunk = decoder.decode(value, { stream: true });
-          for (const m of parser.feed(chunk)) handleMessage(m.data);
-        }
-        for (const m of parser.flush()) handleMessage(m.data);
-        try {
-          reader.releaseLock();
-        } catch {}
-
-        if (linked.signal.aborted || eventStream.isCancelled() || aborted) {
-          throw Object.assign(new Error("Stream aborted"), { name: "AbortError" });
-        }
+        await readSseData(res.body!, linked, eventStream, handleMessage); // openStreamResponse guarantees a body
 
         const toolCalls: ToolCallRecord[] = [];
         for (const [index, c] of calls) {
@@ -1212,9 +1154,11 @@ export class OpenAICompatibleChatProvider implements Provider {
           eventStream.push({ type: "tool_call_complete", toolCall: record });
         }
 
+        // Did the stream really end? Read before the tool-call and refusal overrides below, which also set sawFinishReason.
+        const hasTerminal = completedBody !== undefined || sawFinishReason || sawUsage;
         noteProviderTurn(sessionId, prefix);
         if (toolCalls.length > 0) {
-          finishReason = "tool_calls";
+          finishReason = toolTurnFinishReason(finishReason, toolCalls);
           sawFinishReason = true;
         } else if (refusalSeen) {
           finishReason = "refusal";
@@ -1224,26 +1168,21 @@ export class OpenAICompatibleChatProvider implements Provider {
         // Q-23: terminal-state guard — fail `incomplete_stream` (retryable)
         // instead of resolving empty success. Skipped for non-SSE payloads
         // (JSON mocks) where no SSE frames were ever seen.
-        const hasTerminal = completedBody !== undefined || sawFinishReason || sawUsage;
-        const cleanThinkingEarly = thinking.replace(/\n{3,}/g, "\n\n").trim();
-        const ctEntry = Object.entries(responseHeaders).find(([k]) => k.toLowerCase() === "content-type");
-        const isSSEPayload =
-          (typeof ctEntry?.[1] === "string" && ((ctEntry[1] as string).includes("event-stream") || (ctEntry[1] as string).startsWith("text/"))) ||
-          framesSeen > 0 ||
-          badFrames > 0;
-        if (!hasTerminal && !text && !cleanThinkingEarly && toolCalls.length === 0 && isSSEPayload) {
-          eventStream.fail(incompleteStreamError(prefix, clean));
-          return;
-        }
-        // Refuse unfinished tool calls whose args never parsed (output_item.done
-        // never arrived): truncated `{raw}` must not reach the executor (pi
-        // api/openai-responses-shared.ts:766 rule). Non-SSE mocks preserve legacy.
-        if (!hasTerminal && toolCalls.some((c) => hasTruncatedToolArguments(c.arguments)) && isSSEPayload) {
-          eventStream.fail(incompleteStreamError(prefix, clean));
+        const cleanThinking = thinking.replace(/\n{3,}/g, "\n\n").trim();
+        const incomplete = incompleteStreamCheck({
+          provider: prefix,
+          modelId: clean,
+          hasTerminal,
+          empty: !text && !cleanThinking && toolCalls.length === 0,
+          unfinishedCall: hasUnfinishedToolCall(calls.values(), toolCalls),
+          responseHeaders,
+          framesSeen,
+          badFrames,
+        });
+        if (incomplete) {
+          eventStream.fail(incomplete);
           return;
         }
-
-        const cleanThinking = cleanThinkingEarly;
         const finalThoughtSignature = toolCalls.map((c) => c.thoughtSignature).find(Boolean);
         const finalResponse = new AgentResponse({
           text,
@@ -1264,21 +1203,9 @@ export class OpenAICompatibleChatProvider implements Provider {
         eventStream.push({ type: "done", delta: "", usage, finishReason, responseId });
         eventStream.end(finalResponse);
       } catch (err: unknown) {
-        const raw = err instanceof Error ? err : new Error(String(err));
-        const isAbort =
-          linked.signal.aborted ||
-          eventStream.isCancelled() ||
-          (raw as { name?: string }).name === "AbortError";
-        eventStream.fail(
-          isAbort ? Object.assign(new Error("Stream aborted"), { name: "AbortError" }) : raw
-        );
+        failStream(eventStream, linked, err);
       } finally {
-        try {
-          options?.signal?.removeEventListener("abort", forwardUserAbort);
-        } catch {}
-        try {
-          removeStreamCancel();
-        } catch {}
+        release();
       }
     })();
 

package/src/providers/openai.ts

@@ -44,7 +44,6 @@ import type { ProviderContext, ContentPart } from "../types/message.ts";
 import type { StandardToolDeclaration, ToolCallRecord } from "../types/tool.ts";
 import type { TokenUsage } from "../types/core.ts";
 import { AssistantMessageEventStream } from "../streaming/event-stream.ts";
-import { SSEParser } from "../streaming/sse-parser.ts";
 import { AgentResponse } from "../types/response.ts";
 import { getApiKey, getEnv } from "../utils/env.ts";
 import { buildSessionHeaders, setHeaderCaseInsensitive } from "../utils/headers.ts";
@@ -60,7 +59,14 @@ import {
   redactedHeaders,
   readErrorPayload,
   incompleteStreamError,
-  hasTruncatedToolArguments,
+  hasUnfinishedToolCall,
+  linkStreamAbort,
+  withAbortTimer,
+  readSseData,
+  incompleteStreamCheck,
+  failStream,
+  toolArgumentsJson,
+  toolTurnFinishReason,
   combinedSignal,
   timeoutFor,
   hasAuthHeader,
@@ -340,7 +346,7 @@ async function fullHistoryInput(context: ProviderContext): Promise<ResponseInput
             id: normalizeResponsesItemId(itemId),
             call_id: normalizeResponsesIdPart(callId),
             name: part.name,
-            arguments: JSON.stringify(part.arguments || {}),
+            arguments: toolArgumentsJson(part.arguments),
           });
         } else if (part.type === "text" && part.text) {
           texts.push(part.text);
@@ -776,23 +782,7 @@ export class OpenAIResponsesProvider implements Provider {
     const startTime = Date.now();
     const clean = cleanModelId(model);
 
-    const linked = new AbortController();
-    const forwardUserAbort = () => {
-      try {
-        linked.abort((options?.signal as { reason?: unknown })?.reason);
-      } catch {
-        try {
-          linked.abort();
-        } catch {}
-      }
-    };
-    if (options?.signal?.aborted) forwardUserAbort();
-    else options?.signal?.addEventListener("abort", forwardUserAbort, { once: true });
-    const removeStreamCancel = eventStream.onCancel(() => {
-      try {
-        linked.abort();
-      } catch {}
-    });
+    const { linked, release } = linkStreamAbort(eventStream, options?.signal);
 
     (async () => {
       try {
@@ -827,21 +817,7 @@ export class OpenAIResponsesProvider implements Provider {
         // Transient failures (429/5xx/network) are retried here, before the
         // first byte; once the stream is live it is never retried.
         const res = await openStreamResponse({
-          fetchOnce: async () => {
-            const timer =
-              streamTimeoutMs && streamTimeoutMs > 0
-                ? setTimeout(() => {
-                    try {
-                      linked.abort();
-                    } catch {}
-                  }, streamTimeoutMs)
-                : undefined;
-            try {
-              return await (options?.fetch ?? fetch)(url, this.requestInit(body, options, apiKey, sessionId, linked.signal));
-            } finally {
-              if (timer) clearTimeout(timer);
-            }
-          },
+          fetchOnce: () => withAbortTimer(linked, streamTimeoutMs, () => (options?.fetch ?? fetch)(url, this.requestInit(body, options, apiKey, sessionId, linked.signal))),
           onHttpError: (r, text) => this.throwIfError(r.status, url, text, clean, headersToRecord(r.headers)),
           noBodyError: () => toConciseProviderError(new Error("OpenAI streaming response had no body"), "openai", clean),
           retry: retryOptionsFrom(options),
@@ -855,9 +831,6 @@ export class OpenAIResponsesProvider implements Provider {
         const responseMeta: Record<string, unknown> & { status: number; statusText: string; headers: Record<string, string> } = { status: res.status, statusText: res.statusText, headers: responseHeaders };
         eventStream.push({ type: "start", raw: { request: rawRequest } } as never);
 
-        const parser = new SSEParser();
-        const reader = res.body!.getReader(); // openStreamResponse guarantees a body
-        const decoder = new TextDecoder();
         let text = "";
         let thinking = "";
         // Tool calls keyed by output_index: { itemId, callId, name, startArgs, deltaArgs }.
@@ -871,18 +844,6 @@ export class OpenAIResponsesProvider implements Provider {
         let framesSeen = 0;
         let responseId: string | undefined;
         let completedBody: unknown = undefined;
-        let aborted = false;
-        linked.signal.addEventListener(
-          "abort",
-          () => {
-            aborted = true;
-            try {
-              void reader.cancel();
-            } catch {}
-          },
-          { once: true }
-        );
-
         const handleMessage = (data: string): void => {
           if (!data || data === "[DONE]") {
             if (data === "[DONE]") sawFinishReason = sawFinishReason || completedBody !== undefined || sawUsage;
@@ -1035,26 +996,7 @@ export class OpenAIResponsesProvider implements Provider {
           }
         };
 
-        while (true) {
-          if (linked.signal.aborted || eventStream.isCancelled()) {
-            try {
-              await reader.cancel();
-            } catch {}
-            throw Object.assign(new Error("Stream aborted"), { name: "AbortError" });
-          }
-          const { done, value } = await reader.read();
-          if (done) break;
-          const chunk = decoder.decode(value, { stream: true });
-          for (const m of parser.feed(chunk)) handleMessage(m.data);
-        }
-        for (const m of parser.flush()) handleMessage(m.data);
-        try {
-          reader.releaseLock();
-        } catch {}
-
-        if (linked.signal.aborted || eventStream.isCancelled() || aborted) {
-          throw Object.assign(new Error("Stream aborted"), { name: "AbortError" });
-        }
+        await readSseData(res.body!, linked, eventStream, handleMessage); // openStreamResponse guarantees a body
 
         const toolCalls: ToolCallRecord[] = [];
         for (const c of calls.values()) {
@@ -1070,9 +1012,11 @@ export class OpenAIResponsesProvider implements Provider {
           eventStream.push({ type: "tool_call_complete", toolCall: record });
         }
 
+        // Did the stream really end? Read before the tool-call and refusal overrides below, which also set sawFinishReason.
+        const hasTerminal = completedBody !== undefined || sawFinishReason || sawUsage;
         noteProviderTurn(sessionId, "openai");
         if (toolCalls.length > 0) {
-          finishReason = "tool_calls";
+          finishReason = toolTurnFinishReason(finishReason, toolCalls);
           sawFinishReason = true;
         } else if (refusalSeen && finishReason !== "length") {
           finishReason = "refusal";
@@ -1085,26 +1029,21 @@ export class OpenAIResponsesProvider implements Provider {
         // Skipped for non-SSE payloads (e.g. JSON mocks in delegation tests)
         // where no SSE frames were ever seen — those preserve legacy empty
         // success rather than failing the caller on a mock artifact.
-        const hasTerminal = completedBody !== undefined || sawFinishReason || sawUsage;
-        const cleanThinkingEarly = thinking.replace(/\n{3,}/g, "\n\n").trim();
-        const ctEntry = Object.entries(responseHeaders).find(([k]) => k.toLowerCase() === "content-type");
-        const isSSEPayload =
-          (typeof ctEntry?.[1] === "string" && ((ctEntry[1] as string).includes("event-stream") || (ctEntry[1] as string).startsWith("text/"))) ||
-          framesSeen > 0 ||
-          badFrames > 0;
-        if (!hasTerminal && !text && !cleanThinkingEarly && toolCalls.length === 0 && isSSEPayload) {
-          eventStream.fail(incompleteStreamError("openai", clean));
-          return;
-        }
-        // Refuse unfinished tool calls whose args never parsed (output_item.done
-        // never arrived): truncated `{raw}` must not reach the executor (pi
-        // api/openai-responses-shared.ts:766 rule). Non-SSE mocks preserve legacy.
-        if (!hasTerminal && toolCalls.some((c) => hasTruncatedToolArguments(c.arguments)) && isSSEPayload) {
-          eventStream.fail(incompleteStreamError("openai", clean));
+        const cleanThinking = thinking.replace(/\n{3,}/g, "\n\n").trim();
+        const incomplete = incompleteStreamCheck({
+          provider: "openai",
+          modelId: clean,
+          hasTerminal,
+          empty: !text && !cleanThinking && toolCalls.length === 0,
+          unfinishedCall: hasUnfinishedToolCall(calls.values(), toolCalls),
+          responseHeaders,
+          framesSeen,
+          badFrames,
+        });
+        if (incomplete) {
+          eventStream.fail(incomplete);
           return;
         }
-
-        const cleanThinking = cleanThinkingEarly;
         const finalResponse = new AgentResponse({
           text,
           thinking: cleanThinking || undefined,
@@ -1124,21 +1063,9 @@ export class OpenAIResponsesProvider implements Provider {
         eventStream.push({ type: "done", delta: "", usage, finishReason, responseId });
         eventStream.end(finalResponse);
       } catch (err: unknown) {
-        const raw = err instanceof Error ? err : new Error(String(err));
-        const isAbort =
-          linked.signal.aborted ||
-          eventStream.isCancelled() ||
-          (raw as { name?: string }).name === "AbortError";
-        eventStream.fail(
-          isAbort ? Object.assign(new Error("Stream aborted"), { name: "AbortError" }) : raw
-        );
+        failStream(eventStream, linked, err);
       } finally {
-        try {
-          options?.signal?.removeEventListener("abort", forwardUserAbort);
-        } catch {}
-        try {
-          removeStreamCancel();
-        } catch {}
+        release();
       }
     })();