fix(agent): keep the model reply when a server run ends with a failed tool

When a tool failed and the model then answered, the server executor still
marked the run failed and threw the tracker's English "I couldn't complete
the requested operation" template, which landed under a Chinese reply. The
browser runtimes dropped that template in 66af696 for exactly this
reason: the model saw the failed result and replied to it, and the failed
call stays visible on its tool card. It also made every documented
fallback impossible to complete — probe_media failing and
finalize_uploaded_asset proceeding with ingest defaults — because only a
successful retry of the same tool resolves a failure.

turnDisposition no longer takes a failure flag; the tracker still records
outcomes but never decides the terminal status.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Author: 0xsline

CHANGELOG.md

@@ -29,6 +29,9 @@ The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and
 - **`probe_media` now runs the ffprobe bundled with the app instead of the e2b cloud sandbox** — on a machine without an E2B key the tool used to fail with "e2b sandbox is not configured" even though import, previews and export QA were already probing locally. Uploads and bundled assets are read in place; a public URL is fetched through the SSRF-safe transport into a temp file first.  
   **`probe_media` 改用应用自带的 ffprobe，不再依赖 e2b 云沙箱**——没有配置 E2B key 的机器上，这个工具之前会报「e2b sandbox is not configured」，而导入、预览、导出质检其实早就在本地跑 ffprobe 了。上传文件和内置素材直接就地读取；公开 URL 先经过防 SSRF 的抓取落到临时文件再探测。
 
+- **A server-side agent run no longer ends with "I couldn't complete the requested operation" under the model's own reply** — when a tool failed and the model then answered, the run was still marked failed and an English template was appended beneath a Chinese answer. The reply is the outcome; the failed call stays visible on its tool card, matching the in-browser runtime. This also lets the documented fallbacks (probe fails → finalize with ingest defaults) complete.  
+  **服务端 Agent 运行不再在模型回复下方追加「I couldn't complete the requested operation」**——工具失败后模型已经作答，整轮却仍被判为失败，并在中文回复下面贴一段英文模板。现在以模型的回复为准，失败的调用仍在其工具卡片上可见，与浏览器端运行时一致；也让文档约定的降级路径（探测失败 → 用导入默认值 finalize）能正常完成。
+
 - **Media-pool ratio badges snap to the ratio people actually name** — a 427×240 trailer read "427:240" because the badge reduced the exact pixel fraction; encoders round to codec-friendly sizes, so a frame within 2% of 16:9, 4:3, 1:1, 3:2, 5:4 or 21:9 (and their portrait forms) now shows that name, a small exact fraction such as 7:5 stays, and anything else shows a proportion like 2.40:1. Canvas sizes are exact and unchanged.  
   **素材池比例角标按人们常说的比例显示**——427×240 的预告片之前显示「427:240」，因为角标直接约分了像素分数；编码器会把尺寸凑成编码友好的数值，所以现在与 16:9、4:3、1:1、3:2、5:4、21:9（及其竖版）偏差在 2% 内的都显示该名称，7:5 这类小分数保留，其余显示为 2.40:1 这样的比例。画布尺寸精确，不受影响。
 

server/agent-runs/acceptance-loop.ts

@@ -86,13 +86,19 @@ export function acceptanceInstructions(enabled: boolean): string {
     : '';
 }
 
-export type TurnDisposition = 'continue' | 'completed' | 'failed' | 'max-tokens';
-export function turnDisposition(
-  hitMaxTokens: boolean,
-  continued: boolean,
-  hasUnresolvedFailure = false,
-): TurnDisposition {
-  if (hitMaxTokens) return hasUnresolvedFailure ? 'failed' : 'max-tokens';
-  if (continued) return 'continue';
-  return hasUnresolvedFailure ? 'failed' : 'completed';
+export type TurnDisposition = 'continue' | 'completed' | 'max-tokens';
+/**
+ * A turn that ends without tool calls is complete even when a tool failed earlier in
+ * the run. The model saw the failed result in its own context and replied to it, so
+ * its reply — in the interface language — is the outcome; the failed call stays
+ * visible on its own tool card. This used to return 'failed' and the executor then
+ * threw the tracker's English "I couldn't complete the requested operation" template
+ * underneath the model's reply, which is exactly what the browser runtimes stopped
+ * doing (66af696b). It also broke every documented fallback: probe_media failing and
+ * finalize_uploaded_asset proceeding with ingest defaults could never complete, since
+ * only a successful retry of the same tool resolves a failure.
+ */
+export function turnDisposition(hitMaxTokens: boolean, continued: boolean): TurnDisposition {
+  if (hitMaxTokens) return 'max-tokens';
+  return continued ? 'continue' : 'completed';
 }

server/agent-runs/browser-tool.ts

@@ -22,6 +22,7 @@ export interface ActivationState {
   current: ToolActivation;
   tail: Promise<void>;
   followupText: string | null;
+  /** Records tool outcomes for the run; it never decides the terminal status (see turnDisposition). */
   toolFailures: ToolFailureTracker;
   acceptance: AcceptanceLoopState;
   repeatGuardNote?: string;

server/agent-runs/executor.ts

@@ -348,18 +348,11 @@ async function executeRunTurns(
       return;
     }
     messages = outcome.messages;
-    const disposition = turnDisposition(
-      outcome.hitMaxTokens,
-      outcome.continued,
-      plan.activation.toolFailures.hasUnresolved,
-    );
+    const disposition = turnDisposition(outcome.hitMaxTokens, outcome.continued);
     if (disposition === 'continue') continue;
     if (disposition === 'max-tokens') {
       pushRunEvent(run, 'max-tokens', { turn: turn + 1 });
     }
-    if (disposition === 'failed') {
-      throw new Error(plan.activation.toolFailures.report());
-    }
     if (disposition === 'completed') {
       const acceptance = decideAcceptanceAfterTurn(plan.activation.acceptance);
       plan.activation.acceptance = acceptance.state;

server/agent-runs/executor.verify.ts

@@ -1,4 +1,5 @@
 import assert from 'node:assert/strict';
+import { readFileSync } from 'node:fs';
 import { APICallError } from 'ai';
 import { ASK_MODE_TOOL_SCHEMAS } from '../../src/agent/ask-mode-tools';
 import { TOOL_SCHEMAS } from '../../src/agent/tools';
@@ -326,14 +327,19 @@ console.log('server agent executor message verification passed');
 // instead of feeding truncated text back into the next turn.
 assert.equal(turnDisposition(false, true), 'continue');
 assert.equal(turnDisposition(false, false), 'completed');
-assert.equal(turnDisposition(false, true, true), 'continue',
-  'an unresolved tool failure may continue only while the model is retrying');
-assert.equal(turnDisposition(false, false, true), 'failed',
-  'completion is rejected while a tool failure remains unresolved');
-assert.equal(turnDisposition(true, true, true), 'failed',
-  'a token cutoff cannot turn an unresolved tool failure into completion');
 assert.equal(turnDisposition(true, true), 'max-tokens', 'output cutoff wins over pending tool calls');
 assert.equal(turnDisposition(true, false), 'max-tokens');
+// An unresolved tool failure is not a disposition input: the model already saw the
+// failed result and replied to it, and that reply is the run's outcome. Failing the run
+// here put the tracker's English template under a Chinese reply (a probe_media that
+// could not run, then a complete answer, then "I couldn't complete the requested
+// operation"), and made the documented probe→finalize fallback impossible to complete.
+assert.equal(turnDisposition.length, 2, 'the disposition takes no failure flag');
+assert.doesNotMatch(
+  readFileSync(new URL('./executor.ts', import.meta.url), 'utf8'),
+  /toolFailures\.report\(\)/,
+  'the executor never surfaces the failure-report template',
+);
 
 console.log('server executor turn-disposition checks passed');