newmark-agent 0.5.4 → 0.5.8

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -43,9 +43,6 @@ const child_process_1 = require("child_process");
43
43
  const agentKernelDiagnostics_1 = require("../core/agentKernelDiagnostics");
44
44
  const chat_messages_1 = require("../providers/chat-messages");
45
45
  const providers_1 = require("../providers");
46
- // Keep provider requests below the release-harness/user-visible command
47
- // deadline. A provider that does not answer must produce one bounded error;
48
- // it must not restart the same request through every Windows transport.
49
46
  // Provider responses are intentionally unbounded. User cancellation, transport
50
47
  // errors, and tool-specific limits remain the only automatic stop conditions.
51
48
  const DEFAULT_PROVIDER_REQUEST_TIMEOUT_MS = 0;
@@ -939,7 +936,7 @@ class LLMProvider {
939
936
  * `provider_adapters_v2` context flag. Request serialization and SSE
940
937
  * normalization are delegated to the shared provider adapters while the
941
938
  * transport orchestration (loopback node-http, fetch -> node-http fallback,
942
- * 120s/30s timeouts) and the 4xx Chat -> Responses downgrade stay here.
939
+ * cancellation-only streaming and the 4xx Chat -> Responses downgrade) stay here.
943
940
  * The emitted request body and StreamToken stream are byte-equivalent to
944
941
  * the legacy inlined path.
945
942
  */
@@ -1082,9 +1079,9 @@ class LLMProvider {
1082
1079
  }
1083
1080
  /**
1084
1081
  * Loopback-aware transport injected into adapter `execute`. Streaming
1085
- * requests retain the fetch-to-node fallback for transport failures, while
1086
- * a local deadline is returned directly so one request cannot become a
1087
- * second Windows fallback request.
1082
+ * requests retain the fetch-to-node fallback for transport failures. They
1083
+ * have no response deadline; only caller cancellation or a concrete
1084
+ * transport/provider failure may end the request.
1088
1085
  */
1089
1086
  buildProviderAdapterTransport() {
1090
1087
  return async (request, signal) => {
@@ -1096,7 +1093,9 @@ class LLMProvider {
1096
1093
  forwardAbort();
1097
1094
  else
1098
1095
  signal?.addEventListener('abort', forwardAbort, { once: true });
1099
- const effectiveTimeout = this.effectiveRequestTimeout(120000);
1096
+ // Streaming provider responses are intentionally unbounded. Only
1097
+ // caller cancellation or an explicit provider failure may end them.
1098
+ const effectiveTimeout = 0;
1100
1099
  const timer = effectiveTimeout > 0
1101
1100
  ? setTimeout(() => abort.abort(providerTimeoutError(effectiveTimeout)), effectiveTimeout)
1102
1101
  : undefined;
@@ -1247,7 +1246,9 @@ class LLMProvider {
1247
1246
  forwardAbort();
1248
1247
  else
1249
1248
  signal?.addEventListener('abort', forwardAbort, { once: true });
1250
- const effectiveTimeout = this.effectiveRequestTimeout(120000);
1249
+ // Streaming provider responses are intentionally unbounded. Do not turn
1250
+ // silence into an empty-response failure.
1251
+ const effectiveTimeout = 0;
1251
1252
  const timeout = effectiveTimeout > 0
1252
1253
  ? setTimeout(() => abort.abort(providerTimeoutError(effectiveTimeout)), effectiveTimeout)
1253
1254
  : undefined;
@@ -1285,10 +1286,14 @@ class LLMProvider {
1285
1286
  }
1286
1287
  const decoder = new TextDecoder();
1287
1288
  let buffer = '';
1288
- let currentToolCall = null;
1289
+ const toolCalls = new Map();
1290
+ const toolCallOrder = [];
1291
+ let syntheticToolIndex = 0;
1292
+ let lastToolIndex = 0;
1289
1293
  let currentReasoningContent = '';
1290
1294
  let contentPolicyBlocked = false;
1291
1295
  let emittedContent = false;
1296
+ let explicitCompletion = false;
1292
1297
  const streamSignal = signal || new AbortController().signal;
1293
1298
  while (true) {
1294
1299
  const { done, value } = await (0, providers_1.readProviderStreamChunk)(reader, streamSignal);
@@ -1302,8 +1307,10 @@ class LLMProvider {
1302
1307
  if (!trimmed.startsWith('data: '))
1303
1308
  continue;
1304
1309
  const data = trimmed.slice(6);
1305
- if (data === '[DONE]')
1310
+ if (data === '[DONE]') {
1311
+ explicitCompletion = true;
1306
1312
  continue;
1313
+ }
1307
1314
  try {
1308
1315
  const json = JSON.parse(data);
1309
1316
  if (json.usage)
@@ -1324,27 +1331,54 @@ class LLMProvider {
1324
1331
  }
1325
1332
  if (delta.tool_calls) {
1326
1333
  for (const tc of delta.tool_calls) {
1327
- if (tc.id) {
1328
- if (currentToolCall) {
1329
- yield { type: 'tool_call', text: '', toolCall: currentToolCall, reasoningContent: currentReasoningContent || undefined };
1330
- }
1331
- currentToolCall = { id: tc.id, name: tc.function?.name || '', arguments: tc.function?.arguments || '' };
1332
- }
1333
- else if (tc.function?.arguments && currentToolCall) {
1334
- currentToolCall.arguments += tc.function.arguments;
1334
+ const rawIndex = Number(tc.index);
1335
+ const index = Number.isInteger(rawIndex) && rawIndex >= 0
1336
+ ? rawIndex
1337
+ : (tc.id ? syntheticToolIndex++ : lastToolIndex);
1338
+ lastToolIndex = index;
1339
+ let call = toolCalls.get(index);
1340
+ if (!call && (tc.id || tc.function?.name)) {
1341
+ call = { id: tc.id || '', name: tc.function?.name || '', argumentParts: [] };
1342
+ toolCalls.set(index, call);
1343
+ toolCallOrder.push(index);
1335
1344
  }
1345
+ if (!call)
1346
+ continue;
1347
+ if (tc.id && !call.id)
1348
+ call.id = tc.id;
1349
+ if (tc.function?.name && !call.name)
1350
+ call.name = tc.function.name;
1351
+ if (tc.function?.arguments)
1352
+ call.argumentParts.push(tc.function.arguments);
1336
1353
  }
1337
1354
  }
1338
1355
  }
1339
1356
  catch { /* skip malformed JSON */ }
1340
1357
  }
1341
1358
  }
1342
- if (currentToolCall && currentToolCall.arguments) {
1343
- yield { type: 'tool_call', text: '', toolCall: currentToolCall, reasoningContent: currentReasoningContent || undefined };
1359
+ if (toolCallOrder.length) {
1360
+ for (const index of toolCallOrder) {
1361
+ const call = toolCalls.get(index);
1362
+ if (!call)
1363
+ continue;
1364
+ yield {
1365
+ type: 'tool_call',
1366
+ text: '',
1367
+ toolCall: {
1368
+ id: call.id,
1369
+ name: call.name,
1370
+ arguments: (0, providers_1.assembleCompatibleToolArguments)(call.argumentParts),
1371
+ },
1372
+ reasoningContent: currentReasoningContent || undefined,
1373
+ };
1374
+ }
1344
1375
  }
1345
1376
  else if (!emittedContent && contentPolicyBlocked) {
1346
1377
  yield { type: 'text', text: '[Error] Content policy refusal (content_filter).' };
1347
1378
  }
1379
+ else if (!explicitCompletion && !emittedContent && !currentReasoningContent) {
1380
+ yield { type: 'text', text: '[LLM Error] GitHub Models stream ended before an explicit completion.' };
1381
+ }
1348
1382
  }
1349
1383
  finally {
1350
1384
  reader?.releaseLock();
package/dist/main.js CHANGED
@@ -77,6 +77,7 @@ const compat_1 = require("./core/compat");
77
77
  const dshCompatibility_1 = require("./core/dshCompatibility");
78
78
  const mcpManager_1 = require("./core/mcpManager");
79
79
  const workEventCoalescer_1 = require("./core/workEventCoalescer");
80
+ const terminalOutputBuffer_1 = require("./core/terminalOutputBuffer");
80
81
  const cli_help_1 = require("./cli-help");
81
82
  const APP_NAME = 'Newmark Agent';
82
83
  const APP_ID = 'ai.newmark.agent';
@@ -3020,6 +3021,10 @@ else {
3020
3021
  activePromptLeases.set(promptLeaseKey, (activePromptLeases.get(promptLeaseKey) || 0) + 1);
3021
3022
  activePromptWorkspaces.set(promptLeaseWorkspaceKey, (activePromptWorkspaces.get(promptLeaseWorkspaceKey) || 0) + 1);
3022
3023
  const targetConversation = target.conversationId;
3024
+ // 发送命令时锁定输入框选择的模型:接受命令后的整个运行过程都以该
3025
+ // 模型为准(唯一例外是显式的不可用回退,且回退会以结构化事件同步
3026
+ // 到前端输入框下方的选择区,而不是隐藏的参数回退)。
3027
+ const requestedModel = agent.model;
3023
3028
  const options = {
3024
3029
  mode: agent.mode,
3025
3030
  model: agent.ensureUsableModelSelection(),
@@ -3027,6 +3032,21 @@ else {
3027
3032
  inputMode: agent.inputMode,
3028
3033
  engine: agent.engine,
3029
3034
  };
3035
+ if (options.model && requestedModel && options.model !== requestedModel) {
3036
+ const usableConfig = agent.activeModelConfig();
3037
+ broadcastAgentWorkEvent({
3038
+ id: `model-fallback-${Date.now()}-${Math.random().toString(16).slice(2)}`,
3039
+ conversationId: target.conversationId,
3040
+ type: 'status',
3041
+ content: `[Model fallback] ${requestedModel} unavailable; switched to ${options.model}.`,
3042
+ mode: agent.modeName(),
3043
+ model: options.model,
3044
+ timestamp: new Date().toISOString(),
3045
+ workspaceId: target.workspaceId,
3046
+ workspaceKey: target.workspaceKey,
3047
+ fallback: { from: requestedModel, to: options.model, providerId: usableConfig?.provider_id || agent.activeDeployment()?.providerId },
3048
+ });
3049
+ }
3030
3050
  const queueMode = agent.inputMode === 'guide' ? 'steer' : 'followUp';
3031
3051
  let result;
3032
3052
  if (wslBackendEnabled()) {
@@ -4432,10 +4452,16 @@ else {
4432
4452
  }
4433
4453
  });
4434
4454
  const ptySessions = new Map();
4435
- const sendTerminalData = (sessionId, session, text) => {
4436
- session.buffer = `${session.buffer}${text}`.slice(-256 * 1024);
4455
+ const terminalOutput = new terminalOutputBuffer_1.TerminalOutputBuffer((sessionId, text) => {
4456
+ const session = ptySessions.get(sessionId);
4457
+ if (session)
4458
+ session.buffer = terminalOutput.history(sessionId);
4437
4459
  if (mainWindow && !mainWindow.isDestroyed())
4438
4460
  mainWindow.webContents.send('pty:data', sessionId, text);
4461
+ });
4462
+ const sendTerminalData = (sessionId, session, text) => {
4463
+ terminalOutput.push(sessionId, text);
4464
+ session.buffer = terminalOutput.history(sessionId);
4439
4465
  };
4440
4466
  electron_1.ipcMain.handle('pty:spawn', async (_event, shellId) => {
4441
4467
  const sessionId = (0, crypto_1.randomUUID)().slice(0, 8);
@@ -4457,7 +4483,9 @@ else {
4457
4483
  ptySessions.set(sessionId, session);
4458
4484
  proc.onData(text => sendTerminalData(sessionId, session, text));
4459
4485
  proc.onExit(event => {
4486
+ terminalOutput.flush(sessionId);
4460
4487
  ptySessions.delete(sessionId);
4488
+ terminalOutput.close(sessionId);
4461
4489
  if (mainWindow && !mainWindow.isDestroyed()) {
4462
4490
  mainWindow.webContents.send('pty:exit', sessionId, event.exitCode);
4463
4491
  }
@@ -4515,7 +4543,7 @@ else {
4515
4543
  const session = ptySessions.get(sessionId);
4516
4544
  if (!session)
4517
4545
  return { buffer: '' };
4518
- return { buffer: session.buffer };
4546
+ return { buffer: terminalOutput.history(sessionId) || session.buffer };
4519
4547
  });
4520
4548
  const terminalOwnerFor = (conversationId, actorId) => (0, terminalTakeover_1.normalizeTerminalTakeoverOwner)({
4521
4549
  backend: wslBackendEnabled() ? 'wsl' : (process.platform === 'win32' ? 'windows' : process.platform),
@@ -5206,6 +5234,21 @@ else {
5206
5234
  // make an ordinary minimize click disappear from the taskbar.
5207
5235
  win?.minimize();
5208
5236
  });
5237
+ electron_1.ipcMain.handle('glass:captureBackdrop', async (event, requestedSize) => {
5238
+ const image = await event.sender.capturePage();
5239
+ const sourceSize = image.getSize();
5240
+ const width = Math.max(1, Math.min(sourceSize.width, Math.round(Number(requestedSize?.width) || sourceSize.width)));
5241
+ const height = Math.max(1, Math.min(sourceSize.height, Math.round(Number(requestedSize?.height) || sourceSize.height)));
5242
+ const resized = sourceSize.width === width && sourceSize.height === height
5243
+ ? image
5244
+ : image.resize({ width, height, quality: 'good' });
5245
+ return {
5246
+ bytes: resized.toJPEG(82),
5247
+ mimeType: 'image/jpeg',
5248
+ width,
5249
+ height,
5250
+ };
5251
+ });
5209
5252
  electron_1.ipcMain.handle('app:maximize', () => {
5210
5253
  const win = electron_1.BrowserWindow.getFocusedWindow() || mainWindow;
5211
5254
  if (win?.isMaximized())
package/dist/preload.js CHANGED
@@ -6,6 +6,7 @@ contextBridge.exposeInMainWorld('api', {
6
6
  startupWaitForBackend: () => ipcRenderer.invoke('startup:waitForBackend'),
7
7
  startupUiReady: (payload) => ipcRenderer.invoke('startup:uiReady', payload),
8
8
  startupUiFailed: (payload) => ipcRenderer.invoke('startup:uiFailed', payload),
9
+ captureLiquidBackdrop: (size) => ipcRenderer.invoke('glass:captureBackdrop', size),
9
10
  onStartupStatus: (callback) => {
10
11
  ipcRenderer.on('startup:status', (_event, payload) => callback(payload));
11
12
  },
@@ -105,6 +105,8 @@ class ChatCompletionsAdapter {
105
105
  let contentPolicyBlocked = false;
106
106
  let emittedContent = false;
107
107
  let emittedTool = false;
108
+ let emittedReasoning = false;
109
+ let explicitCompletion = false;
108
110
  try {
109
111
  while (true) {
110
112
  const { done, value } = await (0, provider_events_1.readProviderStreamChunk)(reader, signal);
@@ -118,8 +120,10 @@ class ChatCompletionsAdapter {
118
120
  if (!trimmed.startsWith('data: '))
119
121
  continue;
120
122
  const data = trimmed.slice(6);
121
- if (data === '[DONE]')
123
+ if (data === '[DONE]') {
124
+ explicitCompletion = true;
122
125
  continue;
126
+ }
123
127
  let json;
124
128
  try {
125
129
  json = JSON.parse(data);
@@ -135,13 +139,18 @@ class ChatCompletionsAdapter {
135
139
  if ((0, provider_events_1.isContentPolicyBlocked)(json))
136
140
  contentPolicyBlocked = true;
137
141
  const choices = Array.isArray(json.choices) ? json.choices : [];
138
- const delta = choices[0]?.delta;
142
+ const choice = choices[0];
143
+ if (choice?.finish_reason !== undefined && choice.finish_reason !== null)
144
+ explicitCompletion = true;
145
+ const delta = choice?.delta;
139
146
  if (!delta)
140
147
  continue;
141
148
  if (delta.reasoning_content) {
142
149
  const reasoning = this.extractText(delta.reasoning_content);
143
- if (reasoning)
150
+ if (reasoning) {
151
+ emittedReasoning = true;
144
152
  yield { type: 'reasoning.summary.delta', delta: reasoning };
153
+ }
145
154
  }
146
155
  const textDelta = this.extractText(delta.content);
147
156
  if (textDelta) {
@@ -190,7 +199,7 @@ class ChatCompletionsAdapter {
190
199
  type: 'tool_call.completed',
191
200
  id: currentToolCall.id,
192
201
  name: currentToolCall.name,
193
- arguments: currentToolCall.argumentParts.join(''),
202
+ arguments: (0, provider_events_1.assembleCompatibleToolArguments)(currentToolCall.argumentParts),
194
203
  };
195
204
  }
196
205
  }
@@ -198,6 +207,10 @@ class ChatCompletionsAdapter {
198
207
  yield { type: 'response.failed', error: '[Error] Content policy refusal (content_filter).' };
199
208
  return;
200
209
  }
210
+ if (!explicitCompletion && !emittedContent && !emittedTool && !emittedReasoning) {
211
+ yield { type: 'response.failed', error: '[LLM Error] Chat stream ended before an explicit completion.' };
212
+ return;
213
+ }
201
214
  yield { type: 'response.completed' };
202
215
  }
203
216
  finally {
@@ -17,12 +17,25 @@ export declare function providerStreamTimeoutError(timeoutMs: number): Error;
17
17
  * Read one SSE chunk with both user cancellation and an inactivity deadline.
18
18
  * Cancelling the reader is important: rejecting the race alone leaves the
19
19
  * provider socket alive and lets later requests accumulate behind it.
20
+ *
21
+ * timeoutMs defaults to 0 (no stream idle deadline), matching the request-
22
+ * level DEFAULT_PROVIDER_REQUEST_TIMEOUT_MS = 0 and the Android client's
23
+ * readTimeout(0) / SSE_IDLE_TIMEOUT_MS = 0L. A caller that still wants an
24
+ * inactivity cap passes an explicit positive value (the recovery verify
25
+ * passes 50ms to prove reader cancellation).
20
26
  */
21
27
  export declare function readProviderStreamChunk(reader: ReadableStreamDefaultReader<Uint8Array>, signal: AbortSignal, timeoutMs?: number): Promise<ReadableStreamReadResult<Uint8Array>>;
22
28
  export declare function parseProviderSse(raw: string): Array<{
23
29
  event?: string;
24
30
  data: string;
25
31
  }>;
32
+ /**
33
+ * Compatible gateways may stream function arguments as JSON deltas or repeat
34
+ * a cumulative snapshot on every SSE frame. Keep the normal incremental form
35
+ * when it parses, then fall back to snapshot folding. Returning malformed
36
+ * concatenated snapshots would erase the model's correction at tool parsing.
37
+ */
38
+ export declare function assembleCompatibleToolArguments(parts: string[]): string;
26
39
  /**
27
40
  * Detect content-policy refusals across the provider failure shapes.
28
41
  * Mirrors `LLMProvider.contentPolicyBlocked` semantics exactly so the
@@ -5,6 +5,7 @@ exports.providerAbortError = providerAbortError;
5
5
  exports.providerStreamTimeoutError = providerStreamTimeoutError;
6
6
  exports.readProviderStreamChunk = readProviderStreamChunk;
7
7
  exports.parseProviderSse = parseProviderSse;
8
+ exports.assembleCompatibleToolArguments = assembleCompatibleToolArguments;
8
9
  exports.isContentPolicyBlocked = isContentPolicyBlocked;
9
10
  exports.normalizeProviderUsage = normalizeProviderUsage;
10
11
  exports.estimateRequestTokens = estimateRequestTokens;
@@ -46,8 +47,14 @@ function providerStreamTimeoutError(timeoutMs) {
46
47
  * Read one SSE chunk with both user cancellation and an inactivity deadline.
47
48
  * Cancelling the reader is important: rejecting the race alone leaves the
48
49
  * provider socket alive and lets later requests accumulate behind it.
50
+ *
51
+ * timeoutMs defaults to 0 (no stream idle deadline), matching the request-
52
+ * level DEFAULT_PROVIDER_REQUEST_TIMEOUT_MS = 0 and the Android client's
53
+ * readTimeout(0) / SSE_IDLE_TIMEOUT_MS = 0L. A caller that still wants an
54
+ * inactivity cap passes an explicit positive value (the recovery verify
55
+ * passes 50ms to prove reader cancellation).
49
56
  */
50
- async function readProviderStreamChunk(reader, signal, timeoutMs = 30_000) {
57
+ async function readProviderStreamChunk(reader, signal, timeoutMs = 0) {
51
58
  if (signal.aborted)
52
59
  throw providerAbortError(signal);
53
60
  let timer;
@@ -56,9 +63,14 @@ async function readProviderStreamChunk(reader, signal, timeoutMs = 30_000) {
56
63
  onAbort = () => reject(providerAbortError(signal));
57
64
  signal.addEventListener('abort', onAbort, { once: true });
58
65
  });
59
- const timeoutPromise = new Promise((_, reject) => {
60
- timer = setTimeout(() => reject(providerStreamTimeoutError(timeoutMs)), timeoutMs);
61
- });
66
+ // timeoutMs <= 0 means no inactivity deadline (unlimited). setTimeout with
67
+ // 0 would fire on the next tick, so we only arm the timer for positive
68
+ // values and race a never-settling promise otherwise.
69
+ const timeoutPromise = timeoutMs > 0
70
+ ? new Promise((_, reject) => {
71
+ timer = setTimeout(() => reject(providerStreamTimeoutError(timeoutMs)), timeoutMs);
72
+ })
73
+ : new Promise(() => undefined);
62
74
  try {
63
75
  return await Promise.race([reader.read(), abortPromise, timeoutPromise]);
64
76
  }
@@ -94,6 +106,45 @@ function parseProviderSse(raw) {
94
106
  }
95
107
  return events;
96
108
  }
109
+ /**
110
+ * Compatible gateways may stream function arguments as JSON deltas or repeat
111
+ * a cumulative snapshot on every SSE frame. Keep the normal incremental form
112
+ * when it parses, then fall back to snapshot folding. Returning malformed
113
+ * concatenated snapshots would erase the model's correction at tool parsing.
114
+ */
115
+ function assembleCompatibleToolArguments(parts) {
116
+ const nonEmpty = (parts || []).map(String).filter(part => part && part !== 'null');
117
+ if (!nonEmpty.length)
118
+ return '{}';
119
+ const isJsonObject = (value) => {
120
+ try {
121
+ const parsed = JSON.parse(value);
122
+ return !!parsed && typeof parsed === 'object' && !Array.isArray(parsed);
123
+ }
124
+ catch {
125
+ return false;
126
+ }
127
+ };
128
+ const incremental = nonEmpty.join('');
129
+ if (isJsonObject(incremental))
130
+ return incremental;
131
+ let compatible = '';
132
+ for (const incoming of nonEmpty) {
133
+ if (!compatible)
134
+ compatible = incoming;
135
+ else if (incoming === compatible)
136
+ continue;
137
+ else if (incoming.startsWith(compatible))
138
+ compatible = incoming;
139
+ else if (compatible.startsWith(incoming))
140
+ continue;
141
+ else
142
+ compatible += incoming;
143
+ }
144
+ if (isJsonObject(compatible))
145
+ return compatible;
146
+ return [...nonEmpty].reverse().find(isJsonObject) || compatible;
147
+ }
97
148
  /**
98
149
  * Detect content-policy refusals across the provider failure shapes.
99
150
  * Mirrors `LLMProvider.contentPolicyBlocked` semantics exactly so the
@@ -155,6 +155,7 @@ class ResponsesAdapter {
155
155
  const key = `${String(payload.item_id || '')}:${String(payload.summary_index || 0)}`;
156
156
  const delta = this.extractText(payload.delta);
157
157
  if (delta) {
158
+ emittedContent = true;
158
159
  reasoningSummaries.set(key, (reasoningSummaries.get(key) || '') + delta);
159
160
  yield { type: 'reasoning.summary.delta', delta };
160
161
  }
@@ -185,7 +186,7 @@ class ResponsesAdapter {
185
186
  calls.set(key, {
186
187
  id: String(item.call_id || item.id || key),
187
188
  name: String(item.name || ''),
188
- arguments: String(item.arguments || ''),
189
+ argumentParts: item.arguments ? [String(item.arguments)] : [],
189
190
  emitted: false,
190
191
  });
191
192
  }
@@ -193,9 +194,10 @@ class ResponsesAdapter {
193
194
  }
194
195
  if (eventType === 'response.function_call_arguments.delta') {
195
196
  const key = String(payload.item_id || payload.call_id || payload.output_index || '');
196
- const call = calls.get(key) || { id: String(payload.call_id || key), name: String(payload.name || ''), arguments: '', emitted: false };
197
+ const call = calls.get(key) || { id: String(payload.call_id || key), name: String(payload.name || ''), argumentParts: [], emitted: false };
197
198
  const delta = String(payload.delta || '');
198
- call.arguments += delta;
199
+ if (delta)
200
+ call.argumentParts.push(delta);
199
201
  calls.set(key, call);
200
202
  yield { type: 'tool_call.arguments.delta', id: call.id, delta };
201
203
  continue;
@@ -207,19 +209,21 @@ class ResponsesAdapter {
207
209
  const call = calls.get(key) || {
208
210
  id: String(item.call_id || item.id || key),
209
211
  name: String(item.name || ''),
210
- arguments: String(item.arguments || ''),
212
+ argumentParts: item.arguments ? [String(item.arguments)] : [],
211
213
  emitted: false,
212
214
  };
213
215
  call.id = String(item.call_id || call.id);
214
216
  call.name = String(item.name || call.name);
215
- call.arguments = typeof item.arguments === 'string' ? item.arguments : call.arguments;
217
+ if (typeof item.arguments === 'string' && item.arguments)
218
+ call.argumentParts.push(item.arguments);
216
219
  if (!call.emitted) {
217
220
  call.emitted = true;
221
+ const argumentsJson = (0, provider_events_1.assembleCompatibleToolArguments)(call.argumentParts);
218
222
  yield { type: 'tool_call.started', id: call.id, name: call.name };
219
- if (call.arguments && call.arguments !== '{}') {
220
- yield { type: 'tool_call.arguments.delta', id: call.id, delta: call.arguments };
223
+ if (argumentsJson !== '{}') {
224
+ yield { type: 'tool_call.arguments.delta', id: call.id, delta: argumentsJson };
221
225
  }
222
- yield { type: 'tool_call.completed', id: call.id, name: call.name, arguments: call.arguments };
226
+ yield { type: 'tool_call.completed', id: call.id, name: call.name, arguments: argumentsJson };
223
227
  }
224
228
  calls.set(key, call);
225
229
  }
@@ -252,9 +256,19 @@ class ResponsesAdapter {
252
256
  yield { type: 'response.failed', error: '[LLM Error] Responses stream ended before response.completed.' };
253
257
  }
254
258
  else if (!emittedContent && calls.size === 0) {
255
- yield { type: 'response.failed', error: '[Error] Empty Responses stream.' };
259
+ yield { type: 'response.failed', error: '[Error] Provider returned an empty response.' };
256
260
  }
257
261
  else {
262
+ // Some compatible Responses providers omit output_item.done but still
263
+ // complete the response. Preserve that valid tool activity and fold
264
+ // cumulative argument snapshots before handing it to the kernel.
265
+ for (const call of calls.values()) {
266
+ if (call.emitted)
267
+ continue;
268
+ const argumentsJson = (0, provider_events_1.assembleCompatibleToolArguments)(call.argumentParts);
269
+ yield { type: 'tool_call.started', id: call.id, name: call.name };
270
+ yield { type: 'tool_call.completed', id: call.id, name: call.name, arguments: argumentsJson };
271
+ }
258
272
  yield { type: 'response.completed' };
259
273
  }
260
274
  }
package/dist/server.js CHANGED
@@ -1298,6 +1298,27 @@ async function handleApi(req, res, body) {
1298
1298
  mobileJson(res, await hostedConversationUiAction({ workspaceId, conversationId }, action, String(params.value || ''), params));
1299
1299
  return;
1300
1300
  }
1301
+ case '/api/queue-action': {
1302
+ // Desktop-renderer queue mutations. The mobile endpoint requires a
1303
+ // pairing token; the renderer shares the same process and is trusted,
1304
+ // so it gets a dedicated unauthenticated route that reuses the same
1305
+ // GUI runtime-pool queueAction path (update/delete/reorder/guide).
1306
+ const params = JSON.parse(body || '{}');
1307
+ const workspaceId = String(params.workspaceId || '');
1308
+ const conversationId = String(params.conversationId || '');
1309
+ const action = String(params.action || '');
1310
+ const allowed = new Set(['queue_enqueue', 'queue_update', 'queue_delete', 'queue_reorder', 'queue_toggle_pause', 'queue_guide']);
1311
+ if (!workspaceId || !conversationId || !allowed.has(action)) {
1312
+ jsonResponse(res, { error: 'workspaceId, conversationId, and a valid queue action are required' }, 400);
1313
+ return;
1314
+ }
1315
+ if (!hostedConversationUiAction) {
1316
+ jsonResponse(res, { error: 'This action requires the GUI-hosted runtime pool' }, 409);
1317
+ return;
1318
+ }
1319
+ jsonResponse(res, await hostedConversationUiAction({ workspaceId, conversationId }, action, String(params.value || ''), params));
1320
+ return;
1321
+ }
1301
1322
  case '/api/mobile/conversation-rename': {
1302
1323
  const params = JSON.parse(body || '{}');
1303
1324
  const workspaceId = String(params.workspaceId || '');