@qwen-code/qwen-code 0.23.5-preview.0 → 0.24.0

package/bundled/qc-helper/docs/features/_meta.ts

@@ -31,4 +31,5 @@ export default {
   'scheduled-tasks': 'Scheduled Tasks',
   goals: 'Goals',
   tips: 'Contextual Tips',
+  'omni-media-policies': 'Omni Media Policies',
 };

package/bundled/qc-helper/docs/features/goals.md

@@ -16,13 +16,15 @@ A Goal keeps Qwen Code working across turns until a stated condition is met. Set
 
 Creating, editing, or resuming a Goal requires a trusted workspace (`/trust`). Headless usage is covered in [Headless Mode](./headless.md#run-a-persistent-goal).
 
-Once a Goal has billed a turn, the footer pill and every status card show what it has spent against the window it is allowed, as `1.2k/30.0m`. The figure counts the model calls the Goal makes in its own turns; subagents and the verifier's own checks are not included. The window is set by [`model.goalTokenBudget`](../configuration/settings.md); resuming a Goal that has spent its window grants another one on top of what it has already spent, so the figure reads `30.0m/60.0m` rather than starting over. A Goal with no budget shows only what it has spent. A Goal that has not billed a turn yet shows no figures at all.
+Once a Goal has billed a turn, the footer pill and every status card show what it has spent against the window it is allowed, as `1.2k/30.0m`. The figure counts the model calls in the Goal's own turns, its direct foreground subagents, and its verifier and checkpoint checks. Nested and background subagents, unrelated side queries, cron turns, and notification turns are excluded. Failed checks that return no usable result cannot be counted, including earlier successful batches when a later checkpoint batch fails. Checks performed during recovery count as new calls; restoring already recorded usage does not count it again. The window is set by [`model.goalTokenBudget`](../configuration/settings.md); resuming a Goal that has spent its window grants another one on top of what it has already spent, so the figure reads `30.0m/60.0m` rather than starting over. A Goal with no budget shows only what it has spent. A Goal that has not billed a turn yet shows no figures at all.
+
+When turn or active-time budgets are configured, status cards also show the corresponding usage against its budget, for example `3/20 turns` and `12m 3s/30m`. Headless text spells these out as `3 of 20 turns` and `12m 3s of 30m active`. Zero usage remains hidden, and paused or stopped Goals retain these figures.
 
 Two more windows are available and off by default: [`model.goalMaxTurns`](../configuration/settings.md) caps how many turns a Goal may finish, including user-driven Goal turns, and [`model.goalMaxActiveMinutes`](../configuration/settings.md) caps wall time while it remains active in a running process, including waits and idle time between turns. Time while the Goal is paused, blocked or stopped does not count against the second one, nor does downtime across a restart; a process that is merely suspended is still charged. Whichever window runs out first, the Goal gets the same single wind-down turn to hand off and then stops until you resume it, and resuming authorizes another window on top of what has already been used. Only the window that ran out moves; the others stay where they were, and the `-1` opt-out likewise only takes a ceiling off a Goal that has already spent it. The time window is read between turns rather than enforced by a timer, so a Goal can run well past it before it stops: a turn already running is never interrupted, and a window that runs out while no turn is running is not noticed until the next turn ends. Active time is counted between recorded transitions, so time in a turn a restart interrupted is not charged. Changes to either setting take effect after restart. Both windows arm only on a Goal created after the change, so bounding a Goal already on the record means replacing it with `/goal set` -- a new Goal at revision 1, with its turn, token and active-time meters reset and its evidence window starting at the replacement, so evidence the old Goal recorded is no longer citable -- or clearing it and starting again. A resume or an edit never adds a ceiling that was not armed at creation.
 
 Each turn the session takes on its own reports what the Goal has spent so far, how many turns are behind it, and — unless the Goal runs unbounded — the window it is allowed. Every such turn except the final wind-down hand-off also carries standing instructions to re-check the workspace rather than trust earlier turns' reports, to work toward the end state the objective asks for, to do something different when the previous turn changed nothing (from the second turn on, once there is a previous turn to judge), and to check every requirement against citable evidence before proposing that the Goal is done.
 
-A long Goal periodically compresses the evidence it has recorded into checkpoint claims with a side model check, so later turns and the verifier still have it to cite. Each model call the check makes is bounded by [`model.goalCheckpointTimeoutSeconds`](../configuration/settings.md), 180 seconds by default. A reply wrapped in a markdown fence is read without the fence. If its claims overrun the aggregate byte budget, include a claim over the per-claim character limit, number more than one checkpoint may hold, cite an id that was not in the request, or change the proof kind of a source they cite, it makes one corrective model call that names what was wrong, and both calls share that ceiling. A reply that is not a JSON object holding a non-empty `claims` array gets no corrective call, and neither does one with any malformed claim, even beside usable ones: an extra key at either level, an unrecognised `proofKind`, an empty claim, or a `sourceRefs` list that is empty, holds a non-string or empty id, repeats an id, or holds more than 32 ids. A check that does not finish in time is abandoned as inconclusive; it counts toward the checkpoint stall limit only when the evidence window has overflowed, while a non-overflowing check preserves the streak and retries on a later turn. A check after a stall on an overflowing window does not resend the same request: it sends the evidence in batches of 24 records, or 12 after a second stall, one model call per batch, each bounded by that ceiling on its own. A window that still has room, and the check a resumed session replays at startup, are still sent in one call, since neither can spend a stall. Each batch's claims are carried into the next, and only the last batch's claims are kept. A batch that fails ends the check the way a failed single call does, and the recorded failure names the batch, as in `batch 2/5: …`. A check after a stall can therefore take up to five times the ceiling, or nine times after a second stall, and a message you send meanwhile waits for it; the whole span counts as active time, so it is charged against `model.goalMaxActiveMinutes` like any other wait. Batching helps when the model could not answer for the whole window in time or with usable claims; a checkpoint that comes back with the full 32 claims and still leaves evidence behind still needs a narrower objective. The calls are streamed, so the per-request transport timeout bounds only connect and first response, and the ceiling itself stops at the stream guards' 15-minute lifetime cap because past that the guard, not the setting, ends the check. That 15-minute limit on the setting is fixed, and raising the stream guard's own cap does not lift it.
+A long Goal periodically compresses the evidence it has recorded into checkpoint claims with a side model check, so later turns and the verifier still have it to cite. Usage from completed checks, including corrective retries and batches, counts toward the Goal's token budget. Each model call the check makes is bounded by [`model.goalCheckpointTimeoutSeconds`](../configuration/settings.md), 180 seconds by default. A reply wrapped in a markdown fence is read without the fence. If its claims overrun the aggregate byte budget, include a claim over the per-claim character limit, number more than one checkpoint may hold, cite an id that was not in the request, or change the proof kind of a source they cite, it makes one corrective model call that names what was wrong, and both calls share that ceiling. A reply that is not a JSON object holding a non-empty `claims` array gets no corrective call, and neither does one with any malformed claim, even beside usable ones: an extra key at either level, an unrecognised `proofKind`, an empty claim, or a `sourceRefs` list that is empty, holds a non-string or empty id, repeats an id, or holds more than 32 ids. A check that does not finish in time is abandoned as inconclusive; it counts toward the checkpoint stall limit only when the evidence window has overflowed, while a non-overflowing check preserves the streak and retries on a later turn. A check after a stall on an overflowing window does not resend the same request: it sends the evidence in batches of 24 records, or 12 after a second stall, one model call per batch, each bounded by that ceiling on its own. A window that still has room, and the check a resumed session replays at startup, are still sent in one call, since neither can spend a stall. Each batch's claims are carried into the next, and only the last batch's claims are kept. A batch that fails ends the check the way a failed single call does, and the recorded failure names the batch, as in `batch 2/5: …`. A check after a stall can therefore take up to five times the ceiling, or nine times after a second stall, and a message you send meanwhile waits for it; the whole span counts as active time, so it is charged against `model.goalMaxActiveMinutes` like any other wait. Batching helps when the model could not answer for the whole window in time or with usable claims; a checkpoint that comes back with the full 32 claims and still leaves evidence behind still needs a narrower objective. The calls are streamed, so the per-request transport timeout bounds only connect and first response, and the ceiling itself stops at the stream guards' 15-minute lifetime cap because past that the guard, not the setting, ends the check. That 15-minute limit on the setting is fixed, and raising the stream guard's own cap does not lift it.
 
 A failing checkpoint shows up before it stops the Goal. While an active Goal's stall streak runs, the footer pill switches to `checkpoint N/3 stalled` on its own; once the Goal pauses or stops, the pill shows that status instead. The web shell's Goal status strip shows the count whatever the status. Whenever a terminal Goal status card is rendered, for example by `/goal` or by a pause, resume or verifier card, it shows how many consecutive checks have stalled out of the three the Goal allows, together with the last failure while one is recorded; the web shell's Goals dialog and headless `/goal` text output show the same line, while the web shell's transcript cards for Goal events show only the stop reason, and the model sees both fields when it reads the Goal. A check that fails while the window still has room is shown too, without spending a stall, but only while the Goal is active, or when that failure is itself what stopped the Goal, as with a checkpoint request too large to send. A checkpoint stop for any other reason clears the failure and keeps the streak, and a completed Goal shows no checkpoint line. The failure is kept as a single line with control characters removed. A Goal stopped by three stalled checkpoints names what the last one ran into. A check that could not fit the window within the checkpoint claim bounds, whether a full claim list that still left evidence behind or claims over the claim count or size budget, means the objective produces more evidence than one window holds, so narrow it. An answer that could not be folded into claims means the checkpoint model is not returning the structured output it is asked for, and narrowing the objective does not fix that. A check that never answered can mean an unreachable or rate-limited provider, a check that did not finish within `model.goalCheckpointTimeoutSeconds`, or an error in the check itself; the recorded failure says which. Resuming after any of the three starts a fresh evidence window.
 

package/bundled/qc-helper/docs/features/omni-fixed-policies-preset.json

@@ -0,0 +1,242 @@
+// ============================================================================
+// Omni 固定策略(fixedPolicy)预设 —— 覆盖设计文档 4.1–4.9 全部 9 条
+// ----------------------------------------------------------------------------
+// 用法:把 "omni" 对象合并进工作区 .qwen/settings.json(或用户级
+// ~/.qwen/settings.json)。settings 解析支持 JSONC 注释。
+//
+// 说明:
+//  - 本预设只占用 "omni.processing" 子树;omni.enabled 等其他键按需在
+//    你自己的 settings 里设置(omni 未启用时 fixedPolicies 不会运行)。
+//  - 长链超时:4.1/4.4 的 ASR 链可能跑几十分钟(81 分钟电影 ≈ 28 个
+//    转写分段),务必一并拷贝 policyTools 里的 runtime.timeoutMs 放大项。
+//  - 4.1–4.4 依赖 S7 的 memory.* 条件命名空间(子图可见性:抽音轨→转写
+//    的转写产物挂在派生音频版本下,对视频版本同样可见)。
+//  - 执行顺序由 priority 保证(大者先跑):抽轨/预处理(90–100)→
+//    链内条件分支(70–90)→ 独立降级(50–60)。priority 只决定同一资源
+//    上的匹配顺序;跨模态互不影响。
+//  - 已知限制:4.5 无法表达"判定为文档/文字类图片"(DSL 无图像内容
+//    分类字段),本预设实现为"大图才 OCR"(既省 token 又是最常见文档
+//    形态);需要别的口径时自行改 when。
+// ============================================================================
+{
+  "omni": {
+    "processing": {
+      "fixedPolicies": {
+        // ---- 4.1 长视频(>30min)抽音轨 → ASR(三步链,优先级 100→90→80)----
+        // 设计:长视频的语音内容整体转为带时间戳的文本,模型不再"看"全片。
+        // 链式触发:抽轨产物 origin 为 policy,由下一步接力;reprocessMedia
+        // 让产物重新进入匹配。memory.hasTranscript 门控:已有 ASR 结果(包括
+        // 模型此前自己转写的)就不再重复触发。
+        "long-video-extract-audio": {
+          "priority": 100,
+          "mediaTypes": ["video"],
+          "origins": ["user", "tool"],
+          "when": [
+            "all",
+            [">", ["field", "resource.durationMs"], 1800000],
+            ["==", ["field", "memory.hasTranscript"], 0]
+          ],
+          "toolName": "omni_extract_audio",
+          "arguments": {
+            "format": "wav",
+            "sampleRateHz": 16000,
+            "channels": 1
+          },
+          "output": { "reprocessMedia": true, "source": "keep" }
+        },
+        // 音轨超过 ASR 入参上限(100MB)时先压码率;省略大音轨,小音轨接力。
+        "long-video-audio-downsample": {
+          "priority": 90,
+          "mediaTypes": ["audio"],
+          "origins": ["policy"],
+          "when": [">", ["field", "resource.sizeBytes"], 104857600],
+          "toolName": "omni_downsample_audio",
+          "arguments": {
+            "bitrateKbps": 24,
+            "sampleRateHz": 16000,
+            "channels": 1
+          },
+          "output": { "reprocessMedia": true, "source": "omit" }
+        },
+        "long-video-transcribe": {
+          "priority": 80,
+          "mediaTypes": ["audio"],
+          "origins": ["policy"],
+          "when": ["<=", ["field", "resource.sizeBytes"], 104857600],
+          "toolName": "omni_transcribe_audio",
+          "output": { "source": "omit" }
+        },
+
+        // ---- 4.4 长音频(>30min)直接 ASR(一步,与 4.1 共用转写条件)----
+        "long-audio-transcribe": {
+          "priority": 90,
+          "mediaTypes": ["audio"],
+          "origins": ["user", "tool"],
+          "when": [
+            "all",
+            [">", ["field", "resource.durationMs"], 1800000],
+            ["==", ["field", "memory.hasTranscript"], 0]
+          ],
+          "toolName": "omni_transcribe_audio",
+          // source 保留:转写文本内联投递,原音频是否一并投递由你决定
+          // (长音频通常体积也大,想省 token 就改 "omit" 只留转写)。
+          "output": { "source": "keep" }
+        },
+
+        // ---- 4.2 大图(宽或高>2000px)降采样 --------------------------------
+        // tokenBudget 档位按计费 token 说话(small=256 tokens,按 28px
+        // patch 网格对齐),比裸 maxDimension 更贴近模型计费语义。
+        "large-image-downsample": {
+          "priority": 60,
+          "mediaTypes": ["image"],
+          "origins": ["user", "tool", "policy"],
+          "when": [
+            "any",
+            [">", ["field", "resource.width"], 2000],
+            [">", ["field", "resource.height"], 2000]
+          ],
+          "toolName": "omni_downsample_image",
+          "arguments": { "tokenBudget": "normal" },
+          "output": { "source": "omit" }
+        },
+
+        // ---- 4.3 上下文紧张时抽关键帧 -------------------------------------
+        // 视频本身装不进剩余上下文时,改用 8 帧小档位关键帧代替。
+        "tight-context-keyframes": {
+          "priority": 60,
+          "mediaTypes": ["video"],
+          "origins": ["user", "tool"],
+          "when": [
+            ">",
+            ["field", "resource.estimatedTokenCount"],
+            ["field", "session.availableContextTokens"]
+          ],
+          "onConditionUnavailable": "skip",
+          "toolName": "omni_extract_keyframes",
+          "arguments": { "maxFrames": 8, "frameTokenBudget": "small" },
+          "output": { "source": "omit" }
+        },
+
+        // ---- 4.5 文档/文字图片 OCR(已知限制:按大图近似判定)--------------
+        "document-image-ocr": {
+          "priority": 50,
+          "mediaTypes": ["image"],
+          "origins": ["user", "tool"],
+          "when": [
+            "all",
+            [
+              "any",
+              [">", ["field", "resource.width"], 2000],
+              [">", ["field", "resource.height"], 2000]
+            ],
+            ["==", ["field", "memory.hasOcr"], 0]
+          ],
+          "toolName": "omni_ocr_image",
+          // source 保留:OCR 文本进上下文的同时,图片仍由 4.2 降采样后投递,
+          // 模型同时拿到"准确文字层 + 压缩视觉层"。
+          "output": { "source": "keep" }
+        },
+
+        // ---- 4.6 大视频(最长边>1920)降分辨率 ------------------------------
+        "large-video-downscale": {
+          "priority": 50,
+          "mediaTypes": ["video"],
+          "origins": ["user", "tool", "policy"],
+          "when": [
+            "any",
+            [">", ["field", "resource.width"], 1920],
+            [">", ["field", "resource.height"], 1920]
+          ],
+          "toolName": "omni_downscale_video",
+          "arguments": { "maxHeight": 480, "fps": 10 },
+          "output": { "source": "omit" }
+        },
+
+        // ---- 4.7/4.8/4.9 超限兜底(>10MB,与上游 inline 上限对齐)-----------
+        // 与 transport guard 的关系:guard 是无条件的系统兜底(永远启用、
+        // 不可删);这三条是更早、更聪明的预处理——图片按 token 档位降到
+        // 小档,音频直接转写成文本,视频抽帧——比 guard 的机械降码率保留
+        // 更多语义。处理后仍超限的由 guard 接力。
+        "oversize-image-downsample": {
+          "priority": 40,
+          "mediaTypes": ["image"],
+          "origins": ["user", "tool", "policy"],
+          "when": [">", ["field", "resource.sizeBytes"], 10485760],
+          "toolName": "omni_downsample_image",
+          "arguments": { "tokenBudget": "small", "quality": 85 },
+          "output": { "source": "omit" }
+        },
+        "oversize-audio-transcribe": {
+          "priority": 40,
+          "mediaTypes": ["audio"],
+          "origins": ["user", "tool", "policy"],
+          "when": [
+            "all",
+            [">", ["field", "resource.sizeBytes"], 10485760],
+            ["==", ["field", "memory.hasTranscript"], 0]
+          ],
+          "toolName": "omni_transcribe_audio",
+          "output": { "source": "omit" }
+        },
+        "oversize-video-keyframes": {
+          "priority": 40,
+          "mediaTypes": ["video"],
+          "origins": ["user", "tool", "policy"],
+          "when": [">", ["field", "resource.sizeBytes"], 10485760],
+          "toolName": "omni_extract_keyframes",
+          "arguments": { "maxFrames": 16, "frameTokenBudget": "normal" },
+          "output": { "source": "omit" }
+        }
+      },
+
+      // ---- 工具级配套:ASR 链的超时放大(30 分钟不足以转写长片)----------
+      // transcribe 的 settings.maxInputBytes 放大到 100MB,与链内 downsample
+      // 分支的 100MB 阈值一致(默认 10MB 会在第二步就拒绝)。
+      //
+      // modelAccess.enabled: 默认所有 policy 工具对模型隐藏(只能被
+      // fixedPolicies 触发)。打开后模型可以基于已投递媒体的 resourceId
+      // 自行调用——即"渐进式理解":先看降级产物,需要细节时自己裁剪/
+      // 抽帧/转写更高保真证据。系统提示会列出已开放的工具。
+      // 只开放"取证"类工具;纯降级工具(downsample/downscale)保持
+      // operator-only,模型没有主动降低证据质量的理由。
+      "policyTools": {
+        "omni_transcribe_audio": {
+          "settings": { "maxInputBytes": 104857600 },
+          "runtime": { "timeoutMs": 1800000 },
+          "modelAccess": { "enabled": true }
+        },
+        "omni_extract_audio": {
+          "runtime": { "timeoutMs": 1800000 },
+          "modelAccess": { "enabled": true }
+        },
+        "omni_downsample_audio": {
+          "runtime": { "timeoutMs": 1800000 }
+        },
+        "omni_clip_video": {
+          "runtime": { "timeoutMs": 1800000 },
+          "modelAccess": { "enabled": true }
+        },
+        "omni_clip_audio": {
+          "modelAccess": { "enabled": true }
+        },
+        "omni_extract_keyframes": {
+          "runtime": { "timeoutMs": 1800000 },
+          "modelAccess": { "enabled": true }
+        },
+        "omni_understand_video_segments": {
+          "runtime": { "timeoutMs": 1800000 },
+          "modelAccess": { "enabled": true }
+        },
+        "omni_ocr_image": {
+          "modelAccess": { "enabled": true }
+        },
+        "omni_caption_image": {
+          "modelAccess": { "enabled": true }
+        }
+      },
+
+      // ---- 链式派生预算:默认 1GiB/根对长片抽轨偏紧,放宽到 4GiB ---------
+      "limits": { "maxDerivedBytesPerRoot": 4294967296 }
+    }
+  }
+}