feat(plugins): add source-aware evidence and recording timelines

.env.example

@@ -22,6 +22,9 @@ OPENAI_REASONING_EFFORT=high
 # 功能开关：保存后对新任务生效
 ARTEMIS_VIDEO_ENABLED=false
 ARTEMIS_RECORDING_ENABLED=false
+# 留档编码与质量；自动模式只有客户端确认 HEVC 可播放时才优先 H.265，否则使用 H.264。
+# JSON 配置 agent.recording 和任务 recording 可覆盖。截图分辨率不受此配置影响。
+ARTEMIS_RECORDING_CONFIG={"codec":"auto","max_long_edge":1280,"fps":15,"h265_kbps":1000,"h264_kbps":1500,"computer_scope":"screen","minimum_disk_mb":256,"hevc_playback":false}
 ARTEMIS_OUTPUTTER_ENABLED=true
 ARTEMIS_STEP_SUMMARIZER_ENABLED=true
 # 默认将任务截图长边压至 2048；false 保留设备原始截图。长边值可在 config/artemis.jsonc 的 agent.screenshot 中覆盖。

apps/admin_console/routers/media.py

@@ -115,6 +115,74 @@ async def get_session_video(session_id: str):
 
 
 def _get_session_video_sync(session_id: str):
+    from apps.admin_console.services.evidence_service import plugin_sources
+    from apps.admin_console.services.report_export import session_videos
+
+    legacy = _get_legacy_session_video_sync(session_id)
+    sources = []
+    if legacy.get("video_segments") or legacy.get("video_url"):
+        segments = legacy.get("video_segments") or []
+        # Pauses create sibling resume directories. Only collect files already
+        # attached to this task; unrelated recordings never enter its timeline.
+        if segments and all("offset_ms" in segment for segment in segments):
+            combined = {}
+            for path, _ in session_videos({"id": session_id}):
+                for segment in media_service.resolve_video_segments(
+                    media_service.path_to_video_url(path)
+                ):
+                    if "offset_ms" in segment:
+                        combined[segment["url"]] = segment
+            if combined:
+                segments = sorted(combined.values(), key=lambda segment: segment["offset_ms"])
+        sources.append(
+            {
+                "id": "android",
+                "kind": "android",
+                "label": "安卓",
+                "target_ids": ["android"],
+                "status": "ready",
+                "message": "",
+                "segments": [
+                    {**s, "start": float(s.get("offset_ms", s["start"] * 1000)) / 1000}
+                    for s in segments
+                ],
+                "url": legacy.get("video_url"),
+            }
+        )
+    sources += plugin_sources(session_id)
+    legacy["sources"] = sources
+    legacy["session_timeline"] = any(source["segments"] for source in sources)
+    row = dict(session_repo.get_session_by_id(session_id) or {})
+    legacy["session_duration"] = max(
+        max(
+            0.0,
+            float(row.get("end_time") or row.get("start_time") or 0)
+            - float(row.get("start_time") or 0),
+        ),
+        max(
+            (s["start"] + s["duration"] for source in sources for s in source["segments"]),
+            default=0,
+        ),
+    )
+    ready = next((s for s in sources if s.get("segments") or s.get("url")), None)
+    if ready:
+        legacy.update(
+            status="ready",
+            has_video=True,
+            video_url=ready.get("url") or ready["segments"][0]["url"],
+            video_segments=ready["segments"],
+        )
+    elif any(s.get("status") == "recording" for s in sources):
+        legacy.update(status="processing", retry_after_ms=1000)
+    elif sources:
+        legacy.update(
+            status="failed",
+            message="；".join(str(s.get("message") or "录像不可用") for s in sources),
+        )
+    return legacy
+
+
+def _get_legacy_session_video_sync(session_id: str):
     video_rec_map = session_repo.get_video_recordings_map()
     video_idx = media_service.build_video_index()
     row = session_repo.get_session_by_id(session_id)

apps/admin_console/services/evidence_service.py

@@ -0,0 +1,151 @@
+"""Read task-owned plugin evidence without executing tools or converting media."""
+
+import json
+import math
+from pathlib import Path
+import re
+
+from artemis.config import TRACES_PATH
+
+
+def task_root(session_id: str) -> Path:
+    if not re.fullmatch(r"[A-Za-z0-9_-]{1,100}", session_id):
+        raise ValueError("无效任务编号")
+    root = TRACES_PATH.resolve() / session_id
+    if root.resolve() != root:
+        raise ValueError("任务证据目录不能是符号链接")
+    return root
+
+
+def evidence_file(root: Path, relative: str) -> Path | None:
+    path = root / relative
+    try:
+        if (
+            path.is_file()
+            and path.resolve() == path.absolute()
+            and path.resolve().is_relative_to(root.resolve())
+        ):
+            return path
+    except OSError:
+        return None
+    return None
+
+
+def plugin_sources(session_id: str) -> list[dict]:
+    root = task_root(session_id)
+    path = evidence_file(root, "evidence/recordings/recordings.json")
+    if not path or path.stat().st_size > 4 * 1024 * 1024:
+        return []
+    try:
+        data = json.loads(path.read_text(encoding="utf-8"))
+        if data.get("version") != 1 or not isinstance(data.get("sources"), list):
+            return []
+        from apps.admin_console.services.media_service import MediaService
+
+        sources = []
+        for item in data["sources"]:
+            if not isinstance(item, dict) or not isinstance(item.get("id"), str):
+                continue
+            source = {**item, "segments": []}
+            for segment in item.get("segments", []):
+                if not isinstance(segment, dict):
+                    continue
+                video = evidence_file(root, str(segment.get("file") or ""))
+                if video is None or video.suffix.lower() != ".mp4":
+                    continue
+                duration = float(segment.get("duration", 0))
+                start = float(segment.get("offset_ms", 0)) / 1000
+                if (
+                    not math.isfinite(duration)
+                    or duration <= 0
+                    or not math.isfinite(start)
+                    or start < 0
+                ):
+                    continue
+                source["segments"].append(
+                    {
+                        **segment,
+                        "url": MediaService.path_to_video_url(video),
+                        "start": start,
+                        "duration": duration,
+                    }
+                )
+            sources.append(source)
+        return sources
+    except (OSError, ValueError, TypeError, KeyError):
+        return []
+
+
+def plugin_files(session_id: str) -> list[tuple[Path, str]]:
+    root = task_root(session_id)
+    files = {}
+    for source in plugin_sources(session_id):
+        for segment in source["segments"]:
+            path = evidence_file(root, segment["file"])
+            if path:
+                files[path] = "任务录屏/电脑/" + path.name
+    for directory in (root / "artifacts" / "plugins", root / "evidence"):
+        if not directory.is_dir():
+            continue
+        for path in directory.rglob("*"):
+            if path.suffix.lower() == ".tmp" or (
+                directory.name == "evidence" and path.suffix.lower() in {".mp4", ".mkv"}
+            ):
+                continue
+            relative = path.relative_to(root).as_posix()
+            checked = evidence_file(root, relative)
+            if checked:
+                files[checked] = (
+                    "任务文件/插件/" + path.relative_to(root / "artifacts" / "plugins").as_posix()
+                    if directory.name == "plugins"
+                    else "证据索引/" + path.relative_to(root / "evidence").as_posix()
+                )
+    return list(files.items())
+
+
+def export_index(session_id: str, report: dict, android_videos: list) -> dict:
+    from apps.admin_console.services.report_export import session_artifacts, session_screenshots
+
+    extras = plugin_files(session_id)
+    recordings = plugin_sources(session_id)
+    root = task_root(session_id)
+    paths = {path.resolve(): name for path, name in extras}
+    for source in recordings:
+        for segment in source["segments"]:
+            segment["file"] = paths.get((root / segment["file"]).resolve())
+            segment.pop("url", None)
+    images = session_screenshots(report)
+    artifacts = session_artifacts(report)
+    files = [(path, "步骤截图/" + name) for path, name in images]
+    files += [(path, "任务录屏/" + name) for path, name in android_videos]
+    files += [(path, "任务文件/安卓/" + name) for path, name in artifacts]
+    files += extras
+    calls = []
+    ledger = evidence_file(root, "evidence/calls.jsonl")
+    if ledger:
+        with ledger.open(encoding="utf-8") as handle:
+            for line in handle:
+                try:
+                    call = json.loads(line)
+                    if not isinstance(call, dict):
+                        continue
+                    for attachment in call.get("attachments", []):
+                        relative = str(attachment.get("path") or "")
+                        attachment["file"] = paths.get((root / relative).resolve())
+                    calls.append(call)
+                except (ValueError, TypeError, AttributeError):
+                    continue  # A partial final line after a crash is not a valid receipt.
+    return {
+        "version": 1,
+        "session_id": session_id,
+        "name": report["name"],
+        "status": report["status"],
+        "verdict": report["verdict"],
+        "steps": report.get("steps", []),
+        "calls": calls,
+        "recordings": recordings,
+        "android_videos": ["任务录屏/" + name for _, name in android_videos],
+        "image_references": {path.stem: "步骤截图/" + name for path, name in images},
+        "files": [{"path": name, "bytes": path.stat().st_size} for path, name in files],
+        "note": "工具执行完成不等于用例通过。录像保留真实时间与缺口；各来源截图不能相互替代。",
+    }

apps/admin_console/services/model_settings_service.py

@@ -175,6 +175,7 @@ class CapabilitySettings(BaseModel):
     )
     video_enabled: bool = False
     recording_enabled: bool = False
+    recording: dict = Field(default_factory=dict)
     outputter_enabled: bool = True
     step_summarizer_enabled: bool = True
     screenshot_limit_enabled: bool = True
@@ -204,6 +205,7 @@ def get_capability_settings() -> CapabilitySettings:
     return CapabilitySettings(
         video_enabled=settings.ARTEMIS_VIDEO_ENABLED,
         recording_enabled=settings.ARTEMIS_RECORDING_ENABLED,
+        recording=load_agent_config().recording.model_dump(),
         outputter_enabled=settings.ARTEMIS_OUTPUTTER_ENABLED,
         step_summarizer_enabled=settings.ARTEMIS_STEP_SUMMARIZER_ENABLED,
         screenshot_limit_enabled=screenshot.enabled,
@@ -212,6 +214,12 @@ def get_capability_settings() -> CapabilitySettings:
 
 
 def save_capability_settings(update: CapabilitySettings) -> CapabilitySettings:
+    from artemis.config.agent import RecordingConfig
+    import json
+
+    recording = RecordingConfig.model_validate(
+        {**load_agent_config().recording.model_dump(), **update.recording}
+    ).model_dump()
     values = {
         "ARTEMIS_ADB_AUTO_CONNECT": update.adb_auto_connect,
         "ARTEMIS_ADB_DISCOVERY_INTERVAL_SECONDS": update.adb_discovery_interval_seconds,
@@ -229,6 +237,8 @@ def save_capability_settings(update: CapabilitySettings) -> CapabilitySettings:
         "ARTEMIS_WORKER_CLEANUP_TIMEOUT_SECONDS": update.worker_cleanup_timeout_seconds,
     }
     persist_environment({key: str(value).lower() for key, value in values.items()})
+    persist_environment({"ARTEMIS_RECORDING_CONFIG": json.dumps(recording)})
+    settings.ARTEMIS_RECORDING_CONFIG = recording
     for key, value in values.items():
         setattr(settings, key, value)
     return get_capability_settings()

apps/admin_console/services/report_export.py

@@ -475,6 +475,24 @@ def plugin_operations(steps: list[dict]) -> list[dict]:
                 for block in result.get("content", [])
                 if isinstance(block, dict) and block.get("type") == "text"
             )
+            source = result.get("source") or payload.get("source") or {}
+            evidence = []
+            if result.get("call_id"):
+                evidence.append("调用：" + str(result["call_id"]))
+            if source:
+                evidence.append(
+                    f"来源：{source.get('kind', '')} · {source.get('label') or source.get('target_id', '')}"
+                )
+            if result.get("execution"):
+                evidence.append(
+                    f"执行：{result['execution']}；副作用：{result.get('side_effect', 'unknown')}；错误码：{result.get('error_code') or '无'}"
+                )
+            for attachment in result.get("attachments") or []:
+                evidence.append(
+                    f"文件：{attachment.get('name')}；SHA-256：{attachment.get('sha256')}"
+                )
+            if evidence:
+                content = "\n".join([content or result.get("message") or "", *evidence])
             operations.append(
                 {
                     "step": step.get("step_number"),
@@ -809,6 +827,8 @@ def session_artifacts(report: dict) -> list[tuple[Path, str]]:
 
 
 def export_session_media(session_id: str, kind: str) -> Path:
+    from apps.admin_console.services.evidence_service import export_index, plugin_files
+
     if kind not in {"screenshots", "video", "all"}:
         raise ValueError("不支持的下载类型")
     with EXPORT_LOCK:
@@ -818,7 +838,10 @@ def export_session_media(session_id: str, kind: str) -> Path:
         images = session_screenshots(report) if kind in {"screenshots", "all"} else []
         videos = session_videos(report) if kind in {"video", "all"} else []
         artifacts = session_artifacts(report) if kind == "all" else []
-        if kind == "video" and not videos:
+        extras = plugin_files(session_id) if kind in {"video", "all"} else []
+        if kind == "video":
+            extras = [(path, name) for path, name in extras if path.suffix.lower() == ".mp4"]
+        if kind == "video" and not videos and not extras:
             raise ValueError("此任务没有已保存的录屏")
         if kind == "screenshots" and not images:
             raise ValueError("此任务没有可下载的步骤截图")
@@ -830,7 +853,8 @@ def export_session_media(session_id: str, kind: str) -> Path:
             "status": report["status"],
             "verdict": report["verdict"],
             "screenshots": [name for _, name in images],
-            "videos": [name for _, name in videos],
+            "videos": ["任务录屏/" + name for _, name in videos]
+            + [name for path, name in extras if path.suffix.lower() == ".mp4"],
             "artifacts": [name for _, name in artifacts],
             "note": "录屏按时间顺序分段；暂停期间未占用设备，也未录制其他任务。",
         }
@@ -842,9 +866,27 @@ def export_session_media(session_id: str, kind: str) -> Path:
             for source, name in images:
                 archive.write(source, f"{prefix}步骤截图/{name}")
             for source, name in videos:
-                archive.write(source, f"{prefix}任务录屏/{name}")
+                archive.write(source, f"{prefix}任务录屏/{name}", compress_type=zipfile.ZIP_STORED)
             for source, name in artifacts:
                 archive.write(source, f"{prefix}任务文件/安卓/{name}")
+            for source, name in extras:
+                archive.write(
+                    source,
+                    prefix + name,
+                    compress_type=zipfile.ZIP_STORED
+                    if source.suffix.lower() == ".mp4"
+                    else zipfile.ZIP_DEFLATED,
+                )
+            if kind == "all":
+                archive.writestr(
+                    prefix + "证据索引.json",
+                    json.dumps(
+                        export_index(session_id, report, videos),
+                        ensure_ascii=False,
+                        indent=2,
+                        default=str,
+                    ),
+                )
             archive.writestr(
                 prefix + "文件清单.json", json.dumps(manifest, ensure_ascii=False, indent=2)
             )
@@ -853,6 +895,8 @@ def export_session_media(session_id: str, kind: str) -> Path:
 
 
 def export_batch(batch_id: str, progress: Callable[[int, int, str], None] | None = None) -> Path:
+    from apps.admin_console.services.evidence_service import export_index, plugin_files
+
     with EXPORT_LOCK:
         batch = store.get("batch", identifier(batch_id))
         if not batch:
@@ -901,9 +945,28 @@ def export_batch(batch_id: str, progress: Callable[[int, int, str], None] | None
                 for source, name in session_screenshots(report):
                     archive.write(source, f"{prefix}/步骤截图/{name}")
                 for source, name in session_videos(report):
-                    archive.write(source, f"{prefix}/任务录屏/{name}")
+                    archive.write(
+                        source, f"{prefix}/任务录屏/{name}", compress_type=zipfile.ZIP_STORED
+                    )
                 for source, name in session_artifacts(report):
                     archive.write(source, f"{prefix}/任务文件/安卓/{name}")
+                for source, name in plugin_files(report["id"]):
+                    archive.write(
+                        source,
+                        f"{prefix}/{name}",
+                        compress_type=zipfile.ZIP_STORED
+                        if source.suffix.lower() == ".mp4"
+                        else zipfile.ZIP_DEFLATED,
+                    )
+                archive.writestr(
+                    f"{prefix}/证据索引.json",
+                    json.dumps(
+                        export_index(report["id"], report, session_videos(report)),
+                        ensure_ascii=False,
+                        indent=2,
+                        default=str,
+                    ),
+                )
                 if progress:
                     progress(number, len(records), f"已打包任务 {number}/{len(records)}")
         if progress:

apps/admin_console/services/task_import.py

@@ -8,8 +8,9 @@ import zipfile
 from openpyxl import Workbook, load_workbook
 from openpyxl.styles import Alignment, Font, PatternFill
 from openpyxl.utils import get_column_letter
-from pydantic import BaseModel, ConfigDict, Field
+from pydantic import BaseModel, ConfigDict, Field, field_serializer
 from artemis.plugins.models import ExecutionTarget, PluginSelection
+from artemis.config.agent import RecordingConfig
 
 HEADERS = (
     "用例编号",
@@ -38,6 +39,7 @@ class TaskOptions(BaseModel):
     explorer_mode: str | None = Field(default=None, pattern="^(flash|pro|ultra)$")
     locked_app_package: str | None = Field(default=None, min_length=1, max_length=200)
     record_video: bool | None = None
+    recording: RecordingConfig | None = None
     analyze_video: bool | None = None
     request_timeout_seconds: float | None = Field(default=None, gt=0, le=3600)
     max_retries: int | None = Field(default=None, ge=0, le=10)
@@ -51,6 +53,11 @@ class TaskOptions(BaseModel):
     plugins: list[PluginSelection] | None = None
     execution_target: ExecutionTarget | None = None
 
+    @field_serializer("recording", when_used="unless-none")
+    def serialize_recording(self, value: RecordingConfig) -> dict:
+        # A per-task partial override must not replace other saved recording defaults.
+        return value.model_dump(exclude_unset=True)
+
 
 class TaskCase(BaseModel):
     model_config = ConfigDict(extra="forbid")

apps/admin_console/services/task_queue_service.py

@@ -586,6 +586,10 @@ class TaskQueueService:
         env["ARTEMIS_TASK_INGRESS"] = str(task_item.get("ingress", "frontend"))
         env["ARTEMIS_TASK_WORKER"] = "1"
         overrides = task_item.get("runtime_overrides") or {}
+        if overrides.get("recording") is not None:
+            env["ARTEMIS_TASK_RECORDING_CONFIG"] = json.dumps(overrides["recording"])
+        else:
+            env.pop("ARTEMIS_TASK_RECORDING_CONFIG", None)
         for option, variable in {
             "model": "OPENAI_MODEL",
             "reasoning_effort": "OPENAI_REASONING_EFFORT",
@@ -869,6 +873,10 @@ class TaskQueueService:
         registered at recording start; remuxing that file is all that is
         needed to publish the full video.
         """
+        from apps.admin_console.services.evidence_service import task_root
+        from artemis.plugins.recording import recover_interrupted_recordings
+
+        await recover_interrupted_recordings(task_root(str(sess_id)))
         rec_info = session_repo.get_video_recording_for_session(sess_id)
         rec_status = (rec_info or {}).get("status")
         if rec_status == "ready":
@@ -1890,6 +1898,14 @@ class TaskQueueService:
         recordings published.
         """
         recovered = 0
+        from artemis.config import TRACES_PATH
+        from artemis.plugins.recording import recover_interrupted_recordings
+
+        # This hook runs in the startup worker thread, outside the server event loop.
+        for manifest in TRACES_PATH.glob("*/evidence/recordings/recordings.json"):
+            root = manifest.parent.parent.parent
+            if asyncio.run(recover_interrupted_recordings(root)):
+                recovered += 1
         try:
             pending = session_repo.get_unfinalized_video_recordings()
         except Exception as exc:

apps/showcase_ui/src/app/components/agent-stream/agent-stream.component.html

@@ -809,6 +809,7 @@ limitations under the License.
                                   <div class="action-card-header" (click)="toggleActionExpanded('action-' + block.id + '-' + itemIdx, $event)">
                                     <span class="material-symbols-outlined action-icon" [class.is-failed]="isActionFailed(item.data, block.data)">{{ getActionIcon(item.data) }}</span>
                                     <span class="action-title">{{ getActionTitle(item.data) }}</span>
+                                    <button class="evidence-source" [title]="evidenceSourceLabel(item.data) + ' · 查看对应录像'" (click)="openEvidenceVideo(item.data, block.data, $event)">{{ evidenceSourceLabel(item.data) }}</button>
                                     @if (isActionFailed(item.data, block.data)) {
                                       <span class="action-status-badge is-failed">
                                         <span class="material-symbols-outlined status-icon">error</span>
@@ -1050,7 +1051,7 @@ limitations under the License.
                                   </div>
                                 }
                                 <!-- Generic Tool Calls (Text format for note tools, video analyzer, and other sub-agents) -->
-                                @if (shouldShowTool(item.data, block.data) && !isDeviceActionTool(item.data)) {
+                                @if (shouldShowTool(item.data, block.data) && !isEvidenceTool(item.data)) {
                                   <div class="step-generic-tool"
                                     [class.is-running]="isToolLineRunning(item, block)"
                                     [class.is-note-tool]="isNoteTool(item.data)"
@@ -1091,9 +1092,11 @@ limitations under the License.
                                     </div>
                                   </div>
                                 }
-                                <!-- Action Card format (For Device Actions & ADB Commands) -->
-                                @if (shouldShowTool(item.data, block.data) && isDeviceActionTool(item.data)) {
+                                <!-- Device actions and plugin calls share evidence cards with explicit sources. -->
+                                @if (shouldShowTool(item.data, block.data) && isEvidenceTool(item.data)) {
                                   <div class="step-android-action-card tool-action-card"
+                                    [class.plugin-tool-card]="getPluginToolView(item.data)"
+                                    [class.is-auxiliary]="getPluginToolView(item.data)?.auxiliary"
                                     [class.is-running]="isItemRunning(item, block)"
                                     [class.has-no-details]="isAdbCommandTool(item.data) ? !getAdbCommandLine(item.data) : (!getToolTargetText(item.data) && !getToolInputText(item.data) && !getToolCoords(item.data) && !getToolAnalysisText(item.data) && !getToolGenericDetails(item.data) && getToolExtraParams(item.data).length === 0)"
                                     [class.is-failed]="isToolFailed(item.data)"
@@ -1102,6 +1105,10 @@ limitations under the License.
                                     <div class="action-card-header" (click)="toggleActionExpanded('tool-' + block.id + '-' + itemIdx, $event)">
                                       <span class="material-symbols-outlined action-icon" [class.is-failed]="isToolFailed(item.data)">{{ getToolIcon(item.data) }}</span>
                                       <span class="action-title">{{ getToolTitle(item.data) }}</span>
+                                      <button class="evidence-source" [title]="evidenceSourceLabel(item.data) + ' · 查看对应证据'" (click)="openEvidenceVideo(item.data, block.data, $event)">{{ evidenceSourceLabel(item.data) }}</button>
+                                      @if (getPluginToolView(item.data); as plugin) {
+                                        <span class="plugin-call-kind">{{ plugin.kind }} · {{ plugin.status }}</span>
+                                      }
                                       @if (isToolFailed(item.data)) {
                                         <span class="action-status-badge is-failed">
                                           <span class="material-symbols-outlined status-icon">error</span>
@@ -1114,7 +1121,7 @@ limitations under the License.
                                       </span>
                                     </div>
                                     <!-- Compact preview body (when collapsed) -->
-                                    @if (!isActionExpanded('tool-' + block.id + '-' + itemIdx) && (isAdbCommandTool(item.data) ? getAdbCommandLine(item.data) : (getToolTargetText(item.data) || getToolInputText(item.data) || getToolCoords(item.data) || getToolAnalysisText(item.data) || getToolGenericDetails(item.data)))) {
+                                    @if (!getPluginToolView(item.data) && !isActionExpanded('tool-' + block.id + '-' + itemIdx) && (isAdbCommandTool(item.data) ? getAdbCommandLine(item.data) : (getToolTargetText(item.data) || getToolInputText(item.data) || getToolCoords(item.data) || getToolAnalysisText(item.data) || getToolGenericDetails(item.data)))) {
                                       <div class="action-card-body">
                                         <!-- ADB command details -->
                                         @if (isAdbCommandTool(item.data)) {
@@ -1196,7 +1203,21 @@ limitations under the License.
                                               </div>
                                             }
                                           }
-                                          @if (!isAdbCommandTool(item.data)) {
+                                          @if (getPluginToolView(item.data); as plugin) {
+                                            @for (param of plugin.args; track param.key) {
+                                              <div class="grid-item">
+                                                <span class="grid-label">{{ param.key }}</span>
+                                                <span class="grid-value code-pill">{{ param.value }}</span>
+                                              </div>
+                                            }
+                                            @if (plugin.result) {
+                                              <div class="grid-item full-width">
+                                                <span class="grid-label">实际结果</span>
+                                                <span class="grid-value plugin-result">{{ plugin.result }}</span>
+                                              </div>
+                                            }
+                                          }
+                                          @if (!isAdbCommandTool(item.data) && !getPluginToolView(item.data)) {
                                             @if (getToolTargetText(item.data)) {
                                               <div class="grid-item">
                                                 <span class="grid-label">目标</span>
@@ -1251,7 +1272,7 @@ limitations under the License.
                                             }
                                             @if (getStepPostImageUrl(block.data, item.data)) {
                                               <div class="screenshot-column">
-                                                <span class="screenshot-label">操作后</span>
+                                                <span class="screenshot-label">{{ getPluginToolView(item.data)?.auxiliary ? '观察截图' : '操作后' }}</span>
                                                 <div class="image-wrapper" (click)="openImageModal(getStepPostImageUrl(block.data, item.data)!, null, $event)">
                                                   <div class="img-relative-box">
                                                     <img [src]="getStepPostImageUrl(block.data, item.data)" alt="操作后截图" />

apps/showcase_ui/src/app/components/agent-stream/agent-stream.component.scss

@@ -4053,3 +4053,16 @@ $stream-row-height: 25px;
   th { background: #dbeafe; }
   .md-para-gap { height: 8px; }
 }
+
+.evidence-source { max-width: 180px; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; padding: 2px 7px; border: 1px solid #cbd5e1; border-radius: 6px; background: #f1f5f9; color: #475569; font-size: 11px; cursor: pointer; }
+
+.plugin-tool-card {
+  .action-card-header { flex-wrap: wrap; }
+  .plugin-call-kind { font-size: 10px; color: #64748b; }
+  .plugin-result { white-space: pre-wrap; overflow-wrap: anywhere; }
+  &.is-auxiliary:not(.is-expanded):not(.is-failed):not(.is-running) {
+    background: #f8fafc;
+    border-color: #e2e8f0;
+    box-shadow: none;
+  }
+}

apps/showcase_ui/src/app/components/agent-stream/agent-stream.component.ts

@@ -162,6 +162,7 @@ import {
   extractCheckerResult,
   renderMarkdownToHtml
 } from '../../utils/markdown-parser.util';
+import { getPluginToolView } from '../../utils/plugin-tool-view.util';
 
 import {
   getActionObject,
@@ -182,7 +183,8 @@ import {
   getStepPostImageUrl,
   isReportStatusAction,
   getReportStatusValue,
-  getReportStatusExplanation
+  getReportStatusExplanation,
+  actionEvidenceSource
 } from '../../utils/action-formatter.util';
 
 import {
@@ -1612,6 +1614,17 @@ export class AgentStreamComponent implements AfterViewInit {
     }
   }
 
+  public evidenceSourceLabel(item: any): string { return actionEvidenceSource(item).label; }
+
+  public openEvidenceVideo(item: any, step: any, event: Event): void {
+    event.stopPropagation();
+    const started = this.agentService.currentSession()?.start_time;
+    const value = item?.timestamp || item?.start_time || step?.timestamp;
+    const timestamp = typeof value === 'number' ? value : Date.parse(value) / 1000;
+    if (!started || !Number.isFinite(timestamp)) return;
+    this.agentService.openVideoPlayer(undefined, undefined, undefined, Math.max(0, timestamp - started), undefined, actionEvidenceSource(item).id);
+  }
+
   public getScreenRecordingButtonTitle(): string {
     if (this.agentService.isCurrentSessionRunning()) {
       return this.agentService.currentSession()?.recording_status === 'recording' ? "任务正在运行（录屏中）" : "任务正在运行";
@@ -1637,6 +1650,12 @@ export class AgentStreamComponent implements AfterViewInit {
     return isDeviceActionTool(tool);
   }
 
+  public getPluginToolView = getPluginToolView;
+
+  public isEvidenceTool(tool: any): boolean {
+    return isDeviceActionTool(tool) || getPluginToolView(tool) !== null;
+  }
+
   public isFailureAnalyzerActionTool(tool: any): boolean {
     return isDeviceActionTool(tool);
   }
@@ -2033,11 +2052,12 @@ export class AgentStreamComponent implements AfterViewInit {
   }
 
   public getToolIcon(tool: any): string {
-    return getToolIcon(tool);
+    const plugin = getPluginToolView(tool);
+    return plugin ? (plugin.auxiliary ? 'visibility' : 'extension') : getToolIcon(tool);
   }
 
   public getToolTitle(tool: any): string {
-    return getToolTitle(tool);
+    return getPluginToolView(tool)?.title || getToolTitle(tool);
   }
 
   public getToolTargetText(tool: any): string {

apps/showcase_ui/src/app/components/floating-video-player/floating-video-player.component.html

@@ -95,6 +95,22 @@ limitations under the License.
         </button>
       </div>
     </header>
+    @if (sources().length && !agentService.isVideoMinimized()) {
+      <div class="recording-sources">
+        <label for="recording-source">录像来源</label>
+        <select id="recording-source" [ngModel]="agentService.selectedRecordingSource()" (ngModelChange)="selectSource($event)">
+          @for (source of sources(); track source.id) {
+            <option [value]="source.id">{{ source.label }}{{ source.status === 'failed' ? '（录制异常）' : '' }}</option>
+          }
+        </select>
+      </div>
+      @if (agentService.recordingPlaybackMessage()) {
+        <div class="recording-notice" role="status">{{ agentService.recordingPlaybackMessage() }}</div>
+      }
+      @if (!agentService.activeVideoUrl()) {
+        <div class="recording-notice">此来源暂无可播放录像，请查看步骤截图或日志。</div>
+      }
+    }
     <!-- Video Main Content Area -->
     @if (!agentService.isVideoMinimized()) {
       <div class="video-body">
@@ -332,8 +348,10 @@ limitations under the License.
                 @if (agentService.playerMode() === 'video' && agentService.recordingPlaybackStatus() === 'ready' && !agentService.isVideoLoading() && agentService.activeVideoUrl(); as videoUrl) {
                   @if (!videoLoadError()) {
                     <div class="video-container">
+                        @if (gapMessage()) { <div class="recording-gap" role="status">{{ gapMessage() }}</div> }
                       <video
                         #videoRef
+                          [style.visibility]="gapMessage() ? 'hidden' : 'visible'"
                         [src]="currentVideoUrl()"
                         playsinline
                         (loadedmetadata)="onLoadedMetadata()"

apps/showcase_ui/src/app/components/floating-video-player/floating-video-player.component.scss

@@ -1017,3 +1017,42 @@
     }
   }
 }
+.recording-sources {
+  display: flex;
+  gap: 8px;
+  align-items: center;
+  padding: 8px 12px;
+  background: #f8fafc;
+  border-bottom: 1px solid #e2e8f0;
+  font-size: 12px;
+  color: #475569;
+
+  select {
+    min-width: 0;
+    flex: 1;
+    border: 1px solid #cbd5e1;
+    border-radius: 6px;
+    padding: 4px;
+    background: white;
+  }
+}
+
+.recording-notice {
+  padding: 8px 12px;
+  background: #fffbeb;
+  color: #92400e;
+  font-size: 12px;
+}
+
+.recording-gap {
+  position: absolute;
+  inset: 0;
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  padding: 24px;
+  color: #475569;
+  background: #f1f5f9;
+  text-align: center;
+  z-index: 2;
+}

apps/showcase_ui/src/app/components/floating-video-player/floating-video-player.component.spec.ts

@@ -1,6 +1,6 @@
 import { ComponentFixture, TestBed } from '@angular/core/testing';
 import { signal } from '@angular/core';
-import { AgentService } from '../../services/agent.service';
+import { AgentService, RecordingSource, VideoSegment } from '../../services/agent.service';
 import { StepReplayFrame } from '../../core/models/stream.model';
 import { FloatingVideoPlayerComponent } from './floating-video-player.component';
 
@@ -20,7 +20,7 @@ describe('FloatingVideoPlayerComponent video markers', () => {
       index: 0, stepNumber: 1, title: '步骤 1: 点击设置', imageUrl: 'step.png',
       actionText: '点击设置', timestamp: start + 5
     };
-    return {
+    const service = {
       isVideoWindowOpen: signal(true),
       isVideoMinimized: signal(false),
       recordingPlaybackStatus: signal('ready'),
@@ -29,14 +29,30 @@ describe('FloatingVideoPlayerComponent video markers', () => {
       activeVideoTitle: signal('测试录像'),
       isVideoLoading: signal(false),
       activeVideoUrl: signal('/test.mp4'),
-      activeVideoSegments: signal([{ url: '/test.mp4', start: 0, duration: 20, width: 400, height: 800 }]),
+      activeVideoSegments: signal<VideoSegment[]>([{ url: '/test.mp4', start: 0, duration: 20, width: 400, height: 800 }]),
+      recordingSources: signal<RecordingSource[]>([]),
+      selectedRecordingSource: signal(''),
+      usesSessionTimeline: signal(false),
+      recordingSessionDuration: signal(0),
+      recordingPlaybackMessage: signal(''),
+      selectRecordingSource: jasmine.createSpy('selectRecordingSource'),
       currentSessionStepFrames: signal([frame]),
       currentSession: signal({ session_id: 'test', start_time: start }),
       currentSessionId: signal('test'),
-      videoSeekRequest: signal(null),
+      videoSeekRequest: signal<{ seconds: number; requestId: number; sourceId?: string } | null>(null),
       stepSeekRequest: signal(null),
       consumeVideoAutoplay: jasmine.createSpy('consumeVideoAutoplay').and.returnValue(false)
     };
+    service.selectRecordingSource.and.callFake((id: string) => {
+      const source = service.recordingSources().find(item => item.id === id);
+      if (!source) return false;
+      if (service.selectedRecordingSource() === id) return true;
+      service.selectedRecordingSource.set(id);
+      service.activeVideoSegments.set(source.segments);
+      service.activeVideoUrl.set(source.url || source.segments[0]?.url || '');
+      return true;
+    });
+    return service;
   }
 
   beforeEach(() => {
@@ -58,6 +74,7 @@ describe('FloatingVideoPlayerComponent video markers', () => {
       seeking: { configurable: true, get: () => seeking },
       ended: { configurable: true, get: () => false },
       duration: { configurable: true, get: () => 20 },
+      readyState: { configurable: true, get: () => 1 },
       currentTime: {
         configurable: true, get: () => time,
         set: (value: number) => { time = value; seeking = true; }
@@ -72,6 +89,7 @@ describe('FloatingVideoPlayerComponent video markers', () => {
       video.dispatchEvent(new Event('play'));
       return Promise.resolve();
     });
+    video.dispatchEvent(new Event('loadedmetadata'));
   });
 
   function finishSeek(): void {
@@ -189,6 +207,7 @@ describe('FloatingVideoPlayerComponent video markers', () => {
 
   it('resumes after seeking a recording without a segment manifest', () => {
     agentService.activeVideoSegments.set([]);
+    agentService.usesSessionTimeline.set(true);
     fixture.detectChanges();
     video.dispatchEvent(new Event('loadedmetadata'));
     component.togglePlay();
@@ -201,6 +220,115 @@ describe('FloatingVideoPlayerComponent video markers', () => {
     expect(play).toHaveBeenCalledTimes(1);
   });
 
+  function setSources(computerStart = 0): void {
+    agentService.recordingSources.set([
+      { id: 'android', label: '安卓', kind: 'android', target_ids: ['android'], status: 'ready',
+        segments: [{ url: '/test.mp4', start: 0, duration: 20, width: 400, height: 800 }] },
+      { id: 'screen', label: '电脑', kind: 'computer', target_ids: ['window-1'], status: 'ready',
+        segments: [{ url: '/computer.mp4', start: computerStart, duration: 20, width: 1280, height: 720 }] }
+    ]);
+    agentService.usesSessionTimeline.set(true);
+    agentService.recordingSessionDuration.set(40);
+    agentService.selectRecordingSource('android');
+    fixture.detectChanges();
+  }
+
+  it('handles a seek request once without seeking back on play or pause', () => {
+    const seek = spyOn(component, 'seekToSessionTime').and.callThrough();
+    agentService.videoSeekRequest.set({ seconds: 5, requestId: 1 });
+    fixture.detectChanges();
+    finishSeek();
+    time = 9;
+    component.togglePlay();
+    fixture.detectChanges();
+    expect(time).toBe(9);
+    component.togglePlay();
+    fixture.detectChanges();
+    expect(time).toBe(9);
+    expect(seek).toHaveBeenCalledOnceWith(5, false);
+  });
+
+  it('defers a seek command until the recording is ready', () => {
+    const seek = spyOn(component, 'seekToSessionTime');
+    agentService.recordingPlaybackStatus.set('processing');
+    agentService.videoSeekRequest.set({ seconds: 5, requestId: 2 });
+    fixture.detectChanges();
+    expect(seek).not.toHaveBeenCalled();
+    agentService.recordingPlaybackStatus.set('ready');
+    fixture.detectChanges();
+    component.onLoadedMetadata();
+    expect(seek).toHaveBeenCalledOnceWith(5, false);
+  });
+
+  it('switches to the requested source, waits for seeked and allows later manual source changes', () => {
+    setSources(2);
+    component.togglePlay();
+    agentService.videoSeekRequest.set({ seconds: 7, requestId: 3, sourceId: 'window-1' });
+    fixture.detectChanges();
+    play.calls.reset();
+    video.dispatchEvent(new Event('loadedmetadata'));
+    expect(time).toBe(5);
+    expect(play).not.toHaveBeenCalled();
+    finishSeek();
+    expect(component.currentTime()).toBe(7);
+    expect(paused).toBeFalse();
+    expect(play).toHaveBeenCalledTimes(1);
+    component.selectSource('android');
+    fixture.detectChanges();
+    expect(agentService.selectedRecordingSource()).toBe('android');
+    video.dispatchEvent(new Event('loadedmetadata'));
+    finishSeek();
+    expect(time).toBe(7);
+    expect(agentService.selectedRecordingSource()).toBe('android');
+  });
+
+  it('preserves task time when switching into a gap and advances only on explicit play', () => {
+    setSources(20);
+    time = 12;
+    component.onTimeUpdate();
+    component.togglePlay();
+    component.selectSource('screen');
+    fixture.detectChanges();
+    play.calls.reset();
+    video.dispatchEvent(new Event('loadedmetadata'));
+    expect(component.currentTime()).toBe(12);
+    expect(component.gapMessage()).toContain('未录制');
+    expect(paused).toBeTrue();
+    expect(play).not.toHaveBeenCalled();
+    component.togglePlay();
+    finishSeek();
+    expect(component.currentTime()).toBe(20);
+    expect(component.gapMessage()).toBe('');
+    expect(paused).toBeFalse();
+  });
+
+  it('retains a paused source switch and resumes with one play click', () => {
+    setSources();
+    time = 6;
+    component.onTimeUpdate();
+    component.selectSource('screen');
+    fixture.detectChanges();
+    video.dispatchEvent(new Event('loadedmetadata'));
+    finishSeek();
+    expect(component.currentTime()).toBe(6);
+    expect(paused).toBeTrue();
+    expect(play).not.toHaveBeenCalled();
+    component.togglePlay();
+    expect(play).toHaveBeenCalledTimes(1);
+    expect(paused).toBeFalse();
+  });
+
+  it('does not play another source when the requested source has no recording', () => {
+    setSources();
+    agentService.videoSeekRequest.set({ seconds: 5, requestId: 4, sourceId: 'missing-window' });
+    fixture.detectChanges();
+    component.togglePlay();
+    expect(component.gapMessage()).toContain('来源没有录像');
+    expect(component.currentTime()).toBe(5);
+    expect(paused).toBeTrue();
+    expect(play).not.toHaveBeenCalled();
+  });
+
   it('resets a rejected play request so a single subsequent click can resume', async () => {
     play.and.callFake(() => {
       paused = false;

apps/showcase_ui/src/app/components/floating-video-player/floating-video-player.component.ts

@@ -25,6 +25,7 @@ import {
   ViewChild,
   HostListener,
   effect,
+  untracked,
   OnDestroy
 } from '@angular/core';
 
@@ -33,7 +34,7 @@ import { AgentService } from '../../services/agent.service';
 import { StepReplayFrame } from '../../core/models/stream.model';
 import { drawActionCoordinatesOnOverlay } from '../../utils/image-overlay.util';
 import { getActionIcon } from '../../utils/action-formatter.util';
-import { locateTimelineTime, sessionTimeToTimelineTime } from '../../utils/recording-timeline.util';
+import { locateCoveredSessionTime, locateTimelineTime, sessionTimeToTimelineTime } from '../../utils/recording-timeline.util';
 import { buildVideoStepMarkers, VideoStepMarker } from '../../utils/video-step-markers.util';
 
 @Component({
@@ -61,6 +62,11 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
   public isMuted = signal<boolean>(false);
   public isTheaterMode = signal<boolean>(false);
   public videoLoadError = signal<boolean>(false);
+  public gapMessage = signal('');
+  private sourceSwitch: { time: number; play: boolean } | null = null;
+  private lastHandledVideoSeekRequestId: number | null = null;
+  private gapCanAdvance = true;
+  public sources = computed(() => this.agentService.recordingSources?.() || []);
   public liveStreamUrl = signal<string>('/api/stream/device-live');
   public liveStreamError = signal<boolean>(false);
   public activeSegmentIndex = signal<number>(0);
@@ -71,13 +77,14 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
   private pendingLocalTime: number | null = null;
   private pendingAutoplay = false;
   private pendingAbsoluteSeek: number | null = null;
+  private waitingForMetadata = true;
   private seekPlaybackIntent: boolean | null = null;
   private playRequestId = 0;
 
   // Step Replay state signals
   public activeStepIndex = signal<number>(0);
   public isStepPlaying = signal<boolean>(false);
-  public viewMode = signal<'pre' | 'post'>('pre'); // 'pre' = action preview with coords, 'post' = action result
+  public viewMode = signal<'pre' | 'post'>('post'); // default to the observed result
   public isCardHovered = signal<boolean>(false);
   public isCardPinned = signal<boolean>(false);
   public isCardVisible = computed(() => this.isCardHovered() || this.isCardPinned());
@@ -87,10 +94,14 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
 
   public stepFrames = computed(() => this.agentService.currentSessionStepFrames());
   public videoStepMarkers = computed(() => buildVideoStepMarkers(
-    this.stepFrames(),
+    this.stepFrames().filter(frame => {
+      const source = this.sources().find(s => s.id === this.agentService.selectedRecordingSource?.());
+      return !source || source.target_ids.includes(frame.sourceId || 'android');
+    }),
     this.agentService.currentSession()?.start_time,
     this.agentService.activeVideoSegments(),
-    this.duration()
+    this.duration(),
+    this.agentService.usesSessionTimeline?.() || false
   ));
   public totalStepFrames = computed(() => this.stepFrames().length);
   public currentStepFrame = computed<StepReplayFrame | null>(() => {
@@ -136,12 +147,23 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
         const url = this.agentService.activeVideoUrl();
         const segments = this.agentService.activeVideoSegments();
         this.videoLoadError.set(false);
+        this.waitingForMetadata = true;
         this.activeSegmentIndex.set(0);
-        this.currentTime.set(0);
-        this.duration.set(segments.reduce((sum, segment) => sum + segment.duration, 0));
+        const switching = this.sourceSwitch;
+        this.sourceSwitch = null;
+        this.currentTime.set(switching?.time || 0);
+        this.gapMessage.set('');
+        this.gapCanAdvance = true;
+        this.duration.set(this.agentService.usesSessionTimeline?.()
+          ? Math.max(this.agentService.recordingSessionDuration?.() || 0, ...segments.map(s => s.start + s.duration))
+          : segments.reduce((sum, segment) => sum + segment.duration, 0));
         this.isPlaying.set(false);
         this.pendingLocalTime = null;
         this.pendingAutoplay = false;
+        if (switching) {
+          this.pendingAbsoluteSeek = switching.time;
+          this.pendingAutoplay = switching.play;
+        }
         this.seekPlaybackIntent = null;
         this.playRequestId++;
       },
@@ -157,7 +179,7 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
         if (curSessionId !== lastTrackedSessionId) {
           lastTrackedSessionId = curSessionId;
           this.activeStepIndex.set(0);
-          this.viewMode.set('pre');
+          this.viewMode.set('post');
           this.pauseStepPlay();
         } else if (total > 0 && this.activeStepIndex() >= total) {
           this.activeStepIndex.set(total - 1);
@@ -186,15 +208,42 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
 
     effect(() => {
       const request = this.agentService.videoSeekRequest();
-      if (!request) return;
-      this.pendingAbsoluteSeek = request.seconds;
-      if (this.agentService.recordingPlaybackStatus() !== 'ready') return;
-      const video = this.videoRef?.nativeElement;
-      if (video && video.readyState >= 1) {
-        const target = this.pendingAbsoluteSeek;
-        this.pendingAbsoluteSeek = null;
-        this.seekToSessionTime(target, true);
-      }
+      const status = this.agentService.recordingPlaybackStatus();
+      const sources = this.sources();
+      if (!request || status !== 'ready' || request.requestId === this.lastHandledVideoSeekRequestId) return;
+      // A seek is a command, not a binding to playback/source state. Reading
+      // those signals in this effect would replay the command after play/pause
+      // or a manual source selection.
+      untracked(() => {
+        this.lastHandledVideoSeekRequestId = request.requestId;
+        this.pendingAbsoluteSeek = request.seconds;
+        if (request.sourceId && sources.length) {
+          const source = sources.find(s => s.id === request.sourceId || s.target_ids.includes(request.sourceId!));
+          if (!source) {
+            this.pendingAbsoluteSeek = null;
+            this.pendingLocalTime = null;
+            this.pendingAutoplay = false;
+            this.seekPlaybackIntent = null;
+            this.playRequestId++;
+            this.gapCanAdvance = false;
+            this.currentTime.set(request.seconds);
+            this.gapMessage.set('此步骤的来源没有录像，请查看步骤截图或日志。');
+            this.videoRef?.nativeElement.pause();
+            this.isPlaying.set(false);
+            return;
+          }
+          if (source.id !== this.agentService.selectedRecordingSource?.()) {
+            this.sourceSwitch = { time: request.seconds, play: this.isPlaying() };
+            this.agentService.selectRecordingSource(source.id);
+            return;
+          }
+        }
+        const video = this.videoRef?.nativeElement;
+        if (video && video.readyState >= 1 && !this.waitingForMetadata) {
+          this.pendingAbsoluteSeek = null;
+          this.seekToSessionTime(request.seconds, this.isPlaying());
+        }
+      });
     });
 
     effect(() => {
@@ -305,6 +354,7 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
   public onLoadedMetadata(): void {
     const v = this.videoRef?.nativeElement;
     if (v) {
+      this.waitingForMetadata = false;
       const segments = this.agentService.activeVideoSegments();
       if (!segments.length) {
         this.duration.set(v.duration || 0);
@@ -313,14 +363,6 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
       v.playbackRate = this.playbackRate();
       v.loop = this.isLooping() && !segments.length;
       v.muted = this.isMuted();
-      if (this.pendingAbsoluteSeek !== null) {
-        const target = this.pendingAbsoluteSeek;
-        this.pendingAbsoluteSeek = null;
-        this.seekToSessionTime(target, true);
-        // A seek into another finalized segment changes the video source. Let
-        // that segment's metadata event apply pendingLocalTime.
-        if (this.pendingLocalTime !== null) return;
-      }
       const autoplay = this.agentService.consumeVideoAutoplay();
       if (autoplay) {
         this.isMuted.set(true);
@@ -328,6 +370,14 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
       }
       const shouldPlay = this.pendingAutoplay || autoplay;
       this.pendingAutoplay = false;
+      if (this.pendingAbsoluteSeek !== null) {
+        const target = this.pendingAbsoluteSeek;
+        this.pendingAbsoluteSeek = null;
+        this.seekToSessionTime(target, shouldPlay);
+        // Same-source seeks resume on seeked; cross-segment seeks resume after
+        // the destination metadata and seeked. Never play the old frame here.
+        return;
+      }
       if (this.pendingLocalTime !== null) {
         const localTime = Math.max(0, Math.min(v.duration || this.pendingLocalTime, this.pendingLocalTime));
         this.pendingLocalTime = null;
@@ -339,6 +389,7 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
   }
 
   public onTimeUpdate(): void {
+    if (this.gapMessage()) return;
     const v = this.videoRef?.nativeElement;
     if (v) {
       const segments = this.agentService.activeVideoSegments();
@@ -375,6 +426,14 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
     this.isPlaying.set(false);
     const segments = this.agentService.activeVideoSegments();
     if (segments.length && this.activeSegmentIndex() < segments.length - 1) {
+      const current = segments[this.activeSegmentIndex()];
+      const next = segments[this.activeSegmentIndex() + 1];
+      if (this.agentService.usesSessionTimeline?.() && next.start - current.start - current.duration > 0.05) {
+        this.currentTime.set(current.start + current.duration);
+        this.gapCanAdvance = true;
+        this.gapMessage.set('此处录像存在中断。点击播放可前往下一段，或选择其他步骤。');
+        return;
+      }
       this.pendingLocalTime = 0;
       this.pendingAutoplay = true;
       this.activeSegmentIndex.update(index => index + 1);
@@ -400,6 +459,12 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
    * Playback actions
    */
   public togglePlay(): void {
+    if (this.gapMessage()) {
+      if (!this.gapCanAdvance) return;
+      const next = this.agentService.activeVideoSegments().find(s => s.start > this.currentTime());
+      if (next) this.seek(next.start, true);
+      return;
+    }
     const v = this.videoRef?.nativeElement;
     if (!v) return;
     if (this.pendingLocalTime !== null) {
@@ -462,7 +527,13 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
    */
   public seekToSessionTime(sessionSeconds: number, autoplay = false): void {
     const segments = this.agentService.activeVideoSegments();
-    this.seek(sessionTimeToTimelineTime(segments, sessionSeconds), autoplay);
+    this.seek(this.agentService.usesSessionTimeline?.() ? sessionSeconds : sessionTimeToTimelineTime(segments, sessionSeconds), autoplay);
+  }
+
+  public selectSource(id: string): void {
+    if (id === this.agentService.selectedRecordingSource?.()) return;
+    this.sourceSwitch = { time: this.currentTime(), play: this.isPlaying() };
+    if (!this.agentService.selectRecordingSource(id)) this.sourceSwitch = null;
   }
 
   public seekToVideoMarker(marker: VideoStepMarker): void {
@@ -484,7 +555,23 @@ export class FloatingVideoPlayerComponent implements OnDestroy {
     if (!v) return;
     const target = Math.max(0, Math.min(this.duration(), seconds));
     const segments = this.agentService.activeVideoSegments();
-    const location = locateTimelineTime(segments, target);
+    // URL-only legacy recordings have no coverage manifest. Preserve their
+    // native video seeking instead of treating every timestamp as a gap.
+    const sessionTimeline = this.agentService.usesSessionTimeline?.() && segments.length > 0;
+    const location = sessionTimeline ? locateCoveredSessionTime(segments, target) : locateTimelineTime(segments, target);
+    if (sessionTimeline && !location) {
+      this.playRequestId++;
+      this.pendingLocalTime = null;
+      this.pendingAutoplay = false;
+      this.seekPlaybackIntent = null;
+      v.pause();
+      this.isPlaying.set(false);
+      this.currentTime.set(target);
+      this.gapCanAdvance = true;
+      this.gapMessage.set('此时间段未录制（尚未开始、暂停或录制中断），可查看步骤截图与日志。');
+      return;
+    }
+    this.gapMessage.set('');
     const shouldPlay = autoplay || this.pendingAutoplay || (this.seekPlaybackIntent ?? !v.paused);
     if (!location) {
       this.seekWithinVideo(target, shouldPlay);

apps/showcase_ui/src/app/components/model-settings/capability-settings.component.ts

@@ -4,6 +4,7 @@ import { HttpClient } from '@angular/common/http';
 import { finalize } from 'rxjs';
 
 interface Capabilities {
+  recording?: { codec: string; max_long_edge: number; fps: number; h265_kbps: number; h264_kbps: number; computer_scope: string; minimum_disk_mb: number; hevc_playback: boolean };
   adb_auto_connect: boolean;
   adb_discovery_interval_seconds: number;
   adb_discovery_timeout_seconds: number;
@@ -33,6 +34,20 @@ interface Capabilities {
         <span>视频分析<small>允许智能体分析任务画面。运行期间会临时录制，完成后不保留录像（除非开启录屏）。</small></span></label>
       <label><input type="checkbox" role="switch" [(ngModel)]="config.recording_enabled" [disabled]="busy() || loading() || !loaded()" />
         <span>保存录屏<small>保留任务录像以便复盘，可单独开启，不要求启用视频分析。</small></span></label>
+      @if (config.recording; as recording) {
+        <details class="execution-settings">
+          <summary>录像留档设置</summary>
+          <p>默认低码率留档，不影响模型观察截图。仅录实际使用的来源；屏幕模式包含该屏幕其他可见内容。</p>
+          <div class="number-fields">
+            <label>编码<select [(ngModel)]="recording.codec" [disabled]="busy() || !loaded()"><option value="auto">自动选择兼容编码</option><option value="h265">优先 H.265</option><option value="h264">H.264</option></select></label>
+            <label>电脑录制范围<select [(ngModel)]="recording.computer_scope" [disabled]="busy() || !loaded()"><option value="screen">当前控制的屏幕</option><option value="window">仅目标窗口</option></select></label>
+            <label>最长边（像素）<input type="number" min="320" max="4096" [(ngModel)]="recording.max_long_edge" [disabled]="busy() || !loaded()" /></label>
+            <label>最高帧率<input type="number" min="5" max="60" [(ngModel)]="recording.fps" [disabled]="busy() || !loaded()" /></label>
+            <label>H.265 目标码率（kbps）<input type="number" min="128" max="10000" [(ngModel)]="recording.h265_kbps" [disabled]="busy() || !loaded()" /></label>
+            <label>H.264 目标码率（kbps）<input type="number" min="128" max="20000" [(ngModel)]="recording.h264_kbps" [disabled]="busy() || !loaded()" /></label>
+          </div>
+        </details>
+      }
       <label><input type="checkbox" role="switch" [(ngModel)]="config.outputter_enabled" [disabled]="busy() || loading() || !loaded()" />
         <span>最终报告生成<small>根据执行记录生成任务结果报告。</small></span></label>
       <label><input type="checkbox" role="switch" [(ngModel)]="config.step_summarizer_enabled" [disabled]="busy() || loading() || !loaded()" />
@@ -68,7 +83,7 @@ interface Capabilities {
     summary { cursor: pointer; font-weight: 600; color: #334155; }
     .number-fields { display: grid; grid-template-columns: 1fr 1fr; gap: 12px 16px; }
     .number-fields label { display: block; padding: 0; color: #334155; }
-    .number-fields input { box-sizing: border-box; display: block; width: 100%; height: auto; margin-top: 8px; padding: 9px; border: 1px solid #cbd5e1; border-radius: 8px; font: inherit; }
+    .number-fields input, .number-fields select { box-sizing: border-box; display: block; width: 100%; height: auto; margin-top: 8px; padding: 9px; border: 1px solid #cbd5e1; border-radius: 8px; font: inherit; }
     @media(max-width: 480px) { .number-fields { grid-template-columns: 1fr; } }`]
 })
 export class CapabilitySettingsComponent implements OnInit {
@@ -83,6 +98,9 @@ export class CapabilitySettingsComponent implements OnInit {
   }
   save(): void {
     if (this.busy() || this.loading() || !this.loaded() || !this.valid()) return;
+    if (this.config.recording) {
+      this.config.recording.hevc_playback = !!document.createElement('video').canPlayType('video/mp4; codecs="hvc1"');
+    }
     this.busy.set(true); this.message.set(''); this.error.set('');
     this.http.post<Capabilities>('/api/system/capability-settings', this.config).pipe(finalize(() => this.busy.set(false))).subscribe({
       next: config => { this.config = config; this.changed.emit(config); this.message.set('功能设置已保存，新任务将使用此配置。'); },
@@ -91,6 +109,12 @@ export class CapabilitySettingsComponent implements OnInit {
   }
   valid(): boolean {
     const c = this.config;
+    const r = c.recording;
+    if (r && (!['auto', 'h264', 'h265'].includes(r.codec) || !['screen', 'window'].includes(r.computer_scope)
+      || !Number.isInteger(r.max_long_edge) || r.max_long_edge < 320 || r.max_long_edge > 4096
+      || !Number.isInteger(r.fps) || r.fps < 5 || r.fps > 60
+      || !Number.isInteger(r.h265_kbps) || r.h265_kbps < 128 || r.h265_kbps > 10000
+      || !Number.isInteger(r.h264_kbps) || r.h264_kbps < 128 || r.h264_kbps > 20000)) return false;
     return Number.isInteger(c.task_timeout_seconds) && c.task_timeout_seconds >= 30 && c.task_timeout_seconds <= 86400
       && c.observation_timeout_seconds > 0 && c.observation_timeout_seconds <= 300
       && Number.isInteger(c.observation_max_attempts) && c.observation_max_attempts >= 1 && c.observation_max_attempts <= 5

apps/showcase_ui/src/app/core/models/stream.model.ts

@@ -225,6 +225,8 @@ export interface CheckerResult {
 }
 
 export interface StepReplayFrame {
+  sourceId?: string;
+  sourceLabel?: string;
   index: number;
   stepNumber: number;
   rawStepNumber?: number;

apps/showcase_ui/src/app/services/agent.service.spec.ts

@@ -230,6 +230,17 @@ describe('AgentService recording finalization lifecycle', () => {
     (service as any).videoWaitStartedAt = Date.now();
     service.activeVideoUrl = signal<string | null>(null);
     service.activeVideoSegments = signal<any[]>([]);
+    service.recordingSources = signal([]);
+    service.selectedRecordingSource = signal('');
+    service.usesSessionTimeline = signal(false);
+    service.recordingSessionDuration = signal(0);
+    service.activeVideoTitle = signal('');
+    service.isVideoWindowOpen = signal(false);
+    service.isVideoMinimized = signal(false);
+    service.videoSeekRequest = signal(null);
+    service.currentSessionId = signal<string | null>('session-1');
+    service.sessions = computed(() => (service as any).rawSessions());
+    (service as any).videoSeekRequestId = 0;
     service.isVideoLoading = signal(false);
     service.recordingPlaybackStatus = signal('idle');
     service.recordingPlaybackMessage = signal('');
@@ -293,6 +304,89 @@ describe('AgentService recording finalization lifecycle', () => {
     expect(service.isVideoLoading()).toBeFalse();
     expect(service.recordingPlaybackMessage()).toBe('ffmpeg failed');
   });
+
+  it('publishes task-time sources and selects a computer recording by its target alias', () => {
+    const android = {
+      id: 'android', label: '安卓', kind: 'android', target_ids: ['android'], status: 'ready',
+      segments: [{ url: '/videos/android.mp4', start: 0, duration: 10, width: 720, height: 1280 }]
+    };
+    const computer = {
+      id: 'screen-1', label: '电脑', kind: 'computer', target_ids: ['window-1', 'window-2'], status: 'ready',
+      message: '已使用兼容编码',
+      segments: [{ url: '/videos/computer.mp4', start: 12, duration: 10, width: 1280, height: 720 }]
+    };
+    const service = createVideoService({
+      session_id: 'session-1', status: 'ready', has_video: true,
+      video_url: android.segments[0].url, video_segments: android.segments,
+      sources: [android, computer], session_timeline: true, session_duration: 25
+    });
+
+    (service as any).requestSessionVideo('session-1', 1);
+    expect(service.recordingSources()).toEqual([android, computer]);
+    expect(service.selectedRecordingSource()).toBe('android');
+    expect(service.usesSessionTimeline()).toBeTrue();
+    expect(service.recordingSessionDuration()).toBe(25);
+
+    expect(service.selectRecordingSource('window-2')).toBeTrue();
+    expect(service.selectedRecordingSource()).toBe('screen-1');
+    expect(service.activeVideoUrl()).toBe('/videos/computer.mp4');
+    expect(service.activeVideoSegments()).toBe(computer.segments);
+    expect(service.recordingPlaybackMessage()).toBe('已使用兼容编码');
+    expect(service.selectRecordingSource('screen-1')).toBeTrue();
+    expect(service.activeVideoSegments()).toBe(computer.segments);
+    expect(service.selectRecordingSource('missing-source')).toBeFalse();
+    expect(service.selectedRecordingSource()).toBe('screen-1');
+    expect(service.activeVideoUrl()).toBe('/videos/computer.mp4');
+  });
+
+  it('keeps a failed source explicit without retaining another source video', () => {
+    const service = createVideoService({});
+    service.activeVideoUrl.set('/videos/android.mp4');
+    service.recordingSources.set([{
+      id: 'screen-1', label: '电脑', kind: 'computer', target_ids: ['window-1'], status: 'failed',
+      segments: [], message: '录制进程未产生画面'
+    }]);
+
+    expect(service.selectRecordingSource('window-1')).toBeTrue();
+    expect(service.activeVideoUrl()).toBeNull();
+    expect(service.activeVideoSegments()).toEqual([]);
+    expect(service.recordingPlaybackMessage()).toBe('录制进程未产生画面');
+  });
+
+  it('clears the previous task source and seek command when opening another task normally', () => {
+    const service = createVideoService({});
+    const load = spyOn<any>(service, 'requestSessionVideo');
+    service.recordingSources.set([{
+      id: 'screen-old', label: '电脑', kind: 'computer', target_ids: ['window-old'], status: 'ready',
+      segments: [{ url: '/videos/old.mp4', start: 0, duration: 10, width: 1280, height: 720 }]
+    }]);
+    service.selectRecordingSource('screen-old');
+    service.usesSessionTimeline.set(true);
+    service.recordingSessionDuration.set(10);
+    service.requestVideoSeek(6, 'window-old');
+
+    service.openVideoPlayer('session-2');
+
+    expect(service.videoSeekRequest()).toBeNull();
+    expect(service.recordingSources()).toEqual([]);
+    expect(service.selectedRecordingSource()).toBe('');
+    expect(service.usesSessionTimeline()).toBeFalse();
+    expect(service.recordingSessionDuration()).toBe(0);
+    expect(service.activeVideoSegments()).toEqual([]);
+    expect(service.activeVideoUrl()).toBeNull();
+    expect(load).toHaveBeenCalledOnceWith('session-2', 2);
+  });
+
+  it('replaces the previous task seek with the new source-specific command', () => {
+    const service = createVideoService({});
+    spyOn<any>(service, 'requestSessionVideo');
+    service.requestVideoSeek(6, 'window-old');
+    const oldRequestId = service.videoSeekRequest()!.requestId;
+
+    service.openVideoPlayer('session-2', undefined, undefined, 14, undefined, 'window-new');
+
+    expect(service.videoSeekRequest()).toEqual({ seconds: 14, sourceId: 'window-new', requestId: oldRequestId + 1 });
+  });
 });
 
 describe('AgentService video analysis seeking', () => {

apps/showcase_ui/src/app/services/agent.service.ts

@@ -39,6 +39,18 @@ export interface VideoSegment {
   duration_ms?: number;
 }
 
+export interface RecordingSource {
+  id: string;
+  label: string;
+  kind: string;
+  target_ids: string[];
+  status: string;
+  message?: string;
+  codec?: string;
+  segments: VideoSegment[];
+  url?: string;
+}
+
 export type RecordingPlaybackStatus =
   | 'idle'
   | 'live'
@@ -64,6 +76,9 @@ export interface GuidanceEvent {
 }
 
 interface SessionVideoResponse {
+  sources?: RecordingSource[];
+  session_timeline?: boolean;
+  session_duration?: number;
   session_id: string;
   status?: 'processing' | 'ready' | 'failed' | 'unavailable';
   has_video: boolean;
@@ -272,12 +287,27 @@ export class AgentService {
   public isVideoMinimized = signal<boolean>(false);
   public activeVideoUrl = signal<string | null>(null);
   public activeVideoSegments = signal<VideoSegment[]>([]);
+  public recordingSources = signal<RecordingSource[]>([]);
+  public selectedRecordingSource = signal<string>('');
+  public usesSessionTimeline = signal(false);
+  public recordingSessionDuration = signal(0);
+
+  public selectRecordingSource(id: string): boolean {
+    const source = this.recordingSources().find(s => s.id === id || s.target_ids?.includes(id));
+    if (!source) return false;
+    if (this.selectedRecordingSource() === source.id) return true;
+    this.selectedRecordingSource.set(source.id);
+    this.activeVideoSegments.set(source.segments);
+    this.activeVideoUrl.set(source.url || source.segments[0]?.url || null);
+    this.recordingPlaybackMessage.set(source.message || '');
+    return true;
+  }
   public activeVideoTitle = signal<string>('');
   public isVideoLoading = signal<boolean>(false);
   public recordingPlaybackStatus = signal<RecordingPlaybackStatus>('idle');
   public recordingPlaybackMessage = signal<string>('');
   public shouldAutoplayVideo = signal<boolean>(false);
-  public videoSeekRequest = signal<{ seconds: number; requestId: number } | null>(null);
+  public videoSeekRequest = signal<{ seconds: number; requestId: number; sourceId?: string } | null>(null);
   public stepSeekRequest = signal<{ index: number; requestId: number } | null>(null);
   public playerMode = signal<'video' | 'steps'>('video');
   public streamResetEvent = signal<LLMStreamResetEventData | null>(null);
@@ -1838,7 +1868,8 @@ export class AgentService {
     videoUrl?: string,
     title?: string,
     seekSeconds?: number,
-    stepIndex?: number
+    stepIndex?: number,
+    sourceId?: string
   ): void {
     const targetSessionId = sessionId || this.currentSessionId();
     const session = this.sessions().find(s => s.session_id === targetSessionId);
@@ -1854,7 +1885,12 @@ export class AgentService {
     this.shouldAutoplayVideo.set(false);
     this.recordingPlaybackMessage.set('');
     this.activeVideoSegments.set([]);
-    if (Number.isFinite(seekSeconds)) this.requestVideoSeek(Number(seekSeconds));
+    this.recordingSources.set([]);
+    this.selectedRecordingSource.set('');
+    this.usesSessionTimeline.set(false);
+    this.recordingSessionDuration.set(0);
+    if (Number.isFinite(seekSeconds)) this.requestVideoSeek(Number(seekSeconds), sourceId);
+    else this.videoSeekRequest.set(null);
     if (Number.isFinite(stepIndex)) {
       this.playerMode?.set('steps');
       this.requestStepSeek(Number(stepIndex));
@@ -1941,6 +1977,9 @@ export class AgentService {
           return;
         }
         const status = res.status || (res.has_video && res.video_url ? 'ready' : 'unavailable');
+        this.recordingSources.set(res.sources || []);
+        this.usesSessionTimeline.set(!!res.session_timeline);
+        this.recordingSessionDuration.set(res.session_duration || 0);
         if (status === 'ready' && res.video_url) {
           this.cancelVideoRetry();
           this.isVideoLoading.set(false);
@@ -1948,6 +1987,7 @@ export class AgentService {
           this.recordingPlaybackMessage.set('');
           this.activeVideoUrl.set(res.video_url);
           this.activeVideoSegments.set(res.video_segments || []);
+          this.selectedRecordingSource.set(res.sources?.find(s => s.segments.length || s.url)?.id || '');
           this.playerMode?.set('video');
           this.rawSessions.update((list) =>
             list.map((s) => s.session_id === sessionId
@@ -2041,10 +2081,11 @@ export class AgentService {
     return shouldAutoplay;
   }
 
-  public requestVideoSeek(seconds: number): void {
+  public requestVideoSeek(seconds: number, sourceId?: string): void {
     if (!Number.isFinite(seconds)) return;
     this.videoSeekRequest.set({
       seconds: Math.max(0, seconds),
+      sourceId,
       requestId: ++this.videoSeekRequestId
     });
   }

apps/showcase_ui/src/app/utils/action-formatter.util.spec.ts

@@ -2,13 +2,109 @@ import {
   getStepPreImageUrl,
   getStepPostImageUrl,
   extractStepReplayFrames,
+  getActionTitle,
+  getActionCoords,
   getActionTargetText,
   isActionFailed,
   getActionErrorMessage,
   extractActionExtraParams
 } from './action-formatter.util';
 
+describe('plugin replay marker descriptions', () => {
+  it('does not parse digits in real hashed click and observation names as coordinates', () => {
+    const click = { action: 'plug_96c522c2e2_click', plugin_id: 'open-computer-use', args: { app: 'Demo' } };
+    const observe = { action: 'plug_5f6ec6ac87_get_app_state', plugin_id: 'open-computer-use', args: { app: 'Demo' } };
+    expect(getActionTitle(click)).toBe('点击');
+    expect(getActionTitle(observe)).toBe('观察电脑窗口');
+    expect(getActionCoords(click)).toBe('');
+    expect(getActionCoords(observe)).toBe('');
+    expect(getActionTargetText(observe)).toBe('Demo');
+    expect(getActionCoords({ ...click, args: { app: 'Demo', x: 100, y: 200 } })).toBe('[100, 200]');
+  });
+
+  it('uses the original tool name from the trace rather than its dynamic binding name', () => {
+    const trace = { name: 'plug_96c522c2e2_click', type: 'tool',
+      payload: { plugin_id: 'desktop', tool_name: 'double_click', args: { x: 70, y: 80 } } };
+    expect(getActionTitle(trace)).toBe('双击');
+    expect(getActionCoords(trace)).toBe('[70, 80]');
+  });
+
+  it('counts a Flash action and its tool trace as one frame using their shared call id', () => {
+    const callId = 'tool-call-1';
+    const action = { action: 'plug_96c522c2e2_click', call_id: callId, plugin_id: 'open-computer-use',
+      source_id: 'computer-window', args: { app: 'Demo', x: 100, y: 200 } };
+    const trace = { type: 'tool', name: action.action, trace_id: 'different-trace-uuid', timestamp: 109,
+      payload: { plugin_id: action.plugin_id, call_id: callId, source_id: action.source_id,
+        args: action.args, post_image_name: 'computer-shot' } };
+    const frames = extractStepReplayFrames([{ step_id: 'step-4', step_number: 4, timestamp: 109,
+      action_taken: action, generic_tools: [trace], post_image_name: 'computer-shot' }]);
+    expect(frames.length).toBe(1);
+    expect(frames[0].sourceId).toBe('computer-window');
+    expect(frames[0].postImageUrl).toBe('/images/computer-shot');
+    expect(frames[0].actionText).toBe('点击 ([100, 200])');
+    expect(frames[0].rawStepNumber).toBe(4);
+  });
+
+  it('preserves repeated same-source actions when they have different call ids', () => {
+    const trace = (callId: string, timestamp: number) => ({ type: 'tool', name: 'plug_96c522c2e2_click',
+      trace_id: `trace-${callId}`, timestamp, payload: { plugin_id: 'desktop', call_id: callId,
+        source_id: 'computer-window', args: { x: 100, y: 200 }, post_image_name: `shot-${callId}` } });
+    const first = trace('one', 101);
+    const second = trace('two', 102);
+    const frames = extractStepReplayFrames([{ step_id: 'multi', step_number: 1, timestamp: 101,
+      action_taken: first, generic_tools: [first, second] }]);
+    expect(frames.length).toBe(2);
+    expect(frames.map(frame => frame.timestamp)).toEqual([101, 102]);
+  });
+});
+
 describe('action-formatter.util screenshot chaining', () => {
+  for (const update of [
+    { pre_image_name: null },
+    { pre_screenshot: null },
+    { last_execution_result: { pre_image_valid: false } }
+  ]) {
+    it(`clears invalidated before-image aliases in merged step updates: ${JSON.stringify(update)}`, () => {
+      const frames = extractStepReplayFrames([
+        { type: 'step_recorded', data: {
+          step_id: 'cross-window', step_number: 1, timestamp: 100,
+          pre_image_name: 'window-a', pre_screenshot: 'legacy-window-a',
+          action_taken: { action: 'click' }
+        } },
+        { type: 'step_updated', data: {
+          step_id: 'cross-window', post_image_name: 'window-b', ...update
+        } },
+        { type: 'step_updated', data: { step_id: 'cross-window', summary: 'Completed' } }
+      ]);
+      expect(frames.length).toBe(1);
+      expect(frames[0].preImageUrl).toBeNull();
+      expect(frames[0].postImageUrl).toBe('/images/window-b');
+      expect(frames[0].imageUrl).toBe('/images/window-b');
+    });
+  }
+
+  it('retains a valid before-image when an update omits image fields', () => {
+    const frames = extractStepReplayFrames([
+      { step_id: 'same-window', step_number: 1, timestamp: 100,
+        pre_image_name: 'before', action_taken: { action: 'click' } },
+      { step_id: 'same-window', summary: 'Completed', post_image_name: 'after' }
+    ]);
+    expect(frames[0].preImageUrl).toBe('/images/before');
+    expect(frames[0].postImageUrl).toBe('/images/after');
+  });
+
+  it('keeps distinct windows from the same plugin as distinct evidence and replay frames', () => {
+    const first = { type: 'tool', trace_id: 'w1', name: 'plug_click', timestamp: 101,
+      payload: { plugin_id: 'desktop', source_id: 'window-1', post_image_name: 'one' } };
+    const second = { type: 'tool', trace_id: 'w2', name: 'plug_click', timestamp: 102,
+      payload: { plugin_id: 'desktop', source_id: 'window-2', post_image_name: 'two' } };
+    const step = { step_id: 'multi', step_number: 1, timestamp: 101, action_taken: first, generic_tools: [first, second] };
+    expect(getStepPostImageUrl(step, first)).toBe('/images/one');
+    expect(getStepPostImageUrl(step, second)).toBe('/images/two');
+    const frames = extractStepReplayFrames([step]);
+    expect(frames.map(frame => frame.sourceId)).toEqual(['window-1', 'window-2']);
+    expect(frames.map(frame => frame.postImageUrl)).toEqual(['/images/one', '/images/two']);
+  });
   it('keeps plugin observations separate from the phone image chain', () => {
     const desktop = { type: 'tool', trace_id: 'desktop', name: 'plug_observe', timestamp: 1001,
       payload: { plugin_id: 'desktop-control', post_image_name: 'desktop-screen' } };

apps/showcase_ui/src/app/utils/action-formatter.util.ts

@@ -17,6 +17,7 @@
 import { ActionParam, StepReplayFrame } from '../core/models/stream.model';
 import { extractNumbersFromCoordinateValue, isPureDirectionString, parseSequenceCoordinates, unwrapTraceAction } from './image-overlay.util';
 import { cleanErrorMessage, joinTargetDescriptions } from './tool-formatter.util';
+import { getPluginToolView } from './plugin-tool-view.util';
 
 /**
  * Safely parse JSON string if it looks like an object/array, otherwise return original value
@@ -166,6 +167,8 @@ export function getActionIcon(action: any): string {
 export function getActionTitle(action: any): string {
   const act = getActionObject(action);
   if (!act) return "操作";
+  const plugin = getPluginToolView(act);
+  if (plugin) return plugin.title;
   const name = (act.name || act.action || '').toLowerCase();
   switch (name) {
     case 'tap':
@@ -233,6 +236,11 @@ export function getActionTargetText(action: any): string {
   const name = (act.name || act.action || '').toLowerCase();
   const args = act.args && typeof act.args === 'object' ? act.args : {};
 
+  if (getPluginToolView(act)) {
+    return act.target_text || act.target_description || args.target_text || args.target_description
+      || act.app_name || act.app || args.app_name || args.app || '';
+  }
+
   if (name.includes('app')) {
     const rawApp = act.app_name || act.package_name || act.app || args.app_name || args.package_name || args.app || '';
     const app = rawApp ? (rawApp.charAt(0).toUpperCase() + rawApp.slice(1)) : '';
@@ -321,6 +329,7 @@ export function getActionCoords(action: any): string {
   if (!act) return '';
   const args = act.args && typeof act.args === 'object' ? act.args : {};
   const name = (act.name || act.action || '').toLowerCase();
+  const plugin = getPluginToolView(act) !== null;
 
   // Check sequence first if action is sequence or sequence arg is present
   const isSequenceAction = name === 'click_sequence' || name === 'tap_sequence' || Boolean(act.sequence || args.sequence || act.normalized_sequence || args.normalized_sequence);
@@ -344,15 +353,18 @@ export function getActionCoords(action: any): string {
   }
 
   const normCoords = extractNumbersFromCoordinateValue(act.normalized_coordinates || args.normalized_coordinates);
+  const point = [act.x ?? args.x, act.y ?? args.y];
+  const explicitPoint = point.every(value => typeof value === 'number' && Number.isFinite(value)) ? point : null;
   const coords = normCoords ||
+                 explicitPoint ||
                  extractNumbersFromCoordinateValue(act.coordinates) ||
                  extractNumbersFromCoordinateValue(args.coordinates) ||
                  extractNumbersFromCoordinateValue(act.coords) ||
                  extractNumbersFromCoordinateValue(args.coords) ||
                  extractNumbersFromCoordinateValue(act.target) ||
                  extractNumbersFromCoordinateValue(args.target) ||
-                 extractNumbersFromCoordinateValue(args.action) ||
-                 extractNumbersFromCoordinateValue(act.action) ||
+                 extractNumbersFromCoordinateValue(plugin ? null : args.action) ||
+                 extractNumbersFromCoordinateValue(plugin ? null : act.action) ||
                  extractNumbersFromCoordinateValue(args.gesture) ||
                  extractNumbersFromCoordinateValue(act.gesture);
 
@@ -613,6 +625,29 @@ function pluginSource(item: any): string | null {
   return act?.plugin_id || payload?.plugin_id || payload?.action?.plugin_id || result?.plugin_id || null;
 }
 
+function toolCallIdentity(item: any): string {
+  const act = getActionObject(item);
+  const payload = safeParseJson(act?.payload);
+  const result = safeParseJson(act?.result || payload?.result);
+  return String(act?.call_id || payload?.call_id || result?.call_id || act?.trace_id || '');
+}
+
+export function actionEvidenceSource(item: any): { id: string; label: string } {
+  const act = getActionObject(item);
+  const payload = safeParseJson(act?.payload) || {};
+  const result = safeParseJson(act?.result || payload?.result) || {};
+  const source = act?.source || payload?.source || payload?.action?.source || result?.source;
+  const plugin = pluginSource(item);
+  const android = !item || isAndroidAction(act) || isReportStatusAction(act);
+  const recording = act?.recording || result?.recording;
+  const recordingLabel = recording?.status === 'recording' ? (recording.scope === 'screen' ? ' · 屏幕录制中' : ' · 窗口录制中') : recording?.status === 'failed' ? ' · 录像异常' : '';
+  const labels: Record<string, string> = { computer: '电脑', android: '安卓', terminal: '终端', api: '接口', file: '文件', other: '插件' };
+  return {
+    id: act?.source_id || payload?.source_id || payload?.action?.source_id || result?.source_id || (plugin ? `plugin:${plugin}` : android ? 'android' : 'auxiliary'),
+    label: (source ? (labels[source.kind] || '插件') + (source.label ? ` · ${source.label}` : '') : (plugin ? `插件 · ${plugin}` : android ? '安卓' : '辅助')) + recordingLabel
+  };
+}
+
 export function extractItemPreImage(item: any): string | null {
   const act = getActionObject(item);
   if (!act) return null;
@@ -792,7 +827,7 @@ export function resolveStepImageMap(stepData: any): Map<any, ResolvedImageEntry>
   }
 
   // Ensure stepData.action_taken cross-resolves to the action images
-  if (stepData.action_taken && lastActionItem && pluginSource(stepData.action_taken) === pluginSource(lastActionItem)) {
+  if (stepData.action_taken && lastActionItem && actionEvidenceSource(stepData.action_taken).id === actionEvidenceSource(lastActionItem).id) {
     const lastEntry = map.get(lastActionItem);
     if (lastEntry) {
       map.set(stepData.action_taken, lastEntry);
@@ -905,19 +940,23 @@ export function extractStepReplayFrames(logsOrSteps: any[]): StepReplayFrame[] {
 
       const mergedStepNum = (stepNum !== null) ? stepNum : existing.step_number;
       const mergedStepId = existing.step_id || step.step_id;
+      // Explicit invalidation must clear both aliases, otherwise an old window
+      // screenshot can reappear through the legacy pre_screenshot fallback.
+      const clearPreImage = safeParseJson(step.last_execution_result)?.pre_image_valid === false ||
+        (!preferExistingAction && (step.pre_image_name === null || step.pre_screenshot === null));
 
       const merged = {
         ...existing,
         ...step,
         step_id: mergedStepId,
         step_number: mergedStepNum,
-        pre_image_name: preferExistingAction
+        pre_image_name: clearPreImage ? null : preferExistingAction
           ? existing.pre_image_name
           : (step.pre_image_name || existing.pre_image_name),
         post_image_name: preferExistingAction
           ? (existing.post_image_name || step.post_image_name)
           : (step.post_image_name || existing.post_image_name),
-        pre_screenshot: preferExistingAction
+        pre_screenshot: clearPreImage ? null : preferExistingAction
           ? existing.pre_screenshot
           : (step.pre_screenshot || existing.pre_screenshot),
         post_screenshot: preferExistingAction
@@ -978,6 +1017,20 @@ export function extractStepReplayFrames(logsOrSteps: any[]): StepReplayFrame[] {
         primaryImg
       });
     }
+    // A Pro step may contain actions from several targets. Each tool retains its
+    // own frame and timestamp instead of borrowing the step's last phone image.
+    const seen = new Set<string>([toolCallIdentity(act)]);
+    for (const tool of stepData.generic_tools || []) {
+      if (!pluginSource(tool)) continue;
+      const key = toolCallIdentity(tool);
+      if (tool === act || (key && seen.has(key))) continue;
+      if (key) seen.add(key);
+      const toolPre = getStepPreImageUrl(stepData, tool);
+      const toolPost = getStepPostImageUrl(stepData, tool);
+      if (!toolPre && !toolPost) continue;
+      visualCandidates.push({ stepData: { ...stepData, timestamp: tool.timestamp || stepData.timestamp }, act: tool,
+        preUrl: toolPre, postUrl: toolPost, primaryImg: toolPost || toolPre! });
+    }
   }
 
   // Sort strictly by original step_number ascending, then by timestamp
@@ -1004,6 +1057,8 @@ export function extractStepReplayFrames(logsOrSteps: any[]): StepReplayFrame[] {
       : stepNum;
 
     frames.push({
+      sourceId: actionEvidenceSource(act).id,
+      sourceLabel: actionEvidenceSource(act).label,
       index: i,
       stepNumber: stepNum,
       rawStepNumber: rawStepNum,

apps/showcase_ui/src/app/utils/plugin-tool-view.util.spec.ts

@@ -0,0 +1,71 @@
+import { actionEvidenceSource, getStepPostImageUrl, getStepPreImageUrl } from './action-formatter.util';
+import { isDeviceActionTool } from './tool-formatter.util';
+import { getPluginToolView } from './plugin-tool-view.util';
+
+describe('plugin evidence timeline cards', () => {
+  const observation = () => ({
+    name: 'plug_abcdef1234_get_app_state', type: 'tool', status: 'success',
+    payload: {
+      plugin_id: 'desktop', tool_name: 'get_app_state', effect: 'read', observation: true,
+      args: { app: 'Demo' }, source_id: 'plugin:desktop:window-1',
+      source: { kind: 'computer', label: 'Demo' }, post_image_name: 'desktop-screen',
+      pre_image_valid: false,
+      result: { plugin_id: 'desktop', execution: 'completed', side_effect: 'none',
+        content: [{ type: 'text', text: '当前窗口已截图' }] }
+    }
+  });
+
+  it('renders real dynamic plugin observations with computer evidence, not Android actions', () => {
+    const tool = observation();
+    expect(isDeviceActionTool(tool)).toBeFalse();
+    expect(getPluginToolView(tool)).toEqual(jasmine.objectContaining({
+      title: '观察电脑窗口', kind: '辅助观察', auxiliary: true,
+      status: '已执行', result: '当前窗口已截图', args: [{ key: 'app', value: 'Demo' }]
+    }));
+    expect(actionEvidenceSource(tool)).toEqual({ id: 'plugin:desktop:window-1', label: '电脑 · Demo' });
+    const step = { pre_image_name: 'phone-before', post_image_name: 'phone-after', generic_tools: [tool] };
+    expect(getStepPreImageUrl(step, tool)).toBeNull();
+    expect(getStepPostImageUrl(step, tool)).toContain('desktop-screen');
+  });
+
+  it('uses the declared write contract even before a tool result arrives', () => {
+    const view = getPluginToolView({ name: 'plug_abcdef1234_click', type: 'tool', status: 'running',
+      payload: { plugin_id: 'desktop', tool_name: 'click', effect: 'write', observation: false,
+        args: { x: 30, y: 40 } } });
+    expect(view).toEqual(jasmine.objectContaining({ title: '点击', kind: '插件操作', auxiliary: false, status: '执行中' }));
+  });
+
+  it('keeps read-only file calls folded without labelling them as screen observations', () => {
+    const view = getPluginToolView({ name: 'plug_abcdef1234_read_file',
+      payload: { plugin_id: 'files', effect: 'read', observation: false, args: { path: 'result.txt' } } });
+    expect(view).toEqual(jasmine.objectContaining({ kind: '辅助读取', auxiliary: true, status: '待返回' }));
+  });
+
+  it('uses stored legacy results and keeps unavailable outcomes uncertain', () => {
+    const view = getPluginToolView({ name: 'plug_abcdef1234_click', type: 'tool', status: 'failed',
+      payload: JSON.stringify({ plugin_id: 'desktop', args: { x: 30 },
+        result: { tool: 'click', execution: 'unknown', side_effect: 'possible', message: '连接中断，需重新观察' } }) });
+    expect(view).toEqual(jasmine.objectContaining({ kind: '插件操作', auxiliary: false,
+      status: '结果不确定', result: '连接中断，需重新观察', args: [{ key: 'x', value: '30' }] }));
+  });
+
+  it('does not guess write intent from a dynamic tool name with missing metadata', () => {
+    expect(getPluginToolView({ name: 'plug_abcdef1234_click', status: 'running' }))
+      .toEqual(jasmine.objectContaining({ kind: '插件调用', auxiliary: true, args: [] }));
+  });
+
+  it('does not treat ordinary Android or model tools as plugin calls', () => {
+    expect(getPluginToolView({ name: 'click', type: 'tool', payload: { args: { x: 30 } } })).toBeNull();
+    expect(getPluginToolView({ name: 'plug_abcdef1234_click', type: 'llm_call' })).toBeNull();
+  });
+
+  it('reflects a completed receipt even when stream updates reuse the tool object', () => {
+    const tool: any = { name: 'plug_abcdef1234_click', status: 'running',
+      payload: { plugin_id: 'desktop', effect: 'write', args: {} } };
+    expect(getPluginToolView(tool)?.status).toBe('执行中');
+    tool.status = 'success';
+    tool.payload.result = { execution: 'completed', message: '已点击' };
+    expect(getPluginToolView(tool)?.status).toBe('已执行');
+    expect(getPluginToolView(tool)?.result).toBe('已点击');
+  });
+});

apps/showcase_ui/src/app/utils/plugin-tool-view.util.ts

@@ -0,0 +1,58 @@
+export interface PluginToolView {
+  title: string;
+  kind: string;
+  auxiliary: boolean;
+  result: string;
+  status: string;
+  args: Array<{ key: string; value: string }>;
+}
+
+function objectValue(value: any): Record<string, any> {
+  if (typeof value === 'string') {
+    try { value = JSON.parse(value); } catch { return {}; }
+  }
+  return value && typeof value === 'object' && !Array.isArray(value) ? value : {};
+}
+
+/** Display plugin calls from their declared contract, never from Android tool names. */
+export function getPluginToolView(tool: any): PluginToolView | null {
+  if (!tool || typeof tool !== 'object' || tool.type === 'llm_call') return null;
+  const payload = objectValue(tool.payload);
+  const result = objectValue(payload['result'] ?? tool.result);
+  const pluginId = payload['plugin_id'] || result['plugin_id'] || tool.plugin_id;
+  if (!pluginId && !/^plug_[a-f0-9]{10}_/.test(tool.name || tool.action || '')) {
+    return null;
+  }
+  const name = String(payload['tool_name'] || result['tool'] || tool.name || tool.action || '插件调用')
+    .replace(/^plug_[a-f0-9]{10}_/, '');
+  const effect = payload['effect'] || (result['side_effect'] === 'none' ? 'read' :
+    ['possible', 'confirmed'].includes(result['side_effect']) ? 'write' : '');
+  const auxiliary = effect !== 'write';
+  const titles: Record<string, string> = {
+    get_app_state: '观察电脑窗口', list_apps: '列出电脑应用',
+    click: '点击', double_click: '双击', drag: '拖动', scroll: '滚动',
+    type_text: '输入文本', press_key: '按键', screenshot: '截取屏幕',
+    save_screenshot: '保存截图', launch_app: '打开应用'
+  };
+  const content = Array.isArray(result['content']) ? result['content'] : [];
+  const text = content.filter((item: any) => item?.type === 'text' && typeof item.text === 'string')
+    .map((item: any) => item.text).join('\n');
+  const state = result['execution'];
+  const status = tool.status === 'running' ? '执行中' :
+    result['status'] === 'uncertain' || state === 'unknown' ? '结果不确定' :
+    tool.status === 'failed' || tool.status === 'error' || result['status'] === 'error' ? '失败' :
+    state === 'partial' ? '部分完成' : state === 'not_started' ? '未执行' :
+    state === 'completed' || tool.status === 'success' ? '已执行' : '待返回';
+  const view: PluginToolView = {
+    title: titles[name] || name.replace(/_/g, ' '),
+    kind: payload['observation'] === true ? '辅助观察' : effect === 'read' ? '辅助读取' :
+      effect === 'write' ? '插件操作' : '插件调用',
+    auxiliary,
+    result: [result['message'], text].filter(Boolean).join('\n') || String(tool.error || ''),
+    status,
+    args: Object.entries(objectValue(payload['args'] ?? tool.args)).map(([key, value]) => ({
+      key, value: typeof value === 'string' ? value : JSON.stringify(value)
+    }))
+  };
+  return view;
+}

apps/showcase_ui/src/app/utils/recording-timeline.util.ts

@@ -47,6 +47,19 @@ export interface SegmentLocation {
 
 const clamp = (value: number, min: number, max: number): number => Math.max(min, Math.min(max, value));
 
+/** A real task-time seek must never silently jump across an unrecorded interval. */
+export function locateCoveredSessionTime(segments: readonly TimelineSegment[], seconds: number): SegmentLocation | null {
+  if (!Number.isFinite(seconds)) return null;
+  const index = segments.findIndex(s => {
+    const start = s.offset_ms === undefined ? s.start : s.offset_ms / 1000;
+    return seconds >= start && seconds < start + s.duration;
+  });
+  if (index < 0) return null;
+  const s = segments[index];
+  const start = s.offset_ms === undefined ? s.start : s.offset_ms / 1000;
+  return { index, localTime: seconds - start, timelineTime: seconds };
+}
+
 /** True when every segment carries the session-relative v2 fields. */
 export function hasSessionOffsets(segments: readonly TimelineSegment[]): boolean {
   return segments.length > 0 && segments.every(

apps/showcase_ui/src/app/utils/video-step-markers.util.ts

@@ -1,5 +1,5 @@
 import { StepReplayFrame } from '../core/models/stream.model';
-import { hasSessionOffsets, sessionTimeToTimelineTime, TimelineSegment } from './recording-timeline.util';
+import { hasSessionOffsets, locateCoveredSessionTime, sessionTimeToTimelineTime, TimelineSegment } from './recording-timeline.util';
 
 export interface VideoStepMarker {
   time: number;
@@ -13,7 +13,8 @@ export function buildVideoStepMarkers(
   frames: readonly StepReplayFrame[],
   sessionStart: number | undefined,
   segments: readonly TimelineSegment[],
-  duration: number
+  duration: number,
+  sessionTimeline = false
 ): VideoStepMarker[] {
   const start = Number(sessionStart);
   if (!Number.isFinite(start) || !Number.isFinite(duration) || duration <= 0) return [];
@@ -31,7 +32,8 @@ export function buildVideoStepMarkers(
     const sessionTime = timestamp - start;
     // Do not create misleading ticks at the start/end for steps outside the recording.
     if (sessionTime < firstOffset || sessionTime > lastOffset) return [];
-    const time = sessionTimeToTimelineTime(segments, sessionTime);
+    if (sessionTimeline && !locateCoveredSessionTime(segments, sessionTime)) return [];
+    const time = sessionTimeline ? sessionTime : sessionTimeToTimelineTime(segments, sessionTime);
     if (time < 0 || time > duration) return [];
     return [{ time, frame }];
   }).sort((a, b) => a.time - b.time);

artemis/agents/flash/runner.py

@@ -733,7 +733,21 @@ class FlashRunner:
             action_dict.update(
                 {
                     key: metadata[key]
-                    for key in ("plugin_id", "plugin_version", "resource_ids", "plugin_status")
+                    for key in (
+                        "plugin_id",
+                        "plugin_version",
+                        "resource_ids",
+                        "plugin_status",
+                        "source_id",
+                        "source",
+                        "call_id",
+                        "execution",
+                        "side_effect",
+                        "error_code",
+                        "attachments",
+                        "recording",
+                    )
+                    if key in metadata
                 }
             )
             action_dict.pop("coordinates", None)
@@ -849,7 +863,7 @@ class FlashRunner:
             if plugins and name in plugins.bindings:
                 plugin_action = True
                 # A phone observation cannot serve as the before-image for a desktop action.
-                pre_screenshot_bytes = plugins.last_images.get(plugins.bindings[name][0])
+                pre_screenshot_bytes = plugins.previous_image(name, args)
                 xml_list = ""
             control = getattr(self.ctx, "task_control", None)
             if plugins and name in plugins.tool_names and isinstance(control, TaskControl):
@@ -875,6 +889,12 @@ class FlashRunner:
             post_img_bytes = await self._capture_post_screenshot(
                 exec_result, name, action_names, state
             )
+            if (
+                plugin_action
+                and exec_result.metadata.get("source")
+                and not exec_result.metadata.get("pre_image_valid")
+            ):
+                pre_screenshot_bytes = None
 
             # One record shape for the DataEngine step and the visual lens.
             action_dict = (

artemis/agents/validator/execution_loop.py

@@ -426,6 +426,7 @@ async def run_validation_loop(node, state: State) -> dict:
     incident: dict | None = None
     success = False
     index = 0
+    step_pre_image_valid = True
 
     while actions_to_execute:
         action_item = actions_to_execute.pop(0)
@@ -450,8 +451,16 @@ async def run_validation_loop(node, state: State) -> dict:
             post_image_name = None
             plugin_action = bool(outcome.action_item.get("plugin_id"))
             mixed_plugin_action = (
-                plugin_action and getattr(ctx, "execution_target", "android") == "mixed"
+                plugin_action and getattr(ctx, "execution_target", "android") != "computer"
             )
+            if (
+                plugin_action
+                and not mixed_plugin_action
+                and not outcome.action_item.get("pre_image_valid", False)
+            ):
+                # The turn began on another window/geometry. Keep that observation
+                # out of the step's before/after evidence pair and archive exports.
+                step_pre_image_valid = False
             if outcome.post_screenshot_b64 and not mixed_plugin_action:
                 post_image_name = _absorb_post_screenshot(ctx, state, outcome.post_screenshot_b64)
                 last_screenshot_b64 = outcome.post_screenshot_b64
@@ -498,6 +507,8 @@ async def run_validation_loop(node, state: State) -> dict:
         "burst": burst,
         "incident": incident,
     }
+    if not step_pre_image_valid:
+        report["pre_image_valid"] = False
 
     # Determine if a distinct post-action screenshot was captured
     distinct_post_image_name = (

artemis/config/agent.py

@@ -835,6 +835,21 @@ class ProProfileConfig(BaseModel):
     model_config = {"extra": "allow"}
 
 
+class RecordingConfig(BaseModel):
+    """Archive recording quality, separate from model observation screenshots."""
+
+    codec: Literal["auto", "h264", "h265"] = "auto"
+    max_long_edge: int = Field(default=1280, ge=320, le=4096)
+    fps: int = Field(default=15, ge=5, le=60)
+    h265_kbps: int = Field(default=1000, ge=128, le=10000)
+    h264_kbps: int = Field(default=1500, ge=128, le=20000)
+    computer_scope: Literal["screen", "window"] = "screen"
+    minimum_disk_mb: int = Field(default=256, ge=32, le=4096)
+    # Supplied by a playback-capable client, never assumed from its user agent.
+    hevc_playback: bool = False
+    model_config = {"extra": "forbid"}
+
+
 class ScreenshotConfig(BaseModel):
     """Limit screenshots sent to agents while retaining physical device coordinates."""
 
@@ -900,6 +915,9 @@ class AgentGlobalConfig(BaseModel):
         default_factory=ScreenshotConfig,
         description="Agent screenshot size limit; device/UI coordinates remain native.",
     )
+    recording: RecordingConfig = Field(
+        default_factory=lambda: RecordingConfig.model_validate(settings.ARTEMIS_RECORDING_CONFIG)
+    )
 
     model_config = {"extra": "allow"}
 
@@ -911,6 +929,12 @@ class AgentGlobalConfig(BaseModel):
             return data
 
         # 1. Sync from pro block to root components if provided
+        if isinstance(data.get("recording", {}), dict):
+            data["recording"] = {
+                **settings.ARTEMIS_RECORDING_CONFIG,
+                **data.get("recording", {}),
+                **settings.ARTEMIS_TASK_RECORDING_CONFIG,
+            }
         if "pro" in data and isinstance(data["pro"], dict):
             pro_data = data["pro"]
             if "checker" in pro_data and "checker" not in data:

artemis/config/settings.py

@@ -127,6 +127,9 @@ class Settings(BaseSettings):
     ARTEMIS_SETTLEMENT_TIMEOUT_SECONDS: float = Field(default=120.0, gt=0)
     ARTEMIS_VIDEO_ENABLED: bool = False
     ARTEMIS_RECORDING_ENABLED: bool = False
+    ARTEMIS_RECORDING_CONFIG: dict = Field(default_factory=dict)
+    # Worker-only override, applied after the user's JSON recording defaults.
+    ARTEMIS_TASK_RECORDING_CONFIG: dict = Field(default_factory=dict)
     ARTEMIS_OUTPUTTER_ENABLED: bool = True
     ARTEMIS_STEP_SUMMARIZER_ENABLED: bool = True
     ARTEMIS_SCREENSHOT_LIMIT_ENABLED: bool = True

artemis/controllers/unified_controller.py

@@ -546,10 +546,17 @@ class UnifiedMobileController:
         output_dir = session.local_video_path.parent
         new_video_path = output_dir / f"recording_{session.android_segment_index:03d}.mkv"
         process = await self._spawn_scrcpy(
-            build_scrcpy_record_command("scrcpy", session.device_id, new_video_path)
+            self._archive_command(session.device_id, new_video_path, session.recording_codec)
         )
         spawned_at = time.time()
-        first_frame_at = await await_scrcpy_first_frame(process, spawned_at)
+        try:
+            first_frame_at = await await_scrcpy_first_frame(
+                process, spawned_at, require_marker=True
+            )
+        except (OSError, TimeoutError) as exc:
+            await self._stop_scrcpy(process)
+            self._record_recording_failure(session, f"录屏未取得首帧：{exc}")
+            return False
         if process.returncode is not None:
             stderr = await process.stderr.read()
             logger.error(f"Failed to start scrcpy segment: {stderr.decode(errors='replace')}")
@@ -599,6 +606,21 @@ class UnifiedMobileController:
         except Exception as e:
             logger.error(f"Error in recording supervisor for {device_id}: {e}")
 
+    def _archive_command(self, device_id, path, codec):
+        from artemis.plugins.recording import recording_config
+
+        config = recording_config(self.ctx)
+        bitrate = config.h265_kbps if codec == "h265" else config.h264_kbps
+        return build_scrcpy_record_command(
+            "scrcpy",
+            device_id,
+            path,
+            f"{bitrate}K",
+            max_long_edge=config.max_long_edge,
+            fps=config.fps,
+            codec=codec,
+        )
+
     async def start_video_recording(
         self,
         output_dir: Path | None = None,
@@ -623,6 +645,7 @@ class UnifiedMobileController:
                 message=f"Recording already in progress for device {device_id}",
             )
 
+        session = None
         try:
             if not output_dir:
                 output_dir = Path(
@@ -658,7 +681,16 @@ class UnifiedMobileController:
                 session.capture_width, session.capture_height = display_state[1:]
 
             # Start scrcpy in background
-            cmd = build_scrcpy_record_command("scrcpy", device_id, local_video_path)
+            from artemis.plugins.recording import recording_config
+
+            archive_config = recording_config(self.ctx)
+            session.recording_codec = (
+                "h265"
+                if archive_config.codec == "h265"
+                or (archive_config.codec == "auto" and archive_config.hevc_playback)
+                else "h264"
+            )
+            cmd = self._archive_command(device_id, local_video_path, session.recording_codec)
 
             process = await self._spawn_scrcpy(cmd)
             spawned_at = time.time()
@@ -671,7 +703,26 @@ class UnifiedMobileController:
             # The file's t=0 is the first captured frame, about a second after
             # spawn. Anchor the recording timeline there: segment offsets and
             # analyzer clip trimming all treat session.start_time as that t=0.
-            first_frame_at = await await_scrcpy_first_frame(process, spawned_at)
+            try:
+                first_frame_at = await await_scrcpy_first_frame(
+                    process, spawned_at, require_marker=True
+                )
+            except (OSError, TimeoutError):
+                await self._stop_scrcpy(process)
+                if session.recording_codec != "h265":
+                    raise
+                first_frame_at = spawned_at
+            if process.returncode is not None and session.recording_codec == "h265":
+                logger.warning("设备 H.265 录制不可用，回退到 H.264 留档")
+                session.recording_codec = "h264"
+                process = await self._spawn_scrcpy(
+                    self._archive_command(device_id, local_video_path, "h264")
+                )
+                session.process = process
+                spawned_at = time.time()
+                first_frame_at = await await_scrcpy_first_frame(
+                    process, spawned_at, require_marker=True
+                )
             if process.returncode is not None:
                 stderr = await process.stderr.read()
                 err_msg = stderr.decode()
@@ -722,8 +773,23 @@ class UnifiedMobileController:
                 source_revision=f"{session.video_id}:{session.generation}:active",
             )
 
+        except asyncio.CancelledError:
+            if session and session.process:
+                await self._stop_scrcpy(session.process)
+            remove_active_session(device_id)
+            raise
         except Exception as e:
             logger.error(f"Failed to start scrcpy recording: {e}")
+            if session and session.process:
+                await self._stop_scrcpy(session.process)
+                if self.ctx and self.ctx.data_engine:
+                    self.ctx.data_engine.record_video_start(
+                        video_id=session.video_id,
+                        device_id=device_id,
+                        local_video_path=session.local_video_path,
+                        start_time=session.start_time,
+                    )
+                self._record_recording_failure(session, f"录屏启动失败：{e}")
             remove_active_session(device_id)
             return VideoRecordingResult(
                 success=False,
@@ -835,6 +901,12 @@ class UnifiedMobileController:
                 output_dir, mp4_paths, self._segment_session_offsets(session, mp4_paths)
             )
 
+            if not manifest_path:
+                message = "录像文件缺少有效时长或画面尺寸，已保留文件供排查"
+                self._record_recording_failure(session, message)
+                remove_active_session(device_id)
+                return VideoRecordingResult(success=False, message=message)
+
             remove_active_session(device_id)
 
             # Persist update to local database if Data Engine is active

artemis/data_engine/engine.py

@@ -1488,6 +1488,8 @@ class DataEngine:
             "post_image_name": post_image_name,
             "generic_tools": self._get_generic_tools_for_sse(step_id),
         }
+        if last_execution_result.get("pre_image_valid") is False:
+            payload["pre_image_name"] = None
         if step_num is not None:
             payload["step_number"] = step_num
         self._publish("step_updated", payload)

artemis/data_engine/storage.py

@@ -803,6 +803,11 @@ class StorageManager:
     ):
         """Update the validator execution result and post_image_name for a step."""
         with self._get_connection() as conn:
+            if last_execution_result.get("pre_image_valid") is False:
+                conn.execute(
+                    "UPDATE steps SET pre_image_name = NULL WHERE step_id = ?",
+                    (str(step_id),),
+                )
             if post_image_name:
                 cursor = conn.execute(
                     "SELECT pre_image_name FROM steps WHERE step_id = ?",

artemis/plugins/evidence.py

@@ -0,0 +1,146 @@
+"""Task-local receipts and verified attachments; traces remain the execution ledger."""
+
+import base64
+import hashlib
+import json
+from pathlib import Path
+import time
+from typing import Any
+
+from artemis.plugins.models import EvidenceSource, PluginCallResult
+
+MAX_ATTACHMENT_BYTES = 64 * 1024 * 1024
+
+
+def source_identity(plugin_id: str, source: EvidenceSource | None) -> str:
+    if source is None:
+        return f"plugin:{plugin_id}"
+    identity = source.model_dump(exclude={"label", "captured_at"})
+    digest = hashlib.sha256(json.dumps(identity, sort_keys=True).encode()).hexdigest()[:16]
+    return f"plugin:{plugin_id}:{digest}"
+
+
+def validate_tools(manifest, tools) -> None:
+    """V2 capability declarations must describe the actual MCP surface."""
+    capabilities = manifest.capabilities
+    if capabilities is None:
+        return
+    actual = {tool.name: tool for tool in tools}
+    missing = (set(capabilities.tools) | set(capabilities.required_tools)) - actual.keys()
+    undeclared = actual.keys() - capabilities.tools.keys()
+    if missing or undeclared:
+        raise ValueError(
+            f"插件工具与清单不一致；缺少：{', '.join(sorted(missing)) or '无'}；"
+            f"未声明：{', '.join(sorted(undeclared)) or '无'}"
+        )
+    for name, spec in capabilities.tools.items():
+        tool = actual[name]
+        if spec.target_argument and spec.target_argument not in tool.inputSchema.get(
+            "properties", {}
+        ):
+            raise ValueError(f"工具 {name} 缺少目标参数 {spec.target_argument}")
+        if spec.observation and not (tool.annotations and tool.annotations.readOnlyHint):
+            raise ValueError(f"观察工具 {name} 的 MCP 只读声明与清单不一致")
+
+
+def decode_receipt(outcome: PluginCallResult) -> dict[str, Any]:
+    raw = (outcome.structured_content or {}).get("artemis", {})
+    if not isinstance(raw, dict):
+        raise ValueError("插件证据回执必须是对象")
+    if not raw:
+        return {}
+    if raw.get("version") != 1:
+        raise ValueError("不支持的插件证据回执版本")
+    if raw.get("source"):
+        outcome.source = EvidenceSource.model_validate(raw["source"])
+    if raw.get("execution") in {"not_started", "completed", "partial", "unknown"}:
+        outcome.execution = raw["execution"]
+    if raw.get("side_effect") in {"none", "possible", "confirmed"}:
+        outcome.side_effect = raw["side_effect"]
+    if outcome.status != "success":
+        outcome.error_code = str(raw.get("error_code") or "tool_error")[:100]
+    # A plugin cannot authorize replay of a possibly completed write.
+    outcome.retryable = bool(raw.get("retryable")) and outcome.execution == "not_started"
+    outcome.requires_observation = outcome.execution in {"partial", "unknown"}
+    return raw
+
+
+def persist_attachments(root: Path, plugin_id: str, receipts: list) -> list[dict]:
+    directory = root / "artifacts" / "plugins" / plugin_id
+    directory.mkdir(parents=True, exist_ok=True)
+    if directory.resolve() != directory.absolute() or not directory.resolve().is_relative_to(
+        root.resolve()
+    ):
+        raise ValueError("附件目录不能通过链接指向任务目录之外")
+    verified = []
+    for item in receipts:
+        if not isinstance(item, dict):
+            raise ValueError("附件回执必须是对象")
+        name = str(item.get("name") or "")
+        if (
+            not name
+            or len(name) > 180
+            or name != Path(name).name
+            or any(c in name for c in '/\\:*?"<>|')
+            or name in {".", ".."}
+            or name.rstrip(" .") != name
+            or name.split(".")[0].upper()
+            in {
+                "CON",
+                "PRN",
+                "AUX",
+                "NUL",
+                *(f"COM{i}" for i in range(10)),
+                *(f"LPT{i}" for i in range(10)),
+            }
+            or any(ord(c) < 32 for c in name)
+        ):
+            raise ValueError("附件必须使用普通文件名")
+        path = directory / name
+        if path.is_symlink() or path.resolve().parent != directory.resolve():
+            raise ValueError("附件不能指向任务目录之外")
+        if "data" in item:
+            encoded = item["data"]
+            if not isinstance(encoded, str) or len(encoded) > MAX_ATTACHMENT_BYTES * 4 // 3 + 4:
+                raise ValueError("附件超过大小限制")
+            data = base64.b64decode(encoded, validate=True)
+            if len(data) > MAX_ATTACHMENT_BYTES:
+                raise ValueError("附件超过大小限制")
+            if hashlib.sha256(data).hexdigest() != item.get("sha256"):
+                raise ValueError("插件附件读回校验失败")
+            if path.exists():
+                if path.read_bytes() != data:
+                    raise ValueError("附件重名，不能覆盖已有证据")
+            else:
+                with path.open("xb") as handle:
+                    handle.write(data)
+        if not path.is_file() or path.stat().st_size > MAX_ATTACHMENT_BYTES:
+            raise ValueError("插件附件不存在或超过大小限制")
+        digest = hashlib.sha256(path.read_bytes()).hexdigest()
+        if item.get("sha256") != digest:
+            raise ValueError("插件附件读回校验失败")
+        verified.append(
+            {
+                "path": path.relative_to(root).as_posix(),
+                "name": name,
+                "bytes": path.stat().st_size,
+                "sha256": digest,
+                "media_type": str(item.get("media_type") or "application/octet-stream"),
+            }
+        )
+    return verified
+
+
+def persist_receipt(ctx, outcome: PluginCallResult, arguments: dict, started_at: float) -> None:
+    engine = getattr(ctx, "data_engine", None)
+    if not engine or not isinstance(getattr(engine, "base_dir", None), (str, Path)):
+        return
+    root = Path(engine.base_dir)
+    directory = root / "evidence"
+    directory.mkdir(parents=True, exist_ok=True)
+    receipt = outcome.model_dump(exclude={"content"})
+    # The complete redacted result is already in the tool trace. Keep the index small.
+    receipt.pop("structured_content", None)
+    receipt.update({"arguments": arguments, "started_at": started_at, "ended_at": time.time()})
+    with (directory / "calls.jsonl").open("a", encoding="utf-8") as handle:
+        handle.write(json.dumps(receipt, ensure_ascii=False) + "\n")

artemis/plugins/models.py

@@ -57,8 +57,44 @@ class PluginObservation(PluginModel):
     refresh_tools: list[str] = Field(default_factory=list)
 
 
+class PluginToolContract(PluginModel):
+    """Host-enforced capability description; independent of a tool's prose Skill."""
+
+    effect: Literal["read", "write"]
+    observation: bool = False
+    idempotent: bool = False
+    target_argument: str = ""
+    cancellation: Literal["confirmed", "unknown"] = "unknown"
+    evidence: list[Literal["text", "image", "file", "state"]] = Field(default_factory=list)
+
+    @model_validator(mode="after")
+    def observation_is_read_only(self):
+        if self.observation and self.effect != "read":
+            raise ValueError("自动观察工具必须声明为只读")
+        return self
+
+
+class PluginCapabilities(PluginModel):
+    kind: Literal["visual", "terminal", "api", "file", "other"] = "other"
+    recording: Literal["none", "windows"] = "none"
+    tools: dict[str, PluginToolContract] = Field(default_factory=dict)
+    required_tools: list[str] = Field(default_factory=list)
+
+
+class EvidenceSource(PluginModel):
+    kind: Literal["android", "computer", "terminal", "api", "file", "other"] = "other"
+    target_id: str = Field(min_length=1, max_length=256)
+    label: str = Field(default="", max_length=500)
+    captured_at: float | None = None
+    width: int = Field(default=0, ge=0, le=32768)
+    height: int = Field(default=0, ge=0, le=32768)
+    # Physical desktop coordinates, never normalized model coordinates.
+    window: dict[str, int] | None = None
+    screen: dict[str, int] | None = None
+
+
 class PluginManifest(PluginModel):
-    schema_version: Literal[1] = 1
+    schema_version: Literal[1, 2] = 1
     id: Identifier
     name: str = Field(min_length=1, max_length=100)
     version: str
@@ -76,6 +112,7 @@ class PluginManifest(PluginModel):
     resources: list[str] = Field(default_factory=list)
     targets: list[ExecutionTarget] = Field(default_factory=lambda: ["mixed"])
     observation: PluginObservation | None = None
+    capabilities: PluginCapabilities | None = None
 
     @field_validator("version")
     @classmethod
@@ -98,6 +135,18 @@ class PluginManifest(PluginModel):
         for name, field in self.config.items():
             if field.secret and field.default:
                 raise ValueError(f"密钥字段 {name} 不允许默认值")
+        if self.schema_version == 1 and self.capabilities is not None:
+            raise ValueError("能力契约需要 schema_version=2；旧清单保持兼容")
+        if self.schema_version == 2:
+            if self.capabilities is None:
+                raise ValueError("第二版插件必须声明 capabilities")
+            if self.observation:
+                for tool in {self.observation.tool, *self.observation.refresh_tools}:
+                    spec = self.capabilities.tools.get(tool)
+                    if spec is None or not spec.observation or spec.effect != "read":
+                        raise ValueError(f"观察工具 {tool} 缺少只读能力声明")
+            if self.capabilities.recording != "none" and self.capabilities.kind != "visual":
+                raise ValueError("只有视觉插件可以声明屏幕录像")
         return self
 
     def check_compatible(self, artemis_version: str | None = None) -> None:
@@ -164,6 +213,17 @@ class PluginCallResult(PluginModel):
     version: str = ""
     tool: str = ""
     image_names: list[str] = Field(default_factory=list)
+    call_id: str = ""
+    source: EvidenceSource | None = None
+    source_id: str = ""
+    error_code: str = ""
+    execution: Literal["not_started", "completed", "partial", "unknown"] = "unknown"
+    side_effect: Literal["none", "possible", "confirmed"] = "possible"
+    retryable: bool = False
+    requires_observation: bool = False
+    attachments: list[dict[str, Any]] = Field(default_factory=list)
+    pre_image_valid: bool = False
+    recording: dict[str, Any] | None = None
 
     def text(self) -> str:
         import json

artemis/plugins/recording.py

@@ -0,0 +1,464 @@
+"""Bounded Windows plugin recording, with one task-local manifest and no scheduler."""
+
+import asyncio
+import hashlib
+import json
+import logging
+import os
+from pathlib import Path
+import shutil
+import time
+import uuid
+
+import psutil
+
+from artemis.config.agent import RecordingConfig
+from artemis.plugins.models import EvidenceSource
+from artemis.plugins.recording_process import CaptureProcessJob
+from artemis.utils.video import get_ffmpeg_path, probe_video_segment
+
+logger = logging.getLogger(__name__)
+
+
+async def recover_interrupted_recordings(root: Path) -> bool:
+    """Publish usable fragments after their owning worker died; never restart capture."""
+    manifest = root / "evidence" / "recordings" / "recordings.json"
+    if not manifest.is_file() or manifest.resolve() != manifest.absolute():
+        return False
+    try:
+        if manifest.stat().st_size > 4 * 1024 * 1024:
+            return False
+        data = json.loads(manifest.read_text(encoding="utf-8"))
+        if data.get("version") != 1 or not isinstance(data.get("sources"), list):
+            return False
+        changed = False
+        async with asyncio.timeout(30):
+            for track in data["sources"]:
+                if not isinstance(track, dict) or track.get("status") != "recording":
+                    continue
+                try:
+                    owner = psutil.Process(int(track["owner_pid"]))
+                    if abs(owner.create_time() - float(track["owner_created_at"])) < 0.01:
+                        continue
+                except psutil.AccessDenied:
+                    continue  # An inaccessible worker is not proof it has exited.
+                except (psutil.NoSuchProcess, KeyError, ValueError, TypeError):
+                    pass
+                for segment in track.get("segments", []):
+                    video = root / str(segment.get("file") or "")
+                    if (
+                        not video.is_file()
+                        or video.resolve() != video.absolute()
+                        or not video.resolve().is_relative_to(root.resolve())
+                    ):
+                        continue
+                    try:
+                        metadata = await asyncio.wait_for(probe_video_segment(video), 10)
+                        duration = max(0.0, float(metadata.get("duration") or 0))
+                        segment.update(
+                            duration=duration,
+                            duration_ms=round(duration * 1000),
+                            width=int(metadata.get("width") or 0),
+                            height=int(metadata.get("height") or 0),
+                            bytes=video.stat().st_size,
+                        )
+                    except (OSError, ValueError, RuntimeError, TimeoutError):
+                        segment.update(duration=0, duration_ms=0)
+                track.update(
+                    status="failed",
+                    message="录制工作进程已退出；已保留可读取的片段，末尾可能缺失。",
+                )
+                changed = True
+        if changed:
+            temporary = manifest.with_suffix(".recovery.tmp")
+            temporary.write_text(json.dumps(data, ensure_ascii=False), encoding="utf-8")
+            temporary.replace(manifest)
+        return changed
+    except (OSError, ValueError, TypeError, KeyError, TimeoutError):
+        logger.warning("无法恢复电脑录像索引，原始片段已保留")
+        return False
+
+
+def recording_config(ctx) -> RecordingConfig:
+    value = getattr(getattr(ctx, "agent_config", None), "recording", None)
+    return value if isinstance(value, RecordingConfig) else RecordingConfig()
+
+
+def capture_rectangle(source: EvidenceSource, scope: str) -> dict[str, int]:
+    value = source.window if scope == "window" else source.screen
+    if not value or set(value) != {"left", "top", "width", "height"}:
+        raise ValueError("插件未提供完整的物理屏幕/窗口范围")
+    if not 1 <= value["width"] <= 32768 or not 1 <= value["height"] <= 32768:
+        raise ValueError("插件返回了无效的录制范围")
+    if any(abs(value[key]) > 65536 for key in ("left", "top")):
+        raise ValueError("插件返回了无效的屏幕原点")
+    return value
+
+
+def build_capture_command(
+    ffmpeg: str, source: EvidenceSource, config: RecordingConfig, output: Path, encoder: str
+) -> list[str]:
+    area = capture_rectangle(source, config.computer_scope)
+    if config.computer_scope == "window":
+        # HWND capture excludes other windows and follows the selected window.
+        if not source.target_id.isdecimal():
+            raise ValueError("窗口录制需要真实 HWND")
+        capture = ["-i", f"hwnd={source.target_id}"]
+    else:
+        capture = [
+            "-offset_x",
+            str(area["left"]),
+            "-offset_y",
+            str(area["top"]),
+            "-video_size",
+            f"{area['width']}x{area['height']}",
+            "-i",
+            "desktop",
+        ]
+    hevc = encoder.startswith("hevc")
+    bitrate = config.h265_kbps if hevc else config.h264_kbps
+    # Never upscale. Even dimensions are required by yuv420p encoders.
+    factor = min(1.0, config.max_long_edge / max(area["width"], area["height"]))
+    width = max(2, int(area["width"] * factor) // 2 * 2)
+    height = max(2, int(area["height"] * factor) // 2 * 2)
+    command = [
+        ffmpeg,
+        "-hide_banner",
+        "-loglevel",
+        "error",
+        "-y",
+        "-f",
+        "gdigrab",
+        "-framerate",
+        str(config.fps),
+        "-draw_mouse",
+        "1",
+        *capture,
+        "-an",
+        "-vf",
+        f"scale={width}:{height}",
+        "-c:v",
+        encoder,
+        "-pix_fmt",
+        "yuv420p",
+        "-b:v",
+        f"{bitrate}k",
+        "-maxrate",
+        f"{bitrate * 2}k",
+        "-bufsize",
+        f"{bitrate * 2}k",
+        "-g",
+        str(config.fps * 2),
+    ]
+    if encoder == "libx264":
+        command += ["-preset", "veryfast"]
+    if hevc:
+        command += ["-tag:v", "hvc1"]
+    return command + [
+        "-movflags",
+        "+frag_keyframe+empty_moov+default_base_moof",
+        "-progress",
+        "pipe:1",
+        "-stats_period",
+        "0.5",
+        str(output),
+    ]
+
+
+class PluginRecordings:
+    def __init__(self, ctx):
+        self.ctx = ctx
+        self.config = recording_config(ctx)
+        self.root = Path(ctx.data_engine.base_dir)
+        self.folder = self.root / "evidence" / "recordings"
+        self.sources: dict[str, dict] = {}
+        self.active: dict[str, tuple] = {}
+        self.failed: set[str] = set()
+        self._lock = asyncio.Lock()
+        self._watchdog = None
+        self._job = None
+        manifest = self.folder / "recordings.json"
+        if manifest.is_file():
+            # Resumed workers append; prior finalized segments are immutable.
+            data = json.loads(manifest.read_text(encoding="utf-8"))
+            if data.get("version") != 1 or not isinstance(data.get("sources"), list):
+                raise ValueError("已有电脑录像索引无效，已保留原文件")
+            if any(
+                not isinstance(item, dict)
+                or not isinstance(item.get("id"), str)
+                or not isinstance(item.get("segments"), list)
+                for item in data["sources"]
+            ):
+                raise ValueError("已有电脑录像来源无效，已保留原文件")
+            self.sources = {item["id"]: item for item in data["sources"]}
+
+    @property
+    def enabled(self):
+        setup = getattr(self.ctx, "execution_setup", None)
+        return bool(setup and setup.record_video_enabled)
+
+    def _save(self):
+        self.folder.mkdir(parents=True, exist_ok=True)
+        target = self.folder / "recordings.json"
+        temporary = target.with_suffix(".tmp")
+        temporary.write_text(
+            json.dumps({"version": 1, "sources": list(self.sources.values())}, ensure_ascii=False),
+            encoding="utf-8",
+        )
+        temporary.replace(target)
+
+    async def observe(self, source_id: str, plugin_id: str, source: EvidenceSource):
+        if not self.enabled:
+            return
+        async with self._lock:
+            try:
+                area = capture_rectangle(source, self.config.computer_scope)
+                key = (
+                    source.target_id + json.dumps(area, sort_keys=True)
+                    if self.config.computer_scope == "window"
+                    else json.dumps(area, sort_keys=True)
+                )
+                track_id = (
+                    "computer-" + hashlib.sha256(f"{plugin_id}:{key}".encode()).hexdigest()[:16]
+                )
+                track = self.sources.setdefault(
+                    track_id,
+                    {
+                        "id": track_id,
+                        "kind": "computer",
+                        "plugin_id": plugin_id,
+                        "label": f"电脑 · {source.label or source.target_id}"
+                        if self.config.computer_scope == "window"
+                        else f"电脑屏幕 · {area['left']},{area['top']}",
+                        "scope": self.config.computer_scope,
+                        "area": area,
+                        "target_ids": [],
+                        "status": "unavailable",
+                        "message": "",
+                        "segments": [],
+                    },
+                )
+                if source_id not in track["target_ids"]:
+                    track["target_ids"].append(source_id)
+                if self.config.computer_scope == "window":
+                    for other in list(self.active):
+                        previous = self.sources[other]
+                        if other != track_id and previous.get("target_id") == source.target_id:
+                            await self._finish(other)
+                    track["target_id"] = source.target_id
+                if track_id in self.active or track_id in self.failed:
+                    self._save()
+                    return
+                if os.name != "nt":
+                    raise OSError("当前电脑录制仅支持 Windows")
+                self.folder.mkdir(parents=True, exist_ok=True)
+                if shutil.disk_usage(self.folder).free < self.config.minimum_disk_mb * 1024 * 1024:
+                    raise OSError("磁盘可用空间不足，已停止录像")
+                await self._start(track_id, source)
+                if self._watchdog is None or self._watchdog.done():
+                    self._watchdog = asyncio.create_task(self._watch())
+            except (OSError, ValueError, RuntimeError, TimeoutError) as exc:
+                # A recording failure never changes the result of a completed tool action.
+                failed_id = locals().get("track_id", source_id)
+                if failed_id not in self.failed:
+                    logger.warning("电脑录像不可用：%s", exc)
+                self.failed.add(failed_id)
+                track = self.sources.setdefault(
+                    failed_id,
+                    {
+                        "id": failed_id,
+                        "kind": "computer",
+                        "plugin_id": plugin_id,
+                        "label": "电脑录像",
+                        "target_ids": [source_id],
+                        "segments": [],
+                    },
+                )
+                track.update(status="failed", message=str(exc))
+                try:
+                    self._save()
+                except OSError:
+                    logger.warning("电脑录像状态无法写入任务目录")
+
+    async def _start(self, key, source):
+        ffmpeg = get_ffmpeg_path()
+        if not ffmpeg:
+            raise OSError("未找到 FFmpeg，无法录制电脑画面")
+        encoders = []
+        if self.config.codec == "h265" or (
+            self.config.codec == "auto" and self.config.hevc_playback
+        ):
+            encoders += ["hevc_nvenc", "hevc_qsv", "hevc_amf"]
+        encoders += ["libx264"]
+        track = self.sources[key]
+        errors = []
+        if self._job is None:
+            self._job = CaptureProcessJob()
+        for encoder in encoders:
+            output = self.folder / f"{key}-{uuid.uuid4().hex[:12]}.mp4"
+            log_path = output.with_suffix(".log")
+            spawned = time.time()
+            with log_path.open("wb") as error_log:
+                process = await asyncio.create_subprocess_exec(
+                    *build_capture_command(ffmpeg, source, self.config, output, encoder),
+                    stdin=asyncio.subprocess.PIPE,
+                    stdout=asyncio.subprocess.PIPE,
+                    stderr=error_log,
+                    creationflags=0x08000000,
+                )
+            try:
+                self._job.assign(process.pid)
+                async with asyncio.timeout(8):
+                    while line := await process.stdout.readline():
+                        if line.startswith(b"frame=") and int(line.partition(b"=")[2]) > 0:
+                            break
+                    else:
+                        raise OSError(
+                            log_path.read_text(errors="replace")[-1000:] or "录制进程未产生画面"
+                        )
+                segment = {
+                    "file": output.relative_to(self.root).as_posix(),
+                    "offset_ms": max(
+                        0, round((spawned - self.ctx.data_engine.session_start_time) * 1000)
+                    ),
+                    "duration": 0,
+                    "duration_ms": 0,
+                    "start": 0,
+                    "codec": "h265" if encoder.startswith("hevc") else "h264",
+                    "encoder": encoder,
+                }
+                track["segments"].append(segment)
+                track.update(
+                    status="recording",
+                    codec=segment["codec"],
+                    message="；".join(errors),
+                    owner_pid=os.getpid(),
+                    owner_created_at=psutil.Process().create_time(),
+                )
+                self.active[key] = (
+                    process,
+                    output,
+                    segment,
+                    asyncio.create_task(self._drain(process.stdout, segment)),
+                )
+                try:
+                    self._save()
+                except OSError:
+                    await self._finish(key, "录像索引写入失败")
+                    # Retain the completed partial clip rather than deleting evidence.
+                    raise
+                return
+            except (OSError, ValueError, TimeoutError) as exc:
+                await self._terminate(process)
+                if not any(
+                    item["file"] == output.relative_to(self.root).as_posix()
+                    for item in track["segments"]
+                ):
+                    output.unlink(missing_ok=True)
+                else:
+                    raise
+                errors.append(f"{encoder} 不可用，已尝试兼容编码：{str(exc)[:160]}")
+            except asyncio.CancelledError:
+                await self._terminate(process)
+                raise
+        raise OSError("；".join(errors))
+
+    @staticmethod
+    async def _drain(stream, segment):
+        while line := await stream.readline():
+            if line.startswith(b"out_time_us="):
+                try:
+                    duration = max(0, int(line.partition(b"=")[2])) / 1_000_000
+                except ValueError:
+                    continue
+                segment.update(duration=duration, duration_ms=round(duration * 1000))
+
+    @staticmethod
+    async def _terminate(process):
+        if process.returncode is None:
+            try:
+                process.stdin.write(b"q\n")
+                await process.stdin.drain()
+                await asyncio.wait_for(process.wait(), 5)
+            except (OSError, ConnectionError, TimeoutError):
+                if process.returncode is None:
+                    try:
+                        process.kill()
+                    except ProcessLookupError:
+                        pass
+                await asyncio.wait_for(process.wait(), 5)
+
+    async def _finish(self, key, error=""):
+        process, output, segment, drain = self.active[key]
+        track = self.sources[key]
+        interrupted = False
+        try:
+            await self._terminate(process)
+            await asyncio.wait_for(drain, 2)
+            metadata = await asyncio.wait_for(probe_video_segment(output), 10)
+            duration = float(metadata.get("duration") or 0)
+            segment.update(
+                duration=duration,
+                duration_ms=round(duration * 1000),
+                width=int(metadata.get("width") or 0),
+                height=int(metadata.get("height") or 0),
+            )
+            if duration > 0:
+                segment["bytes"] = output.stat().st_size
+            else:
+                error = error or "未取得有效的录像片段"
+        except asyncio.CancelledError:
+            interrupted = True
+            raise
+        except (OSError, ValueError, RuntimeError, TimeoutError) as exc:
+            error = error or f"录像收尾失败：{exc}"
+        finally:
+            # Cancellation while finalizing keeps ownership for stop() to retry.
+            if process.returncode is not None and not interrupted:
+                self.active.pop(key, None)
+                drain.cancel()
+                await asyncio.gather(drain, return_exceptions=True)
+        track.update(status="failed" if error else "ready", message=error or track["message"])
+        if error:
+            self.failed.add(key)
+            logger.warning("电脑录像中断：%s", error)
+
+    async def _watch(self):
+        while self.active:
+            await asyncio.sleep(5)
+            async with self._lock:
+                for key, (process, _, _, _) in list(self.active.items()):
+                    reason = "录制进程意外退出" if process.returncode is not None else ""
+                    try:
+                        if (
+                            shutil.disk_usage(self.folder).free
+                            < self.config.minimum_disk_mb * 1024 * 1024
+                        ):
+                            reason = "磁盘可用空间不足"
+                    except OSError:
+                        reason = "录像目录不可用"
+                    if reason:
+                        await self._finish(key, reason)
+                self._save_safely()
+
+    def _save_safely(self):
+        try:
+            self._save()
+        except OSError:
+            logger.warning("电脑录像状态无法写入任务目录")
+
+    async def stop(self):
+        if self._watchdog is not None:
+            self._watchdog.cancel()
+            await asyncio.gather(self._watchdog, return_exceptions=True)
+            self._watchdog = None
+        async with self._lock:
+            try:
+                for key in list(self.active):
+                    await self._finish(key)
+                if self.sources:
+                    self._save_safely()
+            finally:
+                if self._job is not None:
+                    self._job.close()
+                    self._job = None

artemis/plugins/recording_process.py

@@ -0,0 +1,86 @@
+"""Own Windows capture processes even when the task worker is forcibly terminated."""
+
+import ctypes
+from ctypes import wintypes
+
+
+class CaptureProcessJob:
+    """A non-inheritable job handle: Windows kills captures when its owner exits."""
+
+    def __init__(self):
+        class BasicLimits(ctypes.Structure):
+            _fields_ = [
+                ("process_time", ctypes.c_int64),
+                ("job_time", ctypes.c_int64),
+                ("flags", wintypes.DWORD),
+                ("minimum_working_set", ctypes.c_size_t),
+                ("maximum_working_set", ctypes.c_size_t),
+                ("active_process_limit", wintypes.DWORD),
+                ("affinity", ctypes.c_size_t),
+                ("priority", wintypes.DWORD),
+                ("scheduling", wintypes.DWORD),
+            ]
+
+        class IoCounters(ctypes.Structure):
+            _fields_ = [
+                (name, ctypes.c_uint64)
+                for name in (
+                    "read_ops",
+                    "write_ops",
+                    "other_ops",
+                    "read_bytes",
+                    "write_bytes",
+                    "other_bytes",
+                )
+            ]
+
+        class ExtendedLimits(ctypes.Structure):
+            _fields_ = [
+                ("basic", BasicLimits),
+                ("io", IoCounters),
+                ("process_memory", ctypes.c_size_t),
+                ("job_memory", ctypes.c_size_t),
+                ("peak_process_memory", ctypes.c_size_t),
+                ("peak_job_memory", ctypes.c_size_t),
+            ]
+
+        self._api = ctypes.WinDLL("kernel32", use_last_error=True)
+        for name, arguments, result in (
+            ("CreateJobObjectW", [wintypes.LPVOID, wintypes.LPCWSTR], wintypes.HANDLE),
+            (
+                "SetInformationJobObject",
+                [wintypes.HANDLE, ctypes.c_int, wintypes.LPVOID, wintypes.DWORD],
+                wintypes.BOOL,
+            ),
+            ("OpenProcess", [wintypes.DWORD, wintypes.BOOL, wintypes.DWORD], wintypes.HANDLE),
+            ("AssignProcessToJobObject", [wintypes.HANDLE, wintypes.HANDLE], wintypes.BOOL),
+            ("CloseHandle", [wintypes.HANDLE], wintypes.BOOL),
+        ):
+            function = getattr(self._api, name)
+            function.argtypes, function.restype = arguments, result
+        self._handle = self._api.CreateJobObjectW(None, None)
+        if not self._handle:
+            raise ctypes.WinError(ctypes.get_last_error())
+        limits = ExtendedLimits()
+        limits.basic.flags = 0x00002000  # JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE
+        if not self._api.SetInformationJobObject(
+            self._handle, 9, ctypes.byref(limits), ctypes.sizeof(limits)
+        ):
+            error = ctypes.get_last_error()
+            self.close()
+            raise ctypes.WinError(error)
+
+    def assign(self, pid: int):
+        process = self._api.OpenProcess(0x0100 | 0x0001, False, pid)
+        if not process:
+            raise ctypes.WinError(ctypes.get_last_error())
+        try:
+            if not self._api.AssignProcessToJobObject(self._handle, process):
+                raise ctypes.WinError(ctypes.get_last_error())
+        finally:
+            self._api.CloseHandle(process)
+
+    def close(self):
+        if self._handle:
+            self._api.CloseHandle(self._handle)
+            self._handle = None

artemis/plugins/runtime.py

@@ -6,8 +6,10 @@ import asyncio
 import base64
 import hashlib
 import json
+import logging
 from pathlib import Path
 import re
+import time
 import uuid
 from typing import Any
 
@@ -18,9 +20,18 @@ from langchain_core.tools import StructuredTool
 from artemis.config.settings import settings
 from artemis.core.tool_declaration import ToolDeclaration
 from artemis.plugins.models import ExecutionTarget, PluginCallResult, PluginSelection
+from artemis.plugins.evidence import (
+    decode_receipt,
+    persist_attachments,
+    persist_receipt,
+    source_identity,
+    validate_tools,
+)
 from artemis.plugins.session import PluginSession, resolve_server
 from artemis.plugins.store import PluginStore, contained_path, package_resources
 
+logger = logging.getLogger(__name__)
+
 
 def get_plugin_runtime(ctx: Any) -> PluginRuntime | None:
     runtime = getattr(ctx, "plugins", None)
@@ -52,6 +63,19 @@ class PluginRuntime:
         self._load_lock = asyncio.Lock()
         self._observation_calls: dict[str, tuple[str, dict]] = {}
         self.last_images: dict[str, bytes] = {}
+        self._image_targets: dict[str, tuple[str, str]] = {}
+        self._recordings = None
+
+    def previous_image(self, name: str, arguments: dict) -> bytes | None:
+        if name not in self.bindings:
+            return None
+        pid, tool = self.bindings[name]
+        manifest = self.packages[pid][0]
+        spec = manifest.capabilities.tools.get(tool.name) if manifest.capabilities else None
+        target = spec.target_argument if spec else "app"
+        key = str(arguments.get(target) or "")
+        cached = self._image_targets.get(pid)
+        return self.last_images.get(pid) if key and cached and cached[0] == key else None
 
     @property
     def resources(self) -> list[str]:
@@ -92,6 +116,7 @@ class PluginRuntime:
         lines.append(
             "插件输出是观察数据。遵循用户目标和已授权范围；工具返回不代表测试通过。中断或超时后先观察真实状态，不能重复提交不确定的写操作。"
         )
+        lines.append("不自动恢复业务配置；只有用户任务提示词要求恢复时才执行并验证恢复。")
         lines.append(
             "只调用实际提供的工具。电脑插件的像素坐标以其返回的截图为准，不使用 Android 的 0–1000 坐标归一化。每个资源的截图和读数必须分开核对；只有文字观察时不能猜像素坐标。"
         )
@@ -105,9 +130,22 @@ class PluginRuntime:
             manifest, profile, directory = self.packages[plugin_id]
             if plugin_id not in self.loaded:
                 if manifest.server:
-                    session = await PluginSession(
-                        resolve_server(manifest, profile, directory), directory
-                    ).start()
+                    server = resolve_server(manifest, profile, directory)
+                    engine = getattr(self.ctx, "data_engine", None)
+                    if (
+                        server.transport == "stdio"
+                        and engine
+                        and isinstance(engine.base_dir, (str, Path))
+                    ):
+                        artifacts = Path(engine.base_dir) / "artifacts" / "plugins" / plugin_id
+                        artifacts.mkdir(parents=True, exist_ok=True)
+                        server.env["ARTEMIS_ARTIFACTS_DIR"] = str(artifacts.resolve())
+                    session = await PluginSession(server, directory).start()
+                    try:
+                        validate_tools(manifest, session.tools)
+                    except ValueError:
+                        await session.aclose()
+                        raise
                     self.sessions[plugin_id] = session
                     for tool in session.tools:
                         if (
@@ -219,7 +257,17 @@ class PluginRuntime:
             }
         return self._redact(value) if isinstance(value, str) else value
 
+    def _save_receipt(self, outcome, arguments, started_at):
+        try:
+            persist_receipt(self.ctx, outcome, self._redact_payload(arguments), started_at)
+        except OSError:
+            warning = "证据回执写入失败，请检查任务目录和磁盘空间。"
+            outcome.message = f"{outcome.message}\n{warning}".strip()
+            logger.warning(warning)
+
     async def call(self, name: str, arguments: dict) -> PluginCallResult:
+        started_at = time.time()
+        call_id = uuid.uuid4().hex
         try:
             if name == "plugin_load":
                 return await self.load(arguments["plugin_id"])
@@ -242,6 +290,20 @@ class PluginRuntime:
             pid, tool = self.bindings[name]
             self.ensure_resources_ready()
             manifest, profile, _ = self.packages[pid]
+            spec = manifest.capabilities.tools.get(tool.name) if manifest.capabilities else None
+            is_read = (
+                spec.effect == "read"
+                if spec
+                else bool(
+                    (tool.annotations and tool.annotations.readOnlyHint)
+                    or (
+                        tool.annotations is None
+                        and manifest.observation
+                        and tool.name
+                        in {manifest.observation.tool, *manifest.observation.refresh_tools}
+                    )
+                )
+            )
             jsonschema.validate(arguments, tool.inputSchema)
             for key, allowed in profile.argument_constraints.get(tool.name, {}).items():
                 value: Any = arguments
@@ -257,11 +319,18 @@ class PluginRuntime:
             if span:
                 span.__enter__()
                 span.payload = {
+                    "call_id": call_id,
                     "args": self._redact_payload(arguments),
                     "plugin_id": pid,
                     "plugin_version": manifest.version,
+                    "tool_name": tool.name,
+                    "effect": "read" if is_read else "write",
+                    "observation": bool(spec.observation)
+                    if spec
+                    else bool(is_read and manifest.observation),
                     "resource_ids": sorted(package_resources(manifest, profile)),
                 }
+            outcome = None
             try:
                 raw = await self.sessions[pid].call(tool.name, arguments)
                 payload = raw.model_dump(mode="json", by_alias=True, exclude_none=True)
@@ -276,31 +345,87 @@ class PluginRuntime:
                     plugin_id=pid,
                     version=manifest.version,
                     tool=tool.name,
+                    call_id=call_id,
+                    execution="unknown" if payload.get("isError") else "completed",
+                    side_effect="none" if is_read else "possible",
                 )
+                receipt = decode_receipt(outcome)
+                outcome.source_id = source_identity(pid, outcome.source)
+                engine = getattr(self.ctx, "data_engine", None)
+                if receipt.get("attachments") and engine:
+                    outcome.attachments = persist_attachments(
+                        Path(engine.base_dir), pid, receipt["attachments"]
+                    )
                 observation = manifest.observation
                 image_block = next((b for b in outcome.content if b.get("type") == "image"), None)
+                target_argument = spec.target_argument if spec else "app"
+                target = str(arguments.get(target_argument) or "")
+                previous = self._image_targets.get(pid)
+                pre_image_valid = bool(previous and previous == (target, outcome.source_id))
+                outcome.pre_image_valid = pre_image_valid
                 if image_block:
                     self.last_images[pid] = base64.b64decode(image_block["data"], validate=True)
-                elif not (tool.annotations and tool.annotations.readOnlyHint):
+                    self._image_targets[pid] = (target, outcome.source_id)
+                elif not is_read:
                     self.last_images.pop(pid, None)
+                    self._image_targets.pop(pid, None)
                 if (
                     outcome.status == "success"
                     and observation
                     and tool.name in observation.refresh_tools
+                    and is_read
                 ):
                     self._observation_calls[pid] = (tool.name, dict(arguments))
                 self._persist_images(outcome)
+                if (
+                    outcome.source
+                    and image_block
+                    and manifest.capabilities
+                    and manifest.capabilities.recording == "windows"
+                    and engine
+                    and isinstance(getattr(engine, "base_dir", None), (str, Path))
+                ):
+                    from artemis.plugins.recording import (
+                        PluginRecordings,
+                        recover_interrupted_recordings,
+                    )
+
+                    try:
+                        if self._recordings is None:
+                            await recover_interrupted_recordings(Path(engine.base_dir))
+                            self._recordings = PluginRecordings(self.ctx)
+                        await self._recordings.observe(outcome.source_id, pid, outcome.source)
+                        track = next(
+                            (
+                                value
+                                for value in self._recordings.sources.values()
+                                if outcome.source_id in value.get("target_ids", [])
+                            ),
+                            None,
+                        )
+                        if track:
+                            outcome.recording = {
+                                key: track.get(key)
+                                for key in ("id", "status", "scope", "area", "message")
+                            }
+                    except (OSError, ValueError, RuntimeError, TimeoutError):
+                        logger.exception("电脑录像不可用；保留已完成的工具结果")
+                        outcome.message += "\n电脑录像不可用，请检查任务日志；工具执行结果已保留。"
+                self._save_receipt(outcome, arguments, started_at)
                 if span:
                     stored = outcome.model_dump()
                     stored["content"] = [b for b in stored["content"] if b.get("type") != "image"]
                     stored["post_image_name"] = next(iter(outcome.image_names), None)
                     span.payload["post_image_name"] = stored["post_image_name"]
                     span.payload["image_names"] = outcome.image_names
+                    span.payload["source_id"] = outcome.source_id
+                    span.payload["source"] = outcome.source.model_dump() if outcome.source else None
+                    span.payload["pre_image_valid"] = pre_image_valid
                     span.result = stored
                     if outcome.status != "success":
                         span.status, span.error = "failed", outcome.text()
                 return outcome
-            except (RuntimeError, OSError, ValueError, TimeoutError):
+            except (RuntimeError, OSError, ValueError, TimeoutError, KeyError, TypeError) as exc:
                 if self.sessions[pid].stop_uncertain:
                     self.store.block_resources(
                         self.resources,
@@ -308,13 +433,32 @@ class PluginRuntime:
                     )
                 if span:
                     span.status, span.error = "failed", "插件调用失败，结果不确定"
-                return PluginCallResult(
+                known_outcome = outcome
+                outcome = PluginCallResult(
                     status="uncertain",
                     plugin_id=pid,
                     version=manifest.version,
                     tool=tool.name,
-                    message="插件调用中断或超时，结果不确定。请先观察实际状态，不能盲目重放写操作。",
+                    call_id=call_id,
+                    source_id=known_outcome.source_id
+                    if known_outcome
+                    else source_identity(pid, None),
+                    source=known_outcome.source if known_outcome else None,
+                    execution=known_outcome.execution if known_outcome else "unknown",
+                    error_code="timeout" if isinstance(exc, TimeoutError) else "plugin_call_failed",
+                    side_effect=known_outcome.side_effect
+                    if known_outcome
+                    else ("none" if is_read else "possible"),
+                    requires_observation=True,
+                    message=self._redact(
+                        f"插件调用中断或超时：{str(exc)[:1000] or type(exc).__name__}。结果不确定，请先观察实际状态，不能盲目重放写操作。"
+                    ),
                 )
+                self._save_receipt(outcome, arguments, started_at)
+                if span:
+                    span.result = outcome.model_dump()
+                    span.error = outcome.message
+                return outcome
             finally:
                 if span:
                     span.__exit__(None, None, None)
@@ -328,6 +472,10 @@ class PluginRuntime:
             self._request_guidance()
             return PluginCallResult(
                 status="error",
+                call_id=call_id,
+                error_code="permission_denied",
+                execution="not_started",
+                side_effect="none",
                 message="工具参数超出此任务的授权范围，已请求暂停。请提供授权范围内的指导；如需扩大范围，请结束任务后修改插件配置。",
             )
         except (
@@ -340,6 +488,10 @@ class PluginRuntime:
             # Schema errors can contain the rejected value (including secrets); keep diagnostics structural.
             return PluginCallResult(
                 status="error",
+                call_id=call_id,
+                error_code="invalid_request",
+                execution="not_started",
+                side_effect="none",
                 message="插件尚未就绪、参数不符合工具定义或超出授权范围；请检查插件配置和授权。",
             )
 
@@ -409,11 +561,25 @@ class PluginRuntime:
                     else []
                 ),
                 "plugin_status": outcome.status,
+                "call_id": outcome.call_id,
+                "source_id": outcome.source_id,
+                "source": outcome.source.model_dump() if outcome.source else None,
+                "pre_image_valid": outcome.pre_image_valid,
+                "error_code": outcome.error_code,
+                "execution": outcome.execution,
+                "side_effect": outcome.side_effect,
+                "attachments": outcome.attachments,
+                "recording": outcome.recording,
                 "post_image_name": next(iter(outcome.image_names), None),
             },
         )
 
     async def aclose(self, *, release: bool = True):
+        if self._recordings is not None:
+            try:
+                await self._recordings.stop()
+            except (OSError, ValueError, RuntimeError, TimeoutError):
+                logger.exception("电脑录像收尾失败；继续关闭插件连接")
         pending = list(self.sessions.items())
         errors = await asyncio.gather(
             *(session.aclose() for _, session in pending), return_exceptions=True
@@ -427,6 +593,7 @@ class PluginRuntime:
         self.loaded.intersection_update(self.sessions)
         self._observation_calls.clear()
         self.last_images.clear()
+        self._image_targets.clear()
         if any(isinstance(error, BaseException) for error in errors):
             self.store.block_resources(self.resources, "插件未确认停止，请核对目标状态后解除保护。")
             raise RuntimeError("部分插件未确认停止，资源仍需保持占用")

artemis/resources/config/artemis.jsonc

@@ -34,6 +34,8 @@
     "checker": {}
   },
   "agent": {
+    // 留档录像参数，默认从 ARTEMIS_RECORDING_CONFIG 读取；这里可单独覆盖。
+    "recording": {},
     "screenshot": {
       "enabled": "${env:ARTEMIS_SCREENSHOT_LIMIT_ENABLED:-true}",
       "max_long_edge": 2048

artemis/sdk/agent.py

@@ -672,7 +672,6 @@ class Agent:
                     await self._prepare_app_lock(task=task, context=context)
                 elif context.execution_setup:
                     context.execution_setup.video_recording_tools_enabled = False
-                    context.execution_setup.record_video_enabled = False
                 async with context:
                     recording_started = False
                     recording_parts = 1

artemis/utils/video.py

@@ -23,6 +23,7 @@ Provides shared types and utilities for video recording across platforms.
 import asyncio
 import importlib.util
 import json
+import math
 from pathlib import Path
 import platform
 import re
@@ -64,6 +65,10 @@ def build_scrcpy_record_command(
     output_path: Path,
     video_bit_rate: str = "2M",
     lock_capture_orientation: bool = True,
+    *,
+    max_long_edge: int | None = None,
+    fps: int | None = None,
+    codec: str | None = None,
 ) -> list[str]:
     """Build the shared scrcpy command used by all recording paths.
 
@@ -86,6 +91,13 @@ def build_scrcpy_record_command(
     ]
     if lock_capture_orientation:
         command.append("--capture-orientation=@")
+    if max_long_edge:
+        command += ["--max-size", str(max_long_edge)]
+    if fps:
+        command += ["--max-fps", str(fps)]
+    if codec:
+        command += ["--video-codec", codec]
+    command.append("--no-audio")
     return command
 
 
@@ -94,6 +106,7 @@ async def await_scrcpy_first_frame(
     spawned_at: float,
     *,
     timeout: float = SCRCPY_STARTUP_TIMEOUT_SECONDS,
+    require_marker: bool = False,
 ) -> float:
     """Estimate when the first frame of a scrcpy recording reached the host."""
     fallback = spawned_at + SCRCPY_STARTUP_FALLBACK_SECONDS
@@ -106,6 +119,8 @@ async def await_scrcpy_first_frame(
     while True:
         remaining = deadline - time.monotonic()
         if remaining <= 0:
+            if require_marker:
+                raise TimeoutError("scrcpy 未在启动时限内确认开始录制")
             logger.warning(
                 f"scrcpy did not report '{SCRCPY_RECORDING_STARTED_MARKER}' within "
                 f"{timeout:.1f}s; assuming the first frame at spawn + "
@@ -121,6 +136,8 @@ async def await_scrcpy_first_frame(
                 await asyncio.wait_for(process.wait(), timeout=1.0)
             except (AttributeError, ProcessLookupError, TimeoutError):
                 pass
+            if require_marker:
+                raise OSError("scrcpy 输出已关闭，但未确认开始录制")
             return fallback
         if SCRCPY_RECORDING_STARTED_MARKER in line.decode(errors="replace"):
             return max(spawned_at, time.time())
@@ -152,6 +169,7 @@ class RecordingSession(BaseModel):
     video_id: UUID
     device_id: str
     start_time: float
+    recording_codec: str = "h264"
     process: Any = None
     data_engine_start_time: float | None = None
     local_video_path: Path | None = None
@@ -310,7 +328,7 @@ async def normalize_recording_to_mp4(
 
 
 async def remux_recording_to_mp4(source_path: Path, output_path: Path) -> bool:
-    """Atomically publish a fixed-dimension recording segment as MP4."""
+    """Publish only a playable segment; preserve the source on failed finalization."""
     if not source_path.exists() or source_path.stat().st_size == 0:
         return False
     temporary_path = output_path.with_name(f"{output_path.stem}.part{output_path.suffix}")
@@ -336,13 +354,13 @@ async def remux_recording_to_mp4(source_path: Path, output_path: Path) -> bool:
             stdout=asyncio.subprocess.PIPE,
             stderr=asyncio.subprocess.PIPE,
         )
-        _stdout, stderr = await process.communicate()
+        _stdout, stderr = await _communicate_video_process(process, 30)
         valid = (
             process.returncode == 0
             and temporary_path.exists()
             and temporary_path.stat().st_size > 0
         )
-        if valid:
+        if valid and await probe_video_segment(temporary_path):
             temporary_path.replace(output_path)
             return True
 
@@ -350,7 +368,9 @@ async def remux_recording_to_mp4(source_path: Path, output_path: Path) -> bool:
             f"Direct remux failed for {source_path} (code {process.returncode}), "
             "attempting transcode fallback with ffmpeg..."
         )
-        # Fallback to ultrafast re-encoding in case MKV container was truncated or has timestamp irregularities
+        # Re-encode the actual decoded frames with their original presentation
+        # intervals. Never manufacture a duration from wall time, duplicate frames,
+        # or use a constant-frame-rate filter to cover a recording gap.
         fallback_proc = await asyncio.create_subprocess_exec(
             get_ffmpeg_path(),
             "-y",
@@ -358,31 +378,47 @@ async def remux_recording_to_mp4(source_path: Path, output_path: Path) -> bool:
             "+genpts+discardcorrupt",
             "-i",
             str(source_path),
+            "-map",
+            "0:v:0",
+            "-an",
+            "-vf",
+            "setpts=PTS-STARTPTS",
             "-c:v",
             "libx264",
             "-preset",
             "ultrafast",
             "-pix_fmt",
             "yuv420p",
+            "-bf",
+            "0",
+            "-fps_mode",
+            "passthrough",
+            "-enc_time_base",
+            "1:1000000",
             "-movflags",
             "+faststart",
             str(temporary_path),
             stdout=asyncio.subprocess.PIPE,
             stderr=asyncio.subprocess.PIPE,
         )
-        _fout, _ferr = await fallback_proc.communicate()
+        _fout, _ferr = await _communicate_video_process(fallback_proc, 120)
         if (
             fallback_proc.returncode == 0
             and temporary_path.exists()
             and temporary_path.stat().st_size > 0
+            and await probe_video_segment(temporary_path)
         ):
             temporary_path.replace(output_path)
             return True
 
         logger.error(
-            f"Failed to remux recording segment (code {process.returncode}): "
-            f"{stderr.decode(errors='replace')[-2000:]}"
+            f"Recording conversion produced no valid timed video; source preserved: {source_path} "
+            f"(remux={process.returncode}, fallback={fallback_proc.returncode}). "
+            f"{(_ferr or stderr).decode(errors='replace')[-2000:]}"
         )
+    except asyncio.CancelledError:
+        temporary_path.unlink(missing_ok=True)
+        raise
     except Exception as e:
         logger.error(f"Failed to remux recording segment: {e}")
     if temporary_path.exists():
@@ -390,40 +426,85 @@ async def remux_recording_to_mp4(source_path: Path, output_path: Path) -> bool:
     return False
 
 
+async def _communicate_video_process(process, timeout: float):
+    """Bound media inspection/conversion and reclaim its child on cancellation."""
+    try:
+        return await asyncio.wait_for(process.communicate(), timeout)
+    except (asyncio.CancelledError, TimeoutError):
+        if process.returncode is None:
+            try:
+                process.kill()
+            except ProcessLookupError:
+                pass
+        await asyncio.wait_for(process.wait(), 5)
+        raise
+
+
 async def probe_video_segment(video_path: Path) -> dict[str, float | int]:
-    """Read duration and coded dimensions for a finalized segment."""
+    """Validate video timing before admitting a segment into the task timeline."""
     process = await asyncio.create_subprocess_exec(
         get_ffprobe_path(),
         "-v",
         "error",
+        "-select_streams",
+        "v:0",
+        "-read_intervals",
+        "%+#64",
+        "-show_packets",
         "-show_entries",
-        "stream=width,height,duration,codec_type:format=duration",
+        "stream=width,height,duration,start_time,nb_frames,avg_frame_rate:"
+        "format=duration:packet=pts_time,flags",
         "-of",
         "json",
         str(video_path),
         stdout=asyncio.subprocess.PIPE,
         stderr=asyncio.subprocess.PIPE,
     )
-    stdout, _stderr = await process.communicate()
+    stdout, _stderr = await _communicate_video_process(process, 10)
     if process.returncode != 0:
         return {}
     try:
         payload = json.loads(stdout)
         streams = payload.get("streams") or []
-        video_stream = next(
-            (s for s in streams if s.get("width") and s.get("height")),
-            next(
-                (s for s in streams if s.get("codec_type") == "video"),
-                streams[0] if streams else {},
-            ),
-        )
+        video_stream = streams[0] if streams else {}
         duration = float(
-            (payload.get("format") or {}).get("duration") or video_stream.get("duration") or 0
+            video_stream.get("duration") or (payload.get("format") or {}).get("duration") or 0
         )
+        width, height = int(video_stream.get("width") or 0), int(video_stream.get("height") or 0)
+        start = float(video_stream.get("start_time") or 0)
+        frame_count = (
+            int(video_stream.get("nb_frames") or 0) if video_stream.get("nb_frames") != "N/A" else 0
+        )
+        numerator, _, denominator = str(video_stream.get("avg_frame_rate") or "0/1").partition("/")
+        average_fps = float(numerator) / (float(denominator or 1) or 1)
+        pts = [
+            float(p["pts_time"])
+            for p in payload.get("packets", [])
+            if "pts_time" in p and "D" not in str(p.get("flags", ""))
+        ]
+        # Recorder configuration is capped at 60 fps. A generous 240 fps ceiling
+        # still catches malformed microsecond durations without rejecting VFR.
+        if (
+            not math.isfinite(duration)
+            or duration <= 0
+            or not math.isfinite(start)
+            or not math.isfinite(average_fps)
+            or not 0 < width <= 32768
+            or not 0 < height <= 32768
+            or average_fps < 0
+            or average_fps > 240
+            or frame_count < 0
+            or frame_count / duration > 240
+            or not pts
+            or any(not math.isfinite(value) for value in pts)
+            or max(pts) - start > duration + 0.05
+        ):
+            logger.warning(f"Recording segment has invalid video timing: {video_path}")
+            return {}
         return {
             "duration": duration,
-            "width": int(video_stream.get("width") or 0),
-            "height": int(video_stream.get("height") or 0),
+            "width": width,
+            "height": height,
         }
     except (TypeError, ValueError, json.JSONDecodeError):
         return {}

config/artemis.jsonc

@@ -34,6 +34,8 @@
     "checker": {}
   },
   "agent": {
+    // 留档录像参数，默认从 ARTEMIS_RECORDING_CONFIG 读取；这里可单独覆盖。
+    "recording": {},
     "screenshot": {
       "enabled": "${env:ARTEMIS_SCREENSHOT_LIMIT_ENABLED:-true}",
       "max_long_edge": 2048

docs/ai-automation.md

@@ -69,3 +69,9 @@ Flash 和 Pro 默认复用只读 Checker，以任务提示词和用户指导中
 `mobile_run_task(runtime_overrides={"timeout_seconds":1800,"observation_timeout_seconds":60,"observation_max_attempts":3,"device_command_timeout_seconds":10,"action_failure_attempts":3})` 可为单任务覆盖执行保护；其中 `model` 是 API 模型名，外层 `model` 仍是 Flash/Pro 模式。相同字段也支持批次的 `task_config`。全局默认通过网页功能配置或明确授权的 `mobile_config(section="capabilities", changes={"task_timeout_seconds":1800}, persist_defaults=true)` 保存，对后续任务生效，不接受密钥。
 
 任务已经产生最终结果，但 SDK 线程或资源清理仍卡住时，默认最多再等待 30 秒后回收工作进程，避免继续占用队列。全局可用 `ARTEMIS_WORKER_CLEANUP_TIMEOUT_SECONDS` 或功能配置的 `worker_cleanup_timeout_seconds` 调整，任务 JSON 同名字段可覆盖。原有结果与截图保留，未完成保存的录屏尽力恢复。
+
+## 插件证据与留档参数
+
+2.0.0 插件协议与可运行示例见 [开发规则](plugin-development.md)。单任务和批次全部资料包含各来源截图/录像、插件附件及包内证据索引。
+
+通过 `mobile_config(section="capabilities")` 可读取 `recording` 参数；用户授权后可保存全局默认。单任务用 `runtime_overrides={"recording":{"codec":"h264","computer_scope":"window","max_long_edge":1280,"fps":15}}`，批次在 `task_config.recording` 提供同样对象。JSON 配置的 `agent.recording` 覆盖环境默认；不要把录像参数当作模型截图参数。录屏开关继续独立且默认关闭。

docs/plans/2.0.0-plugin-evidence.md

@@ -1,6 +1,6 @@
 # ARTEMIS 2.0.0：插件步骤、录像与开发规范
 
-状态：2026-10-15 用户已确认设计，待实现。本文补充 [2.0.0 总体方案](2.0.0-plugins.md)，不代表运行代码已具备这些能力。本轮只更新设计文档。
+状态：2026-10-15 已按确认方案实现于 2.0.0 开发分支，确定性回归和短时集成验证通过；长程实机验收待完成，未发布。本文补充 [2.0.0 总体方案](2.0.0-plugins.md)。接口、配置和插件接入方式见 [开发规范](../plugin-development.md)。
 
 本次确认覆盖统一步骤、跨端录像、证据导出和插件开发规则。用户对审阅稿补充了三条约束：录像以高效留档为主；第三方插件和 Skill 必须按 ARTEMIS 规则适配；不自动恢复业务环境，恢复需求由任务提示词表达。
 
@@ -31,7 +31,7 @@
 
 - 优先评估 H.265/HEVC，优先利用可用的硬件编码能力；限制分辨率、帧率和目标码率，不默认保存原生高分辨率、高帧率视频。
 - 编码能力、实际录制和网页播放兼容性必须验证；保留 H.264 回退，记录最终采用的编码及回退原因，不因强制 H.265 阻断任务。
-- 实现时以长边不超过 1280、最高 15 fps、H.265 约 1 Mbps / H.264 约 1.5 Mbps 作为初始试验档位。这些是待测候选值，不是已验证或已生效的默认值；根据中文界面可读性、快速动作、CPU/GPU 占用和文件大小确定最终留档默认值。
+- 开发版默认长边不超过 1280、最高 15 fps、H.265 目标 1 Mbps / H.264 1.5 Mbps。Windows 短时窗口录制已验证两种编码及暂停恢复；这些是留档初始默认值，长程负载、快速动作及多屏仍需发布前验证，不据短样本宣称长期性能。
 - 优先录制时直接采用留档参数，避免任务结束或打包下载时再全量转码。已有合适片段尽量不重复编码。
 - 默认只保留一套留档视频，不同时持久保存高清原片和压缩副本。临时转码文件按生命周期清理；用户证据不因清理被删除。
 - 模型观察截图、步骤证据截图与录像参数独立，不能为了降低录像体积而降低模型定位所需画面质量。
@@ -95,14 +95,24 @@
 - 延续 HTTPS、文件校验、平台兼容、安全解压与同版本不可覆盖规则。升级由用户发起，排队、暂停和运行任务固定版本；同一资源整条任务串行。
 - 沿用用户启用时授权的范围，不新增逐动作确认；密钥不能进入普通查询、日志和报告。参数范围检查不宣称是操作系统沙箱。
 
-## 10. 待实施顺序与检查点
+## 10. 实施状态与检查点
 
-- [ ] 定义公共结果/附件/时间契约及各能力声明，补齐错误与副作用语义和兼容迁移。
-- [ ] 完成 Flash/Pro 的来源隔离与步骤分组，验证默认安卓任务调用插件、同插件跨窗口及多图证据。
-- [ ] 实现电脑录像、统一录制生命周期和高效留档参数，实测后确定默认值及兼容回退。
-- [ ] 实现多来源回看、缺口展示、关键节点与播放状态同步。
-- [ ] 完成单任务/批次导出索引、原生附件和各报告一致性。
-- [ ] 按 ARTEMIS 规则适配首发插件与 Skill，交付模板、能力附录及合规检查。
-- [ ] 完成确定性检查、真实环境和长期运行验收，记录实测范围再发布。
+- [x] 定义公共结果/附件/时间契约及各能力声明，补齐错误与副作用语义和兼容迁移。
+- [x] 完成 Flash/Pro 的来源隔离与步骤分组，回归覆盖默认安卓任务调用插件、同插件跨窗口及多图证据。
+- [x] 实现电脑录像、统一录制生命周期和高效留档参数；短时实测 H.265/H.264，编码回退与磁盘不足有确定性回归。
+- [x] 实现多来源回看、缺口展示、关键节点与播放状态同步，已在隔离网页实际验证。
+- [x] 完成单任务/批次导出索引、原生附件和各报告一致性。
+- [x] 按 ARTEMIS 规则适配首发插件与 Skill，交付模板、能力附录及合规检查。
+- [x] 完成确定性检查和短时集成验证，记录实测范围。
+- [ ] 完成长程、多屏、多 DPI 及完整跨端模型任务验收，再发布。
 
-已有静态检查发现两项需要复现验证的风险：Pro 对默认安卓任务调用插件时的图片隔离可能不完整；插件前图缓存按插件标识而非具体目标区分，跨窗口可能引用旧图。它们是待验证项，不应在复现前写成已确认故障或已修复。
+已补回归并修正：Pro 默认安卓任务的插件图像隔离、同插件跨窗口的前图缓存与导出，以及任务录像参数的覆盖优先级。发布前仍需覆盖多 DPI、多屏、长程资源开销和完整跨端模型任务；自动测试及短时烟测不替代这些验收。
+
+### 2026-10-15 验证记录
+
+- 最终隔离测试：Python 2561 通过、2 跳过、8 未选择；Angular 224 通过，生产构建通过。Ruff、格式、质量基线、核心 Pyright 与差异空白检查通过。
+- Windows 实际录制验证了 H.265/H.264、暂停恢复、截图附件保存和录制进程随任务进程退出回收。四个短片段均通过最终媒体探测，未据此推断长程性能。
+- 隔离网页验证了来源切换、录像缺口、播放中跳转继续播放、暂停中跳转保持暂停；最终构建的步骤标记没有重复或将工具哈希误当坐标。演示页面使用合成任务数据，不能作为完整跨端模型任务的通过证据。
+- 开发模板通过实际 MCP 调用、中文文件及哈希校验、调用记录、单任务资料导出、目录和 ZIP 导入及 `--live` 合规检查。
+- Android 烟测时模拟器出现启动器无响应，录制曾返回空文件或异常时长。现已要求真实启动标记、有效媒体片段及时间戳校验；异常样本只恢复实际存在的 0.5 秒画面，不补造持续时间。正常设备上的长时间安卓录制仍待验证。
+- 早期测试曾写入开发环境历史；最终全量测试使用独立应用目录、trace 目录和数据库。既有历史未删除。本轮未重启用户运行中的服务，未安装新版市场插件，未推送或发版。

docs/plugin-development.md

@@ -0,0 +1,89 @@
+# 插件开发与证据协议（2.0.0）
+
+新插件必须适配本文规则，再通过实测进入市场。旧 `schema_version: 1` 仍可读取；新包使用版本 2，并声明 `capabilities`。模型执行、任务队列、资源锁、超时和报告由 ARTEMIS 管理，插件不另建任务调度器，也不替用户决定用例通过。
+
+## 从可运行模板开始
+
+模板见 [evidence-demo](../examples/plugins/evidence-demo/plugin.json)。在安装了项目依赖的环境执行：
+
+```sh
+uv run python scripts/check_plugin.py examples/plugins/evidence-demo
+uv run python scripts/check_plugin.py examples/plugins/evidence-demo --live
+```
+
+`--live` 启动 MCP、检查真实工具定义后关闭，不执行写操作。模板默认使用 PATH 中的 Python；需要指定解释器时传 `--values '{"python":"解释器绝对路径"}'`。
+
+网页「插件 → 高级接入」支持插件目录或 ZIP 文件的完整路径。开发时可直接填模板目录；分发时在仓库根目录执行以下命令，生成根层包含 `plugin.json` 的 `evidence-demo.zip`：
+
+```sh
+uv run python -c "import shutil; shutil.make_archive('evidence-demo', 'zip', 'examples/plugins/evidence-demo')"
+```
+
+导入后在插件设置中确认 `python` 指向已安装 MCP 依赖的解释器，再启用并提交“调用文件证据示例保存一段文本”。在任务步骤查看工具结果，下载「全部资料」查看 `任务文件/插件/evidence-demo/` 和 `证据索引.json`。
+
+静态检查通过不等于市场准入，也不表示业务测试通过。需按下面清单提交实际验证记录。
+
+## 清单与工具能力
+
+`plugin.json` 的安装、版本约束、资源、权限设置沿用 [插件说明](plugins.md)。版本 2 增加：
+
+| 字段 | 规则 |
+| --- | --- |
+| `capabilities.kind` | `visual`、`terminal`、`api`、`file`、`other` |
+| `capabilities.recording` | `none` 或 `windows`；只有视觉插件可声明 Windows 屏幕录制 |
+| `capabilities.tools` | 按真实 MCP 工具名定义；未声明和缺失工具均拒绝加载 |
+| `effect` | `read` / `write`，必须如实声明 |
+| `observation` | 自动观察必须只读；MCP 的 `readOnlyHint` 也必须为 true |
+| `idempotent` | 重复调用是否产生额外效果，不授权自动重试未知结果的写操作 |
+| `target_argument` | 指明资源目标的顶层参数名，必须存在于工具 JSON Schema |
+| `cancellation` | `confirmed` / `unknown`，描述工具承诺；不能代替实际停止核对 |
+| `evidence` | `text` / `image` / `file` / `state` 列表 |
+| `required_tools` | Skill 依赖的工具名，加载时校验 |
+
+MCP 参数必须有类型及合理范围；普通插件不能用不受约束的 shell 代替业务能力。终端插件若声明命令能力，应明确主机、工作目录、权限和命令边界。资源作用域用已有 `resources` 和配置约束声明；这些不是操作系统沙箱。
+
+启动、调用、关闭分别受既有 30/60/15 秒默认预算约束，并服从任务总超时。每个写操作最多派发一次；超时、取消或部分执行后先观察真实结果。远端动作未确认停止时保持资源保护。
+
+## 结果回执
+
+MCP 保留标准 `content` 和 `isError`，在 `structuredContent.artemis` 返回版本 1 回执：
+
+```json
+{
+  "version": 1,
+  "execution": "completed",
+  "side_effect": "confirmed",
+  "retryable": false,
+  "source": {
+    "kind": "computer", "target_id": "真实窗口标识", "label": "控制台",
+    "captured_at": 1792000000.25, "width": 1280, "height": 720,
+    "window": {"left": 0, "top": 0, "width": 1280, "height": 720},
+    "screen": {"left": 0, "top": 0, "width": 1920, "height": 1080}
+  },
+  "attachments": []
+}
+```
+
+- `execution`：`not_started`、`completed`、`partial`、`unknown`；`side_effect`：`none`、`possible`、`confirmed`。错误提供稳定的 `error_code` 和实际原因。只有明确未执行的请求才可声明 `retryable`，宿主仍受预算约束。
+- 截图必须属于本次目标。尺寸和范围使用物理像素，时间为 Unix 秒；`target_id` 不能只是插件名。移动、缩放、换屏、换窗口后重新观察。平台结合目标和几何信息生成来源身份，避免拿旧图做动作前证据。
+- 视觉工具通过标准 MCP image 返回图像；无法截图明确返回原因，不混入其他屏幕或猜坐标。终端返回退出码、输出和工作目录，API 返回状态/业务结果，文件返回校验信息，无须伪造图片。
+- 平台生成 `call_id`、步骤关联、开始结束时间，并写入 trace。插件不能根据一次工具成功自行生成整条用例通过结论。
+
+## 文件、录像与生命周期
+
+stdio 宿主注入 `ARTEMIS_ARTIFACTS_DIR`，插件默认把生成文件写入该目录。附件回执包含普通 `name`、`media_type`、`sha256`，不接受任意宿主绝对路径；远端插件可附带 base64 `data`。单附件不超过 64 MiB，并同时受工具结果总大小限制。禁止路径穿越、符号链接和覆盖同名不同内容。
+
+要求另存桌面时先产生并验证任务副本，再复制用户指定位置，分别报告实际结果。任一写入失败不声称全部成功。平台校验回读字节和 SHA-256，保存到任务索引；单任务和批次下载包含原生附件及插件证据。
+
+录屏默认关闭。打开后，首次取得目标观察便启动已声明来源的录制，后续动作和等待持续录制；暂停收尾，恢复追加片段，任务结束关闭进程。电脑默认记录目标所在屏幕，可改为仅目标窗口。屏幕录像可能包含同屏其他可见内容；窗口截图仍只属于目标窗口，两者范围不同。
+
+留档参数独立于模型截图：默认长边 1280、15 fps、H.265 目标 1000 kbps / H.264 1500 kbps。自动模式只有网页播放能力确认支持 HEVC 才尝试硬件 H.265，否则 H.264；编码不可用则有界回退并保留原因。网页使用真实任务时间轴，缺口不会被拼接抹去。H.265 的实际网页支持取决于浏览器/系统，手动优先 H.265 后无法播放应下载或为后续任务选择 H.264，不在下载时整段转码。
+
+录制故障不能把成功动作改为失败，也不能把缺失证据推断为测试通过。录像文件只保留一套留档版本；磁盘不足停止采集并告警。不自动恢复路由器参数、电脑状态或用户文件；只有任务提示词提出恢复要求时才执行普通恢复步骤。连接关闭、进程回收和锁释放始终必要。
+
+## 提交市场前的验证
+
+- 公共：安装/升级兼容、真实工具与清单一致、参数限制、失败/部分执行、超时取消、版本固定、资源互斥、导出索引、卸载后历史可读、无凭据泄漏。
+- 视觉：截图来源、遮挡、DPI、窗口切换、坐标一致、无图时拒绝盲点；录屏启动/结束/暂停恢复、编码回退、磁盘不足、时间缺口、关键节点、播放状态、中文可读性及长程体积。
+- 文件：中文文件名、覆盖保护、校验失败；终端/API：退出码、业务结果和不确定副作用。
+- Skill 必须只描述真实存在且已适配的能力，不绕过权限、调度和证据协议。保留来源和必要许可证，版本与实测范围明确；未测平台/场景不能写成已通过。

docs/plugins.md

@@ -1,5 +1,11 @@
 # 插件与跨端任务（2.0.0 开发版）
 
+插件开发请阅读 [能力、证据协议与验收规则](plugin-development.md)，可运行模板位于 `examples/plugins/evidence-demo`。新插件使用清单版本 2；版本 1 仍兼容。
+
+步骤卡标明安卓、电脑等来源，可点击查看对应录像；同一步内保留实际工具调用。开启「保存录屏」后，已使用的安卓和电脑来源各自留档，通过播放器切换；真实时间轴保留暂停和录制中断的缺口。录像默认 1280 长边、15 fps，编码参数在功能设置的「录像留档设置」调整；不会降低模型观察截图的分辨率。
+
+单任务「全部资料」和批次资料包含插件截图、录像、调用回执和生成文件，`证据索引.json` 使用包内路径关联附件。电脑截图默认保存任务 trace，明确要求桌面时额外复制并校验。工具执行完成与测试通过仍分别呈现，不自动恢复业务环境。
+
 ARTEMIS 可通过插件连接 stdio 或 Streamable HTTP MCP 服务，并按需加载插件内的 Skill。Flash、Pro、网页任务和批次复用同一套插件执行层。版本仍在开发验证中，市场只收录完成实际验证的安装包。
 
 ## 安装和配置

examples/plugins/evidence-demo/SKILL.md

@@ -0,0 +1,10 @@
+---
+name: evidence-demo
+description: 按任务要求保存文本到任务资料目录，返回文件校验结果。
+---
+
+调用 plugin_load 返回映射后的 write_note 工具。只在用户需要保存文本时使用；不是任意文件读写或命令执行接口。
+
+输入 text 不超过 16 KiB，文件名由工具生成，不能覆盖已有文件。成功后检查回执中的文件名、SHA-256 和执行状态；文件会随单任务全部资料及批次资料导出。
+
+保存成功不代表用例通过。结果不确定时先核对任务文件，不重复写入。不自动恢复或删除用户文件。

examples/plugins/evidence-demo/plugin.json

@@ -0,0 +1,24 @@
+{
+  "schema_version": 2,
+  "id": "evidence-demo",
+  "name": "文件证据示例",
+  "version": "1.0.0",
+  "description": "开发模板：写入任务专属文本文件并返回校验回执。",
+  "artemis_version": ">=2.0.0,<3.0.0",
+  "platforms": ["win32", "linux", "darwin"],
+  "architectures": ["amd64", "arm64"],
+  "server": {"command": "${config.python}", "args": ["${plugin_dir}/server.py"]},
+  "config": {"python": {"label": "含 MCP 依赖的 Python", "default": "python"}},
+  "skill": "SKILL.md",
+  "targets": ["mixed", "computer"],
+  "resources": [],
+  "capabilities": {
+    "kind": "file",
+    "recording": "none",
+    "required_tools": ["write_note"],
+    "tools": {"write_note": {
+      "effect": "write", "observation": false, "idempotent": false,
+      "cancellation": "unknown", "evidence": ["text", "file"]
+    }}
+  }
+}

examples/plugins/evidence-demo/server.py

@@ -0,0 +1,64 @@
+"""Minimal task-owned attachment server. Run through ARTEMIS, not as a shell tool."""
+
+import hashlib
+import os
+from pathlib import Path
+import time
+import uuid
+
+from mcp.server.fastmcp import FastMCP
+from mcp.types import CallToolResult, TextContent, ToolAnnotations
+
+mcp = FastMCP("evidence-demo")
+
+
+@mcp.tool(
+    annotations=ToolAnnotations(readOnlyHint=False, destructiveHint=False, idempotentHint=False)
+)
+def write_note(text: str) -> CallToolResult:
+    """Save at most 16 KiB of user-requested text as a new task attachment."""
+    encoded = text.encode("utf-8")
+    root = os.environ.get("ARTEMIS_ARTIFACTS_DIR")
+    if not root or len(encoded) > 16384:
+        return CallToolResult(
+            isError=True,
+            content=[TextContent(type="text", text="任务目录缺失或内容超过 16 KiB；未写文件。")],
+            structuredContent={
+                "artemis": {
+                    "version": 1,
+                    "execution": "not_started",
+                    "side_effect": "none",
+                    "error_code": "invalid_input",
+                    "retryable": False,
+                }
+            },
+        )
+    directory = Path(root)
+    directory.mkdir(parents=True, exist_ok=True)
+    name = f"note-{uuid.uuid4().hex}.txt"
+    path = directory / name
+    with path.open("xb") as handle:
+        handle.write(encoded)
+    digest = hashlib.sha256(path.read_bytes()).hexdigest()
+    return CallToolResult(
+        content=[TextContent(type="text", text=f"已保存并读回校验：{name}")],
+        structuredContent={
+            "artemis": {
+                "version": 1,
+                "execution": "completed",
+                "side_effect": "confirmed",
+                "retryable": False,
+                "source": {
+                    "kind": "file",
+                    "target_id": "task-artifacts",
+                    "label": "任务文件",
+                    "captured_at": time.time(),
+                },
+                "attachments": [{"name": name, "sha256": digest, "media_type": "text/plain"}],
+            }
+        },
+    )
+
+
+if __name__ == "__main__":
+    mcp.run(transport="stdio")

mcp_server/tools/automation.py

@@ -32,6 +32,7 @@ CAPABILITY_FIELDS = {
     "adb_discovery_timeout_seconds",
     "video_enabled",
     "recording_enabled",
+    "recording",
     "outputter_enabled",
     "step_summarizer_enabled",
     "screenshot_limit_enabled",

mcp_server/tools/task_runner.py

@@ -15,6 +15,7 @@
 """MCP Tool: mobile_run_task."""
 
 import logging
+import json
 import os
 import subprocess
 import sys
@@ -344,7 +345,10 @@ def mobile_run_task(
           device_command_timeout_seconds (10), action_failure_attempts (3),
           worker_cleanup_timeout_seconds (30),
           model (API model name), reasoning_effort, request timeout/retries,
-          recording and video analysis. Unknown fields and credentials are rejected.
+          record_video and video analysis. recording is a partial archive-quality object:
+          codec (auto/h264/h265), computer_scope (screen/window), max_long_edge,
+          fps, h264_kbps, h265_kbps; it does not enable recording on its own.
+          Unknown fields and credentials are rejected.
     """
     if runtime_overrides:
         from apps.admin_console.services.task_import import TaskOptions
@@ -572,6 +576,10 @@ def mobile_run_task(
         env["ARTEMIS_SESSION_ID"] = trace_id
         env["ARTEMIS_TASK_INGRESS"] = "mcp"
         env["ARTEMIS_TASK_WORKER"] = "1"
+        if (runtime_overrides or {}).get("recording") is not None:
+            env["ARTEMIS_TASK_RECORDING_CONFIG"] = json.dumps(runtime_overrides["recording"])
+        else:
+            env.pop("ARTEMIS_TASK_RECORDING_CONFIG", None)
         for option, variable in {
             "model": "OPENAI_MODEL",
             "reasoning_effort": "OPENAI_REASONING_EFFORT",

scripts/check_plugin.py

@@ -0,0 +1,47 @@
+"""Validate a plugin package; --live checks MCP declarations without executing actions."""
+
+import argparse
+import asyncio
+import json
+from pathlib import Path
+
+from artemis.plugins.evidence import validate_tools
+from artemis.plugins.models import PluginManifest, PluginProfile
+from artemis.plugins.session import PluginSession, resolve_server
+from artemis.plugins.store import contained_path
+
+
+async def check(directory: Path, live: bool, values: dict[str, str]) -> None:
+    manifest = PluginManifest.model_validate_json((directory / "plugin.json").read_text("utf-8"))
+    manifest.check_compatible()
+    if manifest.skill:
+        skill = contained_path(directory, manifest.skill)
+        if not skill.is_file() or not skill.read_text("utf-8").strip():
+            raise ValueError("缺少有效的 Skill")
+    if manifest.schema_version != 2:
+        raise ValueError("新市场插件必须适配 schema_version=2")
+    if live and manifest.server:
+        server = resolve_server(manifest, PluginProfile(values=values), directory)
+        session = await PluginSession(server, directory).start()
+        try:
+            validate_tools(manifest, session.tools)
+        finally:
+            await session.aclose()
+    print(
+        f"PASS {manifest.id} {manifest.version}: 清单/兼容性/Skill"
+        + ("/MCP 工具声明" if live else "")
+    )
+    print("此检查不代替实际执行、超时取消、证据导出和视觉能力验收。")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("directory", type=Path)
+    parser.add_argument("--live", action="store_true")
+    parser.add_argument("--values", default="{}", help="普通配置 JSON；不得包含密钥")
+    args = parser.parse_args()
+    asyncio.run(check(args.directory.resolve(), args.live, json.loads(args.values)))
+
+
+if __name__ == "__main__":
+    main()

skills/artemis-automation/SKILL.md

@@ -47,6 +47,10 @@ Windows 启动脚本只提示源码更新。升级使用 `update.bat`（国内
 
 ## 配置和恢复边界
 
+- 插件步骤按实际来源区分，截图不能跨手机/电脑或跨目标窗口作为证据。`mobile_export_report(formats=["all"])` 和批次导出包含插件文件、各来源录像及 `证据索引.json`，索引中的路径相对于任务资料目录。
+- 录像默认关闭。保存录屏后，使用到的支持录制来源自动留档；电脑视频分析尚未提供，单开视频分析不会留下电脑录像。留档默认长边 1280、15 fps、H.265 1000 kbps / H.264 1500 kbps；模型截图参数独立。`mobile_config(section="capabilities")` 的 `recording` 对象或单任务 `runtime_overrides.recording`、批次 `task_config.recording` 可覆盖 `codec`（auto/h264/h265）、`computer_scope`（screen/window）、尺寸、帧率和码率。全局保存仍需用户授权。
+- 电脑屏幕录像包含当前控制屏幕的其他可见内容。任务暂停停止录像，恢复追加片段；查看真实时间轴的无覆盖区间时结合截图/日志，不把旧帧或其他来源当作证据。除任务提示词要求，不自动恢复业务配置。
+
 - 本机常见模拟器在设备刷新、启动及重启 ADB 后自动发现并连接，不必猜固定端口。USB 与 Android SDK 模拟器由 ADB 原生发现。首次连接其他 TCP 设备/无线手机时，在网页“无线 ADB”填写实际地址和端口，成功后程序保存地址并自动重连；无线调试的配对和授权仍由用户完成，端口变化后需重新连接。远程/转发 ADB 不执行本机发现。用 `mobile_workspace` 或 `mobile_diagnose` 检查自动连接结果；多设备时仍先确认用户要测试哪台。默认开关及扫描预算可由 `mobile_config(section="capabilities")` 的 `adb_auto_connect`、`adb_discovery_interval_seconds`（10）、`adb_discovery_timeout_seconds`（5）配置，修改全局默认需要用户授权。
 
 - `mobile_config` 读取非敏感设置。只有用户明确要求保存或修改默认配置时，才传 `changes` 与 `persist_defaults=true`；这些设置影响后续任务，不能为一次临时需求擅自改变以后所有任务。

tests/unit/admin_console/test_automation.py

@@ -583,7 +583,8 @@ async def test_fail_fast_includes_skipped_reports(isolated, monkeypatch):
     assert [j["status"] for j in isolated.jobs(batch["id"])] == ["failed", "skipped"]
     archive = reports.export_batch(batch["id"])
     with zipfile.ZipFile(archive) as z:
-        assert len(z.namelist()) == 6  # two summary files + two files per task
+        assert len(z.namelist()) == 8  # summaries + two reports and evidence index per task
+        assert len([name for name in z.namelist() if name.endswith("证据索引.json")]) == 2
         assert sum(n.endswith(".docx") for n in z.namelist()) == 3
 
 

tests/unit/admin_console/test_plugin_media.py

@@ -0,0 +1,82 @@
+import json
+from unittest.mock import MagicMock
+
+from apps.admin_console.routers import media
+from apps.admin_console.services import evidence_service as evidence
+from apps.admin_console.services import report_export as exports
+
+
+def fixture_files(tmp_path, monkeypatch):
+    monkeypatch.setattr(evidence, "TRACES_PATH", tmp_path)
+    root = tmp_path / "task" / "evidence" / "recordings"
+    root.mkdir(parents=True)
+    video = root / "clip.mp4"
+    video.write_bytes(b"fixture video")
+    (root / "recordings.json").write_text(
+        json.dumps(
+            {
+                "version": 1,
+                "sources": [
+                    {
+                        "id": "pc",
+                        "kind": "computer",
+                        "target_ids": ["window"],
+                        "status": "ready",
+                        "segments": [
+                            {
+                                "file": "evidence/recordings/clip.mp4",
+                                "offset_ms": 5000,
+                                "duration": 2,
+                            }
+                        ],
+                    }
+                ],
+            }
+        ),
+        "utf-8",
+    )
+    return video
+
+
+def test_plugin_media_preserves_session_offsets_and_exports_relative_files(tmp_path, monkeypatch):
+    fixture_files(tmp_path, monkeypatch)
+    monkeypatch.setattr(exports, "session_screenshots", lambda _: [])
+    monkeypatch.setattr(exports, "session_artifacts", lambda _: [])
+    source = evidence.plugin_sources("task")[0]
+    assert source["segments"][0]["start"] == 5
+    index = evidence.export_index(
+        "task", {"name": "case", "status": "completed", "verdict": "未验证"}, []
+    )
+    clip = index["recordings"][0]["segments"][0]
+    assert clip["file"] == "任务录屏/电脑/clip.mp4" and "url" not in clip
+    assert clip["file"] in {item["path"] for item in index["files"]}
+
+
+def test_legacy_url_only_does_not_enable_strict_timeline(monkeypatch):
+    monkeypatch.setattr(
+        media,
+        "_get_legacy_session_video_sync",
+        lambda _: {"status": "ready", "video_url": "/videos/old.mp4", "video_segments": []},
+    )
+    monkeypatch.setattr(evidence, "plugin_sources", lambda _: [])
+    monkeypatch.setattr(media, "session_repo", MagicMock(get_session_by_id=lambda _: {}))
+    response = media._get_session_video_sync("task")
+    assert response["session_timeline"] is False
+    assert response["video_url"] == "/videos/old.mp4"
+
+
+def test_computer_only_recording_can_be_downloaded(tmp_path, monkeypatch):
+    import zipfile
+
+    fixture_files(tmp_path, monkeypatch)
+    monkeypatch.setattr(exports, "EXPORT_ROOT", tmp_path / "exports")
+    monkeypatch.setattr(
+        exports,
+        "session_report",
+        lambda _: {"id": "task", "name": "case", "status": "completed", "verdict": "未验证"},
+    )
+    monkeypatch.setattr(exports, "session_videos", lambda _: [])
+    with zipfile.ZipFile(exports.export_session_media("task", "video")) as archive:
+        index = json.loads(archive.read("文件清单.json"))
+        assert index["videos"] == ["任务录屏/电脑/clip.mp4"]
+        assert archive.getinfo(index["videos"][0]).compress_type == zipfile.ZIP_STORED

tests/unit/core/test_platform.py

@@ -37,8 +37,9 @@ def test_platform_singleton():
     assert p.process.path_separator in (":", ";")
 
 
-def test_linux_platform_paths():
+def test_linux_platform_paths(monkeypatch):
     """Verify Linux platform paths conform to XDG Base Directory specification."""
+    monkeypatch.delenv("ARTEMIS_APP_DIR", raising=False)
     with patch.dict(
         "os.environ",
         {"XDG_CONFIG_HOME": "/tmp/test_xdg_config", "XDG_DATA_HOME": "/tmp/test_xdg_data"},
@@ -61,8 +62,9 @@ def test_darwin_platform_paths():
     assert darwin_p.process.path_separator == ":"
 
 
-def test_windows_platform_paths():
+def test_windows_platform_paths(monkeypatch):
     """Verify Windows platform paths adhere to AppData standards."""
+    monkeypatch.delenv("ARTEMIS_APP_DIR", raising=False)
     with patch.dict(
         "os.environ",
         {

tests/unit/test_plugin_evidence.py

@@ -0,0 +1,396 @@
+"""Plugin contracts and evidence persistence, without devices or external services."""
+
+import base64
+import hashlib
+import json
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, Mock
+from uuid import uuid4
+
+from mcp.types import CallToolResult, ImageContent, TextContent, Tool, ToolAnnotations
+import pytest
+
+from artemis.plugins.evidence import (
+    decode_receipt,
+    persist_attachments,
+    persist_receipt,
+    source_identity,
+    validate_tools,
+)
+from artemis.plugins.models import EvidenceSource, PluginCallResult, PluginManifest
+from artemis.plugins.runtime import PluginRuntime
+
+
+def manifest(**updates):
+    return PluginManifest.model_validate(
+        {
+            "schema_version": 2,
+            "id": "visual-test",
+            "name": "Visual fixture",
+            "version": "1.0.0",
+            "skill": "SKILL.md",
+            "observation": {"tool": "observe", "refresh_tools": ["observe"]},
+            "capabilities": {
+                "kind": "visual",
+                "recording": "windows",
+                "tools": {
+                    "observe": {
+                        "effect": "read",
+                        "observation": True,
+                        "target_argument": "app",
+                    }
+                },
+                "required_tools": ["observe"],
+            },
+            **updates,
+        }
+    )
+
+
+def observation_tool(*, readonly=True, properties=None):
+    return Tool(
+        name="observe",
+        inputSchema={
+            "type": "object",
+            "properties": {"app": {}} if properties is None else properties,
+        },
+        annotations=ToolAnnotations(readOnlyHint=readonly),
+    )
+
+
+def attachment(data=b"verified evidence", name="capture.png"):
+    return {
+        "name": name,
+        "data": base64.b64encode(data).decode(),
+        "sha256": hashlib.sha256(data).hexdigest(),
+        "media_type": "image/png",
+    }
+
+
+def test_schema_one_stays_compatible_but_schema_two_requires_contract():
+    legacy = manifest(schema_version=1, capabilities=None)
+    validate_tools(legacy, [observation_tool(readonly=False)])
+    with pytest.raises(ValueError, match="capabilities"):
+        manifest(capabilities=None)
+    with pytest.raises(ValueError, match="schema_version=2"):
+        manifest(schema_version=1)
+
+
+@pytest.mark.parametrize("effect,observation", [("write", True), ("read", False)])
+def test_auto_observation_requires_readonly_contract(effect, observation):
+    with pytest.raises(ValueError, match="只读"):
+        manifest(
+            capabilities={
+                "kind": "visual",
+                "tools": {"observe": {"effect": effect, "observation": observation}},
+            }
+        )
+
+
+def test_nonvisual_plugin_cannot_request_screen_recording():
+    with pytest.raises(ValueError, match="视觉插件"):
+        manifest(observation=None, capabilities={"kind": "api", "recording": "windows"})
+
+
+def test_manifest_is_checked_against_actual_mcp_tools():
+    validate_tools(manifest(), [observation_tool()])
+    with pytest.raises(ValueError, match="只读声明"):
+        validate_tools(manifest(), [observation_tool(readonly=False)])
+    with pytest.raises(ValueError, match="缺少目标参数"):
+        validate_tools(manifest(), [observation_tool(properties={})])
+    with pytest.raises(ValueError, match="缺少：observe"):
+        validate_tools(manifest(), [])
+    with pytest.raises(ValueError, match="未声明：click"):
+        validate_tools(manifest(), [observation_tool(), Tool(name="click", inputSchema={})])
+
+
+def test_source_identity_invalidates_geometry_but_not_title_or_capture_time():
+    source = EvidenceSource(
+        kind="computer",
+        target_id="123",
+        label="before",
+        captured_at=1,
+        width=800,
+        height=600,
+        window={"left": 0, "top": 0, "width": 800, "height": 600},
+    )
+    identity = source_identity("visual-test", source)
+    assert identity == source_identity(
+        "visual-test", source.model_copy(update={"label": "after", "captured_at": 2})
+    )
+    for change in (
+        {"target_id": "456"},
+        {"width": 400},
+        {"window": {"left": 100, "top": 0, "width": 800, "height": 600}},
+    ):
+        assert identity != source_identity("visual-test", source.model_copy(update=change))
+    assert identity != source_identity("another-plugin", source)
+
+
+@pytest.mark.parametrize(
+    "execution,retryable,observe",
+    [
+        ("unknown", False, True),
+        ("partial", False, True),
+        ("completed", False, False),
+        ("not_started", True, False),
+    ],
+)
+def test_uncertain_writes_cannot_claim_to_be_retryable(execution, retryable, observe):
+    result = PluginCallResult(
+        status="error",
+        structured_content={
+            "artemis": {
+                "version": 1,
+                "execution": execution,
+                "side_effect": "possible",
+                "retryable": True,
+                "error_code": "remote_failure",
+            }
+        },
+    )
+    decode_receipt(result)
+    assert result.execution == execution
+    assert result.retryable is retryable
+    assert result.requires_observation is observe
+    assert result.error_code == "remote_failure"
+
+
+@pytest.mark.parametrize("raw", ["invalid", {"version": 99}])
+def test_rejects_unsupported_receipt_shapes(raw):
+    with pytest.raises(ValueError):
+        decode_receipt(PluginCallResult(status="success", structured_content={"artemis": raw}))
+
+
+def test_attachments_are_read_back_and_never_overwrite_existing_evidence(tmp_path):
+    receipt = attachment()
+    saved = persist_attachments(tmp_path, "visual-test", [receipt])[0]
+    path = tmp_path / saved["path"]
+    assert path.read_bytes() == b"verified evidence"
+    assert saved["sha256"] == receipt["sha256"]
+    assert saved["bytes"] == len(b"verified evidence")
+    assert (
+        persist_attachments(
+            tmp_path, "visual-test", [{k: v for k, v in receipt.items() if k != "data"}]
+        )[0]
+        == saved
+    )
+    with pytest.raises(ValueError, match="不能覆盖"):
+        persist_attachments(tmp_path, "visual-test", [attachment(b"different")])
+    assert path.read_bytes() == b"verified evidence"
+
+
+def test_invalid_hash_does_not_publish_attachment(tmp_path):
+    receipt = attachment()
+    receipt["sha256"] = "0" * 64
+    with pytest.raises(ValueError, match="校验失败"):
+        persist_attachments(tmp_path, "visual-test", [receipt])
+    assert not (tmp_path / "artifacts/plugins/visual-test/capture.png").exists()
+
+
+@pytest.mark.parametrize(
+    "name",
+    [
+        "../outside.png",
+        "..\\outside.png",
+        "C:\\outside.png",
+        "file.png:stream",
+        "CON.png",
+        "LPT1",
+        "file.",
+        "file ",
+        "bad\nname",
+    ],
+)
+def test_attachment_names_cannot_escape_or_alias_windows_files(tmp_path, name):
+    with pytest.raises(ValueError, match="普通文件名"):
+        persist_attachments(tmp_path, "visual-test", [attachment(name=name)])
+
+
+def test_receipt_index_has_references_instead_of_binary_content(tmp_path):
+    result = PluginCallResult(
+        status="success",
+        call_id="call-1",
+        image_names=["image-1"],
+        content=[{"type": "image", "data": "large-binary"}],
+        structured_content={"artemis": {"large": "payload"}},
+    )
+    persist_receipt(
+        SimpleNamespace(data_engine=SimpleNamespace(base_dir=tmp_path)),
+        result,
+        {"app": "fixture"},
+        123,
+    )
+    receipt = json.loads((tmp_path / "evidence/calls.jsonl").read_text(encoding="utf-8"))
+    assert receipt["call_id"] == "call-1"
+    assert receipt["image_names"] == ["image-1"]
+    assert receipt["started_at"] == 123
+    assert "content" not in receipt
+    assert "structured_content" not in receipt
+
+
+@pytest.mark.asyncio
+async def test_receipt_disk_failure_preserves_successful_write_result(monkeypatch):
+    runtime = PluginRuntime.__new__(PluginRuntime)
+    runtime.ctx = None
+    runtime.bindings = {
+        "write": (
+            "visual-test",
+            SimpleNamespace(name="write", inputSchema={"type": "object"}, annotations=None),
+        )
+    }
+    runtime.packages = {
+        "visual-test": (
+            SimpleNamespace(capabilities=None, observation=None, version="1.0.0"),
+            SimpleNamespace(argument_constraints={}, secrets={}),
+            None,
+        )
+    }
+    session = SimpleNamespace(
+        call=AsyncMock(
+            return_value=CallToolResult(
+                content=[TextContent(type="text", text="changed")], isError=False
+            )
+        ),
+        stop_uncertain=False,
+    )
+    runtime.sessions = {"visual-test": session}
+    runtime.ensure_resources_ready = lambda: None
+    runtime._image_targets = {}
+    runtime.last_images = {}
+    runtime._observation_calls = {}
+
+    def unavailable(*_args, **_kwargs):
+        raise OSError("disk full")
+
+    monkeypatch.setattr("artemis.plugins.runtime.persist_receipt", unavailable)
+    outcome = await runtime.call("write", {})
+    session.call.assert_awaited_once()
+    assert outcome.status == "success"
+    assert outcome.execution == "completed"
+    assert outcome.side_effect != "none"
+
+
+@pytest.mark.asyncio
+async def test_v2_observation_without_task_context_never_starts_recording(monkeypatch):
+    runtime = PluginRuntime.__new__(PluginRuntime)
+    runtime.ctx = None
+    runtime.bindings = {"observe": ("visual-test", observation_tool())}
+    runtime.packages = {
+        "visual-test": (
+            manifest(),
+            SimpleNamespace(argument_constraints={}, secrets={}),
+            None,
+        )
+    }
+    runtime.sessions = {
+        "visual-test": SimpleNamespace(
+            call=AsyncMock(
+                return_value=CallToolResult(
+                    content=[ImageContent(type="image", mimeType="image/png", data="aW1hZ2U=")],
+                    structuredContent={
+                        "artemis": {
+                            "version": 1,
+                            "execution": "completed",
+                            "source": {"kind": "computer", "target_id": "123"},
+                        }
+                    },
+                )
+            ),
+            stop_uncertain=False,
+        )
+    }
+    runtime.ensure_resources_ready = lambda: None
+    runtime._image_targets = {}
+    runtime.last_images = {}
+    runtime._observation_calls = {}
+    runtime._recordings = None
+
+    def unexpected_recorder(*_args):
+        pytest.fail("Recording requires an owning task trace")
+
+    monkeypatch.setattr("artemis.plugins.recording.PluginRecordings", unexpected_recorder)
+    outcome = await runtime.call("observe", {"app": "fixture"})
+    assert outcome.status == "success"
+    assert outcome.source.target_id == "123"
+    assert outcome.source_id.startswith("plugin:visual-test:")
+    assert runtime.last_images["visual-test"] == b"image"
+    assert runtime._recordings is None
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "target,pre_valid,cleared",
+    [
+        ("computer", False, True),
+        ("computer", True, False),
+        ("android", False, False),
+        ("mixed", False, False),
+    ],
+)
+async def test_pro_cross_window_evidence_does_not_reuse_another_source_before_image(
+    tmp_path, monkeypatch, target, pre_valid, cleared
+):
+    from artemis.agents.validator import execution_loop
+    from artemis.data_engine.engine import DataEngine
+    from artemis.data_engine.models import ImageRecord, SessionMetadata, StepRecord
+    from artemis.data_engine.storage import StorageManager
+    from apps.admin_console.services import report_export
+
+    storage = StorageManager(tmp_path / "evidence.db", tmp_path)
+    session_id, step_id = uuid4(), uuid4()
+    storage.create_session(SessionMetadata(session_id=session_id, initial_goal="fixture"))
+    storage.create_image(ImageRecord(image_name="window-a"))
+    storage.create_image(ImageRecord(image_name="window-b"))
+    storage.create_step(
+        StepRecord(step_id=step_id, session_id=session_id, step_number=1, pre_image_name="window-a")
+    )
+    engine = DataEngine.__new__(DataEngine)
+    engine.storage = storage
+    engine._run_in_background = lambda fn, *args: fn(*args)
+    engine.get_step_number = lambda _step_id: 1
+    engine._get_generic_tools_for_sse = lambda _step_id: []
+    engine._publish = Mock()
+    engine.record_trace = Mock()
+    engine.get_relative_time = lambda _timestamp: "0s"
+    ctx = SimpleNamespace(execution_target=target, data_engine=engine)
+    node = SimpleNamespace(
+        ctx=ctx,
+        _parse_decisions=lambda _: ([{"action": "plugin-click"}], None),
+        _get_mcp_session=AsyncMock(),
+        _get_initial_screenshot=AsyncMock(return_value=("old-screen", "window-a")),
+    )
+    state = SimpleNamespace(
+        structured_decisions=None, current_step_id=str(step_id), open_incident=None
+    )
+    outcome = execution_loop._ActionOutcome(
+        action_item={
+            "action": "plugin-click",
+            "plugin_id": "visual-test",
+            "source_id": "window-b",
+            "post_image_name": "window-b",
+            "pre_image_valid": pre_valid,
+        },
+        success=True,
+        post_screenshot_b64="new-screen",
+    )
+    monkeypatch.setattr(execution_loop, "_process_action", AsyncMock(return_value=outcome))
+    monkeypatch.setattr(execution_loop, "_absorb_post_screenshot", lambda *_args: "window-b")
+    await execution_loop.run_validation_loop(node, state)
+    step = storage.get_steps(session_id)[0].model_dump()
+    assert (step["pre_image_name"] is None) is cleared
+    if cleared:
+        assert step["post_image_name"] == "window-b"
+        assert engine._publish.call_args.args[1]["pre_image_name"] is None
+        before = tmp_path / "window-a.png"
+        after = tmp_path / "window-b.png"
+        before.write_bytes(b"a")
+        after.write_bytes(b"b")
+        monkeypatch.setattr(
+            report_export,
+            "image_path",
+            lambda name: {"window-a": before, "window-b": after}.get(name),
+        )
+        exported = report_export.session_screenshots({"steps": [step]})
+        assert [path for path, _name in exported] == [after]
+        assert all("操作前" not in name for _path, name in exported)

tests/unit/test_plugin_recording.py

@@ -0,0 +1,391 @@
+"""Recorder lifecycle and archive limits use fake processes, never a live desktop."""
+
+import asyncio
+import json
+from pathlib import Path
+from types import SimpleNamespace
+from unittest.mock import AsyncMock
+
+import pytest
+
+from artemis.config.agent import RecordingConfig
+from artemis.plugins.models import EvidenceSource
+from artemis.plugins import recording
+from artemis.plugins.recording import PluginRecordings, build_capture_command, capture_rectangle
+
+
+def source(**updates):
+    return EvidenceSource(
+        **{
+            "kind": "computer",
+            "target_id": "123",
+            "label": "Fixture window",
+            "width": 800,
+            "height": 600,
+            "window": {"left": 40, "top": 50, "width": 800, "height": 600},
+            "screen": {"left": -1920, "top": 0, "width": 1920, "height": 1080},
+            **updates,
+        }
+    )
+
+
+def context(root, *, enabled=True, **config):
+    return SimpleNamespace(
+        data_engine=SimpleNamespace(base_dir=root, session_start_time=100),
+        agent_config=SimpleNamespace(recording=RecordingConfig(**config)),
+        execution_setup=SimpleNamespace(
+            record_video_enabled=enabled, video_recording_tools_enabled=False
+        ),
+    )
+
+
+class FakeRecorderProcess:
+    def __init__(self, *, produces_frames=True):
+        self.pid = 123456
+        self.returncode = None if produces_frames else 1
+        self.stdout = asyncio.StreamReader()
+        if produces_frames:
+            self.stdout.feed_data(b"frame=1\n")
+        self.stdout.feed_eof()
+        self.stdin = SimpleNamespace(write=self._quit, drain=AsyncMock())
+        self.killed = False
+
+    def _quit(self, _data):
+        self.returncode = 0
+
+    async def wait(self):
+        return self.returncode
+
+    def kill(self):
+        self.killed = True
+        self.returncode = -9
+
+
+@pytest.fixture
+def fake_capture(monkeypatch):
+    state = SimpleNamespace(commands=[], processes=[], fail_hevc=False)
+
+    async def spawn(*command, **_kwargs):
+        state.commands.append(command)
+        encoder = command[command.index("-c:v") + 1]
+        process = FakeRecorderProcess(
+            produces_frames=not (state.fail_hevc and encoder.startswith("hevc"))
+        )
+        state.processes.append(process)
+        Path(command[-1]).write_bytes(b"fake mp4")
+        return process
+
+    local_asyncio = SimpleNamespace(**vars(asyncio))
+    local_asyncio.create_subprocess_exec = spawn
+    monkeypatch.setattr(recording, "asyncio", local_asyncio)
+    monkeypatch.setattr(recording, "os", SimpleNamespace(name="nt", getpid=lambda: 123))
+    monkeypatch.setattr(
+        recording,
+        "CaptureProcessJob",
+        lambda: SimpleNamespace(assign=lambda _pid: None, close=lambda: None),
+    )
+    monkeypatch.setattr(
+        recording, "shutil", SimpleNamespace(disk_usage=lambda _path: SimpleNamespace(free=10**10))
+    )
+    monkeypatch.setattr(recording, "time", SimpleNamespace(time=lambda: 123.0))
+    monkeypatch.setattr(recording, "get_ffmpeg_path", lambda: "fake-ffmpeg")
+    monkeypatch.setattr(
+        recording,
+        "probe_video_segment",
+        AsyncMock(return_value={"duration": 2.5, "width": 1280, "height": 720}),
+    )
+    return state
+
+
+def argument(command, option):
+    return command[command.index(option) + 1]
+
+
+def test_archive_command_keeps_physical_capture_and_limits_encoded_output(tmp_path):
+    command = build_capture_command(
+        "ffmpeg", source(), RecordingConfig(), tmp_path / "video.mp4", "libx264"
+    )
+    assert argument(command, "-offset_x") == "-1920"
+    assert argument(command, "-video_size") == "1920x1080"
+    assert argument(command, "-vf") == "scale=1280:720"
+    assert argument(command, "-framerate") == "15"
+    assert argument(command, "-b:v") == "1500k"
+    assert argument(command, "-g") == "30"
+    assert "-an" in command
+
+
+def test_small_window_is_not_upscaled_and_hwnd_capture_is_explicit(tmp_path):
+    command = build_capture_command(
+        "ffmpeg",
+        source(),
+        RecordingConfig(computer_scope="window", codec="h265"),
+        tmp_path / "video.mp4",
+        "hevc_nvenc",
+    )
+    assert argument(command, "-i") == "hwnd=123"
+    assert argument(command, "-vf") == "scale=800:600"
+    assert argument(command, "-b:v") == "1000k"
+    assert argument(command, "-tag:v") == "hvc1"
+    assert "-offset_x" not in command
+
+
+@pytest.mark.parametrize(
+    "area",
+    [
+        None,
+        {"left": 0},
+        {"left": 0, "top": 0, "width": 0, "height": 600},
+        {"left": 65537, "top": 0, "width": 800, "height": 600},
+    ],
+)
+def test_invalid_capture_geometry_is_not_sent_to_ffmpeg(area):
+    with pytest.raises(ValueError):
+        capture_rectangle(source(screen=area), "screen")
+
+
+def test_window_recording_rejects_non_hwnd_targets(tmp_path):
+    with pytest.raises(ValueError, match="HWND"):
+        build_capture_command(
+            "ffmpeg",
+            source(target_id="window-title"),
+            RecordingConfig(computer_scope="window"),
+            tmp_path / "video.mp4",
+            "libx264",
+        )
+
+
+@pytest.mark.asyncio
+async def test_disabled_recording_does_not_spawn_or_create_evidence(tmp_path, fake_capture):
+    recorder = PluginRecordings(context(tmp_path, enabled=False))
+    await recorder.observe("source-1", "visual-test", source())
+    await recorder.stop()
+    assert fake_capture.commands == []
+    assert not (tmp_path / "evidence").exists()
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "codec,hevc_playback,expected_attempts",
+    [("auto", False, 1), ("auto", True, 4), ("h265", False, 4)],
+)
+async def test_hardware_hevc_failure_falls_back_to_compatible_archive(
+    tmp_path, fake_capture, codec, hevc_playback, expected_attempts
+):
+    fake_capture.fail_hevc = True
+    recorder = PluginRecordings(context(tmp_path, codec=codec, hevc_playback=hevc_playback))
+    try:
+        await recorder.observe("source-1", "visual-test", source())
+        assert len(fake_capture.commands) == expected_attempts
+        assert argument(fake_capture.commands[-1], "-c:v") == "libx264"
+        assert len(recorder.active) == 1
+    finally:
+        await recorder.stop()
+    track = next(iter(recorder.sources.values()))
+    assert track["status"] == "ready"
+    assert len(track["segments"]) == 1
+    assert track["segments"][0]["codec"] == "h264"
+    assert track["segments"][0]["duration_ms"] == 2500
+    assert all(process.returncode is not None for process in fake_capture.processes)
+
+
+@pytest.mark.asyncio
+async def test_same_screen_reuses_recording_and_keeps_both_target_ids(tmp_path, fake_capture):
+    recorder = PluginRecordings(context(tmp_path))
+    try:
+        await recorder.observe("window-1", "visual-test", source())
+        await recorder.observe("window-2", "visual-test", source(target_id="456"))
+        assert len(fake_capture.commands) == 1
+    finally:
+        await recorder.stop()
+    assert next(iter(recorder.sources.values()))["target_ids"] == ["window-1", "window-2"]
+
+
+@pytest.mark.asyncio
+async def test_low_disk_space_records_failure_without_starting_capture(
+    tmp_path, fake_capture, monkeypatch
+):
+    monkeypatch.setattr(recording.shutil, "disk_usage", lambda _path: SimpleNamespace(free=1))
+    recorder = PluginRecordings(context(tmp_path))
+    await recorder.observe("source-1", "visual-test", source())
+    await recorder.stop()
+    assert fake_capture.commands == []
+    track = next(iter(recorder.sources.values()))
+    assert track["status"] == "failed"
+    assert "磁盘" in track["message"]
+
+
+@pytest.mark.asyncio
+async def test_index_write_failure_does_not_leave_capture_processes(
+    tmp_path, fake_capture, monkeypatch
+):
+    recorder = PluginRecordings(context(tmp_path))
+
+    def unavailable():
+        raise OSError("disk full")
+
+    monkeypatch.setattr(recorder, "_save", unavailable)
+    await recorder.observe("source-1", "visual-test", source())
+    await recorder.stop()
+    assert not recorder.active
+    assert all(process.returncode is not None for process in fake_capture.processes)
+    assert next(iter(recorder.sources.values()))["status"] == "failed"
+
+
+@pytest.mark.asyncio
+async def test_probe_failure_during_stop_is_recorded_and_capture_is_reaped(
+    tmp_path, fake_capture, monkeypatch
+):
+    recorder = PluginRecordings(context(tmp_path))
+    await recorder.observe("source-1", "visual-test", source())
+    monkeypatch.setattr(
+        recording, "probe_video_segment", AsyncMock(side_effect=OSError("probe unavailable"))
+    )
+    await recorder.stop()
+    assert not recorder.active
+    assert all(process.returncode is not None for process in fake_capture.processes)
+    track = next(iter(recorder.sources.values()))
+    assert track["status"] == "failed"
+    assert "probe unavailable" in track["message"]
+
+
+@pytest.mark.asyncio
+async def test_stop_retries_process_cleanup_interrupted_in_watchdog(
+    tmp_path, fake_capture, monkeypatch
+):
+    recorder = PluginRecordings(context(tmp_path))
+    await recorder.observe("source-1", "visual-test", source())
+    recorder._watchdog.cancel()
+    await asyncio.gather(recorder._watchdog, return_exceptions=True)
+    key = next(iter(recorder.active))
+    entered = asyncio.Event()
+    attempts = 0
+    terminate = recorder._terminate
+
+    async def interrupted_first(process):
+        nonlocal attempts
+        attempts += 1
+        if attempts == 1:
+            entered.set()
+            await asyncio.Future()
+        await terminate(process)
+
+    monkeypatch.setattr(recorder, "_terminate", interrupted_first)
+    recorder._watchdog = asyncio.create_task(recorder._finish(key))
+    await entered.wait()
+    await recorder.stop()
+    assert attempts == 2
+    assert not recorder.active
+    assert all(process.returncode is not None for process in fake_capture.processes)
+
+
+@pytest.mark.asyncio
+async def test_stop_preserves_clip_when_watchdog_is_cancelled_during_probe(
+    tmp_path, fake_capture, monkeypatch
+):
+    recorder = PluginRecordings(context(tmp_path))
+    await recorder.observe("source-1", "visual-test", source())
+    recorder._watchdog.cancel()
+    await asyncio.gather(recorder._watchdog, return_exceptions=True)
+    key = next(iter(recorder.active))
+    entered = asyncio.Event()
+    attempts = 0
+
+    async def interrupted_probe(_output):
+        nonlocal attempts
+        attempts += 1
+        if attempts == 1:
+            entered.set()
+            await asyncio.Future()
+        return {"duration": 2.5, "width": 1280, "height": 720}
+
+    monkeypatch.setattr(recording, "probe_video_segment", interrupted_probe)
+    recorder._watchdog = asyncio.create_task(recorder._finish(key))
+    await entered.wait()
+    await recorder.stop()
+    assert attempts == 2
+    assert not recorder.active
+    track = recorder.sources[key]
+    assert track["status"] == "ready"
+    assert track["segments"][0]["duration_ms"] == 2500
+
+
+@pytest.mark.asyncio
+async def test_resumed_recorder_appends_and_does_not_overwrite_previous_clips(
+    tmp_path, fake_capture
+):
+    first = PluginRecordings(context(tmp_path))
+    await first.observe("source-1", "visual-test", source())
+    await first.stop()
+    original = next(iter(first.sources.values()))["segments"][0].copy()
+    original_path = tmp_path / original["file"]
+    original_path.write_bytes(b"immutable original clip")
+    second = PluginRecordings(context(tmp_path))
+    try:
+        await second.observe("source-1", "visual-test", source())
+    finally:
+        await second.stop()
+    segments = next(iter(second.sources.values()))["segments"]
+    assert len(segments) == 2
+    assert segments[0] == original
+    assert segments[1]["file"] != original["file"]
+    assert original_path.read_bytes() == b"immutable original clip"
+    manifest = json.loads(
+        (tmp_path / "evidence/recordings/recordings.json").read_text(encoding="utf-8")
+    )
+    assert len(manifest["sources"][0]["segments"]) == 2
+
+
+@pytest.mark.asyncio
+async def test_resized_window_seals_old_geometry_before_new_capture(tmp_path, fake_capture):
+    recorder = PluginRecordings(context(tmp_path, computer_scope="window"))
+    try:
+        await recorder.observe("source-1", "visual-test", source())
+        await recorder.observe(
+            "source-2",
+            "visual-test",
+            source(window={"left": 40, "top": 50, "width": 1000, "height": 700}),
+        )
+        assert len(fake_capture.commands) == 2
+        assert fake_capture.processes[0].returncode is not None
+        assert len(recorder.active) == 1
+    finally:
+        await recorder.stop()
+    assert len(recorder.sources) == 2
+    assert all(track["status"] == "ready" for track in recorder.sources.values())
+
+
+@pytest.mark.asyncio
+async def test_dead_worker_recovery_keeps_partial_clip_and_does_not_restart(
+    tmp_path, fake_capture, monkeypatch
+):
+    recorder = PluginRecordings(context(tmp_path))
+    await recorder.observe("source-1", "visual-test", source())
+    await recorder.stop()
+    track = next(iter(recorder.sources.values()))
+    track.update(status="recording", owner_pid=987654321, owner_created_at=1)
+    track["segments"][0].update(duration=0, duration_ms=0)
+    recorder._save()
+    count = len(fake_capture.commands)
+    assert await recording.recover_interrupted_recordings(tmp_path)
+    recovered = PluginRecordings(context(tmp_path))
+    result = next(iter(recovered.sources.values()))
+    assert result["status"] == "failed"
+    assert result["segments"][0]["duration"] == 2.5
+    assert len(fake_capture.commands) == count
+
+
+@pytest.mark.asyncio
+async def test_live_worker_recovery_never_finalizes_active_capture(tmp_path, fake_capture):
+    import os
+    import psutil
+
+    recorder = PluginRecordings(context(tmp_path))
+    await recorder.observe("source-1", "visual-test", source())
+    track = next(iter(recorder.sources.values()))
+    track.update(owner_pid=os.getpid(), owner_created_at=psutil.Process().create_time())
+    recorder._save()
+    try:
+        assert not await recording.recover_interrupted_recordings(tmp_path)
+        assert track["status"] == "recording"
+    finally:
+        await recorder.stop()

tests/unit/test_plugins.py

@@ -671,6 +671,7 @@ async def test_flash_plugin_images_never_replace_the_mixed_phone_observation(mon
         bindings={name: ("test-plugin", None)},
         tool_names={name},
         last_images={"test-plugin": b"desktop-before"},
+        previous_image=lambda name, args: b"desktop-before",
     )
     monkeypatch.setattr("artemis.plugins.runtime.get_plugin_runtime", lambda ctx: plugins)
     capture = AsyncMock(return_value=("phone-path", b"phone-after", "phone-tree-after"))

tests/unit/test_recording_config.py

@@ -0,0 +1,79 @@
+"""Per-task archive quality overrides outrank JSON and saved environment defaults."""
+
+import json
+
+import pytest
+
+from apps.admin_console.services.task_import import TaskCase, TaskOptions
+from artemis.config.agent import AgentGlobalConfig, load_agent_config
+from artemis.config.settings import settings
+
+
+@pytest.fixture(autouse=True)
+def recording_defaults(monkeypatch):
+    monkeypatch.setattr(settings, "ARTEMIS_RECORDING_CONFIG", {"fps": 10, "h264_kbps": 900})
+    monkeypatch.setattr(settings, "ARTEMIS_TASK_RECORDING_CONFIG", {})
+
+
+def test_json_recording_defaults_override_saved_defaults(tmp_path):
+    config = tmp_path / "artemis.jsonc"
+    config.write_text(json.dumps({"agent": {"recording": {"fps": 20}}}), encoding="utf-8")
+    result = load_agent_config(config).recording
+    assert result.fps == 20
+    assert result.h264_kbps == 900
+    assert result.max_long_edge == 1280
+
+
+def test_task_recording_override_has_highest_priority_and_is_partial(tmp_path, monkeypatch):
+    monkeypatch.setattr(settings, "ARTEMIS_TASK_RECORDING_CONFIG", {"fps": 30})
+    config = tmp_path / "artemis.jsonc"
+    config.write_text(
+        json.dumps({"agent": {"recording": {"fps": 20, "max_long_edge": 1920}}}), encoding="utf-8"
+    )
+    result = load_agent_config(config).recording
+    assert result.fps == 30
+    assert result.max_long_edge == 1920
+    assert result.h264_kbps == 900
+
+
+def test_task_recording_override_applies_without_json_recording_block(monkeypatch):
+    monkeypatch.setattr(settings, "ARTEMIS_TASK_RECORDING_CONFIG", {"fps": 30})
+    assert AgentGlobalConfig().recording.fps == 30
+
+
+def test_invalid_task_recording_override_is_rejected(monkeypatch):
+    monkeypatch.setattr(settings, "ARTEMIS_TASK_RECORDING_CONFIG", {"fps": 500})
+    with pytest.raises(ValueError, match="fps"):
+        AgentGlobalConfig.model_validate({"recording": {"fps": 20}})
+
+
+def test_task_and_batch_serialization_do_not_inject_unset_recording_defaults():
+    options = TaskOptions.model_validate({"recording": {"fps": 30}})
+    assert options.model_dump(exclude_none=True)["recording"] == {"fps": 30}
+    case = TaskCase(name="fixture", goal="fixture", task_config=options)
+    assert case.model_dump()["task_config"]["recording"] == {"fps": 30}
+    # Round trips through stored batches preserve the partial override.
+    restored = TaskCase.model_validate(case.model_dump())
+    assert restored.model_dump()["task_config"]["recording"] == {"fps": 30}
+
+
+def test_worker_env_keeps_task_quality_separate_from_saved_defaults(monkeypatch):
+    from apps.admin_console.services.task_queue_service import TaskQueueService
+    from artemis.runtime.adb_endpoint import AdbEndpoint, AdbTarget
+
+    monkeypatch.setenv("ARTEMIS_RECORDING_CONFIG", '{"fps":10}')
+    monkeypatch.setenv("ARTEMIS_TASK_RECORDING_CONFIG", '{"fps":60}')
+    _, environment = TaskQueueService._build_worker_invocation(
+        {"runtime_overrides": {"recording": {"fps": 30}}},
+        "test-run",
+        None,
+        "fixture",
+        "flash",
+        AdbTarget(AdbEndpoint.local()),
+    )
+    assert json.loads(environment["ARTEMIS_RECORDING_CONFIG"]) == {"fps": 10}
+    assert json.loads(environment["ARTEMIS_TASK_RECORDING_CONFIG"]) == {"fps": 30}
+    _, clean_environment = TaskQueueService._build_worker_invocation(
+        {}, "test-run", None, "fixture", "flash", AdbTarget(AdbEndpoint.local())
+    )
+    assert "ARTEMIS_TASK_RECORDING_CONFIG" not in clean_environment

tests/unit/test_recording_process.py

@@ -0,0 +1,42 @@
+"""Exercise Windows ownership without recording a user screen or running FFmpeg."""
+
+import os
+from pathlib import Path
+import subprocess
+import sys
+import time
+
+import psutil
+import pytest
+
+
+@pytest.mark.skipif(os.name != "nt", reason="Windows capture process ownership")
+def test_worker_exit_terminates_owned_capture_child():
+    project = str(Path(__file__).resolve().parents[2])
+    script = f"""
+import os, subprocess, sys
+sys.path.insert(0, {project!r})
+from artemis.plugins.recording_process import CaptureProcessJob
+job = CaptureProcessJob()
+child = subprocess.Popen([sys.executable, '-c', 'import time; time.sleep(120)'],
+    creationflags=0x08000000, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+job.assign(child.pid)
+print(child.pid, flush=True)
+os._exit(0)
+"""
+    parent = subprocess.run(
+        [sys.executable, "-c", script],
+        capture_output=True,
+        text=True,
+        timeout=15,
+        creationflags=0x08000000,
+        check=True,
+    )
+    pid = int(parent.stdout.strip())
+    for _ in range(50):
+        if not psutil.pid_exists(pid):
+            return
+        time.sleep(0.1)
+    # Only the disposable child created above is eligible for cleanup.
+    psutil.Process(pid).kill()
+    pytest.fail("Owned capture child survived abrupt worker exit")

tests/unit/test_unified_controller_video.py

@@ -714,6 +714,32 @@ async def test_start_recording_anchors_timeline_to_first_frame(mock_ctx, tmp_pat
         remove_active_session("emulator-5554")
 
 
+@pytest.mark.asyncio
+async def test_unconfirmed_recording_start_fails_and_closes_own_process(mock_ctx, tmp_path):
+    controller = UnifiedMobileController(mock_ctx)
+    remove_active_session("emulator-5554")
+    proc = _FakeScrcpyProcess([])
+    with (
+        patch.object(controller, "_spawn_scrcpy", AsyncMock(return_value=proc)),
+        patch.object(controller, "_stop_scrcpy", AsyncMock()) as stop,
+        patch(
+            "artemis.controllers.unified_controller.get_android_display_state",
+            AsyncMock(return_value=None),
+        ),
+        patch(
+            "artemis.controllers.unified_controller.await_scrcpy_first_frame",
+            AsyncMock(side_effect=TimeoutError("no first frame")),
+        ) as first,
+    ):
+        result = await controller.start_video_recording(output_dir=tmp_path)
+    assert not result.success
+    assert "no first frame" in result.message
+    assert first.call_args.kwargs["require_marker"] is True
+    stop.assert_awaited()
+    assert get_active_session("emulator-5554") is None
+    mock_ctx.data_engine.record_video_failure.assert_called_once()
+
+
 @pytest.mark.asyncio
 async def test_next_segment_anchors_at_first_frame(mock_ctx, tmp_path):
     controller = UnifiedMobileController(mock_ctx)

tests/unit/utils/test_video_validation.py

@@ -0,0 +1,206 @@
+"""Reject corrupt recording timelines and recover only actually captured frames."""
+
+import asyncio
+import json
+from pathlib import Path
+import subprocess
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, Mock
+
+import pytest
+
+from artemis.utils import video
+
+
+def metadata(**updates):
+    return {
+        "streams": [
+            {
+                "width": 640,
+                "height": 360,
+                "duration": "1.0",
+                "start_time": "0",
+                "nb_frames": "15",
+                "avg_frame_rate": "15/1",
+            }
+        ],
+        "format": {"duration": "1.0"},
+        "packets": [{"pts_time": "0.0", "flags": "K__"}, {"pts_time": "0.9", "flags": "___"}],
+        **updates,
+    }
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "change",
+    ["microsecond-duration", "pts-outside-duration", "nan-duration", "no-packets", "no-dimensions"],
+)
+async def test_probe_rejects_invalid_timeline_despite_valid_container_exit(monkeypatch, change):
+    payload = metadata()
+    stream = payload["streams"][0]
+    if change == "microsecond-duration":
+        stream.update(duration="0.000375", avg_frame_rate="56000/3", nb_frames="7")
+    elif change == "pts-outside-duration":
+        payload["packets"][-1]["pts_time"] = "3"
+    elif change == "nan-duration":
+        stream["duration"] = "NaN"
+    elif change == "no-packets":
+        payload["packets"] = []
+    else:
+        stream["width"] = 0
+    process = SimpleNamespace(
+        returncode=0, communicate=AsyncMock(return_value=(json.dumps(payload).encode(), b""))
+    )
+    monkeypatch.setattr(video.asyncio, "create_subprocess_exec", AsyncMock(return_value=process))
+    assert await video.probe_video_segment(Path("invalid.mp4")) == {}
+
+
+@pytest.mark.asyncio
+async def test_probe_accepts_vfr_timing_without_inventing_extra_duration(monkeypatch):
+    payload = metadata()
+    payload["streams"][0].update(duration="8", avg_frame_rate="1/4", nb_frames="2")
+    payload["packets"][-1]["pts_time"] = "7.9"
+    process = SimpleNamespace(
+        returncode=0, communicate=AsyncMock(return_value=(json.dumps(payload).encode(), b""))
+    )
+    monkeypatch.setattr(video.asyncio, "create_subprocess_exec", AsyncMock(return_value=process))
+    assert (await video.probe_video_segment(Path("valid.mp4")))["duration"] == 8
+
+
+@pytest.mark.asyncio
+async def test_remux_never_publishes_invalid_output_and_preserves_source(tmp_path, monkeypatch):
+    source, output = tmp_path / "source.mkv", tmp_path / "recording.mp4"
+    source.write_bytes(b"original recording")
+    commands = []
+
+    async def spawn(*command, **_kwargs):
+        commands.append(command)
+        Path(command[-1]).write_bytes(b"nonempty but invalid")
+        return SimpleNamespace(returncode=0, communicate=AsyncMock(return_value=(b"", b"")))
+
+    monkeypatch.setattr(video.asyncio, "create_subprocess_exec", spawn)
+    monkeypatch.setattr(video, "probe_video_segment", AsyncMock(return_value={}))
+    assert not await video.remux_recording_to_mp4(source, output)
+    assert source.read_bytes() == b"original recording"
+    assert not output.exists()
+    assert not output.with_name("recording.part.mp4").exists()
+    assert len(commands) == 2
+    fallback = commands[1]
+    assert fallback[fallback.index("-fps_mode") + 1] == "passthrough"
+    assert "-r" not in fallback
+    assert "setpts=PTS-STARTPTS" in fallback
+
+
+@pytest.mark.asyncio
+async def test_bounded_media_wait_reaps_process_on_timeout():
+    async def stalled():
+        await asyncio.Future()
+
+    process = SimpleNamespace(
+        returncode=None, communicate=stalled, kill=Mock(), wait=AsyncMock(return_value=-9)
+    )
+    with pytest.raises(TimeoutError):
+        await video._communicate_video_process(process, 0.01)
+    process.kill.assert_called_once()
+    process.wait.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_strict_scrcpy_start_rejects_eof_without_recording_marker():
+    reader = asyncio.StreamReader()
+    reader.feed_data(b"INFO: connection initialized\n")
+    reader.feed_eof()
+    process = SimpleNamespace(stdout=reader, wait=AsyncMock(return_value=0))
+    with pytest.raises(OSError, match="未确认开始录制"):
+        await video.await_scrcpy_first_frame(process, 100, require_marker=True)
+
+
+@pytest.mark.asyncio
+async def test_strict_scrcpy_start_rejects_silent_process_timeout():
+    process = SimpleNamespace(stdout=asyncio.StreamReader())
+    with pytest.raises(TimeoutError, match="启动时限"):
+        await video.await_scrcpy_first_frame(process, 100, timeout=0.01, require_marker=True)
+
+
+@pytest.mark.asyncio
+async def test_strict_scrcpy_start_accepts_actual_marker():
+    reader = asyncio.StreamReader()
+    reader.feed_data(b"INFO: Recording started to file: recording.mkv\n")
+    process = SimpleNamespace(stdout=reader)
+    assert await video.await_scrcpy_first_frame(process, 100, require_marker=True) >= 100
+
+
+def run_media(*command):
+    return subprocess.run(command, check=True, capture_output=True, timeout=30).stdout
+
+
+def frame_times(path):
+    payload = json.loads(
+        run_media(
+            video.get_ffprobe_path(),
+            "-v",
+            "error",
+            "-select_streams",
+            "v:0",
+            "-show_frames",
+            "-show_entries",
+            "frame=pts_time",
+            "-of",
+            "json",
+            str(path),
+        )
+    )
+    return [float(frame["pts_time"]) for frame in payload["frames"]]
+
+
+@pytest.mark.asyncio
+async def test_real_remux_recovers_pts_without_filling_missing_capture_time(tmp_path):
+    ffmpeg = video.get_ffmpeg_path()
+    clean = tmp_path / "clean.mkv"
+    corrupt = tmp_path / "corrupt.mp4"
+    restored = tmp_path / "restored.mp4"
+    run_media(
+        ffmpeg,
+        "-hide_banner",
+        "-loglevel",
+        "error",
+        "-y",
+        "-f",
+        "lavfi",
+        "-i",
+        "testsrc2=size=64x96:rate=10:duration=0.7",
+        "-c:v",
+        "libx264",
+        "-bf",
+        "0",
+        "-pix_fmt",
+        "yuv420p",
+        str(clean),
+    )
+    # Preserve original frame PTS while compressing decode timestamps to 1 ms.
+    run_media(
+        ffmpeg,
+        "-hide_banner",
+        "-loglevel",
+        "error",
+        "-y",
+        "-i",
+        str(clean),
+        "-map",
+        "0:v:0",
+        "-c",
+        "copy",
+        "-bsf:v",
+        "setts=dts=N:duration=1",
+        str(corrupt),
+    )
+    assert await video.probe_video_segment(corrupt) == {}
+    captured = frame_times(corrupt)
+    assert await video.remux_recording_to_mp4(corrupt, restored)
+    result = await video.probe_video_segment(restored)
+    assert 0.4 <= result["duration"] <= 0.8
+    recovered = frame_times(restored)
+    assert recovered
+    assert len(recovered) <= len(captured)
+    assert recovered == pytest.approx(captured[: len(recovered)], abs=0.001)
+    assert corrupt.exists()