fixes

README.md

@@ -3,12 +3,13 @@
 Evanescere turns recorded livestreams into suggested, rendered clips:
 
 1. Poll IIS WebDAV for finished recordings.
-2. Wait until file size is unchanged across polling cycles.
+2. Record files visible during the first scan as a manual-only historical baseline.
-3. Download the stable source file to Framework-local storage.
+3. For files first seen later, wait until file size is unchanged across polling cycles.
-4. Remux FLV/H264 to MP4 and extract ASR-ready audio.
+4. Download the stable source file to Framework-local storage.
-5. Transcribe Mandarin audio through a FunASR-compatible API.
+5. Remux FLV/H264 to MP4 and extract ASR-ready audio.
-6. Ask DeepSeek for ranked timeline-aware clip suggestions.
+6. Transcribe Mandarin audio through a FunASR-compatible API.
-7. Generate subtitles, render clips, generate thumbnails with optional local image generation and VTuber overlay, and optionally upload.
+7. Ask DeepSeek for ranked timeline-aware clip suggestions.
+8. Generate subtitles, render clips, generate thumbnails with optional local image generation and VTuber overlay, and optionally upload.
 
 The Framework Arch server is the intended production host. The Hyper-V Arch VM can be used for development, API testing, and database/control-plane work.
 
@@ -28,7 +29,8 @@ Runtime data flow:
 
 ```text
 IIS WebDAV recordings
-  -> scheduler polls file size
+  -> first scan records existing files as manual-only baseline entries
+  -> later scans poll new file sizes
   -> PostgreSQL records video state
   -> Redis queues pipeline job
   -> worker downloads source to configured storage.local_root
@@ -67,8 +69,8 @@ The checked-in `docker-compose.yml` mounts local `./config.toml` there for the A
 ## Services And Ports
 
 - Frontend: `http://localhost:3000`
-- Backend API: `http://localhost:8000`
+- Backend API: `http://localhost:8080`
-- API docs: `http://localhost:8000/docs`
+- API docs: `http://localhost:8080/docs`
 - Redis: `localhost:6379` when exposed by Compose
 - PostgreSQL: only started by Compose when using the `local-db` profile
 
@@ -92,12 +94,14 @@ Run database migrations:
 docker compose run --rm api alembic upgrade head
 ```
 
-Bootstrap existing WebDAV files so old recordings are marked `existing_done` instead of auto-processed:
+The scheduler automatically records files visible during its first scan as `existing_done`. Those historical recordings appear in the UI but are not automatically processed. To establish that baseline manually before starting the scheduler:
 
 ```bash
 docker compose run --rm api evanescere bootstrap-existing
 ```
 
+In the frontend, click `Sync WebDAV` to perform the same safe first-time baseline import or to discover later files. Select any historical FLV row and click its play button to manually queue a test pipeline run.
+
 Useful test commands:
 
 ```bash
@@ -141,7 +145,7 @@ The frontend reads the backend URL from `frontend/public/config.js` at runtime:
 
 ```js
 window.__EVANESCERE_FRONTEND_CONFIG__ = {
-  apiBaseUrl: "http://localhost:8000"
+  apiBaseUrl: "http://192.168.1.44:8080"
 };
 ```
 
@@ -181,11 +185,11 @@ provider = "frame_overlay"
 
 This uses the extracted frame as the background, then programmatically overlays the title and character PNG using Pillow.
 
-To overlay the VTuber character, use a transparent PNG:
+To overlay the VTuber character, place transparent PNG variants in a directory. Evanescere picks one randomly for each thumbnail:
 
 ```toml
 [thumbnail]
-character_overlay_path = "/data/evanescere/assets/vtuber.png"
+character_overlay_dir = "/data/evanescere/assets/"
 character_scale = 0.42
 character_position = "bottom-right"
 ```
@@ -277,6 +281,8 @@ The full commented example lives in `config.example.toml`. These are the keys th
 | `deepseek` | `api_key` | empty | DeepSeek API key. Required when clip suggestion is enabled. |
 | `deepseek` | `model` | `deepseek-v4-pro` | Model used for clip suggestion. |
 | `deepseek` | `temperature` | `0.2` | Sampling temperature for clip suggestion. |
+| `llm_prompt` | `system` | VTuber clip editor prompt | Custom system prompt for clip selection. The required JSON schema is appended by code. |
+| `llm_prompt` | `user` | timestamped transcript prompt | Custom user prompt template. Supports documented placeholders. |
 | `defaults` | `suggest_enabled` | `true` | Initial setting for automatic LLM clip suggestion. |
 | `defaults` | `render_enabled` | `true` | Initial setting for automatic rendering. |
 | `defaults` | `upload_enabled` | `true` | Initial setting for automatic upload after render. For testing, set false. |
@@ -285,11 +291,12 @@ The full commented example lives in `config.example.toml`. These are the keys th
 | `clip` | `min_seconds` | `30` | Minimum LLM clip duration accepted by backend. |
 | `clip` | `max_seconds` | `360` | Maximum LLM clip duration accepted by backend. |
 | `clip` | `transcript_chunk_seconds` | `900` | Transcript seconds sent to DeepSeek per request. |
+| `clip` | `max_candidates_total` | `20` | Maximum suggestions kept across the entire transcript after overlap deduplication. |
 | `thumbnail` | `enabled` | `true` | Enables thumbnail generation during clip render. |
 | `thumbnail` | `provider` | `frame_overlay` | `frame_overlay` or `command`. |
 | `thumbnail` | `width` | `1920` | Final thumbnail width in pixels. |
 | `thumbnail` | `height` | `1080` | Final thumbnail height in pixels. |
-| `thumbnail` | `character_overlay_path` | empty | Optional transparent PNG of the VTuber character. |
+| `thumbnail` | `character_overlay_dir` | `/data/evanescere/assets/` | Directory of transparent VTuber PNG variants. One is selected randomly per thumbnail. |
 | `thumbnail` | `character_scale` | `0.42` | Character overlay height as a fraction of final thumbnail height. |
 | `thumbnail` | `character_position` | `bottom-right` | Character placement. |
 | `thumbnail` | `title_enabled` | `true` | Draws the clip title onto the final thumbnail. |
@@ -324,12 +331,42 @@ docker compose run --rm api
 
 Useful API endpoints for generated files:
 
+- `POST /webdav/scan`
 - `GET /videos/{video_id}/artifacts`
 - `GET /clips/{clip_id}/artifacts`
 - `GET /artifacts/{artifact_id}`
 
 Thumbnail outputs are visible through the artifact endpoints. Upload command payloads include `thumbnail` when a `thumbnail_final` artifact exists.
 
+## Custom LLM Prompt
+
+Edit `[llm_prompt]` in your private `config.toml` to tune clip selection without rebuilding containers:
+
+```toml
+[llm_prompt]
+system = """
+Your custom editorial guidance.
+"""
+
+user = """
+Choose up to {max_candidates} clips between {min_clip_seconds} and {max_clip_seconds} seconds.
+Stream range: {chunk_start_sec}-{chunk_end_sec}
+Transcript:
+{transcript}
+"""
+```
+
+Available user-prompt placeholders:
+
+- `{max_candidates}`
+- `{min_clip_seconds}`
+- `{max_clip_seconds}`
+- `{chunk_start_sec}`
+- `{chunk_end_sec}`
+- `{transcript}`
+
+The backend appends its strict JSON response schema to the system message so custom prompt wording cannot accidentally remove the machine-readable output contract.
+
 ## Debugging
 
 Set verbose logs in `config.toml`:
@@ -352,6 +389,7 @@ docker compose logs -f api
 What to look for:
 
 - Scheduler logs `webdav propfind done`, `new webdav file observed`, and `webdav file stable`.
+- On first startup, scheduler logs `webdav baseline missing` and stores all currently visible recordings as `existing_done`.
 - Worker logs `pipeline start`, `prepare media`, `transcription start`, `clip suggestion start`, `render start`, and `upload start`.
 - ffmpeg command failures include the failing command and stderr tail.
 - DeepSeek logs include transcript chunk bounds, character counts, candidate counts, and usage when returned by the SDK.
@@ -371,4 +409,4 @@ This is intentionally noisy and should usually stay off in production.
 
 - FunASR/ROCm setup is intentionally outside the main Compose file for now. Benchmark the Framework host manually before binding the project to a specific GPU runtime.
 - The uploader adapter is deliberately small. `[upload].adapter = "command"` is enough to integrate a Bilibili uploader later without changing the pipeline core.
-- Existing recordings should be bootstrapped before scheduler-driven production runs, otherwise old stable files may be queued as new work.
+- Existing recordings are automatically baselined on the first scheduler scan. The explicit `evanescere bootstrap-existing` command remains available when you want to establish or refresh the baseline before starting services.

config.example.toml

@@ -74,6 +74,30 @@ model = "deepseek-v4-pro"
 # Lower values improve consistency.
 temperature = 0.2
 
+[llm_prompt]
+# Customize these prompts freely. The backend appends the required JSON response schema.
+system = """
+You are an expert editor for Mandarin VTuber livestream clips.
+Identify moments that work as entertaining standalone clips for Bilibili viewers.
+Prioritize strong reactions, jokes, surprising turns, memorable conversations, and moments with a clear payoff.
+Avoid repetitive stretches, dead air, and segments that require too much missing context.
+Return only the requested JSON object.
+"""
+
+# Available placeholders:
+# {max_candidates}, {min_clip_seconds}, {max_clip_seconds},
+# {chunk_start_sec}, {chunk_end_sec}, {transcript}
+user = """
+Review the timestamped livestream transcript below.
+Choose up to {max_candidates} compelling clip candidates.
+Each clip must be between {min_clip_seconds} and {max_clip_seconds} seconds long.
+Use absolute stream timestamps and give each candidate a concise Chinese title and summary.
+
+Transcript range: {chunk_start_sec}-{chunk_end_sec}
+Transcript:
+{transcript}
+"""
+
 [defaults]
 # Initial automatic pipeline settings. These can later be changed through the API/UI.
 suggest_enabled = true
@@ -91,7 +115,10 @@ min_seconds = 30
 max_seconds = 360
 
 # Transcript seconds per DeepSeek request. Larger chunks use more tokens.
-transcript_chunk_seconds = 900
+transcript_chunk_seconds = 14400
+
+# Maximum number of suggestions kept across the entire transcript.
+max_candidates_total = 20
 
 [thumbnail]
 # Generate thumbnails during clip render.
@@ -104,8 +131,8 @@ provider = "frame_overlay"
 width = 1920
 height = 1080
 
-# Optional transparent PNG of the VTuber character.
+# Directory of transparent VTuber character PNGs. One is selected randomly per thumbnail.
-character_overlay_path = ""
+character_overlay_dir = "/data/evanescere/assets/"
 
 # Character height as a fraction of thumbnail height.
 character_scale = 0.42

docker-compose.yml

@@ -10,7 +10,7 @@ services:
     build: .
     command: uvicorn evanescere.api:app --host 0.0.0.0 --port 8000
     ports:
-      - "8000:8000"
+      - "8080:8000"
     volumes:
       - ./config.toml:/etc/evanescere/config.toml:ro
       - ./storage:/data/evanescere

frontend/public/config.js

@@ -1,3 +1,3 @@
 window.__EVANESCERE_FRONTEND_CONFIG__ = {
-  apiBaseUrl: "http://localhost:8000"
+  apiBaseUrl: "http://192.168.1.44:8080"
 };

frontend/src/App.tsx

@@ -1,6 +1,7 @@
 import {
   Check,
   Clapperboard,
+  FolderSync,
   Play,
   RefreshCw,
   Save,
@@ -19,6 +20,7 @@ import {
   patchSettings,
   renderClip,
   runVideo,
+  scanWebDav,
   uploadClip,
 } from "./api";
 import type { ClipSuggestion, PipelineSettings, TranscriptSegment, Video } from "./types";
@@ -47,7 +49,7 @@ function formatTimeRange(start: number, end: number) {
 }
 
 function statusTone(value: string) {
-  if (["done", "stable", "approved", "auto_approved"].includes(value)) return "good";
+  if (["done", "stable", "existing_done", "approved", "auto_approved"].includes(value)) return "good";
   if (["failed", "error"].includes(value)) return "bad";
   if (["running", "queued", "observing", "pending"].includes(value)) return "busy";
   return "neutral";
@@ -61,6 +63,7 @@ export function App() {
   const [clips, setClips] = useState<ClipSuggestion[]>([]);
   const [loading, setLoading] = useState(false);
   const [error, setError] = useState<string | null>(null);
+  const [notice, setNotice] = useState<string | null>(null);
 
   const selectedVideo = useMemo(
     () => videos.find((video) => video.id === selectedVideoId) ?? null,
@@ -124,6 +127,26 @@ export function App() {
     }
   }
 
+  async function syncWebDav() {
+    setLoading(true);
+    setError(null);
+    setNotice(null);
+    try {
+      const result = await scanWebDav();
+      if (result.baseline_initialized) {
+        setNotice(
+          `Imported ${result.baseline_inserted + result.baseline_marked_existing} existing files as manual-only baseline entries.`,
+        );
+      } else {
+        setNotice(`Observed ${result.observed} files; ${result.newly_stable} newly stable.`);
+      }
+      await refresh();
+    } catch (caught) {
+      setError(caught instanceof Error ? caught.message : "Unknown error");
+      setLoading(false);
+    }
+  }
+
   async function saveSettings() {
     if (!settings) return;
     await withRefresh(() => patchSettings(settings));
@@ -149,6 +172,7 @@ export function App() {
       </header>
 
       {error && <div className="error-strip">{error}</div>}
+      {notice && <div className="notice-strip">{notice}</div>}
 
       <section className="band controls-band">
         <div className="section-title">
@@ -176,10 +200,16 @@ export function App() {
 
       <section className="workbench">
         <aside className="video-list">
+          <div className="list-header">
             <div className="section-title">
               <Clapperboard size={18} />
               <h2>Videos</h2>
             </div>
+            <button type="button" className="icon-button" onClick={() => void syncWebDav()} disabled={loading}>
+              <FolderSync size={16} />
+              Sync WebDAV
+            </button>
+          </div>
           <div className="table-scroll">
             <table>
               <thead>
@@ -301,4 +331,3 @@ function Metric({ label, value }: { label: string; value: string }) {
     </div>
   );
 }
-

frontend/src/api.ts

@@ -1,8 +1,15 @@
-import type { Artifact, ClipSuggestion, PipelineSettings, TranscriptSegment, Video } from "./types";
+import type {
+  Artifact,
+  ClipSuggestion,
+  PipelineSettings,
+  TranscriptSegment,
+  Video,
+  WebDavScanResult,
+} from "./types";
 
 const runtimeApiBase = window.__EVANESCERE_FRONTEND_CONFIG__?.apiBaseUrl;
 
-export const apiBaseUrl = (runtimeApiBase || "http://localhost:8000").replace(
+export const apiBaseUrl = (runtimeApiBase || "http://localhost:8080").replace(
   /\/$/,
   "",
 );
@@ -37,6 +44,10 @@ export function getVideos(): Promise<Video[]> {
   return request<Video[]>("/videos");
 }
 
+export function scanWebDav(): Promise<WebDavScanResult> {
+  return request<WebDavScanResult>("/webdav/scan", { method: "POST" });
+}
+
 export function runVideo(videoId: number) {
   return request(`/videos/${videoId}/run`, { method: "POST" });
 }

frontend/src/styles.css

@@ -120,6 +120,16 @@ button {
   overflow-wrap: anywhere;
 }
 
+.notice-strip {
+  margin-bottom: 14px;
+  border-left: 4px solid #21756b;
+  background: #e9f3f1;
+  color: #155c54;
+  padding: 10px 12px;
+  border-radius: 6px;
+  font-size: 14px;
+}
+
 .band {
   border-top: 1px solid #d8dee6;
   padding: 16px 0;
@@ -190,6 +200,13 @@ button {
   padding-top: 16px;
 }
 
+.list-header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 12px;
+}
+
 .table-scroll {
   margin-top: 12px;
   overflow: auto;
@@ -431,4 +448,3 @@ tr.selected td {
     grid-template-columns: 1fr;
   }
 }
-

frontend/src/types.ts

@@ -59,3 +59,11 @@ export interface Artifact {
   artifact_metadata: Record<string, unknown>;
   created_at: string;
 }
+
+export interface WebDavScanResult {
+  observed: number;
+  newly_stable: number;
+  baseline_initialized: boolean;
+  baseline_inserted: number;
+  baseline_marked_existing: number;
+}

src/evanescere/api.py

@@ -21,7 +21,9 @@ from evanescere.schemas import (
     SettingsRead,
     TranscriptSegmentRead,
     VideoRead,
+    WebDavScanRead,
 )
+from evanescere.services.webdav import WebDavClient, safe_scan_once
 from evanescere.settings_store import get_pipeline_settings, patch_pipeline_settings
 
 configure_logging()
@@ -68,6 +70,12 @@ def list_videos(db: Session = Depends(get_db)) -> list[Video]:
     return list(db.scalars(select(Video).order_by(Video.created_at.desc())).all())
 
 
+@app.post("/webdav/scan", response_model=WebDavScanRead)
+def scan_webdav(db: Session = Depends(get_db)) -> WebDavScanRead:
+    result = safe_scan_once(WebDavClient.from_settings(), db)
+    return WebDavScanRead(**result.__dict__)
+
+
 @app.post("/videos/{video_id}/run", response_model=RunRead)
 def run_video(video_id: int, db: Session = Depends(get_db)) -> PipelineRun:
     video = db.get(Video, video_id)

src/evanescere/cli.py

@@ -5,7 +5,11 @@ from evanescere.db import session_scope
 from evanescere.jobs import enqueue_pipeline, enqueue_render, enqueue_transcribe, enqueue_suggest, enqueue_upload
 from evanescere.logging_config import configure_logging
 from evanescere.models import PipelineRun, Video
-from evanescere.services.webdav import WebDavClient, bootstrap_existing as bootstrap_webdav_existing, scan_once
+from evanescere.services.webdav import (
+    WebDavClient,
+    bootstrap_existing as bootstrap_webdav_existing,
+    safe_scan_once,
+)
 
 app = typer.Typer(no_args_is_help=True)
 configure_logging()
@@ -14,15 +18,21 @@ configure_logging()
 @app.command()
 def bootstrap_existing() -> None:
     with session_scope() as session:
-        count = bootstrap_webdav_existing(WebDavClient.from_settings(), session)
+        result = bootstrap_webdav_existing(WebDavClient.from_settings(), session)
-    typer.echo(f"Marked {count} existing WebDAV recordings as existing_done.")
+    typer.echo(
+        f"Observed {result.observed} recordings; inserted {result.inserted} and marked "
+        f"{result.marked_existing} tracked recordings as existing_done."
+    )
 
 
 @app.command()
 def scan() -> None:
     with session_scope() as session:
-        observed, stable = scan_once(WebDavClient.from_settings(), session)
+        result = safe_scan_once(WebDavClient.from_settings(), session)
-    typer.echo(f"Observed {observed} files; {stable} newly stable.")
+    typer.echo(
+        f"Observed {result.observed} files; {result.newly_stable} newly stable; "
+        f"baseline initialized: {result.baseline_initialized}."
+    )
 
 
 @app.command("run-video")

src/evanescere/config.py

@@ -12,6 +12,25 @@ DEFAULT_CONFIG_PATHS = (
     Path("config.toml"),
 )
 
+DEFAULT_LLM_SYSTEM_PROMPT = """\
+You are an expert editor for Mandarin VTuber livestream clips.
+Identify moments that work as entertaining standalone clips for Bilibili viewers.
+Prioritize strong reactions, jokes, surprising turns, memorable conversations, and moments with a clear payoff.
+Avoid repetitive stretches, dead air, and segments that require too much missing context.
+Return only the requested JSON object.
+"""
+
+DEFAULT_LLM_USER_PROMPT = """\
+Review the timestamped livestream transcript below.
+Choose up to {max_candidates} compelling clip candidates.
+Each clip must be between {min_clip_seconds} and {max_clip_seconds} seconds long.
+Use absolute stream timestamps and give each candidate a concise Chinese title and summary.
+
+Transcript range: {chunk_start_sec}-{chunk_end_sec}
+Transcript:
+{transcript}
+"""
+
 
 class Settings(BaseModel):
     app_env: str = "dev"
@@ -38,6 +57,8 @@ class Settings(BaseModel):
     deepseek_api_key: str | None = None
     deepseek_model: str = "deepseek-v4-pro"
     deepseek_temperature: float = 0.2
+    llm_system_prompt: str = DEFAULT_LLM_SYSTEM_PROMPT
+    llm_user_prompt: str = DEFAULT_LLM_USER_PROMPT
 
     default_suggest_enabled: bool = True
     default_render_enabled: bool = True
@@ -51,12 +72,13 @@ class Settings(BaseModel):
     clip_min_seconds: int = 30
     clip_max_seconds: int = 360
     transcript_chunk_seconds: int = 900
+    clip_max_candidates_total: int = Field(default=20, ge=1)
 
     thumbnail_enabled: bool = True
     thumbnail_provider: str = "frame_overlay"
     thumbnail_width: int = Field(default=1920, ge=320)
     thumbnail_height: int = Field(default=1080, ge=180)
-    thumbnail_character_overlay_path: str | None = None
+    thumbnail_character_overlay_dir: str | None = None
     thumbnail_character_scale: float = Field(default=0.42, gt=0, le=1)
     thumbnail_character_position: str = "bottom-right"
     thumbnail_title_enabled: bool = True
@@ -94,6 +116,7 @@ def flatten_config(raw: dict[str, Any]) -> dict[str, Any]:
     webdav = raw.get("webdav", {})
     funasr = raw.get("funasr", {})
     deepseek = raw.get("deepseek", {})
+    llm_prompt = raw.get("llm_prompt", {})
     defaults = raw.get("defaults", {})
     clip = raw.get("clip", {})
     thumbnail = raw.get("thumbnail", {})
@@ -119,6 +142,8 @@ def flatten_config(raw: dict[str, Any]) -> dict[str, Any]:
         "deepseek_api_key": blank_to_none(deepseek.get("api_key")),
         "deepseek_model": deepseek.get("model", "deepseek-v4-pro"),
         "deepseek_temperature": deepseek.get("temperature", 0.2),
+        "llm_system_prompt": llm_prompt.get("system", DEFAULT_LLM_SYSTEM_PROMPT),
+        "llm_user_prompt": llm_prompt.get("user", DEFAULT_LLM_USER_PROMPT),
         "default_suggest_enabled": defaults.get("suggest_enabled", True),
         "default_render_enabled": defaults.get("render_enabled", True),
         "default_upload_enabled": defaults.get("upload_enabled", True),
@@ -127,11 +152,12 @@ def flatten_config(raw: dict[str, Any]) -> dict[str, Any]:
         "clip_min_seconds": clip.get("min_seconds", 30),
         "clip_max_seconds": clip.get("max_seconds", 360),
         "transcript_chunk_seconds": clip.get("transcript_chunk_seconds", 900),
+        "clip_max_candidates_total": clip.get("max_candidates_total", 20),
         "thumbnail_enabled": thumbnail.get("enabled", True),
         "thumbnail_provider": thumbnail.get("provider", "frame_overlay"),
         "thumbnail_width": thumbnail.get("width", 1920),
         "thumbnail_height": thumbnail.get("height", 1080),
-        "thumbnail_character_overlay_path": blank_to_none(thumbnail.get("character_overlay_path")),
+        "thumbnail_character_overlay_dir": blank_to_none(thumbnail.get("character_overlay_dir")),
         "thumbnail_character_scale": thumbnail.get("character_scale", 0.42),
         "thumbnail_character_position": thumbnail.get("character_position", "bottom-right"),
         "thumbnail_title_enabled": thumbnail.get("title_enabled", True),

src/evanescere/pipeline.py

@@ -108,7 +108,12 @@ def transcribe_video(session: Session, video: Video) -> int:
 
 def suggest_clips(session: Session, video: Video) -> int:
     settings = get_settings()
-    logger.info("clip suggestion start video_id=%s chunk_seconds=%s", video.id, settings.transcript_chunk_seconds)
+    logger.info(
+        "clip suggestion start video_id=%s chunk_seconds=%s max_candidates_total=%s",
+        video.id,
+        settings.transcript_chunk_seconds,
+        settings.clip_max_candidates_total,
+    )
     segments = list(
         session.scalars(
             select(TranscriptSegment)
@@ -120,7 +125,12 @@ def suggest_clips(session: Session, video: Video) -> int:
         raise RuntimeError("Cannot suggest clips without transcript segments.")
     chunks = chunk_transcript(segments, settings.transcript_chunk_seconds)
     logger.info("clip suggestion chunks video_id=%s segments=%s chunks=%s", video.id, len(segments), len(chunks))
-    candidates = DeepSeekClipClient.from_settings().suggest(chunks)
+    candidates = DeepSeekClipClient.from_settings().suggest(
+        chunks,
+        max_candidates_total=settings.clip_max_candidates_total,
+        min_clip_seconds=settings.clip_min_seconds,
+        max_clip_seconds=settings.clip_max_seconds,
+    )
     session.query(ClipSuggestion).filter(ClipSuggestion.video_id == video.id).delete()
     for candidate in candidates:
         session.add(

src/evanescere/scheduler.py

@@ -10,7 +10,7 @@ from evanescere.db import session_scope
 from evanescere.jobs import enqueue_pipeline
 from evanescere.logging_config import configure_logging
 from evanescere.models import PipelineRun, Video
-from evanescere.services.webdav import WebDavClient, scan_once
+from evanescere.services.webdav import WebDavClient, safe_scan_once
 
 configure_logging()
 logger = logging.getLogger(__name__)
@@ -45,9 +45,15 @@ def run_forever() -> None:
     while True:
         try:
             with session_scope() as session:
-                observed, stable = scan_once(client, session)
+                result = safe_scan_once(client, session)
             queued = enqueue_stable_videos()
-            logger.info("scan observed=%s newly_stable=%s queued=%s", observed, stable, queued)
+            logger.info(
+                "scan observed=%s newly_stable=%s baseline_initialized=%s queued=%s",
+                result.observed,
+                result.newly_stable,
+                result.baseline_initialized,
+                queued,
+            )
         except Exception:
             logger.exception("scheduler scan failed")
         time.sleep(settings.webdav_poll_interval_seconds)

src/evanescere/schemas.py

@@ -95,6 +95,14 @@ class SettingsRead(BaseModel):
     bake_subtitles: bool = True
 
 
+class WebDavScanRead(BaseModel):
+    observed: int
+    newly_stable: int
+    baseline_initialized: bool = False
+    baseline_inserted: int = 0
+    baseline_marked_existing: int = 0
+
+
 class ClipCandidate(BaseModel):
     start_sec: float = Field(ge=0)
     end_sec: float = Field(gt=0)
@@ -108,4 +116,3 @@ class ClipCandidate(BaseModel):
 
 class ClipCandidateResponse(BaseModel):
     clips: list[ClipCandidate]
-

src/evanescere/services/llm.py

@@ -13,6 +13,11 @@ from evanescere.schemas import ClipCandidate, ClipCandidateResponse
 
 logger = logging.getLogger(__name__)
 
+RESPONSE_SCHEMA_INSTRUCTION = """\
+Return strict JSON only using this schema:
+{"clips":[{"start_sec":number,"end_sec":number,"title_zh":string,"summary_zh":string,"reason":string,"score":number,"tags":[string],"subtitle_priority":string}]}
+"""
+
 
 @dataclass(frozen=True)
 class TranscriptChunk:
@@ -22,10 +27,20 @@ class TranscriptChunk:
 
 
 class DeepSeekClipClient:
-    def __init__(self, api_key: str, base_url: str, model: str, temperature: float) -> None:
+    def __init__(
+        self,
+        api_key: str,
+        base_url: str,
+        model: str,
+        temperature: float,
+        system_prompt: str,
+        user_prompt: str,
+    ) -> None:
         self.client = OpenAI(api_key=api_key, base_url=base_url)
         self.model = model
         self.temperature = temperature
+        self.system_prompt = system_prompt
+        self.user_prompt = user_prompt
 
     @classmethod
     def from_settings(cls) -> DeepSeekClipClient:
@@ -37,18 +52,52 @@ class DeepSeekClipClient:
             settings.deepseek_base_url,
             settings.deepseek_model,
             settings.deepseek_temperature,
+            settings.llm_system_prompt,
+            settings.llm_user_prompt,
         )
 
-    def suggest(self, chunks: list[TranscriptChunk]) -> list[ClipCandidate]:
+    def suggest(
-        logger.info("deepseek suggestion start chunks=%s model=%s", len(chunks), self.model)
+        self,
+        chunks: list[TranscriptChunk],
+        *,
+        max_candidates_total: int,
+        min_clip_seconds: int,
+        max_clip_seconds: int,
+    ) -> list[ClipCandidate]:
+        logger.info(
+            "deepseek suggestion start chunks=%s model=%s max_candidates_total=%s",
+            len(chunks),
+            self.model,
+            max_candidates_total,
+        )
         candidates: list[ClipCandidate] = []
+        quota = candidate_quota_per_chunk(len(chunks), max_candidates_total)
         for chunk in chunks:
-            candidates.extend(self._suggest_for_chunk(chunk))
+            candidates.extend(
-        ranked = dedupe_and_rank_candidates(candidates)
+                self._suggest_for_chunk(
+                    chunk,
+                    max_candidates=quota,
+                    min_clip_seconds=min_clip_seconds,
+                    max_clip_seconds=max_clip_seconds,
+                )
+            )
+        ranked = dedupe_and_rank_candidates(
+            candidates,
+            min_clip_seconds=min_clip_seconds,
+            max_clip_seconds=max_clip_seconds,
+            max_candidates_total=max_candidates_total,
+        )
         logger.info("deepseek suggestion done raw_candidates=%s ranked_candidates=%s", len(candidates), len(ranked))
         return ranked
 
-    def _suggest_for_chunk(self, chunk: TranscriptChunk) -> list[ClipCandidate]:
+    def _suggest_for_chunk(
+        self,
+        chunk: TranscriptChunk,
+        *,
+        max_candidates: int,
+        min_clip_seconds: int,
+        max_clip_seconds: int,
+    ) -> list[ClipCandidate]:
         logger.info(
             "deepseek chunk request start start=%.1f end=%.1f chars=%s",
             chunk.start_sec,
@@ -62,22 +111,16 @@ class DeepSeekClipClient:
             messages=[
                 {
                     "role": "system",
-                    "content": (
+                    "content": f"{self.system_prompt.rstrip()}\n\n{RESPONSE_SCHEMA_INSTRUCTION}",
-                        "You select interesting short clips from Mandarin livestream transcripts. "
-                        "Return strict JSON only. Prefer complete moments with context, funny reactions, "
-                        "surprising reveals, emotional peaks, or strong standalone discussion."
-                    ),
                 },
                 {
                     "role": "user",
-                    "content": (
+                    "content": render_user_prompt(
-                        "Find up to 5 clip candidates in this transcript chunk. "
+                        self.user_prompt,
-                        "Each clip must be 30-360 seconds and use absolute stream seconds. "
+                        chunk=chunk,
-                        "JSON schema: {\"clips\":[{\"start_sec\":number,\"end_sec\":number,"
+                        max_candidates=max_candidates,
-                        "\"title_zh\":string,\"summary_zh\":string,\"reason\":string,"
+                        min_clip_seconds=min_clip_seconds,
-                        "\"score\":number,\"tags\":[string],\"subtitle_priority\":string}]}.\n\n"
+                        max_clip_seconds=max_clip_seconds,
-                        f"Chunk bounds: {chunk.start_sec:.1f}-{chunk.end_sec:.1f}\n"
-                        f"Transcript:\n{chunk.text}"
                     ),
                 },
             ],
@@ -95,6 +138,27 @@ class DeepSeekClipClient:
         return candidates
 
 
+def render_user_prompt(
+    template: str,
+    *,
+    chunk: TranscriptChunk,
+    max_candidates: int,
+    min_clip_seconds: int,
+    max_clip_seconds: int,
+) -> str:
+    try:
+        return template.format(
+            max_candidates=max_candidates,
+            min_clip_seconds=min_clip_seconds,
+            max_clip_seconds=max_clip_seconds,
+            chunk_start_sec=f"{chunk.start_sec:.1f}",
+            chunk_end_sec=f"{chunk.end_sec:.1f}",
+            transcript=chunk.text,
+        )
+    except KeyError as exc:
+        raise ValueError(f"Unknown placeholder in [llm_prompt].user: {exc.args[0]}") from exc
+
+
 def segment_line(segment: TranscriptSegment) -> str:
     return f"[{segment.start_sec:.1f}-{segment.end_sec:.1f}] {segment.text}"
 
@@ -141,17 +205,31 @@ def overlap_ratio(left: ClipCandidate, right: ClipCandidate) -> float:
     return overlap / shortest
 
 
-def dedupe_and_rank_candidates(candidates: list[ClipCandidate]) -> list[ClipCandidate]:
+def candidate_quota_per_chunk(chunk_count: int, max_candidates_total: int) -> int:
+    if chunk_count <= 0:
+        return 0
+    return max(1, (max_candidates_total + chunk_count - 1) // chunk_count)
+
+
+def dedupe_and_rank_candidates(
+    candidates: list[ClipCandidate],
+    *,
+    min_clip_seconds: int = 30,
+    max_clip_seconds: int = 360,
+    max_candidates_total: int | None = None,
+) -> list[ClipCandidate]:
     valid = [
         candidate
         for candidate in candidates
         if candidate.end_sec > candidate.start_sec
-        and 30 <= candidate.end_sec - candidate.start_sec <= 360
+        and min_clip_seconds <= candidate.end_sec - candidate.start_sec <= max_clip_seconds
     ]
     ranked = sorted(valid, key=lambda item: item.score, reverse=True)
     chosen: list[ClipCandidate] = []
     for candidate in ranked:
         if all(overlap_ratio(candidate, existing) < 0.5 for existing in chosen):
             chosen.append(candidate)
+            if max_candidates_total is not None and len(chosen) >= max_candidates_total:
+                break
     logger.debug("dedupe candidates input=%s valid=%s chosen=%s", len(candidates), len(valid), len(chosen))
     return chosen

src/evanescere/services/thumbnail.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import json
 import logging
+import random
 import shlex
 import subprocess
 from pathlib import Path
@@ -63,11 +64,7 @@ def generate_thumbnail(
         raise RuntimeError(f"Unsupported [thumbnail].provider={settings.thumbnail_provider!r}")
 
     final_path = output_dir / "thumbnail_final.jpg"
-    character_path = (
+    character_path = select_character_overlay(settings.thumbnail_character_overlay_dir)
-        Path(settings.thumbnail_character_overlay_path)
-        if settings.thumbnail_character_overlay_path
-        else None
-    )
     compose_thumbnail(
         background_path=background,
         output_path=final_path,
@@ -90,7 +87,7 @@ def generate_thumbnail(
             "provider": provider,
             "source_frame": str(base_frame),
             "background": str(background),
-            "character_overlay_path": settings.thumbnail_character_overlay_path or "",
+            "character_overlay_path": str(character_path or ""),
         },
     )
     logger.info("thumbnail generation done video_id=%s clip_id=%s output=%s", video.id, clip.id, final_path)
@@ -237,6 +234,24 @@ def overlay_character(image, overlay_path: Path, scale: float, position: str) ->
     image.alpha_composite(overlay, (max(0, x), max(0, y)))
 
 
+def select_character_overlay(overlay_dir: str | None) -> Path | None:
+    if not overlay_dir:
+        return None
+    directory = Path(overlay_dir)
+    if not directory.is_dir():
+        logger.warning("thumbnail character overlay directory missing path=%s", directory)
+        return None
+    candidates = sorted(
+        path for path in directory.iterdir() if path.is_file() and path.suffix.lower() == ".png"
+    )
+    if not candidates:
+        logger.warning("thumbnail character overlay directory has no png files path=%s", directory)
+        return None
+    selected = random.choice(candidates)
+    logger.info("thumbnail character overlay selected path=%s candidates=%s", selected, len(candidates))
+    return selected
+
+
 def load_font(size: int):
     from PIL import ImageFont
 

src/evanescere/services/webdav.py

@@ -5,16 +5,17 @@ from dataclasses import dataclass
 from datetime import UTC, datetime
 from pathlib import PurePosixPath
 from typing import Any
-from urllib.parse import quote, urljoin, urlparse
+from urllib.parse import quote, unquote, urljoin, urlparse
 from xml.etree import ElementTree
 
 import httpx
 from sqlalchemy.orm import Session
 
 from evanescere.config import get_settings
-from evanescere.models import Video
+from evanescere.models import Setting, Video
 
 VIDEO_SUFFIXES = {".flv", ".mp4", ".mkv", ".mov"}
+WEBDAV_BASELINE_KEY = "webdav_initial_baseline"
 logger = logging.getLogger(__name__)
 
 
@@ -26,6 +27,22 @@ class WebDavFile:
     modified_at: str | None = None
 
 
+@dataclass(frozen=True)
+class BootstrapResult:
+    observed: int
+    inserted: int
+    marked_existing: int
+
+
+@dataclass(frozen=True)
+class ScanResult:
+    observed: int
+    newly_stable: int
+    baseline_initialized: bool = False
+    baseline_inserted: int = 0
+    baseline_marked_existing: int = 0
+
+
 class WebDavClient:
     def __init__(
         self,
@@ -100,7 +117,7 @@ def parse_propfind_response(xml_text: str, base_url: str) -> list[WebDavFile]:
         href_path = urlparse(href).path
         if href_path.rstrip("/") == base_path:
             continue
-        filename = PurePosixPath(href_path).name
+        filename = unquote(PurePosixPath(href_path).name)
         if not filename or PurePosixPath(filename).suffix.lower() not in VIDEO_SUFFIXES:
             continue
         resource_type = response.find(".//d:resourcetype", ns)
@@ -129,9 +146,29 @@ def has_stable_size(video: Video, required_equal_samples: int = 2) -> bool:
     return len(sizes) == 1 and next(iter(sizes)) > 0
 
 
-def bootstrap_existing(client: WebDavClient, session: Session) -> int:
+def has_webdav_baseline(session: Session) -> bool:
-    count = 0
+    return session.get(Setting, WEBDAV_BASELINE_KEY) is not None
-    for file in client.list_files():
+
+
+def mark_webdav_baseline(session: Session, result: BootstrapResult) -> None:
+    value = {
+        "initialized_at": datetime.now(UTC).isoformat(),
+        "observed": result.observed,
+        "inserted": result.inserted,
+        "marked_existing": result.marked_existing,
+    }
+    row = session.get(Setting, WEBDAV_BASELINE_KEY)
+    if row is None:
+        session.add(Setting(key=WEBDAV_BASELINE_KEY, value=value))
+    else:
+        row.value = value
+
+
+def bootstrap_existing(client: WebDavClient, session: Session) -> BootstrapResult:
+    files = client.list_files()
+    inserted = 0
+    marked_existing = 0
+    for file in files:
         video = session.query(Video).filter(Video.source_url == file.url).one_or_none()
         if video is None:
             video = Video(
@@ -143,11 +180,35 @@ def bootstrap_existing(client: WebDavClient, session: Session) -> int:
                 processing_status="done",
             )
             session.add(video)
-            count += 1
+            inserted += 1
             logger.debug("bootstrap existing file=%s size=%s", file.filename, file.size_bytes)
+        elif video.ingest_status == "observing" and video.processing_status == "pending":
+            video.size_bytes = file.size_bytes
+            video.ingest_status = "existing_done"
+            video.processing_status = "done"
+            marked_existing += 1
+            logger.debug("bootstrap marked existing video_id=%s file=%s", video.id, file.filename)
+    result = BootstrapResult(
+        observed=len(files),
+        inserted=inserted,
+        marked_existing=marked_existing,
+    )
+    mark_webdav_baseline(session, result)
     session.flush()
-    logger.info("bootstrap existing done inserted=%s", count)
+    logger.info(
-    return count
+        "bootstrap existing done observed=%s inserted=%s marked_existing=%s",
+        result.observed,
+        result.inserted,
+        result.marked_existing,
+    )
+    return result
+
+
+def ensure_initial_baseline(client: WebDavClient, session: Session) -> BootstrapResult | None:
+    if has_webdav_baseline(session):
+        return None
+    logger.info("webdav baseline missing; recording currently visible files as existing_done")
+    return bootstrap_existing(client, session)
 
 
 def scan_once(client: WebDavClient, session: Session) -> tuple[int, int]:
@@ -175,3 +236,17 @@ def scan_once(client: WebDavClient, session: Session) -> tuple[int, int]:
             logger.info("webdav file stable video_id=%s file=%s size=%s", video.id, file.filename, file.size_bytes)
     session.flush()
     return observed, newly_stable
+
+
+def safe_scan_once(client: WebDavClient, session: Session) -> ScanResult:
+    baseline = ensure_initial_baseline(client, session)
+    if baseline is not None:
+        return ScanResult(
+            observed=baseline.observed,
+            newly_stable=0,
+            baseline_initialized=True,
+            baseline_inserted=baseline.inserted,
+            baseline_marked_existing=baseline.marked_existing,
+        )
+    observed, newly_stable = scan_once(client, session)
+    return ScanResult(observed=observed, newly_stable=newly_stable)

tests/test_llm.py

@@ -1,5 +1,11 @@
 from evanescere.schemas import ClipCandidate
-from evanescere.services.llm import dedupe_and_rank_candidates, parse_clip_response
+from evanescere.services.llm import (
+    TranscriptChunk,
+    candidate_quota_per_chunk,
+    dedupe_and_rank_candidates,
+    parse_clip_response,
+    render_user_prompt,
+)
 
 
 def test_parse_clip_response():
@@ -62,3 +68,55 @@ def test_dedupe_and_rank_candidates_filters_short_and_overlapping():
     ranked = dedupe_and_rank_candidates(candidates)
     assert [candidate.title_zh for candidate in ranked] == ["best", "second"]
 
+
+def test_candidate_quota_uses_total_budget_across_chunks():
+    assert candidate_quota_per_chunk(1, 20) == 20
+    assert candidate_quota_per_chunk(4, 20) == 5
+    assert candidate_quota_per_chunk(0, 20) == 0
+
+
+def test_dedupe_respects_configured_duration_and_total_cap():
+    candidates = [
+        ClipCandidate(
+            start_sec=index * 240,
+            end_sec=index * 240 + duration,
+            title_zh=f"clip-{index}",
+            summary_zh="",
+            reason="",
+            score=1 - index / 100,
+        )
+        for index, duration in enumerate([60, 180, 181, 90])
+    ]
+    ranked = dedupe_and_rank_candidates(
+        candidates,
+        min_clip_seconds=30,
+        max_clip_seconds=180,
+        max_candidates_total=2,
+    )
+    assert [candidate.title_zh for candidate in ranked] == ["clip-0", "clip-1"]
+
+
+def test_render_user_prompt_inserts_configured_values():
+    prompt = render_user_prompt(
+        "{max_candidates}|{min_clip_seconds}|{max_clip_seconds}|{chunk_start_sec}|{chunk_end_sec}|{transcript}",
+        chunk=TranscriptChunk(12.25, 42.75, "hello"),
+        max_candidates=20,
+        min_clip_seconds=30,
+        max_clip_seconds=180,
+    )
+    assert prompt == "20|30|180|12.2|42.8|hello"
+
+
+def test_render_user_prompt_rejects_unknown_placeholder():
+    try:
+        render_user_prompt(
+            "{unknown}",
+            chunk=TranscriptChunk(0, 1, "hello"),
+            max_candidates=20,
+            min_clip_seconds=30,
+            max_clip_seconds=180,
+        )
+    except ValueError as exc:
+        assert "Unknown placeholder" in str(exc)
+    else:
+        raise AssertionError("Expected unknown placeholder to fail")

tests/test_thumbnail.py

@@ -0,0 +1,22 @@
+from pathlib import Path
+from unittest.mock import patch
+
+from evanescere.services.thumbnail import select_character_overlay
+
+
+def test_select_character_overlay_returns_none_for_empty_directory(tmp_path: Path):
+    assert select_character_overlay(str(tmp_path)) is None
+
+
+def test_select_character_overlay_filters_png_files(tmp_path: Path):
+    first = tmp_path / "first.PNG"
+    second = tmp_path / "second.png"
+    first.write_bytes(b"")
+    second.write_bytes(b"")
+    (tmp_path / "notes.txt").write_text("ignore me")
+
+    with patch("evanescere.services.thumbnail.random.choice", return_value=second) as choice:
+        assert select_character_overlay(str(tmp_path)) == second
+
+    assert choice.call_args.args[0] == [first, second]
+

tests/test_webdav_stability.py

@@ -33,3 +33,20 @@ def test_parse_propfind_response_filters_video_files():
         WebDavFile("https://host/webdav/recordings/stream.flv", "stream.flv", 123, None)
     ]
 
+
+def test_parse_propfind_response_preserves_encoded_chinese_filename():
+    xml = """<?xml version="1.0"?>
+<D:multistatus xmlns:D="DAV:">
+  <D:response>
+    <D:href>/Hirumi/hirumi-2-06%E6%9C%8801%E6%97%A501%E6%97%B602%E5%88%8641%E7%A7%92.flv</D:href>
+    <D:propstat><D:prop><D:getcontentlength>123</D:getcontentlength></D:prop></D:propstat>
+  </D:response>
+</D:multistatus>"""
+    assert parse_propfind_response(xml, "http://recording.home.arpa/Hirumi") == [
+        WebDavFile(
+            "http://recording.home.arpa/Hirumi/hirumi-2-06%E6%9C%8801%E6%97%A501%E6%97%B602%E5%88%8641%E7%A7%92.flv",
+            "hirumi-2-06月01日01时02分41秒.flv",
+            123,
+            None,
+        )
+    ]