anibon-timestamper-local
Use when generating YouTube timestamps and topic summaries for long livestreams locally on a single GPU (such as NVIDIA Tesla P100 16GB, LM Studio, Ollama) without cloud API costs. For cloud/frontier models (Claude 3.5 Sonnet, Gemini Pro) with parallel subagents, use anibon-timestamper.
How do I install this agent skill?
npx skills add https://github.com/zenithth/anibon-stream-synthesis --skill anibon-timestamper-localIs this agent skill safe to install?
- Gen Agent Trust Hubfail
The skill is a local processing pipeline for YouTube livestreams designed to run on single GPUs. It utilizes system tools like ffmpeg for frame extraction and interacts with local or remote OpenAI-compatible endpoints. Automated security scans flagged potential issues including reverse shell patterns and remote code execution, but manual analysis confirms these are false positives related to local connectivity health checks and standard API communication. A low-risk surface for indirect prompt injection exists because the skill processes untrusted transcript data.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
Anibon Timestamper (Local P100 Edition)
Overview
Local processing pipeline for generating YouTube timestamps and summaries from long livestreams via Dynamic Recursive Rolling Summary state-machines (--mode recursive) or group batching (--mode group), multi-modal context fusion, automatic tag normalization, and two-pass YouTube comment formatting.
When to Use
- When processing livestreams locally on a single GPU (e.g. NVIDIA Tesla P100 16GB, Apple Silicon Metal, LM Studio, Ollama) without incurring cloud API costs.
- For Cloud / Frontier Models: If the user has cloud API access and wants high-concurrency subagent orchestration (Claude 3.5 Sonnet, Gemini 2.0 Pro), use
anibon-timestamperinstead.
Companion Skills & Required Sub-Skills
anibon-world-identity(MANDATORY): Must be loaded whenever the stream covers gaming, anime, Pokémon, Tokusatsu, or pop-culture topics. The agent must verify character names, game titles, and terms against local references (references/INDEX.md,atlas_fgo.db,pokemon.db,garbled_replacements.json) before finalizing timestamps.anibon-livechat-analysis: Download, parse, and align YouTube LiveChat replays, or extract burned-in on-screen chat via vision OCR fallback. Provides dual-side perspective (streamer + watchers), SuperChat highlights, and viewer reaction peaks.whisper-corruption-recovery: Automatically invoked if Whisper transcription contains repetition loops or silence stutter.anibon-local-transcription: Used when YouTube has no subtitles or auto-captions to extract 16kHz audio and run local GPU whisper.cpp.
Pipeline Architecture
Pipeline commands run from the plugin root directory (anibon-stream-synthesis). Shared CLI utilities live in scripts/; skill-specific runners live in skills/anibon-timestamper-local/scripts/.
| Stage | Script / Module | Inputs / Outputs | Description |
|---|---|---|---|
| 0. Noise Cleaning, Music Anchors & Pre-normalization | lib/process/signals/detector.py (auto) | garbled_replacements.json + default_mappings.json → music_anchors.json | Strips ASR speaker markers (>>), sound effect tags ([เพลง], [Applause]), music symbols (♪), extracts music_anchors.json, repetitive stutter loops, and corrects phonetic drift before chunk loading/signal detection. |
| 0.5 Knowledge Self-Reading | lib/process/knowledge_reader.py (auto) | signals.json + references/*.md + DBs → entity_glossary.json | Self-reads domain markdown references matched in signals, queries local SQLite databases (atlas_fgo.db, pokemon.db), and builds chunk-filtered verified entity glossaries. Bypass via --no-knowledge-reader. |
| 1. Preparation & Chunking | scripts/prepare_video.py | YouTube URL → raw_transcript.json, chunks/*.txt | Downloads subtitles and segments audio/transcript into overlapping chunks. Pass --live-chat to download & align chat in one step. |
| 1.5 LiveChat, Mood & Proactive Music Detection | scripts/align_live_chat.py / scripts/patch_music_chunks.py / lib/process/music/ | livechat/, music_anchors.json → mood_555.json, patched chunks/*.txt | Aligns LiveChat replay events to transcript chunks, auto-detects 555 bursts, proactively detects background songs via Stage 0 anchors and chat music density, and patches chunks with [Song Context Briefing] and DSP vocal isolation. |
| 2. Topic Segmentation (Pass 1) | skills/anibon-timestamper-local/scripts/process_chunks_local.py | chunks/, signals.json, Storyboard keyframes (--multimodal-chunks), livechat/ | Detects shifts/continuations, emits timestamps via local LLM/VLM with visual ground truth, LiveChat highlights, and mood tone hints. |
| 2.5 Vision Ground Truth (Optional) | lib/process/vision_verify.py (--vision-verify) | video_360p.mp4 + ambiguous stamps | Inspects on-screen video frames via local vision model for vague proper nouns or [?] tags before Pass 2 assembly. |
| 3. Summarizer & Assembly (Pass 2) | lib/process/summarizer.py (auto) | all_timestamps.txt → anibon_timestamps.md | Clusters timestamps into cohesive macro-thematic parts (typically 4–6 parts) and snaps to transition cues (find_best_boundary). |
| 4. Garbled Collector & Whisper Ground Truth | scripts/whisper_dispatcher.py (auto post-pass) | garbled_notes_raw/ → garbled_notes.json → garbled_replacements.json | Automatically runs after summary: guards against attention collapse loops via detect_repetition_anomaly, slices audio on-the-fly, transcribes phonetic ground truth via local whisper.cpp, and auto-grows shared dictionary. |
| 4.5 Post-Recovery Timestamp Unpacking & Sync | Manual / CLI Runner | recovery_report_*.md → all_timestamps.txt, anibon_timestamps_full.md, anibon_timestamps.md | When speech is recovered from music-polluted or garbled segments, unpacks coarse single-timestamp placeholders into granular reaction/dialogue entries ($\ge 45\text{s}$ spacing) and atomically synchronizes all three output files. |
| 5. Deterministic Parity Gates | lib/process/parity_gates.py (auto post-pass) | all_timestamps.txt, anibon_timestamps.md | Runs front-tier parity checks: scripts/audit_gaps.py (enforces NO GAPS > 10m), scripts/check_sections.py (audits timestamp line counts [target 15–20, max 25 lines/part] and ASR garble patterns), and scripts/validate_mood.py (validates mood-bearing timestamps honour mood_555 verdicts). |
Modular Library Architecture (lib/)
lib/process/livechat.py: End-to-end LiveChat replay downloader (yt-dlp), raw event parser, chunk boundary aligner (align_livechat_feed), and mood pulse detection hook.lib/process/multimodal.py: Storyboard slide unpacking, adaptive scene-change & evaluation-batch keyframe selector (1–16 frames per chunk, 16×9 RGB sidecar caching, Thai deictic cue pinning), OpenAI-compatible multi-image payload construction, and VLM execution.lib/process/knowledge_reader.py: Automated domain knowledge discovery fromsignals.json, regex table/bullet entity parser, SQLite DB enricher (atlas_fgo.db,pokemon.db), andentity_glossary.jsongenerator.lib/process/timestamps.py: Tag normalization (TAG_REMAP), timestamp line sanitization, collision guards ($\ge 45\text{s}$ spacing), window validation.lib/process/lmstudio.py: LM Studio & OpenAI-compatible client (TokenMe, vLLM, Ollama), model resolution (defaults togemma-4-12b-it-qat@q4_k_xlorunsloth/gemma-4-26b-a4b-it@q2_k_x), text completions, Bearer auth, Cloudflare bypass, and multi-modal Vision API (image_urldata URIs for Gemma 4 Vision, Qwen-VL).lib/process/websearch.py: Lightweight DuckDuckGo search with local JSON cache (websearch_cache.json) for zero-cost entity verification.lib/process/prompts.py: Recursive rolling summary and group batching prompts, World Identity reference injection, and search/vision context blocks.lib/process/summarizer.py: Pass 2 local summarizer with semantic & chronological deduplication and macro-thematic part assembly (typically 4–6 parts).lib/process/state.py: Checkpoint persistence (anibon_timestamper_state.json), chunk discovery, and multimodal loaders (LiveChat, 555 mood, activity).
Multi-Layer Anti-Hallucination & Verification Architecture (4-Layer Defense)
Local smaller LLMs (e.g. 12B–27B) lack the vast parametric world knowledge of front-tier models, making them prone to phonetic guessing and hallucinated English titles. anibon-timestamper-local implements a deterministic 4-layer defense pipeline:
-
Layer 1: Two-Stage Signal Detection & Chaining Weights (
signal_detector.py/lib/anibon/signal_detector.py)- Stage 1 (Game Domain Identification):
- Ingests video title, description, and tags from
info.jsonviaextract_metadata_signals(). - Dynamically discovers all reference markdown games in
anibon-world-identity/references/*.mdviadiscover_reference_games(). - Detects game presence via exclusive special words and title signals with chaining weights across sequential chunks.
- Strips common Japanese character names (Yuki, Rin, Ren) and generic gaming standing words (E, C, AP, FTK, E6, Preset) from triggering false game candidates.
- Decays candidate probabilities across chunks to preserve ongoing gameplay context while discarding weak tail noise (
prob < 0.15).
- Ingests video title, description, and tags from
- Chronology & Callback Validation: Cross-checks detected game versions against
references/<Game>.mdto prevent retrospective streamer commentary from corrupting the active update title. - Stage 2 (Entity Scoping via
allowed_sources):- Scopes entity glossary queries and prompt context strictly to
allowed_sourcesreturned from Stage 1 candidates.
- Scopes entity glossary queries and prompt context strictly to
- Zero-Example Prompt Discipline:
- Prompt templates and few-shot examples must NEVER hardcode real game titles (Wuthering Waves, Zelda, Bleach) as static text; use abstract token placeholders (
<Game_A>,<Character_X>) to prevent local models (e.g. Gemma 4 12B) from latching onto them as hallucinations.
- Prompt templates and few-shot examples must NEVER hardcode real game titles (Wuthering Waves, Zelda, Bleach) as static text; use abstract token placeholders (
- Stage 1 (Game Domain Identification):
-
Layer 2: Pre-Inference Buffer Normalization & Scrubbing (
state.py)- Automatically scrubs ASR speaker drift (
ปู่บอท,ลุงบอท,ปู่โบต→ปู่โบ๊ต) during chunk loading inload_chunk_file(). - Applies phonetic mappings from
default_mappings.jsonbefore passing text to the LLM prompt.
- Automatically scrubs ASR speaker drift (
-
Layer 3: Anti-Guessing & Invariant Prompt Directives (
prompts.py)- Enforces
STREAMER IDENTITYinvariant: The streamer is exclusively'ปู่โบ๊ต' (Boat). Never write'ปู่บอท','ลุงบอท', or'ปู่โบต'. - Enforces
ANTI-GUESSING CONSTRAINT: Never arbitrarily translate Thai phonetic loanwords into novel English spellings. If not found inVERIFIED DOMAIN ENTITIES, write phonetic Thai or describe the action in Thai. Never invent novel English proper nouns. - Enforces
POKÉMON NAMING RULE: Pokémon entities must strictly follow'ชื่อไทย (English Name)'(Thai community/official name followed by English name in parentheses). - Enforces
CAVEMAN TIMESTAMP STYLE INVARIANT:- NEVER start timestamp descriptions or section titles with
'ปู่โบ๊ต'. - Dynamic Recent-Verb Blacklist: Extracts opening verbs from
recent_timestampsand explicitly bans reusing them in the next chunk. - No Redundant Title Suffixes: When a whole section covers one main title, name specific characters/events rather than appending
'ใน <Title>'to every line. - Use varied action-first or topic-first verbs:
ส่อง,ชำแหละ,เทียบ,เจาะลึก,เม้าท์มอย,สรุปดราม่า,กางทฤษฎี,รีวิว,พักเบรก,ปิดไลฟ์. - Keep 100% of proper nouns while eliminating redundant conversational padding.
- NEVER start timestamp descriptions or section titles with
- Enforces
-
Layer 4: Stage 5 Deterministic Auto-Sanitization (
process_chunks_local.py)- Automatically executes
sanitize_and_audit_timestamps()during Pass 2 summary assembly. - Deterministically corrects residual speaker misspellings and canonical dictionary patterns (
garbled_replacements.json, e.g.Archen→Arcane,Victor→Viktor). - Lexical Diversity Sanitizer & Gate (
world_sanitizer.py&check_sections.py): Deterministically rotates overused opening verbs (วิเคราะห์,อึ้ง!,พูดถึง) and audits that no single opening verb exceeds 2 consecutive lines or 20% of total stamps. - Audits timestamp lines for unverified capitalized Latin words against whitelisted tags and glossaries, raising deterministic audit warnings for suspect hallucinations.
- Automatically executes
Execution Modes
Mode 1: Recursive Rolling Summary (--mode recursive, Default & Recommended)
Pu Boat's discussions follow organic content flow rather than clock boundaries. A topic may last 3 minutes or 25 minutes.
SAME_TOPIC: If Chunk $N$ continues the ongoing topic, it merges into the rolling summary without emitting a timestamp (0 timestamps).- Long-Form Talk/News Exception: If a monologue or societal debate continues across multiple chunks (>8–10 minutes without a stamp), the model MUST extract a distinct sub-angle or specific policy critique (1 timestamp) rather than leaving long blanks.
- Unbounded Chaining Summary Invariant:
build_recursive_prompt()and state persistence (anibon_timestamper_state.json,recursive_outputs/chunk_XX.json) must pass the fullrolling_summaryandchunk_summarytext without hardcoded character truncation so multi-chunk topic context is never clipped.
TOPIC_SHIFT/ Highlights: When the topic shifts, it flushes previous context and stamps the shift. Emits 1 timestamp by default, 2 MAX per chunk ($\ge$45s apart).- Thematic Part Grouping: Pass 2 summarizer groups timestamps by overarching thematic category into sequential integer parts (
ส่วนที่ 1, 2, 3...typically 4–6 parts total) usingfind_best_boundaryto snap part borders to streamer transition cues (e.g.สรุป... และเตรียมเปลี่ยนไปคุย,เตรียมเริ่มเล่นเกม). Over-splitting into 10–18 micro-parts is disallowed; cohesive topic discussions (e.g. complete series reviews or gaming sessions) remain unified in a single part for viewer navigation convenience. Consecutive same-game grinding is consolidated within 10 minutes while talk/news sub-topics use a 4-minute window.
Mode 2: Fixed Window Groups (--mode group, Alternative)
Combines 4 chunks (~16–20 min) per group with chronological loop-breakers and collision guards.
Core Pipeline & Quick Reference
Quick Commands
0. Denoising & Pre-normalization (Automatic)
Before any signal detection or LLM analysis, clean_transcript_noise() automatically cleans ASR artifacts (>>, [เพลง], ♪, character repetitions), and applies 2000+ Whisper ground truths from garbled_replacements.json and default_mappings.json. No manual invocation needed.
0.5 Knowledge Self-Reading & Entity Glossary (Stage 0.5, Automatic)
Before processing chunks, run_knowledge_discovery() inspects signals.json, self-reads matched markdown references in anibon-world-identity/references/*.md, queries local databases (atlas_fgo.db, pokemon.db), and emits <workspace>/entity_glossary.json. Chunk prompts automatically filter relevant entities (capped to ~300 tokens) into the prompt context.
- To bypass entity glossary discovery: pass
--no-knowledge-reader.
1. Download & Chunk (Skip if chunks/chunk_00.txt exists)
macOS / Linux:
python3 scripts/prepare_video.py "https://www.youtube.com/watch?v=<VIDEO_ID>" \
--workspace "~/youtube_<VIDEO_ID>_workspace" \
--format txt --block 300 --overlap 30
Windows (PowerShell / Command Prompt):
python scripts\prepare_video.py "https://www.youtube.com/watch?v=<VIDEO_ID>" --workspace "$HOME\youtube_<VIDEO_ID>_workspace" --format txt --block 300 --overlap 30
1.5 Download & Align LiveChat (Optional but preferred)
Gives the local model the watchers' chat for its own chunk, so situation + emotion are read from BOTH talks and chat. Requires network; if a VOD has no live_chat subtitle, skip this step and rely on transcript-only emotion.
Automated 1-Step (via prepare_video.py):
python3 scripts/prepare_video.py "https://www.youtube.com/watch?v=<VIDEO_ID>" \
--workspace "~/youtube_<VIDEO_ID>_workspace" \
--format txt --block 300 --overlap 30 --live-chat
Manual / Step-by-Step Execution:
# 1. Download LiveChat replay (network required)
yt-dlp --sub-langs live_chat --write-sub --skip-download "https://www.youtube.com/watch?v=<VIDEO_ID>" -o "<workspace>/%(id)s.%(ext)s"
# 2. Parse to coarse chunks + a seconds-prefixed raw event feed
python3 -X utf8 skills/anibon-livechat-analysis/scripts/parse_live_chat.py <workspace>/<id>.live_chat.json --chunk-minutes 90 --raw-events <workspace>/livechat_events.txt -o <workspace>/livechat_90/
# 3. Align raw feed to transcript chunk windows
python3 -X utf8 scripts/align_live_chat.py --events <workspace>/livechat_events.txt --chunks <workspace>/chunks/ -o <workspace>/livechat/
# 4. Detect Thai-Laugh / Meme Pulses (auto-run by process_chunks_local.py or run manually)
python3 -X utf8 scripts/analyze_555.py \
--livechat ~/youtube_<id>_workspace/livechat/ \
--index ~/youtube_<id>_workspace/livechat/livechat_index.json \
--out ~/youtube_<id>_workspace/mood_555.json
Output: <workspace>/livechat/livechat_chunk_NN.txt + livechat/livechat_index.json and mood_555.json.
Fallback for Edited Livestreams / Missing .live_chat.json (Visual Chat Extraction):
When a livestream has been trimmed or edited in YouTube Studio, YouTube permanently deletes the .live_chat.json chat replay track. Use extract_visual_livechat.py from anibon-livechat-analysis to OCR the burned-in chat overlay:
# 1. Spot-extract visual livechat for target chunk time range
python3 -X utf8 skills/anibon-livechat-analysis/scripts/extract_visual_livechat.py \
--video-id <VIDEO_ID> \
--range HH:MM:SS-HH:MM:SS \
-o <workspace>/visual_livechat_events.txt
# 2. Align visual raw event feed directly to chunk windows
python3 -X utf8 scripts/align_live_chat.py \
--events <workspace>/visual_livechat_events.txt \
--chunks <workspace>/chunks/ \
-o <workspace>/livechat/
[!NOTE] Emoji Dictionary Reference: Refer to
references/anibon_emoji_dictionary.mdfor channel-exclusive emote mapping (:_CunnyBoat:,:_MonkeyBoat:,:_Nerd:,:_shutup:,:Grind:,:_Ripfish:,:_noname:,:_What:,:_WOW:,:_Ahh:,:_Meh:,:_MoneyBoat:,:_Tahaan:,:_KonDee:,:_NeoSlim:,:_Slim:,:_BoatSOM:) and YouTube global emotes to mood pulse weights.
1.9 Verify Model is Resident in VRAM (MANDATORY)
/v1/models lists all models on disk, not those currently loaded in VRAM. If inference starts without an active model, calls will timeout after 60s or return "Model is unloaded.". Always check:
curl -s http://100.115.25.30:1234/api/v1/models | jq '.models[] | select(.loaded_instances | length > 0) | .key'
If empty, explicitly load the recommended profile before starting the run:
curl -s -X POST http://100.115.25.30:1234/api/v1/models/load \
-H "Content-Type: application/json" \
-d '{
"model": "google/gemma-4-12b-qat",
"context_length": 262144,
"eval_batch_size": 1024,
"flash_attention": true,
"offload_kv_cache_to_gpu": true
}'
Smoke-test with --max-chunks 1 before launching a full 50+ chunk run.
2. Run Timestamper Directly (Foreground)
macOS / Linux:
python3 -X utf8 skills/anibon-timestamper-local/scripts/process_chunks_local.py "~/youtube_<VIDEO_ID>_workspace" \
--mode recursive --lang th
Windows (PowerShell):
python -X utf8 skills\anibon-timestamper-local\scripts\process_chunks_local.py "$HOME\youtube_<VIDEO_ID>_workspace" --mode recursive --lang th
[!TIP] Single Comment Byte Limit (
--byte-cap): Pass--byte-capif you need each part strictly under 3,500 bytes for YouTube single-comment pasting. By default, byte-capping is disabled to preserve cohesive macro-thematic grouping (4–6 parts).
2.1 In-Pipeline Vision Verification (Optional)
Inspect on-screen video frames for ambiguous or vague timestamps (deictic pronouns, missing game names, [?] markers) using the local vision model before Pass 2 summary:
python -X utf8 skills\anibon-timestamper-local\scripts\process_chunks_local.py "$HOME\youtube_<VIDEO_ID>_workspace" --mode recursive --lang th --vision-verify
2.2 Multimodal Chunk Analysis with Storyboard Keyframes (Default: Enabled)
Samples high-resolution Storyboard keyframes per chunk directly into the VLM (e.g. Gemma 4 12B Vision, Qwen-VL) alongside the transcript text during Pass 1. Enabled by default via config.json ("multimodal_chunks": true). Eliminates game title misattribution and grounds event/chapter subtitles in real time:
# Enabled by default (Adaptive Scene-Change Mode):
python3 skills/anibon-timestamper-local/scripts/process_chunks_local.py ~/youtube_<VIDEO_ID>_workspace --mode recursive --lang th
# To tune visual sensitivity or evaluation batch size:
python3 skills/anibon-timestamper-local/scripts/process_chunks_local.py ~/youtube_<VIDEO_ID>_workspace \
--eval-batch-size 8192 \
--max-chunk-frames 16 \
--scene-delta-threshold 0.06
# To disable adaptive scene deduplication (fallback to legacy fixed 3-5 frames):
python3 skills/anibon-timestamper-local/scripts/process_chunks_local.py ~/youtube_<VIDEO_ID>_workspace --no-scene-dedup
# To disable all multimodal analysis and run in text-only mode:
python3 skills/anibon-timestamper-local/scripts/process_chunks_local.py ~/youtube_<VIDEO_ID>_workspace --no-multimodal
Adaptive Scene-Change Keyframe Selector (scene_dedup=True):
- Static Screens: Drops static talking-head / waiting screens down to 1–2 frames (saving physical batch compute and prompt tokens).
- Dynamic Action: Automatically expands up to 16 frames per 5-min chunk whenever visual scene transitions occur, bounded by
--eval-batch-size(default8192). - Thai Deictic Cue Pinning: Pins exact visual frames whenever Pu Boat uses deictic pointers (
"ดูอันนี้","ในจอ","ตรงนี้","[?]"). - RGB Sidecar Caching: Automatically persists 16×9 432-byte downsampled buffers (
tile_XXXXXX.rgb) alongside.jpgtiles, eliminating repeatedffmpegsubprocess overhead across comparisons and reruns. - Clean-Slate
--no-resume: Passing--no-resumeautomatically unlinks stalerecursive_outputs/chunk_*.jsonfiles to guarantee clean execution without leftover chunk contamination.
Re-running an Existing Workspace while Preserving Previous Results (e.g. Text-Only → Multimodal):
cd ~/youtube_<VIDEO_ID>_workspace
cp -a . ~/youtube_<VIDEO_ID>_workspace_text_backup
cp anibon_timestamps.md anibon_timestamps_text_only.md
cp all_timestamps.txt all_timestamps_text_only.txt
mv recursive_outputs recursive_outputs_text_only && mkdir -p recursive_outputs
cp anibon_timestamper_state.json anibon_timestamper_state_text_only.json
rm -f anibon_timestamper_state.json all_timestamps.txt anibon_timestamps.md timestamper.pid
3. Detached Background Launch (3 OS Families)
Runs inference detached in background so agent sessions or timeouts do not kill the run.
macOS / Linux (Bash):
./skills/anibon-timestamper-local/scripts/launch_local.sh ~/youtube_<VIDEO_ID>_workspace 100.115.25.30 auto th
Windows (PowerShell):
powershell -ExecutionPolicy Bypass -File "skills\anibon-timestamper-local\scripts\launch_local.ps1" -Workspace "$HOME\youtube_<VIDEO_ID>_workspace" -Endpoint "100.115.25.30"
Windows (CMD Batch):
skills\anibon-timestamper-local\scripts\launch_local.bat "%USERPROFILE%\youtube_<VIDEO_ID>_workspace" 100.115.25.30 auto th
[!NOTE] Endpoint Auto-Probe & Fallback Rule: All runners and Python client (
anibon.lmstudio.resolve_endpoint()) automatically probehttp://100.115.25.30:1234with a 1.0s TCP socket test. If unreachable (indicating execution on localhost itself), they immediately and safely fall back tohttp://127.0.0.1:1234.
[!TIP] Remote / TokenMe Endpoint Support: You can target remote OpenAI-compatible endpoints such as TokenMe (
--endpoint https://api.tokenme.limited/v1orapi.tokenme.limited) by passingTOKENME_API_KEY(orLM_API_KEY) in the environment or--api-key.
- The client sets standard
Authorization: Bearer <KEY>and browserUser-Agentheaders to bypass Cloudflare WAF code 1010.- Remote models like
nvidia/Gemma-4-26B-A4B-NVFP4support multimodal vision inspection via base64image_urlpayloads out-of-the-box.
[!TIP] Configuration (
config.jsonand--config): Numeric constants, token limits, timeouts, temperatures, pruning bounds, and default endpoints are managed viaskills/anibon-timestamper-local/config.jsonand loaded automatically byanibon/config.pywith zero dependencies.
- Secrets (like
TOKENME_API_KEYorLM_API_KEY) belong in.env(which is gitignored).- To customize parameters for a specific run or hardware setup, copy or create a JSON config file and pass
--config /path/to/config.json(or--env-filefor backward compatibility).- Environment variables (
ANIBON_CHUNK_MAX_TOKENS,ANIBON_TEMPERATURE, etc.) seamlessly overrideconfig.jsonvalues without touching files.
- Section Splitting Rule (Line Count over Byte Count):
- Macro-parts (
ส่วนที่ X) are grouped and split strictly by timestamp line count. - Target: 15–20 lines/part. Hard split limit: > 25 lines/part.
- Byte counting is canceled for section splitting decisions (avoiding false alarms from 3-byte Thai UTF-8 text).
- Splitting snaps to the nearest topic transition marker or line midpoint (
len(stamps) // 2).
- Macro-parts (
4. Garbled Collector & Whisper Ground Truth (Automatic Post-Pass)
After anibon_timestamps.md summary assembly finishes, process_chunks_local.py automatically checks for surviving phonetic garbles in <workspace>/garbled_notes_raw/ and invokes whisper_dispatcher.py with local whisper.cpp to slice audio, aligns acoustic segments via resolve_garbled_notes, and updates garbled_replacements.json.
-
Garbled Collector Safety Rules:
- Rejects ASCII patterns
< 5characters ("ON","MDB") to avoid substring corruptions (e.g."Donghua"→"DNexonghua"). - Rejects patterns where
garbled.lower() in target.lower()(e.g."on"in"Nexon"). - Rejects ungrounded LLM targets not found in dictionary, glossary, stream title, or Whisper audio segments.
- Re-cleans both
chunks/*.txtandanibon_timestamps.mdupon updating the dictionary.
- Rejects ASCII patterns
-
To pass video URL for on-the-fly stream slicing without full audio downloads:
python3 -X utf8 scripts/process_chunks_local.py "~/youtube_<VIDEO_ID>_workspace" \ --video-url "https://www.youtube.com/watch?v=<VIDEO_ID>" -
To disable automatic garbled collection pass:
python3 -X utf8 scripts/process_chunks_local.py "~/youtube_<VIDEO_ID>_workspace" \ --no-garbled-collector -
To run garbled collection manually at any time:
python3 scripts/whisper_dispatcher.py ~/youtube_<VIDEO_ID>_workspace \ --video-url "https://www.youtube.com/watch?v=<VIDEO_ID>" --verbose python3 ../cleaning-auto-transcripts/scripts/update_garbled_dictionary.py \ --from-notes ~/youtube_<VIDEO_ID>_workspace/garbled_notes.json \ --workspace ~/youtube_<VIDEO_ID>_workspace
5. World Identity & Fact Verification Audit (Mandatory Pass)
Immediately after generating all_timestamps.txt and anibon_timestamps.md:
- Load
anibon-world-identity: Scan all emitted timestamp titles and descriptions against local knowledge bases:- FGO Servants / Classes: Query
atlas_fgo.db(do NOT confuse Ashiya Douman with Ascalon/Asclepius). - Pokémon Entities: Query
pokemon.dband enforce Thai community/official names first (e.g.แบกซ์แคลิเบอร์ (Baxcalibur)). - Anime / Gacha Titles: Verify against
references/*.md(e.g.Chaos_Zero_Nightmare.md,Genshin_Impact.md,Honkai_Star_Rail.md,Wuthering_Waves.md). - Unfamiliar Indie / Web / Thai Games: If the streamer plays an unindexed game genre (e.g. probability life sims like 1 in 8 Billion, 1 ชีวิต ณ กรุงศรี, หนึ่งในอสงไขย), research and scaffold
references/<Topic>.mdfirst, then reconstruct timestamps directly fromraw_transcript.json. - New Competitive Games (Deadlock, etc.): Cross-reference hero names (Baba / Baba Yaga, Celeste, Abrams) via web search or patch notes.
- FGO Servants / Classes: Query
- Audit Against
raw_transcript.jsonfor 4 Local-Model Distortion Patterns:- Domain Context Bleeding: Check non-gaming talk segments adjacent to gaming segments. Local 12B models often distort general words (e.g.
"บริษัทนั้น"in a food/political news story) into"บริษัทเกม". Verify the actual subject inraw_transcript.json. - Comparative Analogy Promotion: When a secondary title (e.g. Heaven Burns Red, Yu-Gi-Oh! GX, Bleach) appears in a section dedicated to another main topic (e.g. Wuthering Waves), verify whether it was only used as a comparative example (e.g. ensemble cast or companion writing) and frame the timestamp around the primary topic.
- Gameplay vs. Gacha Pulling Transitions: During gameplay segments of gacha games (Nikke, Arknights: Endfield, WuWa), check the transcript for gacha pull keywords (
หลุดเรท,ตู้ถาวร,เพชรหมด,Token Exchange) so gacha sessions are not mislabeled as continuing gameplay mechanics (e.g. factory/belt building). - Dual-Track Desync (Gameplay State vs. Talk Topic): Pu Boat often sets up an in-game action (summoning mobs to fight each other, AFK farms, long crafts, waiting for a boss) and then talks about something unrelated for 5–40 minutes while it runs. Track both every time:
- Gameplay Track: Find the plan statement that explains the current scene. Search the transcript backward (up to ~20 min) for intent verbs (
เดี๋ยวให้มันสู้กัน,รอ,ปล่อยไว้,เสก,ล่อ). Confirm it with storyboard frames: quest tracker text, boss bar name, and the mobs actually present. A creature on screen does NOT mean Boat is fighting it. - Talk Track: Find the topic Boat is actually speaking about (donations, news, Q&A).
- Label Rule: If both tracks run at the same time, write one stamp that names both. Use the talk tag when speech dominates, and keep the game state as context (e.g.
[Talk] เฝ้า <Mob_A> สู้กับ <Mob_B> พร้อมถกประเด็น <Topic>). Never drop either track. Close the gameplay arc with a payoff stamp when it resolves (e.g. loot collected, quest completed).
- Gameplay Track: Find the plan statement that explains the current scene. Search the transcript backward (up to ~20 min) for intent verbs (
- Domain Context Bleeding: Check non-gaming talk segments adjacent to gaming segments. Local 12B models often distort general words (e.g.
- Apply Entity & Chronology Corrections: Correct phonetic drifts, inverted topic order, or LLM hallucinations directly in
anibon_timestamps.md, and verify spacing ($\ge 45\text{s}$ min gap, $\le 600\text{s}$ max gap). - Persist Knowledge:
- Add confirmed phonetic drift patterns (e.g.
เคราสเซียร์นี้แม่→Chaos Zero Nightmare,อัชเฉร้อน→อาชิยะ โดมัน) togarbled_replacements.jsonacross root and skill resources. - Add newly verified franchises/characters to
anibon-world-identity/references/.
- Add confirmed phonetic drift patterns (e.g.
Discipline & Anti-Rationalization
The Iron Rule of Local Processing
NEVER TIME-LOCK TOPICS TO ARBITRARY MINUTE BOUNDARIES
Livestreams flow organically. Use the Recursive Rolling Summary state-machine to detect true content shifts rather than stamping every $N$ minutes.
Rationalization Table
| Rationalization | Reality | Counter-Measure |
|---|---|---|
| "Single chunk is faster to run." | Produces 35+ micro-stamps that clutter comment sections. | Use --mode recursive with rolling summaries. |
| "I can invent tags like [วิเคราะห์] because it fits the content." | Front-tier benchmarks reject non-standard tags; breaks comment parsers. | Normalizer automatically remaps [วิเคราะห์] → [Talk]. |
| "I'll just summarize the whole 3-hour transcript in 1 prompt." | Exceeds context and degrades entity recall on 12B models. | Sequential rolling state-machine keeps prompt under 2k tokens. |
| "Let's write a custom Python script to speed this up." | Ad-hoc scripts break state tracking, resume logic, and encoding. | Strictly use process_chunks_local.py flags. |
| "The timestamps look good enough without checking World Identity." | Hallucinates post-cutoff games or uses bare English Pokémon names without community Thai transliteration. | Strictly invoke /anibon-world-identity gate before final handoff. |
Red Flags — STOP and Reset
- Delivering timestamp output to user without verifying franchise entities against
/anibon-world-identity. - Attempting to run timestamper before transcript recovery pipeline is 100% complete and validated (
whisper-corruption-recovery). - Output file contains 35+ timestamps for a 2-hour stream (micro-stamping symptom).
- Output tags include non-whitelisted words (
[วิเคราะห์],[ชำแหละ],[เปิดตัว]). - Over-splitting topics into excessive micro-parts (>6–8 parts for a stream). Viewer navigation and thematic cohesion take precedence over artificial comment byte limits.
- Script crashes with
UnicodeEncodeError: 'charmap'(forgot-X utf8). - Outro omitted when ending chunk is marked continuation. Always verify final 5 minutes for closing stamp.
- Attributing endgame mechanics ("หอรี", "เฟ้อกว่าเวเนซุเอลา") to a previously discussed game when a new game was booted up.
- Hallucinating unmentioned lore/chapter titles (e.g., "Lostbelt") solely due to pre-training association with a franchise.
- Labeling viewer donation alerts with custom avatars as literal product reviews or figure showcases (
[Talk] โชว์ฟิกเกอร์). - Labeling a stamp
[Gameplay] สู้กับ Xonly because X appears in a frame, without verifying the boss bar, quest HUD, or the streamer's stated plan. Equally wrong: replacing it with a pure[Talk]stamp that erases the running game state.
Pre-Roll Standby / Waiting Screen Invariant
If the opening transcript (chunk 00) contains predominantly [เพลง], sound effect markers, or isolated foreign ASR phonetic tokens (Japanese, Korean, English lyrics) before the streamer speaks the first coherent Thai sentence:
- MUST mark
00:00:00 - [Intro](e.g.00:00:00 - [Intro] เปิดไลฟ์ทักทายยามเย็น...). - Anchor the first topical timestamp strictly to the streamer's opening Thai speech.
- NEVER infer stream topics, reactions, or anime titles from background waiting music (e.g. mistaking opening waiting BGM for "ดูอนิเมะสาวน้อยเวทมนตร์").
Stream Ending / Outro Invariant
In --mode recursive, long unbroken gameplay or concluding sessions must not allow is_continuation=True to swallow the stream outro.
The final 5 minutes of transcript must always be checked for closing tokens (ขอบคุณครับ, ลงไลฟ์, ราตรีสวัสดิ์, เจอกัน, บ๊ายบาย) and emit:
- SuperChat / Birthday Shoutouts: Use
[Donation]for reading donor lists, SuperChats, and member birthday blessings (HH:MM:SS - [Donation] อ่านโดเนท ขอบคุณผู้สนับสนุน และอวยพรวันเกิดสมาชิก). - Closing Sign-off: Use
[Ending]for the final wrap-up statement and next-stream announcements (HH:MM:SS - [Ending] กล่าวปิดสตรีมและแจ้งกำหนดการไลฟ์รอบถัดไป...). Do NOT begin descriptions with'ปู่โบ๊ต'. - Recurring Live Routine Guardrail: Pu Boat frequently schedules a 22:00 Tokusatsu stream ("ไลฟ์โทกุ", "คอนเทนต์โทกุ 3 เรื่อง 3 รส"). Phonetic ASR artifacts (
โทกกู,ทั่วกุ,คอนเทนทุกคึ) must strictly map toโทคุซัตสึ (Tokusatsu), NEVER "Tokyo Ghoul".
Macro-Thematic Part Grouping Invariant (No 3,500-Byte Cap Over-Splitting)
Timestamps are grouped strictly into broad, cohesive macro-thematic parts (typically 4–6 parts for a long stream).
Prioritize viewer navigation and thematic cohesion: keep related discussions (such as an entire series review, media watchparty, or gaming session) unified in a single part rather than arbitrarily slicing into 10–18 micro-parts to fit single-comment byte limits.
assemble_parts() and run_local_summarizer_pass() do not enforce byte caps by default. enforce_byte_cap() is retained as an opt-in utility only if single-comment splitting is explicitly requested.
Essential Curated Timestamps Selection Invariant
When generating timestamps from long livestreams (4–8 hours), follow these operating standards:
- Primary Output Focus: Present the primary result as a curated essential edition (~40–60 timestamps for a 7–8 hour stream, or approximately 8–15 stamps per macro-thematic part).
- Automated Essential Pruning (
prune_essential): For streams yielding >60 timestamps,finalize_and_assemble_timestamps()automatically queries the local model to distill essential pivot points, news, gameplay milestones, and core conclusions down to ~40–60 stamps. The pipeline enforces strict verbatim matching against raw timestamps, guarantees retention of opening and[Ending]timestamps, and falls back cleanly if the LLM call fails.[!NOTE] If
prune_essentialfails to extract matches (e.g. local 12B model emits modified text that fails verbatim matching), the pipeline falls back to the full timestamp list. Reviewers should manually prune micro-talk stamps down to essential milestones before delivery.
- Automated Essential Pruning (
- Essential Selection Criteria:
- Major topic starts or pivot points (Topic Shift / Pivot).
- Official news, announcements, releases, or statistical data (Official News, Releases, Data).
- In-depth analytical conclusions and core takeaways (In-depth Analysis & Core Takeaways).
- Key gameplay milestones (tutorial start, exploring new characters/skills, key item crafting, boss spawns, highlight fights, victory/defeat).
- Stream outro and closing thanks (Stream Outro).
- Pruning Micro-Stamps:
- Prune micro-talk stamps repeating the same talking point within a short window (e.g. if the streamer dissects a point for 5–10 minutes, do not place stamps every 2 minutes).
- Prune casual chat chatter, minor jokes, or banter that does not impact the core narrative.
- Wrap Same-Topic Clusters: Consolidate consecutive stamps that belong to the same topic group into a single comprehensive timestamp.
- Skip Fleeting Micro-Talk: Never create a timestamp for an offhand reaction or passing thought spanning only 2–3 sentences (e.g. praising how a game translated a boss name into Thai).
- Dual-File Storage:
- Save the curated essential timestamps file to
anibon_timestamps_essential.md(oranibon_timestamps.md). - Retain the exhaustive, full chronological timestamp log in
anibon_timestamps_full.md(andall_timestamps.txt) for reference and archival lookup.
- Save the curated essential timestamps file to
Macro-Topic Boundary Snapping Invariant
Major domain/topic transitions (e.g. Thai Politics → Anime/Series review, Gaming → SuperChat Q&A) MUST NOT begin with 1–4 straggler stamps at the tail of an unrelated topic section.
assemble_parts() automatically snaps partition boundaries to transition markers (สรุปการพูดคุย... และเตรียมเปลี่ยนไปคุย, เตรียมเริ่มเล่นเกม, เตรียมเข้าสู่ช่วง) using find_best_boundary(). This ensures viewer-friendly, clean topical grouping without stragglers.
Timestamp Lead-in & Analogy Attribution Invariant
- Lead-in Anchor: The start time of a timestamp must always anchor to the opening sentence of the topic. Never mark the timestamp at or after the keyword appears mid-sentence, ensuring viewers who click the link do not miss critical context spoken 10–20 seconds earlier.
- Off-hand Analogy Rule: Never use the name of a work, person, or place mentioned only briefly in passing (1–2 sentences) as an analogy or rhetorical comparison as the main title of a timestamp.
- Incorrect:
Comparing standards... with Frieren(when the topic is a critique of One Punch Man Season 3 and Frieren was only mentioned in a single comparative sentence). - Correct:
Critiquing One Punch Man Season 3 and J.C.Staff's 'watchable' standard (noted that it cannot compare to major productions like Frieren).
- Incorrect:
World Identity Audit Invariant
NEVER deliver the final timestamp output without performing an anibon-world-identity audit on all proper nouns, game titles, and character names. Thai Whisper outputs are phonetic; local LLMs hallucinate well-known games (Genshin, Dota) when hearing unfamiliar ones (CZN, Deadlock). Local references (references/*.md) and card/entity databases (atlas_fgo.db, pokemon.db) take strict precedence over model memory. process_chunks_local.py automatically audits generated anibon_timestamps.md against <workspace>/entity_glossary.json upon completion to replace confirmed phonetic variants.
Stream Title Token Grounding Invariant
Proper nouns, mod names, and franchise titles in the stream's own YouTube metadata (info.json title, description) are ground-truth anchors.
- Title Token Priority: When an emitted timestamp description contains a phonetic near-match or truncated form of an entity present in the stream title (e.g. ASR transcribed
Vitywhile the title explicitly specifiesVerity, orNaniavsNarnia), the pipeline must deterministically snap to the title's canonical spelling. - Title vs. ASR Phonetic Drift: Stream titles reflect the streamer's deliberate broadcast topic. ASR phonetic drift within the video must never overwrite or degrade entities verified in the title.
LiveChat Dual-Side & Mood Invariants
- READ BOTH SIDES Invariant: When
livechat/exists, prompt context automatically loadsLiveChat Highlights:(prioritizing SuperChats, donations, and direct questions to Pu Boat). The model must evaluate both streamer dialogue and audience reactions to prevent one-sided factual misinterpretations. - HONOUR THE MOOD VERDICT Invariant: When
mood_555.jsonexists, each chunk receivesViewer Reaction:mood classification and Thai tone hints (tone+ suggestedverbs). The model keeps its own verb creativity, but must not flatline high-energy meme peaks (MEME_PULSE,SHOCK_HYPE_PULSE) into dry, detached monologues.
Common Mistakes & Troubleshooting
-
LM Studio Model Eviction:
- Symptom: LM Studio unloads Gemma 4 and reloads another model, causing 60-second stalls.
- Fix: Keep
--model auto.resolve_modelautomatically latches onto the active in-memory model.
-
Gemma 4 Thinking Token Trap:
- Symptom: Output timestamp is blank or truncated.
- Cause: Model used token budget in
reasoning_content. - Fix: Built-in system prompt limits thinking to <2 sentences; regex extractor extracts structured JSON directly from content or reasoning.
-
ASR Phonetic Drift:
- Symptom: Names like "นครโตะ" appear instead of "Naucrate (นอคราเต้)".
- Fix:
process_chunks_local.pyloadsresources/default_mappings.jsonand runs TF-IDF signal detection before prompt generation.
-
Missing Multimodal Projector (
mmproj) on Vision Prompts:- Symptom: Image verification or storyboard inspection responds with "I cannot see an image or video".
- Cause: Model GGUF has multimodal prompt tokens but was loaded into LM Studio /
llama.cppwithout its companionmmproj-*.gguffile. - Fix: Explicitly attach the
mmprojfile in LM Studio model settings orllama-server --mmproj, or switch to a unified vision model likeQwen2.5-VL-7B.
-
Re-running with Modified/Recovered Transcript:
- Symptom: Timestamper finishes in 1 second showing
[skip] chunk_XX (already processed). - Cause: Script caches per-chunk responses in
<workspace>/recursive_outputs/chunk_XX.jsonas well asanibon_timestamper_state.json. - Fix: Pass
--no-resumeand removerecursive_outputs/andanibon_timestamper_state.json.
- Symptom: Timestamper finishes in 1 second showing
-
Outro Swallowed by Continuation Trap:
- Symptom: Output timestamp stops 30–60 minutes before video end, missing streamer's closing remarks.
- Cause:
is_continuation=Truein recursive rolling mode treated the final session as unbroken gameplay, discarding outro. - Fix: Outro invariant — final 5 minutes must always be scanned for
[Ending]/ farewell stamps.
-
Storyboard MHTML with WebP Content-Type:
- Symptom:
prepare_storyboard_slides()extracts 0 slides despitestoryboard.mhtmlbeing downloaded (20MB+). - Cause: YouTube's
yt-dlpoften packages multipart storyboard slides withContent-type: image/webp(RIFF...WEBP) rather thanimage/jpeg. - Fix:
anibon/multimodal.pychecks bothimage/jpegandimage/webp. FFmpeg crops tiles from either format to high-res JPEG output without requiring PIL.
- Symptom:
-
yt-dlpStoryboard--skip-downloadTrap & Empty Slide Directories:- Symptom:
download_storyboard()exits 0 but produces no.mhtmlfile, orprepare_storyboard_slides()returns an emptyslides_clean/directory. - Cause: Unlike
--write-auto-subs --skip-download, storyboard (-f sb0/sb1/sb2/sb3) is the primary format; passing--skip-downloadskips downloading it entirely. Checkingclean_dir.exists()without globbingslide_*.jpgalso false-positives on empty directories. - Fix: Never pass
--skip-downloadwhen downloading-f sb0/sb1/sb2/sb3, and always verifyclean_dir.is_dir() and any(clean_dir.glob("slide_*.jpg")).
- Symptom:
-
Multimodal Endpoint Missing
/v1/chat/completions(KeyError: 'choices'):- Symptom:
call_multimodal()raisesKeyError: 'choices'when passed a barehttp://host:1234endpoint. - Cause:
urllib.requesthits the LM Studio root endpoint instead of/v1/chat/completions. - Fix: Always normalize
urlincall_multimodal()to append/v1/chat/completionsif not already present.
- Symptom:
-
Missing YouTube Storyboards (
sb0/sb1/sb2/sb3Not Available) in Multimodal Mode:
- Symptom:
yt-dlp -f sb0/sb1/sb2/sb3fails withRequested format is not available, leavingframes/slides_clean/empty. - Cause: Certain YouTube livestream VODs do not generate MHTML storyboard streams.
- Fix: Download the lightweight 144p video stream (
-f 160, ~190MB for an 8.5h stream) and synthesize native-compatible 960×540 3×3 storyboard slide sheets directly intoframes/slides_clean/:yt-dlp -f 160 -o "$WORKSPACE/frames/video_144p.mp4" "https://www.youtube.com/watch?v=<VIDEO_ID>" mkdir -p "$WORKSPACE/frames/slides_clean" # Compute interval = total_duration_sec / (350 * 9) or ~1 frame every 10s: ffmpeg -i "$WORKSPACE/frames/video_144p.mp4" -vf "fps=1/10,scale=320:180,tile=3x3" -q:v 3 "$WORKSPACE/frames/slides_clean/slide_%04d.jpg"
-
call_local()Positional Timeout Collision (HTTP 400 Bad Request):- Symptom: Summarizer or Pruning pass fails with
HTTP Error 400: Bad Requestand falls back to heuristics. - Cause: Calling
call_local(ep, model, prompt, max_tokens, temp, TIMEOUT)places the integer timeout into thesystem_promptparameter (arg 6), creating{"role": "system", "content": 180}. - Fix: Always pass timeout via keyword argument
timeout=....lmstudio.call_local()includes anisinstance(system_prompt, (int, float))guard to reroute legacy positional timeouts.
- Symptom: Summarizer or Pruning pass fails with
-
Tokusatsu Stage Show & TTFC Phonetic Traps:
- Stage Show Forms vs. Gacha Games: Stage show original forms (
Chaos Drone,Extra Shadow,Advance) must not be attributed to gacha games (Chaos Zero Nightmare). - ASR "ชิรูบิ": Maps to
ไรเดอร์ชิโนบิ (Kamen Rider Shinobi)(TTFC series). - Actor Memorials & Old Interviews: News regarding
Today is Friday in Californiarefers to actorHiroshi Tsuchida(Sasuke / Kakuranger). ActorShu Watanabe(Eiji Hino / Kamen Rider OOO) must never be confused with anime characters likeGojo Satoru. - Live-Action vs Anime: Chinese Tokusatsu series like Armor Hero (นักล่าอสูรราชันผงาด / โอเมก้าฮอร์น) are live-action Tokusatsu, not anime.
- Stage Show Forms vs. Gacha Games: Stage show original forms (
-
MTP (Multi-Token Prediction) & Speculative Decoding Pauses:
- Pausing a long run to attach MTP speculative decoding or adjust VRAM contexts in LM Studio is fully supported.
- Re-running the identical command automatically discovers
recursive_outputs/chunk_XX.jsonand state files, resuming immediately from the paused chunk without re-billing previous inferences.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/zenithth/anibon-stream-synthesis/anibon-timestamper-local">View anibon-timestamper-local on skillZs</a>