config.example.yaml

# =============================================================================
# MPT Autopilot — config.example.yaml
#
# ONE config file for all stages. Copy it and edit:
#
#     cp config.example.yaml config.yaml
#     cp .env.example .env                # secrets go in .env, never here
#     cp accounts.example.yaml accounts.yaml
#     cp jobs/jobs.example.yaml jobs.yaml
#
# config.yaml, .env, and accounts.yaml are gitignored, so `git pull` will never
# overwrite your settings.
#
# Layout:
#     shared     — telegram_prefix, paths, langs   (read by several stages)
#     pilot:     — mpt refill / mpt init-seen
#     batch:     — mpt batch
#     enricher:  — mpt enrich
#     uploader:  — mpt upload
#     pipeline:  — mpt run
#
# EVERY relative path in this file resolves against THIS FILE's directory, not
# the directory you run `mpt` from. So a cron job can `mpt --config
# /srv/mpt/config.yaml run` from anywhere and get identical behaviour.
#
# A stage only needs its own section. `mpt batch` never reads `enricher:`, so
# you can leave sections you don't use out entirely.
# =============================================================================


# =============================================================================
# SHARED
# =============================================================================

# Prefix on every Telegram alert, e.g. "[mpt-autopilot] Batch done".
# Any section may override it with its own telegram_prefix (useful if you run
# several farms into one chat). TELEGRAM_TOKEN / TELEGRAM_CHAT_ID go in .env —
# leave them empty there to disable alerts entirely.
telegram_prefix: "mpt-autopilot"

# Default directories, shared by all stages. A stage-level key of the same name
# wins over these. Leave any of them out to fall back to per-stage defaults.
paths:
  jobs_dir: "./jobs"        # jobs.yaml / jobs_es.yaml live here
  seen_dir: "./jobs"        # seen.txt / seen_es.txt; shared primary setting for all stages
  exports_dir: "./exports"  # rendered videos (batch writes, enrich reads)
  videos_dir: "./videos"    # finished videos waiting for upload (uploader reads)

# ``mpt:`` is for optional integration with MoneyPrinterTurbo itself.
# Point ``config_path`` at your MPT install's TOML file so MPT-Autopilot
# can reuse the settings you already entered in the WebUI.  The upload-post
# block below reads credentials from there, so you do NOT enter them here.
# ── Language profiles ────────────────────────────────────────────────────────
#
# ONE shared block, because every stage needs the same answer to "what does the
# _es suffix mean?". Before the merge each tool had its own copy and a typo in
# one language config was a real source of broken multi-language runs.
#
# The file_suffix drives file naming everywhere:
#     ""      → jobs.yaml, seen.txt, exports/
#     "_es"   → jobs_es.yaml, seen_es.txt, exports_es/
#
# Only `pilot` reads voices / job_defaults / theme_list; `batch`, `enrich`, and
# `run` only need file_suffix. Delete a language you don't produce.
langs:
  en:
    label: English
    file_suffix: ""          # no suffix → bare filenames

    # Voice rate bounds used by pilot to validate LLM output (voice_rate).
    # Values outside this range are clamped to the nearest bound.
    voice_rate_min: 1.05
    voice_rate_max: 1.20

    # Allowed voice aliases — pilot rejects any LLM-suggested voice not here.
    # Omit for "any voice is fine" (pilot won't filter by name).
    # voices:
    #   - gemini:aoede
    #   - gemini:charon
    #   - gemini:puck

    # Optional theme mode for `mpt refill`: when non-empty, generated topics are
    # constrained to these themes. Omit for free-topic generation.
    # theme_list:
    #   - job
    #   - animal
    #   - computer

    # Written verbatim into every generated job entry. Pilot validates these
    # against MPT constraints; batch passes them through to the MPT API.
    job_defaults:
      video_clip_duration: 3
      video_concat_mode: random
      bgm_type: random
      bgm_volume: 0.15
      paragraph_number: 2
      # Target narration length, e.g. "30-60" or "120+" (seconds). Omit for no
      # limit. Converted to word bounds and injected into the script prompt;
      # paragraph_number is nudged up if needed, never down.
      # duration_range: "30-60"

  es:
    label: Spanish
    file_suffix: "_es"       # → jobs_es.yaml, seen_es.txt, exports_es/

    voice_rate_min: 1.05
    voice_rate_max: 1.15

    # voices:
    #   - gemini:aoede
    #   - gemini:kore
    #   - gemini:puck

    job_defaults:
      video_clip_duration: 4
      video_concat_mode: random
      bgm_type: random
      bgm_volume: 0.15
      paragraph_number: 2


# ``mpt:`` is for optional integration with MoneyPrinterTurbo itself.
# Point ``config_path`` at your MPT install's TOML file so MPT-Autopilot
# can reuse the upload-post credentials you already entered in the MPT WebUI.
# If you set batch.upload_post.* overrides below, those take priority.
mpt:
  config_path: ""  # path to MoneyPrinterTurbo's config.toml (optional)


# =============================================================================
# pilot: — `mpt refill` and `mpt init-seen`
#
# Generates video ideas with an LLM and appends them to the jobs queue.
# LLM_API_KEY / LLM_BASE_URL / LLM_MODEL come from .env.
# =============================================================================
pilot:
  generation:
    # How many new jobs one `mpt refill` adds.
    count: 21

    # Refill only fires when pending jobs drop below this. "Pending" = enabled
    # in the jobs file AND not yet rendered (not in the seen registry).
    # `--force` ignores the threshold.
    threshold: 10

    # Reasoning-model support. Leave the three keys out entirely and nothing is
    # sent — identical behaviour to a plain OpenAI-compatible provider. Only
    # relevant for self-hosted vLLM / SGLang / NVIDIA NIM serving a model that
    # turns "thinking" on by default (DeepSeek-R1, Qwen3-thinking, Inkling):
    # those can spend the whole token budget on a hidden draft and return an
    # empty answer unless you ask for less reasoning or more room.
    #
    # This maps to chat_template_kwargs.reasoning_effort. It does NOT apply to
    # official OpenAI reasoning models (o-series / gpt-5, which take a top-level
    # reasoning_effort) or to Anthropic extended thinking.
    reasoning_enabled: false     # true | false | omit entirely
    # reasoning_effort: "medium" # none|low|medium|high|xhigh|max
    # reasoning_max_tokens: 8192 # extra budget reserved for the hidden draft

  # Directories scanned by `mpt init-seen` for existing .mp4 files, so refill
  # never re-proposes a topic you already rendered. Combined with any --dir
  # flags. Leave empty to always pass --dir at runtime.
  scan_dirs: []

  # Overrides for the shared paths: block, if pilot needs different ones.
  # jobs_dir: "./jobs"
  # seen_dir: "./jobs"


# =============================================================================
# batch: — `mpt batch`
#
# Renders pending jobs through the MoneyPrinterTurbo API.
# =============================================================================
batch:
  # MoneyPrinterTurbo API endpoint (required). Replace with the actual host
  # running MPT — this is where `mpt batch` sends rendering jobs and polls
  # for results. Common values: "http://127.0.0.1:8080" (same machine),
  # "http://<server-ip>:8080" (remote server).
  api_url: "http://127.0.0.1:8080"

  # MoneyPrinterTurbo's own directories (required). These are paths on the
  # machine running MoneyPrinterTurbo, as it sees them.
  mpt_storage: "/root/MoneyPrinterTurbo/storage"           # contains tasks/, cache_videos/
  mpt_songs_dir: "/root/MoneyPrinterTurbo/resource/songs"  # background .mp3 files

  # Where finished videos and their script.json files land. `mpt enrich` reads
  # this same directory, with the language suffix applied (exports_es/ etc.).
  # Defaults to paths.exports_dir.
  # output_dir: "./exports"

  # Dedup registry location comes from the shared `paths.seen_dir` (or
  # `pilot.seen_dir` if you only need an override for pilot). `--lang` derives
  # `seen_es.txt` from that directory automatically.
  # seen_file: "./seen.txt"

  # Direct jobs file path (optional). When set it wins over --lang suffixes and
  # jobs_dir. Usually you want the langs: mechanism instead.
  # jobs: "./jobs_prod.yaml"

  log_file: "./logs/batch.log"
  log_max_mb: 10

  # Give up waiting for one task after this long.
  max_wait_seconds: 2400
  # Give up if a task's progress does not move for this long.
  stuck_threshold_seconds: 3600

  # Retries cover transient trouble (dropped connection, MPT briefly down, a
  # one-off stuck task). They will not fix bad credentials —
  # max_consecutive_failures is the circuit breaker for that.
  max_retries: 3
  retry_delay_seconds: 180
  max_consecutive_failures: 3

  # MoneyPrinterTurbo's cache_videos/ folder grows without bound. Cleanup runs
  # once at the end of every batch when enabled; the interval additionally
  # triggers it every N successful videos during long runs (0 = only at the
  # end). Set cache_cleanup_enabled: false to never touch it.
  cache_cleanup_enabled: true
  cache_cleanup_interval: 6

  # ── Upload via upload-post.com ──────────────────────────────────────────
  #
  # After a video finishes rendering, optionally submit it to
  # [upload-post.com](https://docs.upload-post.com) for cross-posting to
  # YouTube / TikTok / Instagram.  The credentials are read from your
  # existing MoneyPrinterTurbo config.toml (set `mpt.config_path` below),
  # so you do NOT enter them again here.  Omit the keys entirely to fall
  # back to MPT's config, or set them here to override per-run.
  upload_post:
    enabled: false
    # api_key: ""           # override — leave empty to read from MPT config.toml
    # username: ""          # override — leave empty to read from MPT config.toml
    # platforms: ["youtube", "tiktok"]  # override — leave empty to read from MPT
    # youtube_privacy_status: "public"  # public | unlisted | private

  # ── Voice presets ────────────────────────────────────────────────────────
  #
  # MoneyPrinterTurbo picks the voice purely from the shape of `voice_name`: a
  # "provider:voice" prefix like "gemini:puck" for paid providers, or a plain
  # Edge TTS id like "es-ES-ElviraNeural" for the free default. `tts_server`
  # rides along for readability.
  #
  # All 314 free Edge TTS voices are built in — you do NOT list them here.
  # Browse them with:
  #
  #     mpt batch --list-voices es       # every Spanish variant
  #     mpt batch --list-voices gemini   # your own presets
  #
  # and use any alias straight in jobs.yaml:
  #
  #     defaults:
  #       voice: "es_es_elvira"          # es-ES-ElviraNeural (Female)
  #
  # This section is only for EXTRA presets: paid providers, or a bundled voice
  # with a custom rate/volume. Each preset sets tts_server and voice_name
  # together so they cannot drift apart. The grouping keys (gemini, ...) are
  # for readability only — alias names must be unique across the whole section,
  # and an alias here overrides a bundled one of the same name.
  #
  # Paid providers need their key configured on the MoneyPrinterTurbo SERVER;
  # nothing here can supply it. If your fork supports siliconflow / mimo /
  # elevenlabs / azure v2, check its voice.py for the voice_name shape and add
  # a group the same way.
  voices:
    # Same bundled Edge voice, but slower — good for a dense script.
    # edge_overrides:
    #   es_es_elvira:
    #     tts_server: "edge"
    #     voice_name: "es-ES-ElviraNeural-Female"
    #     voice_rate: 0.9

    # Gemini TTS — paid, needs a Gemini key on the MPT server. Full list:
    # https://ai.google.dev/gemini-api/docs/speech-generation
    gemini:
      gemini_puck:
        tts_server: "gemini"
        voice_name: "gemini:puck"
      gemini_aoede:
        tts_server: "gemini"
        voice_name: "gemini:aoede"
      gemini_kore:
        tts_server: "gemini"
        voice_name: "gemini:kore"
      gemini_charon:
        tts_server: "gemini"
        voice_name: "gemini:charon"
      gemini_fenrir:
        tts_server: "gemini"
        voice_name: "gemini:fenrir"
      gemini_leda:
        tts_server: "gemini"
        voice_name: "gemini:leda"

    # A preset may pin rate/volume too:
    # gemini_aoede_fast:
    #   tts_server: "gemini"
    #   voice_name: "gemini:aoede"
    #   voice_rate: 1.3
    #   voice_volume: 1


# =============================================================================
# enricher: — `mpt enrich`
#
# Generates hashtags for rendered videos and writes them into the sidecar json
# the upload stage reads. LLM credentials come from .env.
# =============================================================================
enricher:
  # Directory scanned when --dir/--file is not given. Defaults to the current
  # directory. In a `mpt run` pipeline this is ignored: the exports directory
  # for the language is passed explicitly.
  # videos_dir: "./exports"

  # Target platform — affects both the prompt and the tag hard limit:
  #   youtube    3–5 tags recommended, hard limit 60
  #   tiktok     3–5 tags, hard limit 5
  #   instagram  3–5 tags, hard limit 5
  # `mpt enrich --platform tiktok` overrides this for one run.
  platform: youtube

  # The LLM generates between min_tags and max_tags CONTENT hashtags.
  # always_include counts against max_tags. 3–5 total is current best practice
  # across all major platforms.
  min_tags: 3
  max_tags: 5

  # Tags longer than this (excluding the #) are dropped: nobody searches them.
  max_tag_length: 20

  # Never emitted, whatever the LLM returns. Empty by default — this tool has no
  # opinion about which tags are bad for you.
  # banned_tags:
  #   - "#viral"
  #   - "#foryou"
  banned_tags: []

  # Always prepended, in this order. The LLM is told to skip them, so there are
  # no duplicates or ordering surprises.
  always_include:
    - "#shorts"

  # Maximum videos sent to the LLM in a single request (batch enrichment).
  # 1 = fallback to per-file processing (no batching).
  # Higher values reduce API calls but require a model with a large enough
  # context window. The CLI flags --batch-size and --no-batch override this.
  batch_size: 5

  log_dir: "./logs"
  log_max_mb: 5

  # Set false for reasoning models (o1, o3, o4) that reject `temperature`.
  supports_temperature: true

  # Same three-state reasoning switch as pilot.generation above; see there for
  # the full explanation. Omit entirely to send nothing.
  reasoning_enabled: false
  # reasoning_effort: "medium"
  # reasoning_max_tokens: 8192

  # ── Prompts ──────────────────────────────────────────────────────────────
  # Required. Placeholders available in all three:
  #   {video_subject}  topic (from the filename or script.json)
  #   {platform}       youtube | tiktok | instagram
  #   {min_tags} {max_tags} {max_tag_length}
  #   {excluded_tags}  always_include + banned_tags, comma-joined
  #   {language}       prompt_generate only
  #   {text}           prompt_detect_language only

  # Fallback only: used if the combined call below fails to parse, in which case
  # the tool detects the language first and then generates tags.
  prompt_detect_language: |
    What language is the following text written in or intended for?
    Reply with ONLY the language name in English (e.g. "English", "Spanish", "Russian").
    Do not add any explanation or punctuation — just the language name.
    Text: "{text}"

  # Used when the language is NOT already known (no --lang, no video_language in
  # script.json): detects the language and generates tags in ONE call.
  # Must return {"language": "...", "tags": ["#tag1", ...]}.
  prompt_detect_and_generate: |
    You are a {platform} short-form video SEO specialist.

    Video topic: "{video_subject}"

    Task 1: Detect the language this content is in or intended for.
    Task 2: Generate between {min_tags} and {max_tags} SEO hashtags for {platform}.

    Hashtag rules:
    - Primary tags (2–3): highly specific to THIS video's unique angle
    - Secondary tags (1–2): niche-specific, medium competition
    - All tags in the detected language
    - Lowercase only, no CamelCase
    - Preserve diacritics (e.g. #recuperación, #astronomía)
    - Max {max_tag_length} characters per tag (after #)
    - Tags to EXCLUDE (already included automatically): {excluded_tags}

    Respond ONLY with a JSON object — no markdown, no explanation:
    {{"language": "English", "tags": ["#tag1", "#tag2", "#tag3"]}}

    Few-shot examples:
    topic "ostriches never bury heads" →
    {{"language": "English", "tags": ["#ostrichfacts", "#ostrichmyths", "#birdmyths", "#animalscience"]}}
    topic "nadie puede oír tus gritos en el espacio" →
    {{"language": "Spanish", "tags": ["#silencioenelespacio", "#sonidoenelespacio", "#vacíoespacial", "#curiosidadescósmicas"]}}

  # Used when the language IS known (--lang, or video_language in script.json).
  prompt_generate: |
    You are a {platform} short-form video SEO specialist applying 2026 best practices.

    Video topic: "{video_subject}"
    Target language: {language}
    Target platform: {platform}

    Generate between {min_tags} and {max_tags} CONTENT hashtags.

    ── PRIORITY ORDER (most to least important) ──────────────────────────────────

    PRIMARY tags (generate 2–3):
      Highly specific to THIS video's unique angle.
      Ask yourself: "What would someone type into {platform} search to find exactly this video?"
      Good examples:
        topic "ostriches don't bury heads"    → #ostrichfacts  #ostrichmyths  #birdmyths
        topic "sound impossible in space"     → #silencioenelespacio  #sonidoenelespacio
        topic "opossum involuntarily faints"  → #involuntaryfainting  #opossumfacts
      Bad examples (do NOT generate these):
        ❌ #animals  ❌ #nature  ❌ #wildlife  ❌ #science  ❌ #espacio  ❌ #ciencia

    SECONDARY tags (generate 1–2):
      Describe the broader niche — medium competition level (not mega-popular, not obscure).
      Think: "What community or sub-topic does this video belong to?"
      Good examples: #animalscience  #astronomía  #curiosidadesdeluniverso  #animalbehavior

    ── STRICT FORMATTING RULES ──────────────────────────────────────────────────
    - All hashtags in {language} — never mix languages
    - Every hashtag starts with #, no spaces inside
    - ALL LOWERCASE — no CamelCase, no capitals
    - Preserve native diacritics: #recuperación, #astronomía, #ñoño (NOT stripped)
    - Max {max_tag_length} characters per hashtag (counting only the part after #)
    - No duplicate tags

    ── QUALITY RULES ────────────────────────────────────────────────────────────
    - Relevance over volume — a niche tag beats a mega-tag every time
    - NO off-topic tags (don't add #astronautas to a video about sound in a vacuum)
    - NO awkward keyword-stuffed compounds (#didyouknowostriches, #funfactsaboutanimals)

    Tags to EXCLUDE entirely (they are added automatically — do NOT include them):
    {excluded_tags}

    ── OUTPUT FORMAT ────────────────────────────────────────────────────────────
    Return ONLY a valid JSON array of strings — no markdown fences, no explanation, nothing else.

    Few-shot examples:

    English / youtube / topic "ostriches don't bury heads":
    ["#ostrichfacts", "#ostrichmyths", "#birdmyths", "#animalscience"]

    Spanish / youtube / topic "sound impossible in space":
    ["#silencioenelespacio", "#sonidoenelespacio", "#vacíoespacial", "#curiosidadescósmicas"]

    English / tiktok / topic "opossum involuntarily faints":
    ["#opossumfacts", "#involuntaryfainting", "#animalbiology", "#wildlifefacts"]

    Spanish / instagram / topic "moon fits inside Russia":
    ["#lalunacaberusia", "#curiosidadesdeluniverso", "#astronomía", "#datoscuriosos"]


# =============================================================================
# uploader: — `mpt upload`
#
# Uploads rendered videos to YouTube via youtubeuploader. Per-channel OAuth
# paths live in the SEPARATE accounts.yaml, because that is the file you are
# most likely to keep outside the repository.
# =============================================================================
uploader:
  # The youtubeuploader binary: https://github.com/porjo/youtubeuploader
  # A bare name (e.g. "youtubeuploader") is looked up on PATH.
  uploader_binary: "/root/youtubeuploader/youtubeuploader"

  # Scratch directory for the metaJSON file handed to youtubeuploader before
  # each upload. Created if missing; defaults to ./meta beside this file.
  meta_dir: "./meta"

  # accounts.yaml location. Defaults to accounts.yaml beside THIS file.
  # accounts_file: "./accounts.yaml"

  # Pause between uploads, so YouTube does not see a burst.
  sleep_between_uploads: 5

  # Subfolder inside each account's videos_dir where a video is moved once its
  # upload succeeds.
  uploaded_dir_name: "old_videos"

  # Used when a video has no <name>.json sidecar, or the sidecar omits a field.
  defaults:
    privacy_status: "private"   # private | unlisted | public
    category_id: "22"           # 22 = People & Blogs
    tags: ["shorts"]

    # Where the hashtags from <video>.json land:
    #   tags        hidden metadata tags field only
    #   description visible hashtags appended to the description
    #   both        both fields independently (default)
    hashtag_placement: both

    # YouTube's "AI use" disclosure (containsSyntheticMedia). Default true,
    # because MoneyPrinterTurbo output is AI-generated realistic content. A
    # sidecar can override it per video with "containsSyntheticMedia".
    contains_synthetic_media: true


# =============================================================================
# pipeline: — `mpt run`
#
# `mpt run` chains refill → batch → enrich → upload, once per language. It
# needs no configuration of its own; the only setting is the language→account
# mapping below.
# =============================================================================
pipeline:
  # Which uploader account (a key in accounts.yaml) each language uploads to.
  # Omitted languages use the language code as the account name, so if your
  # accounts are already named "en"/"es" you can delete this whole block.
  # accounts:
  #   en: main-channel
  #   es: spanish-channel

# One join `mpt run` cannot make for you: each account's `videos_dir` in
# accounts.yaml must be the directory batch renders that language into —
# batch.output_dir plus the language's file_suffix. With the settings above:
#
#   batch.output_dir: ./exports    →  en uses ./exports
#                                     es uses ./exports_es
#
# Nothing validates this, because uploading from a folder no batch run produced
# is a legitimate use. `mpt run --dry-run` shows it: upload finding no files
# right after batch listed jobs to render is the symptom of a mismatch.