config.example.yaml

version: 2

# Runtime defaults are intentionally non-destructive. Enable and calibrate a
# page profile before running. `session` enables input for that process; use
# --dry-run to preview. `run` respects both flags. See docs/deployment.md.
runtime:
  active_profiles: []
  poll_interval_seconds: 0.5
  dry_run: true

# Relative paths are resolved next to the private config file, not the repo.
storage:
  data_dir: data
  log_dir: logs
  screenshot_dir: screenshots

answering:
  providers:
    - name: deepseek_flash
      base_url: https://api.deepseek.com
      api_key_env: DEEPSEEK_API_KEY
      models: [deepseek-v4-flash]
      input_mode: text
      timeout_seconds: 20
      max_output_tokens: 8192
      request_json_object: true
      thinking: disabled  # Keep timed Flash responses short.
      reasoning_effort: null
  # Start three independent samples; add up to two if voting is inconclusive.
  minimum_responses: 2
  minimum_agreement: 2  # Winning vote count, not a percentage.
  minimum_confidence: 0.6
  maximum_parallel_requests: 3
  initial_rounds: 3
  maximum_rounds: 5  # Per configured model, within the question deadline.
  auto_apply: false
  prompt_template: null

knowledge:
  enabled: false
  database_path: data/knowledge.db
  question_ocr_enabled: true
  maximum_results: 4
  minimum_score: 0.08
  maximum_context_characters: 5000
  chunk_size_characters: 800
  chunk_overlap_characters: 100

pages:
  - id: example_yuketang
    display_name: Example Yuketang profile
    course_id: example_course
    enabled: false
    monitor_index: 1

    # For formal Windows operation, configure these in the PRIVATE config.
    # Use `autoykt windows` to obtain the selector and calibrated client size.
    # target_window:
    #   title_pattern: '<unique regular expression for the quiz window>'
    #   class_name: '<window class from autoykt windows>'
    #   client_size: [1280, 800]  # Replace with the actual dimensions.
    #   background_monitoring: false  # Read a fully visible background window.
    #   focus_point: null  # Optional calibrated neutral client-relative point.
    target_window: null
    # Calibrate with g, or configure {path, region, threshold} privately.
    # A bound live run requires this page identity and explicit success templates.
    page_guard: null

    # Every region is [x, y, width, height], relative to monitor_index,
    # or to the window CLIENT area when target_window is configured.
    regions:
      detection: [0, 0, 100, 100]
      question: [0, 0, 100, 100]
      answers: [0, 0, 100, 100]
      verification: [0, 0, 100, 100]
      rearm: [0, 0, 100, 100]

    # Multiple visual styles can route to answer or notification workflows.
    triggers:
      - name: single_choice
        path: templates/question.png
        threshold: 0.85
        consecutive_hits: 3
        action: answer
        question_type: single  # Use multiple for the multi-choice title template.

    # Optional private regression images. Empty means no screenshot validation.
    # See docs/configuration.md for labeled active/completed/expired samples.
    calibration_samples: []

    answer_style:
      option_match_threshold: 0.85
      # Match letter shapes across background colors at the calibrated size.
      match_grayscale: false
      # Calibrate colors with b/u and full size with c; see docs/deployment.md.
      # For central-letter templates, include button_size: [width, height].
      # Unknown background colors are reported with selected=null.
      button_colors: null
      selection_timeout_seconds: 2
      option_templates:
        A: templates/options/A.png
        B: templates/options/B.png
        C: templates/options/C.png
        D: templates/options/D.png
      # Used only when an option template cannot be located.
      fallback_positions: {}
      fallback_coordinate_space: monitor  # Use window for bound-window points.

    # Optional click needed to open or focus a question.
    entry_target: null

    # Formal operation requires the real submit button (point OR region).
    submit_target: null
    submit_on_select: false  # Only for confirmed single-choice auto-submit pages.

    question_ready:
      delay_seconds: 0.3
      timeout_seconds: 3
      stable_frames: 2
      maximum_change_ratio: 0.01

    # Formal operation requires explicit completion feedback templates.
    verification:
      success_templates: []
      failure_templates: []
      require_success_template: false  # Bound-window live runs always require it.
      timeout_seconds: 5
      poll_interval_seconds: 0.25
      stable_hits: 2
      minimum_change_ratio: 0.02

    # All stages share total_seconds. Text-only models require usable OCR.
    question_budget:
      total_seconds: 45
      context_seconds: 5
      answer_seconds: 25

    # Only transient model failures are retried, within the same time budget.
    recovery:
      maximum_attempts: 2
      delay_seconds: 1

    # Configure real state images and controls privately after calibration.
    # after_submit entries: {name, when: {path, region, threshold},
    #                        target: {point, coordinate_space}, stable_hits}
    # Each entry can click once, only after explicit success verification.
    page_flow:
      before_question: []  # Match a new-question notice, then click its entry.
      after_submit: []
      loading: []  # Regional templates: [{path, region, threshold}].
      manual: []   # Expired question/session or another state needing help.
      finished: [] # Class ended; stop after two observations.
      transition_timeout_seconds: 10

    # The completed question must change or disappear before retriggering.
    rearm:
      timeout_seconds: 30
      stable_hits: 2
      # Ignore recognized button colors; calibrate against consecutive pages.
      minimum_change_ratio: 0.02

# Personal IDs and tokens are read only from the named environment variables.
notifier:
  enabled: []
  qq:
    onebot_url: http://127.0.0.1:4096
    target_env: AUTOYKT_QQ_TARGET
    access_token_env: AUTOYKT_ONEBOT_ACCESS_TOKEN
  telegram:
    token_env: AUTOYKT_TELEGRAM_TOKEN
    chat_id_env: AUTOYKT_TELEGRAM_CHAT_ID

logging:
  level: INFO