default_quality_mode: auto

start_lanes:
  SIMPLE: A
  MEDIUM_OPS: A
  MEDIUM_CODE: A
  COMPLEX: B
  REASONING: B
  VISION_OCR: A
  VISION_UI: A

quality_mode_map:
  economy: A
  balanced: B
  max: C

lane_order: [A, B, C]

budget_caps:
  warn_pct: 80
  hard_pct: 95
  warn_max_lane: B
  hard_max_lane: A

escalation:
  repeat_prompt_window_sec: 300
  context_tokens_min_lane_b: 32000
  session_ttl_sec: 1800

redis:
  key_prefix: "ai-router:session:"
  ttl_sec: 1800

classifier:
  mode: hybrid
  llm_model: gigachat-classifier
  low_confidence_threshold: 0.6
  timeout_sec: 15
  max_tokens: 64

# Strong Max plan → cheap coder workers/executor → conditional DeepSeek verify
hierarchical:
  enabled: true
  trigger_tiers: [COMPLEX, REASONING]
  header: X-AI-Orchestrate  # off | auto | force
  planner_model: novita-planner
  verifier_model: novita-verifier
  planner_fallback: b-complex
  verifier_fallback: b-complex
  verify_policy: on_fail_or_hard
  verify_input: digests
  synthesize_model: a-medium-code
  plan_confirm: true
  agent_executor_model: a-medium-code
  agent_escalate_model: b-complex
  agent_escalate_after_edit_failures: 2
  skip_synthesize_if_single: true
  max_subtasks: 3
  max_verify_retries: 1
  plan_max_tokens: 1024
  verify_max_tokens: 384
  synthesize_max_tokens: 1024
  worker_max_tokens: 768
  wall_clock_sec: 600
  digest_chars: 300
  call_timeout_sec: 300
  # First turn (kickstart read) can stay longer; after read_file body Coder hangs —
  # tool_loop uses short timeout + escalate_after_read → DeepSeek.
  executor_timeout_sec: 90
  executor_num_retries: 0
  executor_input_chars: 16000
  executor_tool_loop_chars: 12000
  executor_tool_loop_timeout_sec: 45
  executor_newest_tool_chars: 3500
  executor_minimal_context: true
  executor_max_tools: 6
  executor_tool_desc_chars: 120
  executor_tools_full_chars: 8000
  executor_max_tokens: 1024
  executor_use_stream: false
  executor_synthetic_on_timeout: true
  # First turn only: synthetic read_file on plan path if model returns prose.
  # Mid-loop still stops (no fake edits).
  executor_synthetic_mode: kickstart_only
  executor_max_synthetic_continues: 1
  executor_max_edit_failures: 4
  executor_midloop_model_retry: true
  executor_midloop_escalate: true
  # After successful read_file skip DeepSeek (was 408→400 loop) → edit_file
  executor_force_edit_after_read: true
  # After a successful read_file, skip cheap coder — it stalls on fat tool results
  executor_escalate_after_read: true
  # Rewrite read_file on already-read paths → edit (breaks re-read loops)
  executor_rewrite_reread: true
  # When model returns 200 without tool_calls: one escalate retry before kickstart
  executor_no_tools_escalate: true
  # After plan approve: 1) deterministic path index 2) if gaps → cheap find_* model
  path_resolve_enabled: true
  path_resolve_model: a-simple
  path_resolve_timeout_sec: 45
  path_resolve_max_tools: 3
  # DevOps (*.calentiq.com / Bad Gateway): probe live ports before edit
  runtime_probe_enabled: true
  runtime_probe_timeout_sec: 8
  # Optional: ssh host for docker inspect (empty = TCP probe from gateway only)
  runtime_probe_ssh_host: ""
  # Quiet Zed: milestones only (no ctx bars / LiteLLM waits / subtask dumps)
  progress_verbose: false
  show_context_fill: false
  progress_in_content: true
  stream_progress: true
  context_window_tokens: 32768
  plan_input_chars: 48000
  llm_cache_ttl_sec: 3600
  worker_map:
    simple: a-simple
    medium_code: a-medium-code
    hard: b-complex
  worker_map_max:
    hard: c-complex