agent_config.yaml

# =============================================================================
# AGENT CONFIGURATION
# =============================================================================
# This file contains agent capabilities and tool settings.
# Credentials should be stored in .env file.
# For non agent related config, see config.yaml

llm:
  # Model name from src/llms/manifest/models.json
  name: "claude-opus-5-5"
  flash: "claude-sonnet-5-5"
  compaction: "" # blank → same model as flash
  fetch: ""         # blank → same model as flash
  fallback:
    - "deepseek-flash"
    - "glm-5.2"

sandbox:
  # Provider selection: "daytona" (cloud) or "docker" (local).
  # - daytona: Cloud-hosted sandboxes via Daytona API. Each workspace gets an isolated
  #   VM with auto-stop/archive/delete lifecycle. Snapshots are versioned by a hash of
  #   this config (working_dir, Python version, dependencies, apt packages).
  # - docker: Local Docker containers using Dockerfile.sandbox. Useful for development.
  # Override at runtime with SANDBOX_PROVIDER env var.
  provider: "daytona"

  # Daytona cloud sandbox configuration
  daytona:
    base_url: "https://app.daytona.io/api"
    auto_stop_interval: 3600 # 1 hour in seconds
    auto_archive_interval: 604800 # 7 days in seconds — keep workspaces "stopped" (fast restart) longer before moving to cold storage
    auto_delete_interval: 7776000 # 90 days in seconds — total lifetime before a dormant sandbox is garbage-collected
    python_version: "3.12"

    # Hash will be automatically appended based on dependencies
    snapshot_enabled: true # Recommended: Enable snapshot-based sandbox creation
    snapshot_name: "langalpha" # Base name for snapshot (hash will be appended)
    snapshot_auto_create: true # Automatically create snapshot if it doesn't exist

  # Platform-secret catalog — Daytona provider only (ignored under docker/memory).
  # Each entry delivers a backend-owned API key into sandboxes as a managed
  # Secret: the sandbox env var holds an opaque placeholder, and Daytona
  # substitutes the real value into outbound HTTPS request *headers* for the
  # allowlisted hosts only. Declared per deployment (not code); an entry is the
  # opt-in and works in OSS mode too. Uncomment the FMP example to enable —
  # boot then fail-closes unless DAYTONA_SECRET_NAMESPACE and the source env
  # var are set.
  # platform_secrets:
  #   - source_env_var: "FMP_API_KEY"        # read from the host .env
  #     sandbox_env_var: "FMP_API_KEY"       # placeholder-bearing var in the sandbox
  #     name_suffix: "platform-fmp-api-key"  # Secret name = <DAYTONA_SECRET_NAMESPACE>-<suffix>
  #     description: "Platform FMP API key"
  #     hosts: ["financialmodelingprep.com"] # egress substitution allowlist

  # Docker local sandbox configuration
  # Ignored when provider is "daytona". Override with DOCKER_SANDBOX_* env vars.
  docker:
    image: "langalpha-sandbox:latest" # Build with: docker build -f Dockerfile.sandbox -t langalpha-sandbox:latest .
    memory_limit: "4g"
    cpu_count: 2.0
    network_mode: "bridge" # "bridge" (default) or "none" for full isolation
    dev_mode: false # true = bind-mount host dir (fast local dev, files shared instantly)
    # host_work_dir: "/tmp/langalpha-sandbox-work"  # Required when dev_mode is true
    # volumes:  # Extra mounts in "host:container[:ro]" format
    #   - "/data/datasets:/mnt/datasets:ro"

mcp:
  # The servers that ship with LangAlpha are declared by the bundles under
  # plugins/, one Agent Plugins package per group -- see plugins/README.md.
  # This list is yours: servers you add, and overrides for the ones we ship.
  #
  # A name no bundle declares ADDS a server. A name a bundle already declares
  # OVERRIDES it -- the keys you write here are laid over the shipped server
  # and every key you leave out keeps its shipped value. Switching one off is
  # two lines, and changing one field never means restating the command:
  #
  #   - name: "price_data"          # stop launching a server we ship
  #     enabled: false
  #
  #   - name: "yf_price"            # keep ours, show the agent every tool
  #     tool_exposure_mode: "detailed"
  #
  #   - name: "fundamentals"        # keep our description, run your own build
  #     command: "uvx"
  #     args: ["--from", "my-fundamentals-mcp", "serve"]
  #
  # Overriding here rather than editing plugins/ keeps your choice out of a
  # file the next pull overwrites. `env` and `headers` replace the shipped map
  # whole rather than merging key by key, so you can take a variable away as
  # well as add one.
  #
  # Third-party stdio servers: always launch isolated -- uvx/npx with pinned
  # versions -- never from the shared environment, whose dependency pins
  # (including the mcp SDK) will strand them on the next SDK major. Builtin
  # servers under plugins/ are era-proof via _bootstrap and run with
  # `uv run python`.
  servers:
    # A no-op override, here so the shape is in front of you. Flip it to
    # false and price_data stops launching on the next restart.
    - name: "price_data"
      enabled: true

    # Adding a server of your own looks like this:
    # - name: "tavily"
    #   description: "Web search engine for finding current information online"
    #   instruction: "Use for web searches, news, research, and real-time information."
    #   tool_exposure_mode: "summary"
    #   transport: "stdio"
    #   command: "npx"
    #   args: ["-y", "tavily-mcp@latest"]
    #   env:
    #     TAVILY_API_KEY: "${TAVILY_API_KEY}"

  # Tool discovery settings
  tool_discovery_enabled: true
  lazy_load: true # Load tools on-demand
  cache_duration: 300 # Cache tool metadata for 5 minutes

filesystem:
  # Filesystem access configuration for first-class filesystem tools.
  # These tools provide direct file and directory operations without code generation.

  # Working directory inside the sandbox — the root for all agent file operations.
  # Agent sees virtual paths like /results/file.txt which map to {working_directory}/results/file.txt.
  # Must match the home directory of the sandbox user (created in Dockerfile.sandbox).
  # Changing this value triggers automatic workspace migration: existing workspaces will
  # back up files to PostgreSQL and recreate their sandbox with the new working directory
  # on next reconnect (tracked via sandbox_config_hash in workspaces.config JSONB).
  working_directory: "/home/workspace"

  # allowed_directories and denied_directories are auto-derived from working_directory:
  #   allowed: [working_directory, "/tmp"]
  #   denied:  [working_directory/_internal]
  # Override only if you need custom values.

  enable_path_validation: true # Validate paths against allowed_directories

storage:
  # Cloud storage provider for image/chart/file uploads and avator
  # All credentials and settings are loaded from .env file
  # Options: s3, r2, oss, none (to disable uploads)
  provider: "s3"

logging:
  # Logging configuration
  level: "INFO" # DEBUG, INFO, WARNING, ERROR, CRITICAL
  file: "logs/ptc.log"

agent:
  background_auto_wait: false

# Subagent configuration
subagents:
  # List of enabled subagents (available: research, general-purpose, data-prep, equity-analyst, report-builder, plus user-defined)
  enabled:
    - general-purpose
    - research
    - data-prep
    - equity-analyst
    - report-builder

  # User-defined subagents (optional). Each key becomes a subagent name.
  # definitions:
  #   equity-analyst:
  #     description: "Specialized equity research analyst for deep-dive company analysis"
  #     mode: ptc                              # ptc or flash
  #     role_prompt: |
  #       You are a senior equity research analyst...
  #     tools: [execute_code, filesystem, finance, web_search]
  #     skills: [automation]                   # Runtime-loaded via SkillsMiddleware
  #     preload_skills: [creating-financial-models, xlsx]  # Injected into prompt
  #     max_iterations: 15
  #     sections:
  #       workspace_paths: true
  #       tool_guide: true
  #       data_processing: true
  #       visualizations: true

# Compaction Middleware Configuration
# Automatically compacts conversation context when approaching token limits
compaction:
  enabled: true # Enable/disable context compaction
  token_threshold: 120000 # Trigger full compaction when messages exceed this token count
  keep_messages: 10 # Preserve last N messages after compaction
  truncate_args_trigger_messages: 40 # Truncate tool args when >= N messages (Tier 1)
  truncate_args_keep_messages: 10 # Protect last N messages from arg truncation
  truncate_args_max_length: 2000 # Per-arg-value max chars before truncation

# =============================================================================
# TOOL CONFIGURATION
# =============================================================================

# Search API Configuration
# Options: tavily, bocha, serper, exa, parallel
search_api: tavily

# Crawler Configuration
# Web crawling with SafeCrawlerWrapper for circuit breaker and fault tolerance.
# Backend: "scrapling" uses tiered HTTP→browser→stealth fetching.
crawler:
  backend: "router" # Options: "scrapling", "router" (adds PDF/YouTube/X extraction)
  # Stage-level concurrency caps (replace the old max_concurrent_crawls).
  # Tier 1 (curl_cffi) is cheap, Tier 2/3 (browsers) bound RAM at ~400MB each.
  http_concurrency: 20    # Max concurrent Tier-1 HTTP fetches
  browser_concurrency: 6  # Max concurrent Tier-2/3 browser fetches (~2.4GB RAM ceiling)
  page_timeout: 60000 # Page load timeout in ms (overall, all tiers)

  # Circuit breaker settings — applied to both per-host and global infra breakers.
  # Per-host: a failing Reuters cannot poison Wikipedia.
  # Global infra: trips on cross-cutting failures (browser crash, DNS).
  circuit_breaker:
    failure_threshold: 5
    recovery_timeout: 60
    success_threshold: 2

  # Queue limit — admission-control counter. Returns queue_full immediately when
  # in-flight exceeds max_size; no separate slot wait.
  queue:
    max_size: 100

# Web Fetch Configuration
# Settings for web_fetch tool's sitemap-aware content extraction
web_fetch:
  sitemap_enabled: true # Sitemap fetching for URL suggestions
  sitemap_max_urls: 100 # Max URLs to fetch from sitemap
  sitemap_max_examples: 3 # Max example URLs per path prefix in summary
  sitemap_timeout: 10 # Timeout in seconds for sitemap fetch