example.config.yaml

# Universal Agent Runtime (UAR) Configuration
# This file serves as a reference for all available configuration options.
# Configuration is loaded in the following order (highest priority first):
# 1. CLI Arguments (e.g. --port 8080)
# 2. Environment Variables (e.g. UAR_SERVER__PORT=8080)
# 3. Config File (this file)
# 4. Defaults

server:
  # The port to listen on.
  # Default: 1906
  # Env: UAR_SERVER__PORT
  port: 1906

  # The host address to bind to.
  # Default: 0.0.0.0
  # Env: UAR_SERVER__HOST
  host: "0.0.0.0"

security:
  # Whether to require JWT authentication for requests.
  # Default: true
  # Env: UAR_SECURITY__JWT_REQUIRED
  jwt_required: true

  # The secret key used to sign and verify HS256 JWT tokens.
  # Required at startup when jwt_required=true. Generate a deliberate value,
  # for example with: openssl rand -hex 32
  # Env: UAR_SECURITY__JWT_SECRET
  jwt_secret: "replace_with_a_deliberate_secret"

  # Optional registered claims. When set, matching claims are required.
  # Env: UAR_SECURITY__JWT_ISSUER / UAR_SECURITY__JWT_AUDIENCE
  jwt_issuer: null
  jwt_audience: null

  # Reject tokens whose nbf (not-before) time is beyond jsonwebtoken's
  # 60-second default clock-skew allowance.
  # Default: true
  # Env: UAR_SECURITY__JWT_VALIDATE_NBF
  jwt_validate_nbf: true

  # Require the configured X-UAR-Admin-Key on protected settings and reload APIs.
  # Default: true. Set false only on trusted localhost for admin UI without a header.
  # Env: UAR_SECURITY__SETTINGS_MUTATION_AUTH_REQUIRED
  settings_mutation_auth_required: true

  # Replace before startup. Required when settings mutation auth is enabled.
  # Env: UAR_SECURITY__SETTINGS_ADMIN_KEY
  settings_admin_key: "replace-with-a-long-random-secret"

# Governed outbound A2A is disabled when this section has no trusted peers.
# Each endpoint must also appear in the calling agent artifact's a2a.dependencies.
a2a:
  # Stable ID acknowledged by other UAR instances. Required when peers exist.
  instance_id: "uar-west"
  trusted_peers:
    # Exact named-agent JSON-RPC endpoint; HTTPS is required off loopback.
    - instance_id: "uar-east"
      agent_id: "general-purpose"
      endpoint: "https://uar-east.example.com/a2a/agents/general-purpose"
      # A target-issued JWT whose signed `uar_instance_id` claim is exactly
      # this instance_id ("uar-west"). Keep it out of committed production
      # config and set UAR_A2A__TRUSTED_PEERS__0__BEARER_TOKEN in the environment.
      bearer_token: "replace-with-target-issued-jwt"

resilience:
  # Enable rate limiting to prevent abuse.
  # Default: true
  # Env: UAR_RESILIENCE__RATE_LIMIT_ENABLED
  rate_limit_enabled: true

  # Usage quota: allowed requests per second.
  # Default: 5.0
  # Env: UAR_RESILIENCE__REQUESTS_PER_SECOND
  requests_per_second: 5.0

  # Burst size: max requested allowed in a short burst.
  # Default: 10.0
  # Env: UAR_RESILIENCE__BURST_SIZE
  burst_size: 10.0

persistence:
  # The database provider to use. Supported values: "surreal" or "postgres".
  # Default: "surreal"
  # Env: UAR_PERSISTENCE__PROVIDER
  provider: "surreal"

  # Connection string for the primary database.
  # Default: "surrealkv://./data/uar.db"
  # Env: UAR_PERSISTENCE__DATABASE_URL
  database_url: "surrealkv://./data/uar.db"

  # Whether to use an external cache (Redis) for session and token storage.
  # Default: false
  # Env: UAR_PERSISTENCE__EXTERNAL_CACHE_ENABLED
  external_cache_enabled: false

  # Connection string for Redis (only used if external_cache_enabled is true).
  # Env: UAR_PERSISTENCE__REDIS_URL
  # redis_url: "redis://localhost:6379"

# =============================================================================
# LLM CONFIGURATION (liter-llm unified client)
# =============================================================================
# Configure the default LLM provider and model used by UAR agents.
# Uses liter-llm to support 142+ providers with a single unified API.
#
# Environment variable precedence (highest → lowest):
#   1. CLI args  (--llm-model, --llm-api-key, --llm-base-url, …)
#   2. UAR_LLM__* env vars  (e.g. UAR_LLM__MODEL, UAR_LLM__API_KEY)
#   3. Legacy LLM_* env vars  (e.g. LLM_MODEL, LLM_API_KEY — for backward compat)
#   4. Provider-specific keys  (e.g. OPENAI_API_KEY, ANTHROPIC_API_KEY, GROQ_API_KEY)
#   5. This config file  (llm: section below)
#   6. Compiled defaults

llm:
  # Model in "provider/model" format.
  # See https://models.dev for a full list of supported providers and models.
  # Default: "openai/gpt-4o"
  # Env: UAR_LLM__MODEL | LLM_MODEL | --llm-model
  model: "openai/gpt-4o"

  # API key for the selected provider.
  # Can also be set via OPENAI_API_KEY, ANTHROPIC_API_KEY, GROQ_API_KEY, etc.
  # Default: none (required for cloud providers)
  # Env: UAR_LLM__API_KEY | LLM_API_KEY | --llm-api-key
  # api_key: "sk-..."

  # Base URL override — bypasses liter-llm provider auto-detection.
  # Use this to point at local Ollama, LM Studio, or a custom proxy.
  # Default: none (auto-detected from provider name)
  # Env: UAR_LLM__BASE_URL | LLM_BASE_URL | --llm-base-url
  # base_url: "http://localhost:11434"

  # Request timeout in seconds.
  # Default: 60
  # Env: UAR_LLM__TIMEOUT_SECS
  timeout_secs: 60

  # Maximum number of retries on 429 / 5xx errors.
  # Default: 3
  # Env: UAR_LLM__MAX_RETRIES
  max_retries: 3

  # Enable parallel tool calls (where supported by the model).
  # Default: null (let the model decide)
  # Env: UAR_LLM__PARALLEL_TOOL_CALLS
  # parallel_tool_calls: true

  # Enable per-request cost tracking (logged via tracing).
  # Default: false
  # Env: UAR_LLM__COST_TRACKING
  cost_tracking: false

  # Enable OpenTelemetry tracing spans for LLM requests.
  # Default: true
  # Env: UAR_LLM__TRACING
  tracing: true

  # Cooldown duration after transient errors (seconds).
  # Default: none
  # Env: UAR_LLM__COOLDOWN_SECS
  # cooldown_secs: 5

  # Health-check polling interval (seconds).
  # Default: none (no polling)
  # Env: UAR_LLM__HEALTH_CHECK_SECS
  # health_check_secs: 30

  # --- Budget enforcement (Tower middleware) ---
  # budget:
  #   global_limit: 10.00       # Max total USD spend before requests are blocked
  #   per_request_limit: 0.05   # Max USD per single request

  # --- Rate limiting (Tower middleware) ---
  # rate_limit:
  #   requests_per_minute: 60
  #   tokens_per_minute: 100000

  # --- Response caching (Tower middleware) ---
  # cache:
  #   enabled: true
  #   ttl_secs: 300

# =============================================================================
# PROVIDERS (Multi-Provider LLM Support)
# =============================================================================
# Configure multiple LLM providers for per-agent provider selection.
# The global LLM_BASE_URL / LLM_MODEL env vars are always seeded as the
# "default" provider. Additional providers defined here are available for
# agent-level ProviderPolicy selection and fallback chains.
#
# Agents reference providers by ID in their artifact YAML:
#   policy:
#     provider:
#       default: { provider: "groq", model: "llama-3.3-70b-versatile" }
#       fallbacks:
#         - { provider: "openai", model: "gpt-4o-mini" }

providers: []
  # Example: uncomment and configure as needed
  # - id: "openai"
  #   display_name: "OpenAI"
  #   base_url: "https://api.openai.com/v1"
  #   api_key: "sk-..."            # Or use env: OPENAI_API_KEY
  #   protocol: auto               # auto | chat | responses
  #   default_model: "gpt-4o"
  #   enabled: true
  #   models:
  #     - id: "gpt-4o"
  #       context_window: 128000
  #       supports_vision: true
  #       supports_tools: true
  #       max_output_tokens: 16384
  #     - id: "gpt-4o-mini"
  #       context_window: 128000
  #       supports_vision: true
  #       supports_tools: true
  #       max_output_tokens: 16384
  #
  # - id: "groq"
  #   display_name: "Groq"
  #   base_url: "https://api.groq.com/openai/v1"
  #   api_key: "gsk_..."           # Or use env: GROQ_API_KEY
  #   protocol: chat
  #   default_model: "llama-3.3-70b-versatile"
  #   enabled: true
  #   models:
  #     - id: "llama-3.3-70b-versatile"
  #       context_window: 131072
  #       supports_vision: false
  #       supports_tools: true
  #     - id: "deepseek-r1-distill-llama-70b"
  #       context_window: 131072
  #       supports_vision: false
  #       supports_tools: false
  #
  # - id: "openrouter"
  #   display_name: "OpenRouter"
  #   base_url: "https://openrouter.ai/api"
  #   api_key: "sk-or-..."
  #   protocol: chat
  #   default_model: "anthropic/claude-3.5-sonnet"
  #   enabled: true

# =============================================================================
# MULTIMODAL SUPPORT
# =============================================================================

file_processing:
  # Provider for document text extraction.
  # Options: "kreuzberg" (local), "auto", "unstructured", "mistral", "local"
  # "auto" tries providers in order: kreuzberg > unstructured > mistral > local
  # Default: "kreuzberg"
  # Env: UAR_FILE_PROCESSING__PROVIDER
  provider: "kreuzberg"

  # Directory for storing uploaded files before processing.
  # Default: System temp directory + "uar-uploads"
  # Env: UAR_FILE_PROCESSING__UPLOAD_DIR
  upload_dir: "/tmp/uar-uploads"

  # Maximum number of files allowed per prompt.
  # Default: 10
  # Env: UAR_FILE_PROCESSING__MAX_FILES_PER_PROMPT
  max_files_per_prompt: 10

  # Maximum file size per file in bytes.
  # Default: 52428800 (50MB)
  # Env: UAR_FILE_PROCESSING__MAX_FILE_SIZE
  max_file_size: 52428800

  # Maximum total upload size for all files in a single prompt.
  # Default: 104857600 (100MB)
  # Env: UAR_FILE_PROCESSING__MAX_TOTAL_SIZE
  max_total_size: 104857600

  # Allowed MIME types (empty = allow all supported types).
  # Example: ["application/pdf", "image/png", "image/jpeg"]
  # Env: UAR_FILE_PROCESSING__ALLOWED_MIME_TYPES (comma-separated)
  allowed_mime_types: []

# Unstructured.io configuration (hosted or self-hosted)
# Used when file_processing.provider = "unstructured" or "auto"
unstructured:
  # API URL for Unstructured.io service.
  # Hosted: "https://api.unstructured.io/general/v0/general"
  # Self-hosted: Your deployment URL
  # Default: Hosted URL
  # Env: UAR_UNSTRUCTURED__API_URL
  api_url: "https://api.unstructured.io/general/v0/general"

  # API key for the hosted Unstructured.io service.
  # Not required for self-hosted deployments.
  # Env: UAR_UNSTRUCTURED__API_KEY
  # api_key: "your-api-key-here"

# Mistral OCR configuration
# Used when file_processing.provider = "mistral" or "auto"
mistral_ocr:
  # Mistral API key for OCR service.
  # Can reuse LLM_API_KEY if using Mistral for LLM.
  # Env: UAR_MISTRAL_OCR__API_KEY or MISTRAL_API_KEY
  # api_key: "your-mistral-api-key"

  # OCR model to use.
  # Default: "mistral-ocr-latest"
  # Env: UAR_MISTRAL_OCR__OCR_MODEL
  ocr_model: "mistral-ocr-latest"

# Kreuzberg local file processing configuration
# High-performance Rust-powered document intelligence (56+ file formats)
# Used when file_processing.provider = "kreuzberg" or "auto"
kreuzberg:
  # Enable OCR for scanned documents and images.
  # Requires Tesseract installed on the system (brew install tesseract / apt-get install tesseract-ocr)
  # Default: false
  # Env: UAR_KREUZBERG__OCR_ENABLED
  ocr_enabled: false

  # Force OCR even if text layer exists in PDFs.
  # Useful for PDFs with garbled or low-quality text layers.
  # Default: false
  # Env: UAR_KREUZBERG__FORCE_OCR
  force_ocr: false

  # OCR backend to use.
  # Options: "tesseract" (default, most portable), "easyocr", "paddleocr"
  # Note: easyocr and paddleocr require Python dependencies
  # Default: "tesseract"
  # Env: UAR_KREUZBERG__OCR_BACKEND
  ocr_backend: "tesseract"

  # OCR language(s) for text recognition.
  # Tesseract format: "eng" (English), "deu" (German), "eng+deu" (multiple)
  # Language packs must be installed (brew install tesseract-lang / apt-get install tesseract-ocr-deu)
  # Default: "eng"
  # Env: UAR_KREUZBERG__OCR_LANGUAGE
  ocr_language: "eng"

  # PDF rendering DPI for OCR (higher = better quality but more memory).
  # Recommended: 150 for speed, 300 for quality, 450 for fine print
  # Default: 300
  # Env: UAR_KREUZBERG__PDF_DPI
  pdf_dpi: 300

  # Extract tables from documents (PDFs, Word, Excel, etc.).
  # Tables are appended to content in Markdown format.
  # Default: true
  # Env: UAR_KREUZBERG__EXTRACT_TABLES
  extract_tables: true

  # Extract metadata from documents (title, author, pages, keywords).
  # Metadata is included in the processing result.
  # Default: true
  # Env: UAR_KREUZBERG__EXTRACT_METADATA
  extract_metadata: true

  # Output format for extracted content.
  # Options: "markdown" (default, preserves structure), "text" (plain text)
  # Default: "markdown"
  # Env: UAR_KREUZBERG__OUTPUT_FORMAT
  output_format: "markdown"

# Vision/Image processing configuration
vision:
  # Explicit vision model for image understanding.
  # If not set, uses the default LLM model (if vision-capable).
  # Examples: "gpt-4o", "claude-3-5-sonnet", "pixtral-12b"
  # Env: UAR_VISION__MODEL
  # model: "gpt-4o"

  # Auto-detect vision capability from default LLM model.
  # If true and model supports vision, images are passed directly to LLM.
  # Default: true
  # Env: UAR_VISION__AUTO_DETECT
  auto_detect: true

# =============================================================================
# MODEL FILES (Tokenizers, Embeddings)
# =============================================================================

models:
  # Directory containing model files (tokenizer.json, embeddings, etc.).
  # This path is used by the VectorMatcher for loading tokenizer files.
  # Default: "src/uar/runtime/matching/models" (development)
  # Production: "/app/models" (Docker)
  # Env: UAR_MODELS_DIR
  models_dir: "src/uar/runtime/matching/models"

  # Threshold for vector similarity matching in skill classification.
  # Higher values require closer matches (more conservative).
  # Lower values allow looser matches (more inclusive).
  # Range: 0.0 to 1.0
  # Default: 0.75
  # Env: UAR_MODELS__VECTOR_THRESHOLD
  vector_threshold: 0.75

# =============================================================================
# KNOWLEDGE BASES (RAG Document Scoping)
# =============================================================================

knowledge_bases:
  # Default knowledge base - documents go here if no KB specified
  default:
    name: "default"
    description: "Default knowledge base for general documents"
    embedding_provider: "fastembed"
    embedding_model: "BAAI/bge-small-en-v1.5"
    # vector_dimensions: 384  # Automatically determined from model
    file_processor: "kreuzberg"
    chunking:
      strategy: "recursive"
      chunk_size: 512

  # Additional named knowledge bases (optional)
  # named:
  #   technical-docs:
  #     name: "technical-docs"
  #     description: "Technical documentation and API references"
  #     embedding_provider: "openai"
  #     embedding_model: "text-embedding-3-small"
  #     vector_dimensions: 1536
  #     file_processor: "unstructured"
  #     chunking:
  #       strategy: "semantic"
  #       semantic_threshold: 0.75
  #
#   code-examples:
#     name: "code-examples"
#     description: "Code samples and snippets"
#     embedding_provider: "fastembed"
#     embedding_model: "BAAI/bge-small-en-v1.5"
#     file_processor: "kreuzberg"
#     chunking:
#       strategy: "token"
#       chunk_size: 256

# =============================================================================
# INTENT CLASSIFICATION (Skill Matching)
# =============================================================================

intent_classifier:
  # Classification backend to use for skill matching.
  # Options:
  #   - "rules": Keyword/tag matching only (fast, high precision, low recall)
  #   - "tfidf": TF-IDF based matching (pure Rust, no external deps, good balance)
  #   - "hybrid": Rules first, then TF-IDF for remaining (recommended)
  #   - "localembedding": On-device embedding similarity via VectorMatcher
  #   - "llm": LLM-powered skill ranking with TF-IDF fallback
  #   - "wasm": WASM component for custom classifiers (future)
  # Default: "tfidf"
  # Env: UAR_INTENT_CLASSIFIER__BACKEND
  backend: "tfidf"

  # Number of top skill candidates to consider.
  # Higher values increase recall but may reduce precision.
  # Default: 5
  # Env: UAR_INTENT_CLASSIFIER__TOPK
  topk: 5

  # Minimum score threshold to accept a classification.
  # Skills below this threshold may trigger fallback/clarification.
  # Range: 0.0 to 1.0
  # Default: 0.5
  # Env: UAR_INTENT_CLASSIFIER__ACCEPT_THRESHOLD
  accept_threshold: 0.5

  # Minimum margin between top-1 and top-2 scores for confident selection.
  # If margin is too small, the query may be ambiguous.
  # Range: 0.0 to 1.0
  # Default: 0.05
  # Env: UAR_INTENT_CLASSIFIER__MARGIN_THRESHOLD
  margin_threshold: 0.05

  # Path to WASM classifier component (only used when backend = "wasm").
  # The component must implement the intent:classifier WIT interface.
  # Env: UAR_INTENT_CLASSIFIER__WASM_COMPONENT_PATH
  # wasm_component_path: "/path/to/classifier.wasm"

# =============================================================================
# LLM FAILOVER (Runtime Model Failover)
# =============================================================================
# When enabled, UAR automatically switches to a fallback driver if the primary
# LLM driver returns an error. The fallback is configured via the provider
# registry (see `providers:` above).

failover:
  # Enable automatic failover to the fallback model.
  # Default: false
  # Env: UAR_FAILOVER__ENABLED
  enabled: false

  # Strategy to use when selecting among fallback models.
  # Options:
  #   - "priority"        – try fallbacks in order, stop on first success
  #   - "round_robin"     – distribute requests across fallbacks evenly
  #   - "cost_optimized"  – prefer cheapest model that hasn't errored recently
  # Default: "priority"
  # Env: UAR_FAILOVER__STRATEGY
  strategy: "priority"

  # Number of consecutive errors before switching to the fallback driver.
  # Default: 3
  # Env: UAR_FAILOVER__ERROR_THRESHOLD
  error_threshold: 3

  # Seconds to wait before retrying the primary driver after a failover.
  # Set to 0 to keep using the fallback indefinitely.
  # Default: 60
  # Env: UAR_FAILOVER__COOLDOWN_SECS
  cooldown_secs: 60

  # Ordered list of fallback models (used by "priority" strategy).
  # Each entry is a "provider/model" string matching a configured provider.
  # Env: UAR_FAILOVER__FALLBACK_MODELS (comma-separated)
  # fallback_models:
  #   - "anthropic/claude-3-5-haiku-20241022"
  #   - "groq/llama-3.3-70b-versatile"

# =============================================================================
# PROJECT INSTRUCTIONS (Host-owned workspace trust)
# =============================================================================
project_instructions:
  # Optional additional names: CLAUDE.md, GEMINI.md. A matching *.override.md
  # takes precedence over its base file in the same directory.
  file_names: [AGENTS.md]
  root_markers: [.git]
  # Explicit absolute workspace roots only. Empty means no instruction files
  # are read. Requests cannot mark their own workspace trusted.
  trusted_workspaces: []

world_state:
  # A positive time bucket size. Calls within the same bucket do not resend time.
  time_granularity_secs: 60

# =============================================================================
# NATIVE TOOLS (In-Process Tool Execution)
# =============================================================================
# Built-in tools that run directly inside the UAR process without an MCP server.
# Native tools are registered automatically on startup; enable/disable categories
# below to control which tools are exposed to agents.

native_tools:
  # Enable the native tool subsystem.
  # Default: true
  # Env: UAR_NATIVE_TOOLS__ENABLED
  enabled: true

  # ── File system tools ────────────────────────────────────────────────────
  file_read_enabled: true
  file_write_enabled: false   # Opt-in — writes to the host filesystem

  # Maximum file size the read tool will return (KB).
  # Default: 1024 (1 MB)
  # Env: UAR_NATIVE_TOOLS__FILE_MAX_SIZE_KB
  file_max_size_kb: 1024

  # Maximum bytes the file write tool will accept per call (KB).
  # Default: 512
  # Env: UAR_NATIVE_TOOLS__FILE_WRITE_MAX_KB
  file_write_max_kb: 512

  # ── Web fetch tool ───────────────────────────────────────────────────────
  web_fetch_enabled: true

  # Request timeout for the web fetch tool (seconds).
  # Default: 30
  # Env: UAR_NATIVE_TOOLS__WEB_FETCH_TIMEOUT_SECS
  web_fetch_timeout_secs: 30

  # Maximum response body size the web fetch tool will return (KB).
  # Default: 2048 (2 MB)
  # Env: UAR_NATIVE_TOOLS__WEB_FETCH_MAX_SIZE_KB
  web_fetch_max_size_kb: 2048

  # ── Terminal / shell tool ────────────────────────────────────────────────
  # WARNING: Exposing a shell to an LLM is a significant security risk.
  # Only enable in trusted, sandboxed environments.
  terminal_enabled: false

  # Shell binary to use for the terminal tool.
  # Default: "/bin/sh"
  # Env: UAR_NATIVE_TOOLS__TERMINAL_SHELL
  terminal_shell: "/bin/sh"

  # Maximum wall-clock time the terminal tool will wait for a command (seconds).
  # Default: 30
  # Env: UAR_NATIVE_TOOLS__TERMINAL_TIMEOUT_SECS
  terminal_timeout_secs: 30

  # Run terminal commands inside a container/nsjail sandbox.
  # Requires Docker or nsjail to be available on the host.
  # Default: false
  # Env: UAR_NATIVE_TOOLS__TERMINAL_USE_SANDBOX
  terminal_use_sandbox: false

  # ── Session search tool ──────────────────────────────────────────────────
  session_search_enabled: true

  # Maximum results returned by the session search tool per call.
  # Default: 10
  # Env: UAR_NATIVE_TOOLS__SESSION_SEARCH_MAX_RESULTS
  session_search_max_results: 10

# =============================================================================
# SKILL EVOLUTION (Hermes Learning Cycle)
# =============================================================================
# After a qualifying run completes, UAR fires a reflection prompt against the
# run transcript. The LLM analyses the tool calls and suggests new skills (or
# updates/deletions) to codify the learned behaviour for future runs.
#
# This is opt-in and disabled by default — enable it only when you have a
# SkillService configured with at least one writable storage provider.

skill_evolution:
  # Enable the post-run reflection hook.
  # Default: false
  # Env: UAR_SKILL_EVOLUTION__ENABLED | --skill-evolution-enabled
  enabled: false

  # Model to use for the reflection call ("provider/model" format).
  # Falls back to the run's primary model when unset.
  # Env: UAR_SKILL_EVOLUTION__TRIGGER_MODEL | --skill-evolution-model
  # trigger_model: "openai/gpt-4o-mini"

  # Minimum number of tool completions in a run before evolution is considered.
  # Prevents trivial single-turn interactions from spawning skills.
  # Default: 2
  # Env: UAR_SKILL_EVOLUTION__MIN_TOOL_CALLS
  min_tool_calls: 2

  # Maximum new skills that can be created from a single run.
  # Default: 1
  # Env: UAR_SKILL_EVOLUTION__MAX_SKILLS_PER_RUN
  max_skills_per_run: 1

  # Allow the reflection system to update (patch) existing skills.
  # Default: false
  # Env: UAR_SKILL_EVOLUTION__ALLOW_UPDATE
  allow_update: false

  # Allow the reflection system to delete under-used skills.
  # Default: false
  # Env: UAR_SKILL_EVOLUTION__ALLOW_DELETION
  allow_deletion: false

  # Minimum times a skill must have executed before deletion is permitted.
  # Default: 5
  # Env: UAR_SKILL_EVOLUTION__MIN_EXECUTIONS_BEFORE_DELETE
  min_executions_before_delete: 5

  # Custom system-prompt for the reflection call.
  # Uses the built-in Hermes prompt when unset.
  # Env: UAR_SKILL_EVOLUTION__REFLECTION_PROMPT
  # reflection_prompt: |
  #   You are Hermes, the UAR skill librarian. Analyse the run transcript …

# =============================================================================
# ACP SERVER (Agent Communication Protocol)
# =============================================================================
# Exposes a BeeAI / IBM Research ACP endpoint alongside the main HTTP API.
# ACP provides JSON-RPC 2.0 agent introspection, session management, and
# streaming run execution — used by IDEs, debuggers, and the BeeAI platform.
#
# Spec: https://agentcommunicationprotocol.dev

acp:
  # Enable the ACP endpoint.
  # Default: false
  # Env: UAR_ACP__ENABLED
  enabled: false

  # URL path prefix for ACP routes.
  # Default: "/acp"
  # Env: UAR_ACP__PATH
  path: "/acp"

  # Require JWT authentication for ACP requests.
  # Inherits from security.jwt_required when not set.
  # Default: true
  # Env: UAR_ACP__AUTH_REQUIRED
  auth_required: true

  # Session TTL in seconds — ACP sessions are garbage-collected after this.
  # Default: 3600 (1 hour)
  # Env: UAR_ACP__SESSION_TTL_SECS
  session_ttl_secs: 3600