config.yaml.example

# Main Manifold runtime configuration.
#
# This file is the primary source of truth for runtime behavior.
# Environment variables are still supported, but only because YAML values can
# reference them via ${VAR}. Use example.env for the full set of variables that
# this example references.
#
# Separate files:
# - specialists.yaml / specialists.yml: specialist agents and routing rules
# - mcp.yaml / mcp.yml: MCP server definitions

workdir: .
systemPrompt: ""

# Agent runtime limits.
maxSteps: 1000
maxToolParallelism: 0 # 0=unbounded, 1=sequential, >1=capped
outputTruncateBytes: 131072
agentRunTimeoutSeconds: 0
streamRunTimeoutSeconds: 0
workflowTimeoutSeconds: 0

# Logging.
logPath: manifold.log
logLevel: info
logPayloads: false
logRawPrompts: false

# Tool execution policy.
enableTools: true
# When enabled, allowTools is the bootstrap set and agents can discover
# additional registered tools at runtime via tool_search.
autoDiscover: false
# Caps how many non-bootstrap tools a single run may add when autoDiscover is on.
maxDiscoveredTools: 12
allowTools:
  - run_cli
  - apply_patch
  - llm_transform
  - describe_image
  - specialists_infer
  - web_fetch
  - mcp_duckduckgo_search

exec:
  maxCommandSeconds: 1200
  blockBinaries:
    - rm
    - sudo
    - chown
    - chmod
    - dd
    - mkfs
    - mount
    - umount

# Chat summarization. This LLM is independent from the primary runtime LLM so
# plain text summaries remain portable when switching providers.
summary:
  enabled: true
  # Maximum time a single summary-model request may block a chat request.
  # Set to 0 to use the built-in default.
  callTimeoutSeconds: 120
  contextWindowTokens: 32000
  # Optional global plain-text summary trigger budget. Set to 0 to fall back
  # to the configured summary model's context size.
  plainTextContextWindowTokens: 0
  reserveBufferTokens: 25000
  minKeepLastMessages: 4
  maxKeepLastMessages: 12
  maxSummaryChunkTokens: 4096
  llm_client:
    provider: openai
    openai:
      apiKey: "${OPENAI_API_KEY}"
      baseURL: https://api.openai.com/v1
      model: gpt-5-nano
      api: completions
      extraHeaders: {}
      extraParams: {}

# Primary LLM provider configuration.
llm_client:
  # openai | anthropic | google | local
  provider: openai
  openai:
    apiKey: "${OPENAI_API_KEY}"
    baseURL: https://api.openai.com/v1
    model: gpt-5-mini
    # completions | responses
    api: completions
    extraHeaders: {}
    extraParams:
      parallel_tool_calls: true
      prompt_cache_key: manifold-
      reasoning_effort: medium
  anthropic:
    apiKey: "${ANTHROPIC_API_KEY}"
    model: claude-sonnet-4-6
    baseURL: https://api.anthropic.com
    maxTokens: 4096
    extraParams: {}
    promptCache:
      enabled: false
      cacheSystem: true
      cacheTools: true
      cacheMessages: false
  google:
    apiKey: "${GOOGLE_LLM_API_KEY}"
    model: gemini-2.5-pro
    baseURL: https://generativelanguage.googleapis.com/
    timeoutSeconds: 30
    extraParams: {}

# Observability.
obs:
  serviceName: manifold
  serviceVersion: 0.0.1
  environment: local
  # Optional OTLP export endpoint. Leave empty when running without an
  # OpenTelemetry Collector; the dashboard will still use local process
  # telemetry when obs.local.enabled is true.
  # Use bare host:port here. The current OTLP HTTP exporter code appends
  # the default /v1/{traces,metrics,logs} paths for collector endpoints.
  otlp: ""
  # Local process telemetry is enabled by default and backs the dashboard when
  # ClickHouse is not configured or unavailable. It is bounded and resets when
  # agentd restarts.
  local:
    enabled: true
    metricsWindowMinutes: 60
    metricsBucketSeconds: 30
    maxLogs: 5000
    maxTraces: 1000
    maxSpansPerTrace: 256
  clickhouse:
    # Optional persistent telemetry backend. Keep CLICKHOUSE_DSN empty for a
    # self-contained install that does not run ClickHouse.
    dsn: "${CLICKHOUSE_DSN}"
    database: otel
    metricsTable: metrics_sum
    tracesTable: traces
    logsTable: logs
    timestampColumn: TimeUnix
    valueColumn: Value
    modelAttributeKey: llm.model
    promptMetricName: llm.prompt_tokens
    completionMetricName: llm.completion_tokens
    lookbackHours: 24
    timeoutSeconds: 5

# Optional web search backend used by the internal web_search tool.
web:
  searXNGURL: http://localhost:8080

# Optional authentication.
auth:
  enabled: false
  # oidc | oauth2
  provider: oauth2
  issuerURL: ""
  clientID: "${AUTH_CLIENT_ID}"
  clientSecret: "${AUTH_CLIENT_SECRET}"
  redirectURL: http://localhost:32180/auth/callback
  allowedDomains: []
  cookieName: manifold_session
  cookieSecure: false
  cookieDomain: ""
  stateTTLSeconds: 600
  sessionTTLHours: 72
  oauth2:
    authURL: https://github.com/login/oauth/authorize
    tokenURL: https://github.com/login/oauth/access_token
    userInfoURL: https://api.github.com/user
    logoutURL: ""
    logoutRedirectParam: redirect_uri
    scopes:
      - read:user
      - user:email
    providerName: github
    defaultRoles:
      - user
    emailField: email
    nameField: name
    pictureField: avatar_url
    subjectField: id
    rolesField: ""
    disablePKCE: false

# Database backends.
databases:
  # Docker Compose users can keep embedded=false and use pg-manifold through
  # DATABASE_URL. Host-run users can set embedded=true and leave DATABASE_URL
  # empty; agentd will start a bundled local PostgreSQL and wire its DSN into
  # all database-backed stores at startup.
  embedded: false
  embeddedPort: 5433
  embeddedDataDir: "" # defaults to ~/.manifold/embedded-postgres/data
  embeddedVersion: "18"
  embeddedExtensions: # extensions installed into embedded PG (default: pgvector, postgis, pgrouting)
    - pgvector
    - postgis
    - pgrouting
  embeddedExtensionURL: "" # override for pre-built extension package base URL
  defaultDSN: "${DATABASE_URL}"
  chat:
    backend: postgres # memory | auto | postgres
    dsn: "${DATABASE_URL}"
  search:
    backend: postgres # memory | auto | postgres
    dsn: "${DATABASE_URL}"
    index: documents
  vector:
    backend: postgres # memory | auto | postgres | qdrant
    dsn: "${DATABASE_URL}"
    index: embeddings
    dimensions: 1536
    metric: cosine
  graph:
    backend: postgres # memory | auto | postgres
    dsn: "${DATABASE_URL}"

# Embedding service, used by vector features and evolving memory.
embedding:
  baseURL: https://api.openai.com
  model: text-embedding-3-small
  apiKey: "${EMBED_API_KEY}"
  apiHeader: Authorization
  headers: {}
  path: /v1/embeddings
  timeoutSeconds: 30

# Optional defaults for the describe_image tool. Leave blank to use the
# invoking agent/specialist endpoint and model; per-tool-call base_url/model
# arguments still override these values.
imageTool:
  baseURL: ""
  model: ""

# Search -> Synthesis -> Evolve memory.
evolvingMemory:
  enabled: false
  persistDebounceMs: 250
  sessionTTLMinutes: 120
  janitorIntervalMinutes: 15
  storeJanitorIntervalMinutes: 60
  llmClient:
    provider: local
    openai:
      apiKey: ""
      baseURL: http://localhost:11434/v1
      model: qwen-memory
      api: completions
      extraHeaders: {}
      extraParams: {}
  maxSize: 2000
  topK: 4
  windowSize: 20
  enableRAG: true
  reMemEnabled: false
  maxInnerSteps: 3
  model: gpt-4o-mini
  enableSmartPrune: true
  pruneThreshold: 0.97
  relevanceDecay: 0.99
  minRelevance: 0.05
  pruneQualityFloor: 3
  promotionAccessThreshold: 5

# Shared durable memory.
transit:
  enabled: false
  defaultSearchLimit: 10
  defaultListLimit: 100
  maxBatchSize: 100
  enableVectorSearch: true

# Shared belief memory foundations. Phase 0 keeps all runtime behavior disabled
# while the store, schema, and interfaces are available for incremental rollout.
beliefMemory:
  enabled: false
  enableDistillation: false
  enableRetrieval: false
  enableConstraintEnforcement: false
  maxBeliefsPerPrompt: 5
  maxEvidencePerBelief: 3
  defaultConfidence: 0.5
  promotionThreshold: 0.8
  # Optional: blend RAG document retrieval into the belief prompt block as a
  # delimited "Retrieved Evidence (untrusted)" sub-section. Disabled by default;
  # requires the RAG service to be configured (see embedding/RAG settings).
  enableRAGEvidence: false
  maxRAGEvidencePerPrompt: 3
  ragRetrievalK: 8
  ragMinScore: 0.0

# Speech helpers.
tts:
  baseURL: https://api.openai.com/v1
  model: gpt-4o-mini-tts
  voice: alloy

stt:
  baseURL: https://api.openai.com
  model: gpt-4o-mini-transcribe

# Placeholder for future per-project controls.
projects: {}

# Accurate token counting.
tokenization:
  enabled: false
  cacheSize: 1000
  cacheTTLSeconds: 3600
  fallbackToHeuristic: true

# Specialist and MCP configuration live in dedicated files now:
# - specialists.yaml.example
# - mcp.yaml.example