@spec flow_context_provider() :: module() | nil

Configured host context provider for portable flows.

@spec flow_guard_runner() :: module() | nil

Configured host guard runner (a Baton.Flow.GuardRunner) for portable flow LLM nodes.

@spec flow_prompt_resolver() :: module() | nil

Configured host prompt resolver for portable flow LLM nodes.

@spec flow_registry() :: module() | nil

Configured host registry for portable flow action nodes.

@spec inline_threshold_bytes() :: pos_integer()

Encoded result size (bytes) at or below which a result is stored inline on the node row. Larger results spill to result_store/0. Default 32 KB.

@spec llm_client() :: module() | nil

The configured LLM client module used by Baton.Debug.call_llm/3.

Must export complete/2 taking (messages, opts) and returning {:ok, response} or {:error, reason}. Returns nil if unconfigured, in which case call_llm/3 raises a helpful error.

config :baton, llm_client: MyApp.LLM

@spec llm_timeout() :: pos_integer()

Maximum runtime for an LLM step in milliseconds (default: ten minutes).

This is the safe default used by Baton.LLMWorker when an individual worker does not declare its own timeout. Hosts should set their Oban orphan-job recovery window above this value.

@spec max_cache_bytes() :: pos_integer()

Byte budget for the read cache. When an insert would exceed it the cache is flushed wholesale. Default 64 MB.

@spec max_result_bytes() :: pos_integer()

Hard ceiling (bytes) on an encoded result. A step returning a result larger than this fails with {:error, :result_too_large} rather than persisting a blob that would pressure the shared store. Default 16 MB.

@spec oban_name() :: atom()

Configured Oban instance name (default: Oban).

@spec pricing() :: module()

The configured pricing module (a Baton.Pricing implementation).

@spec pubsub() :: atom() | nil

The Phoenix.PubSub server used for live step events.

Returns nil if not configured; Baton.Events no-ops when nil, so the engine runs fine without Phoenix.

@spec rate_limiter() :: module()

The configured Baton.RateLimiter used as a pre-call token-throughput gate by Baton.LLMStep. Defaults to Baton.RateLimiter.Noop (reserves nothing), so the engine runs unchanged until a host opts in.

config :baton, rate_limiter: MyApp.LLM.RateLimiter

@spec repo(atom() | nil) :: module()

Resolve the Ecto repo from Oban's config.

@spec result_cache_enabled?() :: boolean()

Whether the node-local read cache (Baton.ResultCache) is enabled. Off by default — a pure performance opt-in for repeated reads of large results.

@spec result_store() :: module()

The backend module (a Baton.ResultStore implementation) used to store step results too large to keep inline. Defaults to the built-in Postgres backend.

@spec snooze_seconds() :: pos_integer()

Snooze interval (seconds, default 15) used both to park a dependent step at insert and to recheck when a dependency is still pending.

With completion-triggered dispatch (Baton.Reschedule + Baton.RescheduleReporter) the happy path no longer waits this out, so this is essentially a fallback: it only governs latency when a dispatch nudge is lost. A smaller value shortens that worst-case tail at the cost of slightly more frequent rechecks in that rare case.

@spec stale_executing_threshold_seconds() :: pos_integer()

Age (seconds) past which an executing dep is treated as failed (default 600).