Configuration reference

Everything lives in one block, usually in config/initializers/silas.rb (the installer generates a commented starter):

Silas.configure do |c|
  # ...
end

Inference

Option	Default	Meaning
`adapter`	`:ruby_llm`	The inference seam. `:ruby_llm` (any provider RubyLLM supports), or any object responding to `#execute_step` — the eval harness, the chaos suite, and the template’s keyless demo all inject one.
`default_model`	`"claude-sonnet-4-5"`	Used when `agent.yml` doesn’t set `model:`. Must resolve in your installed RubyLLM registry.
`around_model_call`	`nil`	Wrap every model call — e.g. `->(ctx, &call) { RubyLLM::Resilience.chain(:anthropic) { call.() } }`.
`queue_name`	`:default`	Active Job queue for agent turns.

Which provider serves a turn is decided by the model id — OpenRouter, OpenAI-compatible gateways, Bedrock/Vertex/Azure, and local runtimes are all recipes in providers.

The loop

Option	Default	Meaning
`max_steps`	`25`	Hard cap on model calls per turn; `agent.yml` `limits.max_steps` overrides per agent. Hitting it fails the turn (`max_steps`).
`compact_at`	`0.9`	Context compaction trigger. A Float in (0, 1] = fraction of the model’s registry context window; an Integer = absolute token threshold (the form custom adapters need); `nil`/`false` disables. Compaction summarises prior turns into a persisted, exactly-once row — the current turn is never compacted, and replays stay byte-identical.
`isolate_steps`	`true`	Continuation isolation per step — the durability contract. Leave on; specs may disable for inline runs.

Approvals & questions

Option	Default	Meaning
`approval_ttl`	`7.days`	How long a parked approval (or in-doubt invocation) waits before expiring. Parked-forever ghosts are a bug, not a feature.
`ask_question`	`true`	The built-in tool that parks the turn to ask the operator something. Toggling it changes the definitions digest; parked turns resume against their own definitions snapshot, so the flip doesn’t fail them — they just keep their original toolset until they settle.

Memory

Option	Default	Meaning
`memory`	`true`	`false` removes the memory tools entirely (digest note above applies).
`memory_approval`	`:always`	Every `remember` parks a card for a human; `:never` auto-approves.
`memory_injection_limit`	`8`	How many recent memories are injected into each turn; `recall` digs deeper on demand.

Inbox (`/silas/inbox`)

Option	Default	Meaning
`inbox_auth`	deny (404)	The lambda DENIES by rendering (or `head`-ing) and PASSES by not rendering — Devise-compatible: `->(controller) { controller.head :not_found unless controller.current_user&.admin? }`.
`inbox_public_read`	`false`	Read-only for anyone; approve/decline/answer stay gated.
`inbox_actor`	`current_user&.email \\|\\| "inbox"`	Identity recorded on approvals/declines/answers made in the inbox.
`inbox_streaming`	`nil` (auto)	Live Turbo updates when the host has `turbo-rails`; `false` forces the polling fallback.
`model_prices`	`{}`	Override map for cost accounting: `{"model-id" => { in:, out: }}` in units per 1k tokens, 1e6 units = $1 (a $3/MTok rate is `3000`). Everything else prices from RubyLLM’s registry.

JSON API (`/silas/api/v1`)

Option	Default	Meaning
`api_auth`	deny (404)	Same contract as `inbox_auth`.
`api_actor`	`"api"`	Identity recorded on approvals made through the API.
`api_stream_poll_interval`	`0.5`	Seconds between SSE row polls.
`api_stream_max_duration`	`300`	Seconds before an SSE stream closes itself (clients reconnect with `Last-Event-ID`); bounds thread hold.

Channels

Option	Default	Meaning
`slack_signing_secret` / `slack_bot_token`	`credentials.silas.slack.*`	Explicit setters override the credentials path; `nil` disables Slack.
`channel_routes`	`{}`	Which agent an inbound thread wakes: `{ "slack" => { "C0BILLING" => "bookkeeper" }, "email" => { "billing@shop.test" => "bookkeeper" } }`. Unmatched threads wake the root agent. Names are checked at boot against `app/agents/` — a typo fails the deploy, never a webhook. See channels.

Evals

Option	Default	Meaning
`eval_dir`	`"test/agent_evals"`	Where `*_eval.rb` scenarios live (`bin/rails silas:eval`).
`eval_grader`	`nil`	Custom LLM grader for `assert_rubric`; offline, rubric asserts skip rather than fail.

Sandbox

Option	Default	Meaning
`sandbox`	`:none`	Code execution off. `:docker` (interim), or any object responding to `#run` — e.g. `Hermetic.gvisor(image: "python:3.12-slim")`. Configuring one advertises the `run_code` tool automatically.
`sandbox_image` / `sandbox_network` / `sandbox_memory` / `sandbox_cpus` / `sandbox_pids` / `sandbox_workdir` / `sandbox_docker_bin` / `sandbox_timeout`	`nil` / `"none"` / `"512m"` / `"1"` / `256` / `"/workspace"` / `"docker"` / `30`	Docker knobs; inert unless `sandbox = :docker`.

MCP & testing seams

Option	Default	Meaning
`mcp_auth`	deny (404)	Auth for the mounted MCP endpoint (`POST /silas/mcp` — your tools served to MCP clients, approval gates included). Same contract as `api_auth`; bearer-token example in connections.
`mcp_server_host`	`"127.0.0.1"`	Unused. The old in-process TCP server is gone; the endpoint is a mounted route. Retained one release so initializers don’t break — delete it from yours.
`mcp_client_factory`	`nil`	`->(connection) { client }` — inject a fake MCP client per connection in tests.

The remaining accessors (tool_resolver, tool_definitions, definitions_digest, skills, schedules, subagent_*, named_agent_scopes, channel_resolver, instructions_dir, agent_override) are internal wiring seams the Registry fills at boot — test hooks, not app configuration.

Boot guards (fail-loud misconfiguration)

Two checks run at boot and raise in production, warn in development:

No provider key configured while adapter = :ruby_llm — the most common first-run failure, surfaced at boot with the fix instead of dying inside the first turn.
ActiveJob on the in-process :async adapter — it runs a re-enqueued continuation concurrently with the original, double-executing steps and voiding exactly-once. Use Solid Queue (production) or :inline (scripts/demos). bin/rails silas:doctor checks both, everywhere.