Runtime engine manifests (kind: RuntimeEngine)¶
Shared kernel and LLM engine tuning for a deployment — queue depth, dispatch
step budget, streaming, and LLM response cache policy (read/write flags).
LLM endpoints are not configured here — use workspace infra_refs with
standard:openai (or another provider bundle). Offline CI uses
llm_cache replay recorded against a live provider.
Agent and MAS manifests must not reference runtime (no runtime_refs on
agents or MAS).
Schema: docs/schemas/runtime/fragments/runtime-engine.schema.yaml
Bundled ref: standard:runtime-default (library-standard) — queue
depth only; no dispatch cap. Cap the loop with overlay
pkg://mas.library.standard/overlays/with-hardened.yaml (max_steps: 512)
or set spec.engine.max_auto_steps on a RuntimeEngine you pass via
--runtime-ref.
Example¶
apiVersion: infra/v1
kind: RuntimeEngine
metadata:
name: ci-runtime
spec:
engine:
queue_depth: 64
max_auto_steps: 256
stream: false
cache:
read: true
write: true
How refs are resolved¶
Same discovery model as LLM infra, but only from deployment config — never from agent, MAS, overlay, or flavour manifests:
runtime_refs is the parallel of infra_refs — same kind of wiring, different
manifest kind (RuntimeEngine vs LLMProxy / bundles):
| Source | Infra (infra_refs) |
Runtime (runtime_refs) |
|---|---|---|
| CLI | --infra-ref |
--runtime-ref |
| Workspace | config.yaml → infra_refs |
config.yaml → runtime_refs |
| User default | default_infra in ~/.config/mas/config.yaml |
default_runtime in same file |
| Search dirs | infra/, ~/.config/mas/infra/ |
runtime/, ~/.config/mas/runtime/ |
Example workspace file (see library-samples/sample-workspace/config.yaml):
infra_refs:
- standard:openai
runtime_refs:
- standard:runtime-default # optional; omit for package defaults only
Omit runtime_refs entirely to use package kernel/cache/stream defaults only.
RuntimeEngine manifests listed in infra_refs are rejected — use
runtime_refs or --runtime-ref instead.
Precedence¶
When building KernelConfig and LLM cache/stream behaviour:
- Package defaults (
DEFAULT_ENGINE_QUEUE_DEPTH, etc.) - Merged
RuntimeEnginemanifest(s) from the resolution ladder above - CLI flags on
mas-ctl chat(--cache-read,--stream, etc.)
See also¶
- Infrastructure manifests — LLM proxy and middleware (
infra_refs) - user-config.md — workspace and XDG layout
- LLM cache — cache middleware vs cache policy