Plan Agent Machine runtime adapter for Superconscious ReasoningRun workloads

[mdheller]

Purpose

Superconscious now defines the reference governed cognition loop, and SourceOS-Linux/sourceos-spec has canonical reasoning contracts for ReasoningRun, ReasoningEvent, ReasoningReceipt, ReasoningReplayPlan, and ReasoningBenchmark.

Agent Machine should define how a future Superconscious runtime adapter requests machine/runtime/cache placement facts without moving cognition logic into Agent Machine.

Scope

• Add docs describing the Agent Machine adapter contract for reasoning workloads.
• Define what an Agent Machine runtime decision should return to a Superconscious run.
• Map runtime facts to existing Agent Machine concepts: AgentPod, InferenceProvider, cache posture, model residency, activation decision, and runtime receipt.
• Add a fixture or example runtime-plan decision for a no-activation deterministic Superconscious M1 run.
• Preserve M1 inert posture: no provider activation, no host mutation, no network requirement.

Acceptance criteria

• Agent Machine docs explain how Superconscious should request runtime planning.
• Example runtime decision is safe, dry-run, and no-activation.
• The integration clearly preserves Agent Machine authority over runtime placement/activation and Superconscious authority over the reference cognition loop.

Related repos

• SocioProphet/superconscious
• SourceOS-Linux/sourceos-spec
• SocioProphet/agentplane
• SourceOS-Linux/sourceos-devtools