Architecture

Architecture and state ownership

Why ACP is separate

A model predicts content; a harness executes a reasoning/tool loop; a scheduler allocates resources; an identity provider issues trust; and ACP stores durable project context. These are independent components. ACP must remain usable when an individual model, agent worker, UI tab or Jupyter Pod disappears.

Quantum Platform login/programme policy
         | signed AgentPrincipal assertion
         v
ACP API / PostgreSQL -- canonical projects, conversations, messages,
         |              memories, revisions, profiles, task/run references
         +---- approved runtime adapter (Hermes first; others pluggable)
         |                   |
         |                model gateway / local GPU inference
         |
         +---- typed integration requests -> Quantum Platform
                                          -> Slurm/provider scheduler
                                          -> Quantum Workflows results

Authoritative systems

State	Owner
Human account, institution, programme, entitlements	Quantum Platform
ACP external principal identity	signed AgentPrincipal UUID
Projects, conversation/message history, memory revisions and personal profile	ACP PostgreSQL
Runtime-native sessions, checkpoints, personas	runtime-local secondary cache
Portable shared policy/skills	reviewed Git revisions
HPC/QPU allocation and state	Slurm or hardware provider
Scientific data/provenance	Quantum Workflows / science library
Resource files and secrets	protected storage and secret custody

A POSIX UID, SSH username or a caller-supplied tenant field is never sufficient proof of ACP conversation ownership. A model is never an IAM authority. RLS, signed authentication and scoped authorization guard data on the server.

Milestone boundaries

M4a: canonical cross-client personal conversation continuity. M4b: private memory and source/revision lineage. M4c: richer projects and authorised programme rooms. M4d: governed scientific skill references. M4e: portable profile and harness continuity. M4f: browser Guide integration and passive observed status. M4g: terminal/SSH client integration with a trusted pairing broker. A contract written in an ADR is not proof that the endpoint or broker is deployed.

See state architecture, routing, ADR 0008 and ADR 0012.