Concepts

Hard constraints

• One deployment is exactly one organisation. There is no multi-organisation creation, switching or archiving.
• A workspace carries membership authorisation and shared capabilities inside the organisation. It is not bound to a repository.
• Platform operations rights do not imply workspace data rights. Every read and write is authorised by the control plane against the signed-in principal and membership.

A project organises repositories, tasks, knowledge and decisions around one sustained engineering goal inside a workspace — a long-lived scope, not a single execution.

Hard constraints

• A task binds one specification and one sandbox, then enters a running state machine asynchronously.
• The specification is resolved before creation and is immutable while the task runs. It freezes the creator, scope, repository, initial flow snapshot, model and capability versions — and holds no secrets.
• Retry, rework, restart and resume all reuse the same specification. Floating versions are never re-resolved mid-run.

Hard constraints

• A work node is a reusable, independently maintained and versioned agent execution definition. A flow composes several invocations into a DAG by dependency.
• Drafts are editable; a published version and its content digest are immutable. `latest` is a movable alias and is not guaranteed to be the highest version number.
• An asset has exactly one editable authority at any moment. A published release can only be deprecated or yanked — never overwritten or physically deleted.

By default an ordinary user only describes the task. Preflight reuses a marketplace flow when one fits and generates an auditable task-scoped flow when none does. Users can still author flows, or promote a verified generated snapshot into a marketplace draft.

Hard constraints

• One task maps to one sandbox, and a sandbox has exactly one supervisor as the authority for DAG state and scheduling.
• The control plane does not persist node-level state. It keeps task-level lifecycle and final delivery references only.
• Losing the sandbox fails the task. This restriction is a deliberate cost, accepted in exchange for a single failure domain.

Hard constraints

• Every work node is its own agent session, and every attempt uses its own Git worktree rather than editing a shared workspace.
• Execution-level retry handles gateway timeouts, process crashes and command timeouts. Semantic rework handles upstream artefacts a downstream node has proven defective. The two are counted separately.
• Rework always creates a new session, a new attempt and a new artefact version. A failure record is appended to history, never rewritten.

A successful attempt freezes its own Git revision under a dedicated ref, so code remains listable and readable after the worktree is released.

Hard constraints

• Completion is deterministic: every declared artefact exists, the manifest is present and enumerable, the result summary has valid content, and every unit gate passes.
• Artefact identity combines the work node, output name, attempt and content digest. A downstream node cannot modify an upstream version.
• Deterministic build, test, lint and scan steps run as shell commands declared on the work node, judged by exit code.

Hard constraints

• Knowledge is formal cognition that is synthesised, reusable and sourced within an explicit scope. It splits into repository knowledge and project knowledge.
• Memory is the ongoing mechanism that captures, consolidates, recalls and forgets cognition. It is subject-centred, and layered into working, semantic and resident memory.
• They are separate authorities. Memory does not enter the knowledge retrieval plane, tasks consume only the snapshot frozen at creation, and an agent cannot write long-term memory directly.

Content entering long-term knowledge must be traceable, reviewable and explicitly sourced. A complete session is never promoted to knowledge automatically.