State Machine Management

State Machine Management

Intro

A StateMachine defines the valid states of a primitive and the allowed transitions between them. processkit ships default state machines for WorkItem and DecisionRecord; projects override these or define new ones for custom entities.

Overview

Shape

---
apiVersion: processkit.projectious.work/v1
kind: StateMachine
metadata:
  id: SM-workitem
  name: workitem
  created: 2026-04-06T00:00:00Z
spec:
  description: "WorkItem lifecycle with a formal review state."
  initial: backlog
  terminal: [done, cancelled]
  states:
    backlog:
      description: "Not started."
      transitions:
        - to: in-progress
        - to: cancelled
    in-progress:
      description: "Being actively worked on."
      transitions:
        - to: review
        - to: blocked
        - to: cancelled
    blocked:
      description: "Cannot proceed until a blocker is resolved."
      transitions:
        - to: in-progress
        - to: cancelled
    review:
      description: "Awaiting review."
      transitions:
        - to: done
        - to: in-progress
    done:
      description: "Complete."
      transitions: []
    cancelled:
      description: "Abandoned."
      transitions: []
---

Overriding the default

To customize the workitem state machine for a specific project, put an override at context/state-machines/workitem.yaml. The index MCP server and any other validator prefer the project file over the processkit default.

Overrides must:

• Keep the same initial state (or migrate existing data).
• Not remove states that existing entities are currently in.
• Add new transitions only from states that already exist.

Workflow

1. Copy the default from src/primitives/state-machines/<name>.yaml.
2. Edit states, transitions, and guards.
3. Save to context/state-machines/<name>.yaml.
4. Run aibox lint (once it validates state machines) — or audit manually to ensure no existing entity is stranded in a state the new machine doesn't know about.
5. Log state-machine.updated.

Gotchas

Agent-specific failure modes — provider-neutral pause-and-self-check items:

• Cycles in the state graph without explicit handling. A cycle (e.g., in-review → in-progress → in-review) is fine if it's intentional, but every cycle should be marked and have an exit criterion. Unmarked cycles cause infinite loops in process execution.
• Terminal states with outgoing transitions. A "terminal" state by definition has no outgoing edges. If cancelled has a transition to re-opened, then cancelled isn't terminal — rename it. Mislabeled terminals confuse downstream queries filtering by lifecycle stage.
• Missing initial state declaration. Every state machine needs an explicit initial state. Without it, new entities have no lawful starting point and the validator can't reject malformed creations.
• Hand-editing state machine YAML instead of using the MCP tools. Direct edits skip validation. The state-machine primitive enforces "no orphan states" and "all transitions reference declared states" at write time; bypassing it lets invalid graphs land on disk.
• State machine that doesn't match the entity's actual lifecycle. If the schema says a workitem can be cancelled but the state machine has no cancelled state, the two are out of sync and queries break. Schema and state-machine must agree — treat them as one unit when changing either.
• Forgetting to update consumers when transitions change. When you remove a transition (e.g., dropping in-progress → done direct), every workitem currently in in-progress is stranded. Audit current entities before removing transitions and write a Migration if any are affected.
• Adding states without removing dead ones. State machines accumulate cruft over time. When you introduce a new state that supersedes an old one, remove or deprecate the old one. Otherwise the graph becomes a dictionary of every state anyone ever proposed.
• Project-level state machines that diverge silently from defaults. If a project overrides a default state machine (e.g., context/state-machines/workitem.yaml), record the override as a DecisionRecord. Future maintainers need to know why the project's workitem flow differs from processkit's default.

Full reference

Fields

Transition object

transitions:
  - to: <state-name>
    guard: "optional human-readable precondition"
    on_enter: <event-type>        # LogEntry event to emit
    required_role: <role-name>    # who is allowed to make the transition

Guards are advisory unless enforced by an MCP server that knows about them.

Validation

• No cycles from terminal states.
• Every non-terminal state has at least one outgoing transition.
• Every state referenced in transitions must exist in states.
• initial must exist in states.

Machine history

Changes to a state machine are themselves LogEntries (event_type: state-machine.updated) rather than git commits alone — this lets queries reason about "what did the valid states look like on date X?".

Multiple machines for one kind

You can ship multiple state machines for the same primitive kind by using distinct names. Entities opt into a specific machine via spec.state_machine: <name>. Useful when the same kind has fundamentally different lifecycles (e.g. WorkItem with bug-lifecycle vs story-lifecycle).