Define the task execution graph and monotonic capability-state contract

[dinpd]

Parent: #78

Outcome

Define the versioned task identity, capability ceiling, state machine, policy vocabulary, and evidence contract that every AgentPass runtime will use for monotonic task capability enforcement.

Scope

• Task graph and template identifiers.
• Canonical capability-ceiling representation and digest.
• Named capability states and state versions.
• Protected-event triggers and declared capability removals.
• Rules for initial restricted state, expiry, new-task boundaries, and approval behavior.
• Transition evidence without protected payloads.
• JSON Schema, TypeScript types, examples, and policy documentation.
• Relationship to existing job boundaries, intent contracts, data flows, JIT grants, and receipts.

Non-goals

• Executing or persisting transitions.
• Defining a universal identity credential.
• Network-vendor-specific configuration.
• Automatic policy widening from historical evidence.

Acceptance criteria

• A versioned schema represents graph ID, template ID, ceiling digest, state, version, allowed/removed capabilities, and transition policy.
• Validation rejects transitions that widen within the same task.
• Approval and JIT semantics explicitly cannot restore removed capabilities.
• Initial prompts, restored inputs, and transferred protected references can start a task restricted.
• Evidence schema identifies state-before, trigger reference, removed capabilities, acknowledgements, state-after, and outcome without sensitive content.
• Existing manifests remain compatible or fail with an explicit migration error.
• Documentation includes baseline-to-restricted and initially restricted examples.
• Unit/schema tests cover valid narrowing, attempted widening, unknown state, stale version, and malformed transition.

Security considerations

The policy must not trust model-supplied classifications, state versions, ceiling digests, or acknowledgements. Portable records must exclude credentials and protected payloads.