Skip to main content
Amends 0100 and 0155. Builds on the AI runtime of 0156.

​
Context

An AI agent needs instructions, tools, knowledge and examples, and each product needs its own. The owner’s intelligence-plane model (mem L902–912, 2026-10-20) maps the owner document’s agents, skills, intents, plans, capabilities, tools, context and prompts onto the action model. It adds a few primitives and keeps actions as the only unit of behaviour.

​
Decision

  1. Actions stay the only behaviour unit (ADR 0098).
  2. Capability. A capability is an action or a query. The registry is the catalogue.
  3. Tool. A tool is the MCP projection of a capability, generated, never hand-written (ADR 0155).
  4. Intent. An intent is metadata on an action (intents). An intent router maps natural language to an action, for the command palette and for AI triggers.
  5. Skill. defineSkill({ instructions, tools, examples, knowledge }) is loaded on demand, with progressive disclosure, so the agent reads only the skill it needs.
  6. Prompts. definePrompts(...) defines versioned MCP prompts, with evaluations. A tenant may override a prompt within limits.
  7. Context providers. @ContextProvider classes are contributed by modules: facts about a record, a page, a user or a tenant. They are assembled with retrieval and memory under a token budget, and exposed as MCP resources.
  8. Plans. An agent proposes a list of actions with inputs. A user approves the plan at level L2 and above. A plan runs as an ephemeral workflow on the workflow engine (ADR 0152). There is no separate plan engine.
  9. Free and permission-aware. All of it is free code, and every item respects the caller’s permissions (ADR 0155).

​
Consequences

  • A new product adds skills and context providers beside its actions, and the agent finds them through the registry.
  • A plan cannot act beyond what its user may do, because each step is a normal action call.

​
Enforcement

planned — the MCP generator checks each tool against an action; the plan runner refuses a step the approving user lacks permission for.

​
Supersedes and sources

Supersedes: none. Amends: 0100, 0155. Sources: owner intelligence-plane model (stackra-decisions-2026-10-19.md L902–912, 2026-10-20).