RFC-0009 · Design Q&A
Versioned draft snapshot
RFCs, reference catalogs, and evidence reflect Spec 3b2e3e7a114fb826b1445d39a48756b8fff629cd. Counts and pending/status notes belong to that snapshot. Read the dated project update for later progress.
Source repositories are private during staging. Links marked “repository access required” need access to the linked repository.
Q: Should agent declarations be versioned independently of module versions?
A: Agent declarations are versioned as part of their enclosing module (RFC-0014). When a model update changes behavior, the module MUST increment its version and the install plan MUST include a model_update activation entry (RFC-0005 §5a). Model digest is tracked in the model declaration (RFC-0008 §6.20) and pinned in the module lockfile. Independent agent versioning would require a separate registry and resolution protocol that is not warranted; module-level versioning is sufficient.
Where normative: §4 (agent versioning) of RFC-0009
Q: Should recommend and propose be unified?
A: recommend and propose are kept distinct. recommend produces a typed domain instance (e.g., a PlanRecommendation) for human review. propose requests a state transition — it produces a Command proposal, not a data instance. The runtime routes them differently: recommend output is surfaced as a data item; propose output is surfaced as a confirmable action button. Unifying them would conflate data production with action authorization in ways that would complicate the surface layer and policy model.
Where normative: §5 (agent protocols) of RFC-0009
Q: How should multi-agent workflows be declared — as a separate declaration kind, or as a composition of individual agent declarations?
A: Composition of individual agent declarations. A multi-agent workflow is declared as a flow (RFC-0008 §6.7) whose state transitions may each invoke different agents. There is no separate “workflow” declaration kind. This keeps the declaration surface minimal and reuses the existing FlowType state machine model. An agent that orchestrates sub-agents MUST declare each sub-agent invocation as a distinct action step within its flow, with evidence lineage threading through each step.
Where normative: §6 (multi-agent workflows) of RFC-0009
Q: Should confidence thresholds in policy rules be numeric constants or symbolic levels?
A: Numeric constants in normative policy declarations (e.g., confidence >= 0.85). The gnomon.core standard library module provides symbolic aliases as annotation facets: Confidence.HIGH = 0.9, Confidence.MEDIUM = 0.75, Confidence.LOW = 0.5. Authors who prefer symbolic levels import gnomon.core and use confidence >= Confidence.HIGH. The expression language supports both forms; symbolic aliases are resolved to their numeric values at compile time.
Where normative: §7 (policy expression language) of RFC-0009
Q: How does the AI participation contract interact with external AI services?
A: External AI models (third-party LLM providers) MUST be declared using model declarations (RFC-0008 §6.20) with provider, name, digest, and roles fields. They are bound to agents via binding declarations (RFC-0015). The same evidence obligations apply regardless of whether the model is local or remote. RFC-0015 §13 specifically addresses the read-scope constraint for externally dispatched inference.
Where normative: §13 of RFC-0015 (see RFC-0009 §9)