Skip to content

forwarding-policy.md

Latest commit

 

History

History
63 lines (47 loc) · 3.1 KB

forwarding-policy.md

File metadata and controls

63 lines (47 loc) · 3.1 KB

Forwarding Policy

The protocol supports two deployment patterns for inter-task data flow, controlled by compliance.forwarding_policy in each task's envelope output.

Policy Types

Policy Description Risk Level
semantic_forward Next task receives only semantic_payload — structured, auditable semantic extractions. Raw tokens/vectors/embeddings are stripped from the context before the next task runs. Required for HIGH
raw_forward Next task receives the full envelope (all fields). Faster, but audit trail may diverge from what the agent actually consumed. Permitted for LOW/MEDIUM

Key Properties

  • Per-task granularity — each task controls its own forwarding via the envelope it outputs. A fetch task can use raw_forward while the downstream classification task switches to semantic_forward.

  • Monotonic enforcement — once any task in a crew sets semantic_forward, subsequent tasks cannot downgrade to raw_forward. The ContextMixin._resolve_forwarding_policy() tracks this boundary in flow state. Violations are overridden with a logged warning.

  • Full persistence regardless — the callback persists the complete envelope (with all artifact metadata) to the backend API, even when the next task only sees semantic_payload. Nothing is lost for audit — raw artifacts are always in the backend for debugging.

  • Envelope-driven — the forwarding_policy is a first-class field in the ComplianceBlock, not a prompt instruction. The EnvelopeBuilder.set_risk_level(HIGH) auto-sets semantic_forward; set_risk_level(LOW) auto-sets raw_forward. Tasks can override via their YAML-instructed envelope output.

Healthcare Example (Mixed Forwarding)

sensor_task  (raw_forward)      →  output.raw = full envelope (all observations)
                                        ↓  situation_task sees everything
situation_task (semantic_forward) →  output.raw = {"semantic_payload": [...]} ONLY
                                        ↓  decision_task sees only semantic data
decision_task (semantic_forward)  →  only semantic_payload visible

The semantic boundary is at the classification step (situation_task). Before it, raw data flows freely for ingestion. After it, only structured semantic extractions are visible to decision-making tasks — enforced structurally, not by prompt.

Recommendation Example (Raw-Forward Throughout)

profile_task  (raw_forward)  →  search_agent sees EVERYTHING
search_task   (raw_forward)  →  personalize_agent sees EVERYTHING

No semantic boundary. All agents consume full aggregated context.

Task Preamble Injection

The ContextMixin._get_forwarding_preamble() method generates a constraint instruction from the flow-level envelope's forwarding policy. This preamble is passed to crew.kickoff(inputs=...) and interpolated into task descriptions via {_forwarding_preamble} placeholders in YAML. For Semantic-Forward flows, tasks receive an explicit instruction to read only semantic_payload. For Raw-Forward, the preamble is empty.