What
Expand intercepted records beyond payload + raw response to include concise intent and execution context. The new fields make each intercept easier to understand and more useful for debugging, auditing, and downstream verification decisions.
Why
Today an intercept row only describes what was sent and received. Consumers also need to know why the call was made and where in the agent's execution it happened, without exposing private chain-of-thought.
Scope
- Add fields for intent / goal and minimal sanitized reasoning context.
- Add key execution metadata (stage, action / tool, timing, model + provider when applicable).
- Keep backward compatibility for existing rows and existing readers.
Suggested field set (initial)
intent_label — short goal string.intent_context — brief, sanitized context.execution_stage — e.g. intercept | handoff | eval.tool_or_action — logical action name.model_id / provider — when present.trace_tags — small {str: str} map for filtering.
Design constraints
- Additive-only schema evolution first (no breaking changes to existing rows / readers).
- No sensitive chain-of-thought text. Reasoning context must be policy-safe and operational.
- All new fields optional with safe defaults.
Done when
What
Expand intercepted records beyond payload + raw response to include concise intent and execution context. The new fields make each intercept easier to understand and more useful for debugging, auditing, and downstream verification decisions.
Why
Today an intercept row only describes what was sent and received. Consumers also need to know why the call was made and where in the agent's execution it happened, without exposing private chain-of-thought.
Scope
Suggested field set (initial)
intent_label— short goal string.intent_context— brief, sanitized context.execution_stage— e.g.intercept | handoff | eval.tool_or_action— logical action name.model_id/provider— when present.trace_tags— small{str: str}map for filtering.Design constraints
Done when