Appearance
Decision Records
The Decision Context Record (DCR) is the single source of truth for one runtime decision — created at intake and updated in place only until it reaches a terminal status. It is immutable per Agent Run: a retry never edits an existing record, it creates a brand-new one.
Fields
| Field | Description |
|---|---|
decision_id | Unique per Agent Run |
tenant_id | Resolved tenant — see Security |
case_id, task_type, user_role | What was requested |
agent_id, agent_version, model_provider, model_version | Who produced the output |
input_context_hash | Deterministic hash of the input context |
retrieved_sources | The sources available for evidence checking |
agent_output | The structured output submitted (null until received) |
policy_results | Every policy evaluated, pass/fail |
evidence_results | Per-claim support assessment |
risk_score | Computed runtime risk |
status | intake_rejected | awaiting_agent_output | awaiting_human_review | blocked | finalized |
human_action, reviewer_id, review_notes, final_output | Set once a reviewer acts |
runtime_events | Every step recorded against this record |
Status lifecycle
text
awaiting_agent_output ──► (evaluated) ──┬─► finalized (auto-approved)
├─► awaiting_human_review ──► finalized
└─► blocked ──► finalized
intake_rejected (terminal, gateway validation failed before any agent ran)API
GET /runtime/decision-records
GET /runtime/decision-records/{id}
GET /runtime/decision-records/{id}/export
GET /runtime/decision-records/{id}/replayexport wraps the record with an exported_at timestamp for audit handoff. replay returns the single-run Decision Replay for this record.
SDK usage
ts
const record = await runtime.getDecisionRecord(decisionId);
const all = await runtime.listDecisionRecords("finalized");