Agentic run execution fields — Oshun Platform

Current POST /v1/agentic/runs/execute persistence boundary. The route runs every plan through the governed boundary (runGovernedToolPlan): each reached tool call is grant-resolved through the capability broker, recorded as a pending attempt before dispatch, and then replaced by its authoritative outcome telemetry — ok, partial, error, or denied — carrying redacted arguments, the result or error, the resolved grant, tool-produced evidence, effects, and measured usage. Measured cost is charged to the run's ledger before the next admission, so a later tool is never admitted against a balance an earlier call already spent. Pause, cancel, and kill block later dispatch; kill and budget terminations are appended to the durable lifecycle store alongside the terminated envelope and the denied attempt. The reading rule for every consumer of this store: a persisted run is the authoritative record of what executed — every reached tool call is grant-resolved, recorded, charged, and persisted, and a field that has nothing to carry (a tool that produced no evidence) is an honest empty list, never a fabricated one.

Runtime sources#

apps/oshun/bff/src/agentic/runs-route.ts
apps/oshun/bff/src/agentic/run-lifecycle-store.ts
libs/oshun/agentic-studio/src/runs/governed-orchestrator.ts
libs/oshun/agentic-studio/src/runs/dispatcher.ts
libs/oshun/agentic-studio/src/runs/executor.ts

Exercised by#

apps/oshun/bff/src/agentic/runs-route.test.ts

Field contract#

Field	Current producer	Current behavior
`status`	governed boundary, runs-route finalizer, or run control	Refreshed at every tool boundary; pause/cancel/kill block later dispatch and a fully executed plan becomes completed.
`completedAtUnixSeconds`	runs-route finalizer or terminal run control	Populated for a completed or otherwise terminal run.
`updatedAtUnixSeconds`	runs-route finalizer or terminal run control	Populated when this route finalizes or terminates the envelope.
`toolCalls`	runGovernedToolPlan	A pending attempt is persisted before dispatch, then replaced by ok, partial, error, or denied telemetry with redacted arguments, result/error, grant, evidence, effects, and measured usage.
`capabilityAuditTrail`	PortfolioCapabilityBroker through runGovernedToolPlan	The complete immutable hash chain is retained with the run, including grant request/policy creation plus every use, result, denial, revocation, and attempted bypass reached by this execution.
`evidenceTrail`	server-owned tool result observer through runGovernedToolPlan	Real tool-produced evidence is appended to the call and the run trail; tools that produce no evidence retain an honest empty list.
`costLedger`	runGovernedToolPlan measured-usage accounting	Success, partial, error, retry, and denied policy charges update total, per-stage, per-tool, and category counters before the next admission; mutable budget meters consume the same measured category usage.
`termination auditEvent`	admitToolCall/kill	Kill/budget termination events are appended to the durable lifecycle store while the terminated envelope and denied attempt are persisted.

Claim boundary#

The legacy runGuardedToolPlan/dispatchGuardedToolCall primitive remains a lower-level seam and still does not manufacture telemetry; the executed route uses runGovernedToolPlan.
The lifecycle store preserves boundary-produced fields and control audit events; it does not fabricate tool evidence or usage.
toolCalls is write-ahead, not reconstructed: the pending attempt is persisted before dispatch and only then upgraded to its ok, partial, error, or denied outcome.
The capability audit trail is an immutable hash chain retained with the run — grant request and policy creation plus every use, result, denial, revocation, and attempted bypass this execution reached.

Contract: scripts/audit/machine-adjacent-doc-contract.json. Regenerate with pnpm generate:audit:machine-adjacent-docs.