# Implementation status

Separate available capabilities from future work.

This documentation reflects the local 0.4 source implementation reviewed on **27 September 2026**, with the standalone Jev SDK and its examples updated on **25 September 2026**. Code availability, offline tests, and production qualification are different claims.

## Available in the SDK

- Shared engine for ordinary and streaming execution.
- Typed messages, events, calls, and results.
- Schema inference and validation for Python tools.
- Execution-path permissions, explicit approval callbacks, and guardrails.
- Conversation snapshots, artifact storage, and result spilling.
- Skills, MCP, lifecycle hooks, middleware, and extensions.
- Open Knowledge Format bundles from local folders and git URLs, with keyword search, concept reads, and link traversal; no embeddings, and attested computations are not executed.
- SQLite durable state, persisted approval, replay policies, and explicit recovery.
- Incident recording, export, verification, and offline playback on the local SQL path.
- Standalone [Jev decision SDK](../decisions/): typed Choice, Score, and Noul questions, validated results, bounded retries and deadlines, cancellation, and client lifecycle management.
- Offline Jev routing, document-classification, and answer-review examples, including harness-web entries.

## Integrations that need qualification

| Path | Scope |
| --- | --- |
| Live model providers | Actual credentials, model IDs, rate limits, and tool calling must be tested |
| PostgreSQL | Shared SQL implementation exists; release needs live service validation |
| Temporal | Workflow/activity implementation exists; release needs service-backed validation |
| Temporal flight recording | Explicitly unsupported in this milestone |
| MCP endpoints | Transport, server version, credentials, and individual tools require validation |
| Docker / Seatbelt | Guarantees depend on the selected backend being available and qualified on the host |
| LangSmith | Upload and remote tracing require service validation |
| Jev decisions | Official SDK adapter tested with mock HTTP transport; live service access, decision quality, and application thresholds still need qualification |

The existence of a backend class is not proof of production readiness.

## Current web demo limitations

harness-web is a local development application. Its conversations, run IDs, and messages are retained in server memory. Generated files persist, but restarting the server does not restore web sessions.

The public City of Agents site is static and independent from that application. Its animation is an explanation of the architecture.

## Planned mechanisms

The standalone Jev adapter is implemented. Automatic agent-loop observation, persisted decision assessments, approval enforcement, and automatic interventions are later milestones. The current answer-review example runs separately after a completed agent result and does not change its status.

The existing roadmap still includes symbolic result references and provenance; decision validity and selective refresh; evidence-bound action previews; conditional compensation; generated regression scenarios; speculative execution; and reusable workflow recipes.

These are proposals, not public SDK guarantees. The flight recorder is an initial implementation of the incident-debugging portion of that roadmap; isolated executable reruns remain outside its offline playback contract.

## Guarantees we do not claim

There is no universal rollback, arbitrary semantic verification, exactly-once external effect, or deterministic rerun of live model calls. Application adapters and external systems determine which stronger guarantees can be established for a bounded workflow.

See [Approvals & recovery](../recovery/) and [Flight recorder](../flight-recorder/) for the exact recovery and inspection boundaries.