0.4.0 reduces the work needed to maintain a task while preserving source authority, required context, evidence checks and recovery behavior. It adds explicit management assessments, combined task preparation, optional concise responses, and examples for host orchestration and reviewed execution reports.
npm install -g @originoneai/agent-work-runtime@0.4.0
# or, in a Python virtual environment
python -m pip install agent-work-runtime==0.4.0
Changes
- Source diagnostics include a file, a line/column when known or a field pointer, the violated rule and a repair direction. Human, JSON and MCP outputs use the same underlying outcome. Evidence summaries preserve source and verification differences.
work preparereturns readiness, required context and management requirements together. Hosts acknowledge the context actually delivered before recording a checkpoint or completing work; missing capabilities retain the older workflow.- Management is assessed from observed coordination needs. Persistent waits, handoffs, separate subtasks, unresolved dependencies and uncertain outcomes require continued tracking. Time and rework thresholds trigger reassessment; unknown facts remain unknown. An upgrade preserves identity and history and does not lower completion requirements or automatically downgrade later.
- CLI and MCP support an explicitly selected summary response for task preparation and transitions. Full output remains the default. Summaries retain required context, identities, versions and recovery links; failures and unknown outcomes retain their full results. Full durable operation receipts remain queryable.
- MCP exposes structured task creation and plan changes with dependency checks and request recovery. The host orchestration example handles ready work, bounded parallel execution and interruption. Agents plan the work; hosts own model execution, scheduling and concurrency limits.
- The host workflow example collects actual AWR-managed command results, source snapshots, logs and artifact hashes into reviewed completion reports. Explicit review is required; exit code zero alone cannot complete a task. Changed, missing or uncertain evidence is rejected. Lost-response recovery reuses the original execution and evidence, including reports larger than 64 KiB.
- Managed execution identity recognizes equivalent filesystem directories across platforms, including Windows native path prefixes, while still rejecting a different project directory or changed command.
See the workflow contract, host workflow guide and host example.
Measured behavior
The full-workflow benchmark records 30 synthetic runs: combined preparation reduces tool calls by 18–27% and returned text by 3.4–4.7% against the equivalent older workflow.
The summary-response comparison records 45 synthetic runs: summary responses reduce cumulative returned text by 6.8–9.2% against full prepared responses, with the same number of calls. These are separate comparisons with recorded source identities, not additive savings or estimates of actual model tokens or billing. Context caching and incremental context reuse are not part of this release.
Compatibility and upgrade
Database schema remains 4. Keep the original project directory, authoritative sources, runtime history and previous executable; retain a matching backup and stop active writers before switching a pinned binary. Follow the upgrade and rollback procedure. Negotiate capabilities when using the new host helpers with an older binary.
The release provides macOS 15+ arm64 and Intel x64, Linux x64 with glibc 2.39+, and Windows x64 packages. npm launchers require Node 22.14+; Python launchers require Python 3.9+. Keep npm optional dependencies enabled. Native executables can run without either launcher runtime. Packages contain Apache-2.0 and applicable dependency notices; manifests and checksums bind their source and binary identity.
Stdio MCP exposes 29 tools; shared HTTP adds authorized project discovery for 30. Existing commands retain their full-output defaults. The shared service continues to isolate projects and clients and preserve sessions, waits and operation recovery across reconnects. Hosts remain responsible for model execution and user interaction. Platform, installation and fixture checks are distinct from a host's own application or business acceptance.
