Skip to content

v0.2.0 - Worker-first MCP architecture for long-running parallel Codex tasks

Choose a tag to compare

@aresyn aresyn released this 20 Jun 13:21
· 3 commits to main since this release

v0.2.0 - Worker-first MCP architecture for long-running parallel Codex tasks

This release is a major architecture update for Codex Control Plane MCP. It
focuses on reliability, stability, and predictable agent behavior when several
long-running Codex turns are running in parallel across different projects.

Highlights

  • Added a central worker architecture with client, worker, observe, and
    inline execution modes.
  • Added durable scheduling for long-running tasks with queue state, worker
    heartbeats, resource locks, per-thread locks, and concurrency limits.
  • Made status/read endpoints passive and bounded by default so polling does not
    create new live app-server work.
  • Added a self-describing MCP contract through codexMcpGuide,
    codex_get_agent_contract, tool annotations, startup flows, and guide hashes.
  • Improved agent guidance, diagnostics, recovery loop guards, and structured
    error next steps so agents know whether to poll, wait, diagnose, repair, or
    stop.
  • Strengthened terminal reconciliation, queue/slot/lock release, and worker
    command behavior for long-running and interrupted turns.
  • Added consistent agent-safe redaction for public status, diagnostics, search,
    runtime, lifecycle, and worker command responses.
  • Added external MCP client tooling for live testing without restarting Codex
    Desktop sessions.

Why it matters

OpenClaw, Hermes, and other MCP clients can now treat the server as a durable
control plane: submit work, poll status, follow agentGuidance, and avoid
duplicate retries while the central worker owns Codex app-server execution. The
new architecture is designed for multi-project workloads where several heavy
turns may run for minutes or hours.

Compatibility

The public API remains backward compatible. Existing tools are still available,
while new fields are additive. Agents should prefer codex_submit_task for new
work, codex_health_summary on startup, and codex_get_agent_contract or the
tools/list guide for discovery.

Validation

  • python -m pytest -q
  • python -m compileall -q openclaw_codex_mcp codex_control_plane_mcp tests scripts
  • git diff --check
  • python .\scripts\mcp_live_smoke.py --scenario protocol --timeout-seconds 30
  • external MCP client baseline and diagnostics scenarios
  • python -m build
  • python -m twine check dist/*
  • clean virtualenv wheel install and console script smoke