aa-blinov/agent-orchestrator
v0.4.3MIT
Route bounded coding tasks to local coding agents with per-chat call limits and usage reporting.
Changelog
All notable changes to Agent Orchestrator. Versions are ordered newest first.
[Unreleased]
[0.4.3] - 2026-11-12
Fixed
- Detect structured Antigravity and Codex failures even when their CLI exits with status zero.
- Return tool validation and execution errors using MCP's
isErrorresult field. - Run MCP tool calls concurrently so the server can receive cancellation notifications during long agent calls.
- Stop cancelled child process groups and suppress responses for cancelled requests.
[0.4.2] - 2026-11-12
Added
- Added an enforced per-chat budget of four agent CLI calls and three distinct peer agents, keyed by a stable
session_id. - Added a local Codex plugin package with bundled MCP configuration and an agent-routing skill.
- Added server-level routing instructions for cost-aware delegation.
- Added selectable MCP hosts through
AGENT_ORCHESTRATOR_HOST. - Excluded the active host from peer tools to prevent recursive delegation.
- Renamed the project and bridge to host-neutral names.
- Added a shared JSON configuration for built-in and custom CLIs, preferred models, and per-task model profiles.
- Added separate preferred models for research and implementation roles.
- Published the Codex plugin package with a reusable routing skill.
- Set the maximum consultation and implementation call duration to two hours.
0.1.0 (2026)
Added
- Added a local MCP bridge for consultation and delegated implementation with Antigravity, Claude Code, and OpenCode.
- Added separate Git worktrees for implementation results and reviewable diffs.