Millrace

v0.15.8 Breaking

This release includes breaking changes for platform teams planning a safe upgrade.

Published 3mo AI Agents & Assistants

✓ No known CVEs patched

✓ No known CVEs patched in this version

Topics

agent-loop agentic-ai-development agentic-coding agentic-engineering agentic-framework agentic-workflow

+7 more

autonomous-agents governed-autonomy governed-loop harness-engineering loop-engineering runtime-engine runtime-engineering

Affected surfaces

breaking_upgrade auth

Summary

AI summary

Added version commands, daemon log monitoring, Codex reasoning effort config, and bulk watcher policy enforcement.

Full changelog

0.15.8 hardens upgrade/config surfaces and serializes bulk watcher root-spec
intake behind the v1 one-open-closure-target policy.

Added millrace --version and millrace version for package-version
visibility in installed and module-entrypoint environments.
Added millrace run daemon --monitor-log PATH to write basic monitor output
to a file while keeping stdout monitor mode independently selectable.
Added first-class Codex model_reasoning_effort config on
runners.codex and per-stage stages.<stage>.model_reasoning_effort, with
compile, request, runner artifact, and run-inspection visibility.

Bulk watcher intake now respects the v1 one-open-closure-target policy as
queue backpressure: unrelated root specs remain queued while the active root
lineage runs or reaches Arbiter, status reports the deferred-root count, and
direct stale-state recovery preserves the open closure target.
stages.<stage> and runners.codex.permission_by_stage now accept learning
stages such as professor.
Workspace doctor now resolves the selected mode against workspace-local
deployed assets, matching runtime compile/startup behavior for custom modes.
millrace upgrade can intentionally localize removed managed assets through
--localize-removed or --localize-removed-from, and runtime asset
manifests now ignore cache artifacts such as __pycache__ and *.pyc.
Runtime shutdown now clears process_running, status suppresses stale
process_running truth when no active ownership lock exists, and
clear-stale-state reports applied when it only clears that stale process
bit.