Second release of acr-runtime. 45 commits since v0.1.0.
Highlights
- Per-role tiered routing —
spawn_agent_with_escalation(): start cheap, only pay for a higher quality tier if the cheap attempt fails review. - Paired trajectory auditing —
acr skills audit-trajectory: a real LLM judge compares baseline vs. candidate skill on the same live objective before proposing a promotion. - Dashboard
/settingspage — configure Claude/OpenAI API keys from the browser, live in-process, no restart. - A third dashboard theme (Observatory) — designed by running ACR's own trial-and-error skill-evolution loop against a local Ollama model.
- Two full audits (security, architecture, tests, dashboard, docs; the second added performance) with real fixes: a critical approve-flow crash, a judge false-positive risk, a CSRF gap, a skill-ID path-traversal risk, a prompt-injection surface in the trajectory judge, a
.envtest-isolation gap, and a SQL-side rewrite of usage aggregation for a route polled every 2 seconds. - Ko-fi support link +
.github/FUNDING.yml.
Full details in docs/ARCHITECTURE.md.
pip install --upgrade acr-runtime