QitOS is the torch-flavor framework for agent researchers.
Prototype methods, run benchmarks, and inspect long-horizon trajectories on one AgentModule + Engine kernel with built-in qita observability.
QitOS core is the small framework. Product-grade applications and showcase agents live in qitos-zoo, including planned apps such as qitos-coder and qitos-cyber-agent.
Quickstart · Tutorial Track · Benchmarks · CLI Reference · Changelog · Chinese README
Build multi-round tool Agents on one AgentModule + Engine kernel, with Session
pause, restore, fork and handoff. Configure budgets in YAML, select durable memory
and explicit compaction policies, and inspect or export canonical Trajectory with
qita. Extend tools, providers, context, stores, sinks and sandboxes. The framework
owns execution correctness; application strategy and model behavior remain with
the application and provider.
- Documentation links now reject unpublished branch targets, verify immutable first-party source objects, and keep Chinese pages on Chinese routes; the memory guide's broken consumer link is repaired.
- Agent Design Lab now includes six independently installable EN/zh teaching projects, custom
agent_factorycomposition, durable skill revisions and live-driven runtime fixes. Python 3.10/3.12 installed and real-Docker mechanism checks pass; real-model task failures remain explicit. See the evidence and limitations and implementation ledger. This is not a package release or whole-lab live qualification. - Python 3.10 model cleanup compatibility: every owned resource is attempted, with safe numeric cleanup diagnostics and primary exception priority preserved.
- R1 local integration and all five review repairs are complete: reasoning, accurate request/cleanup facts, handoff ownership and YAML budget/loss wiring; restoration retains original conversation facts.
- Memdir recall and closed-exchange compaction, canonical Read/Edit ToolResult, consistent Observation, strict snapshot paging and atomic export are available through the user workflow.
- Bilingual tutorials include complete runnable files, and API Reference binds to exact implementation source. See migration guidance for compatibility changes.
Current R1 status and remote synchronization is the sole status entry; the V5 roadmap retains subsequent scope. R1 completion does not complete V5. Offline framework qualification does not imply live-model qualification, a package release or a documentation deployment.
pip install qitos selects a published PyPI version, not an identity for unreleased R1.
Follow Installation, then the
credential-free Quickstart.
Real providers use agent.yaml, CredentialRef and an explicit resolver.
Docker file tools fail closed when unavailable; they do not fall back to host execution.
- Eight learning units: custom Agents, parallel tools, Sessions, context/memory, sandbox/artifacts, multi-agent work, qita and third-party extensions.
- Complete teaching files and example directory.
- Migration, limits and troubleshooting.
- Contributing and architecture.
- CHANGELOG, historical engineering progress and G5 evidence.
Advanced AgentModule.run() and historical trace compatibility remain supported.