Skip to content

agent

diapason.agents.hybrid.skillorchestra.agent

SkillOrchestraAgent — the Diapason harness entry point.

A faithful port of the SkillOrchestra eval orchestrator (arXiv:2602.19672, orchestration/eval_frames.py). The agent runs the multi-round search -> reasoning -> answer loop in :mod:.orchestrator, using the verbatim eval_orchestrator prompts, the StageSkillHandbook + RoutingStrategy machinery, real Python subprocess execution, and a model-alias pool collapsed onto the cell's local/cloud pair.

Three things the original needs that this environment does not have, and how each is handled (see README.md in this package for the full note):

  • Learned handbook — produced offline by the explore->learn->select pipeline. With no handbook the orchestrator runs routing_strategy = "none" (the original's baseline mode). Point method_cfg.handbook_path at a StageSkillHandbook JSON to enable skill routing.
  • FAISS wiki retriever — the search tool POSTs to it when method_cfg.retriever_url is set; otherwise it falls back to Anthropic web_search.
  • 6+ model pool — the alias tiers collapse onto the cell's local + cloud models; override per alias with method_cfg.model_pool.

SWE-bench cells are out of scope for the original (it is a QA orchestrator). They run the cloud backbone through the shared mini SWE agent loop instead.

SkillOrchestraAgent

Bases: LocalCloudAgent

Inference-time skill-aware orchestrator. See module docstring.