tools
diapason.agents.hybrid.skillorchestra.tools
¶
The three SkillOrchestra tools: search, enhance_reasoning (code), answer.
Faithful port of orchestration/eval_frames.py:call_tool — same worker
prompts, same extraction, same Python subprocess execution. Two deltas,
both forced by the Diapason environment and documented inline:
search— the original POSTs to a FAISS wiki retriever service. We honormethod_cfg.retriever_urland POST the exact same payload when it's set; with no retriever configured we fall back to Anthropic's server-sideweb_searchtool so the stage still grounds.- in-tool correctness check — the original
answertool LLM-judges the prediction against the gold answer insidecall_tool. Diapason scores with its own harness judge downstream, so we only return the prediction; no gold answer is threaded in.
anthropic_tools()
¶
The 3 orchestrator tools in Anthropic input_schema shape.
gemini_tools()
¶
The 3 orchestrator tools in Gemini function-declaration shape.
openai_tools()
¶
The 3 orchestrator tools in OpenAI function shape.
run_answer(agent, spec, *, context_str, problem, max_tokens=40000)
¶
Generate the final answer with spec and extract the prediction.
The original branches the prompt by model family: Qwen-3 / Qwen-math
get a \boxed{} system prompt; GPT-5 / Claude (and we extend this
to every other model) get the <think>/<answer> instruction. The
in-tool LLM correctness check is dropped — Diapason scores
downstream.
run_code(agent, spec, *, context_str, problem, bash_timeout_s=60)
¶
Generate self-contained Python with spec, execute it, return stdout.
Mirrors the original worker prompt and subprocess.run(['python', ...],
timeout=60) verbatim. Execution failures yield empty exec_result
rather than raising — the orchestrator learns the model can't code.
run_search(agent, spec, *, context_str, problem, retriever_url=None, topk=150, web_search_max_uses=5, search_backend='provider', tavily_max_results=5)
¶
Write a search query with spec, then retrieve documents.
Query generation is the original verbatim worker prompt. Retrieval:
if retriever_url is set we POST the original /retrieve payload;
otherwise we fall back to Anthropic web_search (the documented
Diapason substitution for the missing FAISS wiki index).