workflows/stateful_react_agent/profiles/simple.yaml

# Deterministic single-agent baseline: keep five recent tool results and never
# call an LLM to summarize history.
llm:
  model: ${OPENAI_MODEL}
  api_key: ${OPENAI_API_KEY}
  base_url: ${OPENAI_BASE_URL}
  # Not 0.0 — see the note in this workflow's benchmark.yaml. At temperature 0
  # sglang takes the argmax path and a reasoning runaway becomes a deterministic
  # deadlock; top_p/top_k are inert until temperature leaves 0.
  temperature: 1.0
  top_p: 0.95
  extra_body:
    top_k: 20
  max_tokens: 32768

agent:
  main_max_turns: 200
  tool_timeout_s: 900
  llm_timeout_s: 600
  reasoning_only_timeout_s: 120
  reasoning_only_max_tokens: 16384
  logical_call_timeout_s: 900
  thinking_in_history: true
  # Optional PER-TURN cap on the reasoning replayed for one assistant turn; it
  # also enables thinking history when the boolean above is omitted. An explicit
  # thinking_in_history: false always takes precedence. This is not a budget
  # over the whole history — total reasoning still grows with turn count, which
  # is compaction's job to bound. Applies to the `tag` and `reasoning_content`
  # thinking formats; native content-block reasoning (Anthropic extended
  # thinking, OpenAI Responses) is always replayed verbatim, signatures
  # included, because the provider validates it server-side.
  # thinking_history_max_tokens: 8192
  # `off` only blanks older tool-result bodies, keeping the latest five. It
  # never summarizes user/assistant history, so very long runs can still fill
  # the context window.
  keep_last_k: 5
  context_compaction: "off"
  compaction_spill: false
  max_input_tokens: 229376
  agent_tools: [web_search, web_fetch, bash, download_file, grep_search, glob_search, recover_result]
  web_search_impl: original
  web_fetch_impl: original