Hermes + DeepSeek V4 Flash: a one-line reasoning-effort throttle

The Stack

Tested Against

Prerequisites

• Hermes Agent installed
• An OpenRouter API key (only to actually run the agent)

Steps

1. 1

   Set the model and a default reasoning effort, then validate

   Write config.yaml with the deepseek/deepseek-v4-flash slug and agent.reasoning_effort: high (Hermes' default is medium). At runtime you switch per task with /reasoning xhigh or /reasoning none, no restart. CI parses the YAML and asserts reasoning_effort is a valid level.

   cat > config.yaml <<'YAML'
   model:
     provider: openrouter
     model: deepseek/deepseek-v4-flash
   agent:
     reasoning_effort: high
   YAML
   ruby -ryaml -e '
   allowed = ["none","minimal","low","medium","high","xhigh"]
   c = YAML.load_file("config.yaml")
   m = c["model"] || {}; a = c["agent"] || {}
   abort "BAD: model not deepseek/deepseek-v4-flash" unless m["model"] == "deepseek/deepseek-v4-flash"
   re = a["reasoning_effort"].to_s
   abort "BAD: reasoning_effort not a valid level: #{re}" unless allowed.include?(re)
   puts "config OK: deepseek/deepseek-v4-flash + reasoning_effort=" + re + " is a valid level"
   '

2. 2

   Throttle per task at runtime (the model step, not checked by CI)

   /reasoning xhigh for the hard one, /reasoning none for a quick lookup. xhigh can multiply output tokens, so use it deliberately; keep stable prefixes consistent to hit DeepSeek's cached-input discount. The reasoning itself is fenced.