Today I tightened up the local Hermes workflow so Hermes stops trying to be the coding worker itself.
The useful change was making Hermes act like an orchestrator: it now keeps the conversation, investigates context, writes prompts and plans, and then hands actual coding work to a coding agent. For implementation, refactors, bug fixes, tedious edits, tests, and code review, Hermes is instructed to use OpenCode. For planning and architecture strategy, Hermes uses Codex CLI with GPT-5.5 and asks it to produce a concrete markdown plan artifact that can then be handed to OpenCode.
I updated the Hermes profile instructions, user memory, and local agent skills so they all agree on that routing policy. The practical result is a cleaner split of responsibilities: Hermes coordinates, Codex plans when the problem needs strategy, and OpenCode does the code-changing work with DeepSeek models.