dsh-llama-responses

Every time a DeepSeek Harness sub-agent is spawned, it routes through the official API and consumes tokens on top of what the main agent already spends. dsh-llama-responses splits that routing: the main agent keeps the DeepSeek official model for quality, while all sub-agents are dispatched through the OpenAI Responses protocol (/v1/responses) to a local llama.cpp llama-server instance, driving per-task inference costs close to zero.

The adapter handles multi-turn memory, streaming function-call output, usage accounting, and graceful interruption. The most critical design choice is context isolation. subagent_fork is patched to the spawn provider, so a sub-agent only ever sees the instruction text the parent agent explicitly writes into the delegation call. No previous turns, no intermediate reasoning, no parent-session history ever leaks into the sub-agent context. This keeps the full context window available for actual task content and closes the information-leakage channel between agents.

Ideal for developers who rely heavily on dsh sub-agents, have a GPU capable of running llama.cpp models locally, and want to cut inference costs without sacrificing main-agent quality. A bundled delegation SKILL enforces hard limits such as one active sub-agent at a time and self-contained instructions, keeping multi-agent coordination predictable and simple.