custom-first-control-prompt

In DeepSeek Harness, the system prompt a model sees on each request is usually built from a built-in persona and tool guidance. If you want to inject a fixed deployment-level instruction or reference examples before every ordinary conversation, directly rewriting the request path or session often pollutes logs, breaks prefix-cache reuse, and is hard to maintain. custom-first-control-prompt is designed for exactly this: it intercepts and prepends content on the llm/stream request path, with zero session-log writes, so the injected material exists only in the model-visible request and never appears in the chat transcript.

The plugin's core capability has two layers. The first is ordered system-prompt sections: each section registers through the standard system-prompt assembly waterfall, supports variable interpolation and scope shadowing, and can be positioned by order (for example, ahead of the persona) to create a deployment-level prompt prefix. The second is reference conversation history: configured user/assistant texts are built into real alternating message objects, deep-frozen and prepended to every ordinary request. Because the content renders byte-identically, prefix caches stay reusable, reducing token costs; meanwhile session history remains clean, so compaction, forks, and other mechanisms are unaffected.

It suits teams or advanced users who need fine-grained deployment-level prompt control—especially those wanting to inject few-shot examples, fixed behavior rules, or a consistent override of the default persona—while also caring about KV-cache stability and log cleanliness. The panel provides visual config editing and LLM-listener debugging without hand-written patches, though config changes require a web restart before they apply to new requests.