dsh-weneedfirst

Chain-of-thought output from LLMs tends to drift mid-conversation: the language shifts, the persona wavers, and the opening of each reasoning block is unpredictable. dsh-weneedfirst tackles that style inconsistency head-on by injecting a single global system-prompt section that pins three things down at once—the model should adopt a role that fits the conversation, think in English throughout, and open every reasoning block with a configurable phrase (default: We need). There is no build step, no runtime state, and no side effects; the plugin contributes exactly one prompt section to the systemPrompt registry and nothing else.

Its core capabilities are role anchoring, English-only reasoning, and a uniform opening phrase, all enforced globally so they apply regardless of which tool or conversation context is active. Because the injection targets the systemPrompt service, it activates in any profile that composes a system prompt with no per-chat configuration. Want a different opener? Change the openingPhrase value in your profile's cordis.patch.yml and restart the web app—that is the entire configuration surface.

Who will benefit most? If you use DeepSeek Harness for long-chain reasoning, analytical notes, or code reviews and want the thinking to read like a single, steady author rather than a montage of disconnected monologues, this plugin is the minimal ruler: one configurable string, zero dependencies, MIT licensed.