Source dive: DeepSeek Harness quietly relies on Pi libraries, with a 46-char system prompt
solyarisoftware · x · 2026-09-26
A developer digging through DeepSeek Harness (DSH) source code found it directly depends on earendil-works' pi-ai and pi-tui libraries (the pi-tui dependency was removed in August), plus a dedicated dsh-llm-pi-ai package. pi-ai handles provider catalogs, OpenAI/Anthropic protocol adaptation, reasoning and compatibility params, and context overflow handling — the vendor-facing layer is effectively Pi's work for now, likely to be replaced later.
Comparing system prompts on the same DeepSeek-V41-Flash/High setup: DSH minimal mode is just 46 characters — literally "You are a helpful software engineer assistant." — while Pi's default 4-tool prompt is 2,741 chars (1,784 even with all tools off), roughly 59.6x larger. Conclusion: Harness's minimal mode really does run close to a bare model (tool JSON schemas not counted).
More from coding & agent
- Skip pptx: Web-Based AI Slides Look Better, But You Still Need PowerPoint for the Boss — lxfater · 2026-09-26
- Personal Agents Will Be Interchangeable; Personal Context Is the Real Moat — vaibhavbetter · 2026-09-26
- Matt Pocock: your CODING_STANDARDS.md should be empty for only 5 minutes — mattpocockuk · 2026-09-26
- Custom benchmark: 35B Qwen3.6 scores 95% vs 53% for 120B GPT-OSS on coding agent — pauliusztin · 2026-09-26
- Installing DeepSeek Harness on a Muse cloud VM to unlock Chinese web search — op7418 · 2026-09-26
- Coded animation lands in a single AI one-shot, no assets or examples needed — eschadiol · 2026-09-26