pi-dsh-minimal: a community extension that pins DeepSeek V4 to the official DSH surface

On 2026-08-15, independent contributor Averyyy announced pi-dsh-minimal (v0.3.1) in Pi Discussion #8179. The extension is a one-line install (pi install npm:pi-dsh-minimal) that remaps pi onto DeepSeek Harness’s official minimal surface — system prompt You are a helpful software engineer assistant. plus only the persistent bash and str_replace_editor tools — for the first request of every new session (and the first request after every compaction), then restores pi’s full tool catalog for subsequent turns. The trigger fires automatically when the selected model id contains deepseek-v4-pro or deepseek-v4-flash, with /dsh unmatch to opt a model out per session.

The motivation, in the author’s words: V4 Pro “overfits the first-request surface.” On the official minimal schema the first thinking line tends to open with We need… or I need…; on pi’s default richer catalog it often falls into Let me…, and the trajectory stays put after that opening line. The author measured 69 of 113 tasks solved on the full DeepSWE v1.1 benchmark (61.1%) with opencode-go/deepseek-v4-flash:max and the v0.3.1 defaults — about 8 percentage points over the official Flash leaderboard’s 53%, and within 2 percentage points of V4 Pro’s official 63% at roughly a quarter of the cost. The README itself flags caveats: single sample, and the agent container had network access during measurement (7 of 113 runs fetched upstream repos — excluding them entirely lands at 61.3%, worst case at 57.5%, still above the leaderboard). v0.3.0 also removed an earlier v0.2.x “weak-routing + mode-boost” profile after the author’s own A/B measured no lift on the same harness.

pi-dsh-minimal is an early-stage community project — Averyyy states explicitly that it is “not affiliated with or endorsed by DeepSeek” or pi. The Discussion thread is where compatibility notes against new pi releases and new DeepSeek model ids will land first, and the maintainer invites reports of mismatched schemas against the upstream dsh-anchored-standard reference. Treat the benchmark numbers as the author’s own measurements on a single sample, not an independent result.