Making DeepSeek Flash think deeply without pro prices

What it does

This is a DeepSeek Harness (dsh) agent preset for the opencode-go provider’s deepseek-v4-flash model. It replaces the default bare system prompt with a static WEAK_FLASH persona that mandates task classification, progress review, anti-digression guardrails, deep architectural thinking, and a decision loop before any code is written. The author claims this pushes the model from shallow “ghost mode” into methodical “god mode” on complex build tasks.

The interesting bit

The trick isn’t a bigger model or more tokens—it’s stricter prompt engineering. The preset borrows research from dsh-router-standard, strips out three dynamic-injection mechanisms that break under dsh rc.6, and collapses the guidance into a static persona that still handles cross-platform shell switching.

Key highlights

• Targets the cheap deepseek-v4-flash endpoint; non-Flash models fall back to default logic
• Claims a roughly 13× jump in planning depth (2.9万字 → 37.5万字) on a 3D diesel-engine simulation benchmark
• Includes self-testing hooks that reportedly reached zero errors across 236 meshes and 49 parts
• Auto-switches between bash and pwsh for Linux/macOS/Windows without user intervention
• Avoids fragile dsh rc.6 session-event hooks by statically merging prompts instead of injecting them at runtime

Caveats

• Only validated on dsh rc.6+ with the opencode-go provider; other toolchains are untested
• The dramatic quality claims come from a single complex task (4-stroke diesel 3D simulation), not a broad benchmark suite
• macOS and Windows support rely on platform-detection logic rather than explicit testing; Linux is the only explicitly “实测” platform

Verdict

Grab it if you’re already running dsh with opencode-go and want senior-level planning from the cheapest Flash tier. Everyone else can treat it as a well-documented prompt-engineering case study.