Tricking DeepSeek V4 Pro into Minimal mode, keeping Standard tools

What it does This is a two-phase DeepSeek Harness preset that treats the first model request as a controlled experiment. It feeds DeepSeek V4 Pro the real Minimal tool pair—bash and str_replace_editor—and strips the usual workspace digest and skill-catalog reminders. After the first durable tool/call or assistant/message, the session promotes to a resident catalog that includes discovery tools and anything the model unlocks on demand. You get the Minimal preset’s “We need…” reasoning spark without permanently giving up Standard tools.

The interesting bit The author systematically isolated three levers—tool schema, output budget, and injected reminders—and found the Minimal tool schema alone anchors the trajectory at default token budgets. Because dumping the full Standard catalog immediately after promotion regresses the model back to “Let me…” phrasing, heavier tools like web_search stay one dev_tool_search call away rather than sitting in the initial catalog.

Key highlights

• Three self-contained modes: the base Anchored Standard, plus Zero-Anchored and Whoami variants that spend an extra model call on a synthetic anchor turn.
• Project2 V4.1b scores hit 98 and 99 on ability with DeepSeek V4 Pro, alongside near-zero “let me” usage.
• Promotion state is derived from durable session events, so resumes and reloads do not reset the phase.
• The bootstrap catalog uses the persistent PTY bash; the sandboxed Standard bash is disabled to avoid registry collisions, with pwsh remaining available post-promotion on Windows.
• Shared plugins live in shared/ and are materialized into each mode directory by a sync script; invariants are enforced by a zero-dependency test suite and a headless verification runner.

Caveats

• DeepSeek Harness is an explicit developer preview that permits breaking changes; this preset targets 0.1.0-rc.5 and observed that some prebuilt profiles ignore the optional first-request output cap.
• The 98/99 Project2 scores are reproducible evidence for this specific task, not a claim of universal improvement across models or workloads.
• Developed and tested on Windows with Node.js 24; cross-platform behavior is not documented.

Verdict Grab it if you are tuning DeepSeek V4 Pro inside DeepSeek Harness and obsess over first-request reasoning trajectories. Everyone else can safely ignore this very specific piece of configuration engineering.