dsh-bestthink-standard

English | 中文

Best-Thinking + Standard mode — a DeepSeek Harness (DSH) agent preset that enters the measured best-thinking trajectory (the RL-aligned minimal condition) on the very first request, then auto-promotes to the full Standard 25-tool catalog with workspace context restored.

Compatible baseline: DeepSeek Harness 0.1.0-rc.5 / commit 47f9438. License: MIT (ported from xiaobright/dsh-anchored-standard; agent.cordis.yml derives from the official standard preset — see preset/NOTICE).

---

Why

A frontier reasoning model's behavior is strongly conditioned by the first request's system prompt and visible tool catalog — not by its raw capability. Measured on the Project2 maintenance benchmark (DeepSeek V4 Pro, reasoning_effort=max):

Condition	Ability
official minimal (1-sentence prompt, 2 tools)	99 / 96
two-phase anchored (minimal first, full catalog after first tool call)	98 / 99
official standard (full 25-tool catalog from the start)	91

The first request commits the session to a trajectory; expanding the tool catalog after the first durable tool call perturbs at most one reasoning block and never flips the mode. So the anchor is free: you get the best-thinking first turn and the full Standard capability.

How it works

Three anchoring factors lock the first request into the RL-aligned minimal form, uniformly for every model (no per-model special-casing — sessions that switch models mid-flight can never receive mismatched settings):

1. Pristine system prompt — the persona row is the minimal sentence with complete: true + includeRuntimeContext: false, so the assembled system prompt is exactly You are a helpful software engineer assistant. (harness identity, tool guidance, and plugin sections are all suppressed).
2. Tight output budget — the first request is capped at maxTokens=1024 (the measured anchor value); the cap is explicitly stripped after promotion so it can never leak into later steps.
3. Narrow tool surface + stripped injections — the first request sees only shell + read; skill-catalog / AGENTS.md injected messages are filtered out during bootstrap and naturally return after promotion.

First-turn quick-action guide — the first wave of thinking is the only wave under the bootstrap budget, and a complex task's reasoning can blow past 1024 and end the first step empty. A fixed near-field user message (the strongest measured guidance position) is injected once on the first turn — hard-forbidding first-turn design reasoning — and stops automatically after promotion:

First turn: call exactly one tool NOW (read a file or run a command). Do NOT design, plan, or analyze in this turn — no architecture, no implementation details. All design and planning happens in later steps, after the tool result.

Promotion — after the first durable tool/call or assistant/message (promoteOn: either), the full Standard catalog, the normal output budget, and the workspace context (AGENTS.md / skill catalog) all return. The system prompt never changes again (at most one prefix-cache miss).

No tool-wrapping abstractions (the official code preset's run_code layer measured worse: 92 vs 91-99, plus a new failure surface).

Install

# 1. copy the preset directory into your user preset root
Copy-Item -Recurse <repo>\preset $HOME\.dsh\.agent-presets\bestthink-standard

# 2. restart DeepSeek Harness (presets are scanned at startup)
# 3. start a NEW session and pick 「最佳思维链标准模式」 (the preset.yml name)

The install directory name (bestthink-standard) is the preset id.

Usage

• First turn: minimal best-thinking trajectory (1-sentence system, shell + read, 1024 budget, no injected context, one quick-action guide).
• After promotion: full Standard 25-tool catalog, normal budget, AGENTS.md / skill catalog restored.
• No special prompts needed — the minimal sentence is injected by the preset, not something you type.

Configuration

Key	Default	Meaning
shellTools	[bash, pwsh]	candidate platform shells
commonTools	[read]	first-turn core tool
promoteOn	either	tool-call / assistant-message / either
bootstrapMaxTokens	1024	first-request output budget (explicitly stripped after promotion)
firstTurnGuideText	built-in	first-turn quick-action guide ('' disables; a custom string replaces the default)
suppressedContextSources	[skill-catalog, agent-instructions]	injected message kinds stripped during bootstrap ([] disables)

Verification

1. Unit tests: node --test (Node ≥ 18) — promotion modes, maxTokens cap/release, injection strip/release, guide one-shot behavior.
2. Structure (export a session JSONL): first request/header must show maxTokens=1024, ≤2 tools (shell + read), system = exactly the minimal sentence, and one source.kind: plugin guide message; after the first tool call the header must show the full catalog with no 1024 residue.
3. Trajectory (node scripts/trace-stats.mjs <session.jsonl>): English reasoning traces should show we-dominant, let me ≈ 0, and a single visible reply per turn (the minimal-trajectory fingerprint).

Development history

• The first version included a task-router (spec/react/weak persona selection, ported from dsh-router-standard). It was removed after measurement: the classifier cannot see the first user message at assembly time (user/message events persist only after the agent/pre-step waterfall, while system-prompt/assemble runs before it), so first requests always fell into the weak persona whose long guidance text inflated first-turn reasoning past the 1024 budget — empty first steps. Minimal is the optimum: one unconditional anchor, no routing.
• The bootstrap budget was briefly made model-adaptive, then unified back to 1024: agent.options.model is snapshotted at session creation, so GUI model switching produces a stale snapshot; any per-model logic risks mismatched budgets/guides. One uniform setting is immune.

License & acknowledgements

MIT. This project ports and composes:

• xiaobright/dsh-anchored-standard (MIT) → tool-bootstrap.mjs mechanics and the minimal persona row
• yjh051108/dsh-router-standard (MIT) → measurement methodology and the near-field guidance insight (no code retained)
• official DeepSeek Harness standard preset (rc.5 / 47f9438) → base composition

See preset/NOTICE for full attribution. Project2 measurements come from xiaobright/modeltest (V4.1b, frozen). Not affiliated with DeepSeek; measurements are environment-specific and not a general benchmark.