DeepSeek V4 Flash has strong reasoning capabilities under the reasoning: max setting, but it also exhibits instability: it tends to overthink simple tasks, underthink complex tasks, and drift during a session. The dsh-flash-godmode plugin aims to convert this unstable thinking process of the Flash model into reproducible capability through a routing mechanism.

Plugin Overview

This plugin is a workflow plugin in the DeepSeek Harness (DSH) ecosystem, intended to provide a “god mode” configuration for the V4 Flash model. Its core approach is to use specific personality anchoring and guidance strategies to address inefficiency on simple tasks and decision ambiguity on complex tasks. The plugin is maintained by Cavan-Ou and is released under the MIT license.

Core Features

The plugin mainly includes the following three core mechanisms:

  1. w7 personality anchoring
    This is a measured optimal personality configuration for the Flash model. It consists of three elements:

    • Neutral identity: avoids causing the model to carry overly strong preset positions.
    • Classify-act instructions: forces the model to determine the task type before acting.
    • Review/anti-runaway anchor: uses specific instructions to prevent the model from wasting turns on environment checks (such as echo, whoami) or divergent thinking.

  2. Complexity-based guidance dispatch
    The plugin decides the guidance strategy for the first-turn request based on task complexity:

    • Simple tasks: receive fast-convergence guidance and directly obtain results.
    • Complex tasks: receive deep guidance for decision closure and perform in-depth analysis.
    • Trigger conditions: text length exceeds 120, or architecture keywords are included (an author-authored heuristic). For example, a filename such as analyze.py would be treated as a complex task.

  3. First-turn tool anchoring
    On the first request, the model can only see a core tool set (read/write/edit + Shell). The full tool catalog is exposed only after one persistent tool call has been completed. This helps the model focus on the current task.

  4. agent/inbox/spliced capture path
    This is the key technical path for implementing “first-turn guidance.” In DSH’s headless configuration (rc.6), the standard user/message event often arrives only after the initial prompt assembly, making it impossible to influence the first request. This plugin listens to the agent/inbox/spliced event, which carries the task text before assembly, ensuring that guidance decisions can be applied to the first request.

Installation and Dependencies

The plugin is a headless-native single-plugin form, requires no additional installation script, and can be installed directly through the DSH command line.

Prerequisites:
* DSH version 0.1.x (using the headless profile is required).
* Any provider compatible with the openai-completions protocol (verified in the opencode-go channel, using deepseek-v4-flash and reasoning: max).

Installation command:

dsh plugin --profile headless add github:Cavan-Ou/dsh-flash-godmode

Typical Usage

After installation, the plugin automatically detects the Flash model and applies the configuration.

When running any task, the stderr output includes scheduling decision information for verifying whether the plugin is active:

dsh --profile headless "简单任务：运行 python3 -c 'print(1+1)' 并报告输出" 2>&1 | grep godmode

Example expected output:

[godmode] taskLen=NN complex=false guide=fast mode=weak

In this mode, simple tasks are routed to mode=weak together with guide=fast, causing the model to converge quickly; complex tasks, in contrast, trigger guide=deep.

Notes

  • The complexity heuristic is author-defined: the complexity classification logic in the plugin (text length > 120 or architecture keywords) is defined by the author and is not a built-in DSH standard. If a filename contains an architecture keyword (such as analyze.py), it will be treated as a complex task.
  • Fallback handling: if a future DSH version makes the agent/inbox/spliced capture path stop working, the plugin automatically falls back to using only the raw w7 personality (without the guidance strategy). Tasks can still run in that case, but the complexity dispatch feature is lost.
  • Verification environment: the measurement data referenced in the plugin (such as the 0 tool failure rate) comes from the official DeepSeek API; this plugin was verified on the opencode-go channel (which uses the same weights but has a different service architecture).

Summary

By integrating w7 personality, tool anchoring, and complexity-based guidance into a single plugin, dsh-flash-godmode addresses stability issues with the V4 Flash model in DSH headless scenarios. It is especially suitable for developers who need to call the Flash model reliably in production environments or automated scripts.

Project address: GitHub