Introduction¶

In DSH or agent development, if the system prompt prefix changes for every request due to variable content such as the working directory, the provider-side prompt cache based on prefix matching may not be easily reused. dsh-cache-stabilizer is an MIT-licensed plugin for DeepSeek Harness, maintained by dongsheng123132. It performs two main tasks: first, relocating the working directory from known default persona sentences in DSH to the runtime-context snapshot; and second, normalizing the order of object-keys within tool schemas. The plugin also provides a /cache command to report cache read tokens, uncached input tokens, and cache write tokens returned by the provider.

What is this¶

dsh-cache-stabilizer is designed for cache-prefix stabilization and evidence-based cache metrics. It does not implement a second cache layer, does not proxy model responses, does not freeze the tool catalog, and does not reuse stale context. It only performs semantic-preserving prefix stabilization during the current DSH request construction phase and displays cache evidence through provider metrics.

This plugin requires Node.js >=22.

Core Features¶

Relocating cwd from Default Persona Sentences¶

The plugin relocates the working directory from known exact sentences in the DSH standard/headless coding persona out of the system prompt prefix and places it into the runtime-context snapshot. This allows different projects to share the same system-prompt prefix, while each request still receives the correct cwd.

There is a clear boundary here: only exact sentences used by the DSH standard or headless coding persona will be migrated. If a custom persona mentions {{cwd}} in another form, the plugin will not modify it, as blindly moving arbitrary natural language could alter its meaning.

Normalizing Tool Schema Key Order¶

The plugin normalizes the order of object-keys inside tool schemas. This change serves prefix stability but does not change the tool set itself.

Adding the /cache Command¶

The plugin adds the /cache command. This is a human-only command used to report provider’s persistent metrics:

• cacheReadTokens
• inputTokens (uncached)
• cacheWriteTokens

/cache does not fabricate a cache hit out of thin air. DeepSeek provider caching is automatic and relies on precise prefix matching starting from token zero; cache storage and eviction remain controlled by the provider.

Installation and Usage¶

Install under the web profile:

dsh plugin --profile web add dsh-cache-stabilizer

If using a custom profile, replace web with the corresponding profile name.

Typical Usage¶

1. Install the plugin:

dsh plugin --profile web add dsh-cache-stabilizer

2. Restart DSH, send a few messages, then enter in a command-supporting client:

/cache

If you need to disable an optimization in a specific profile, use a profile patch configuration:

- id: dsh-cache-stabilizer
  config:
    relocateCwd: false
    canonicalizeTools: false

Applicable Scenarios and Notes¶

Suitable for developers using the DSH standard/headless coding persona, who wish to reduce the impact of variable content on system prompt prefixes, and want to view provider cache evidence.

Note the following before use:

• The plugin runs with the current DSH process permissions; check the source code and license before installing.
• This plugin does not create a new cache layer, nor does it change the provider’s cache storage and eviction strategy.
• If the custom persona’s {{cwd}} is not in the exact form of known default sentences in DSH, the plugin will not automatically migrate it.
• /cache only reports metrics given by the provider and does not infer or generate non-existent cache hits.

Development and Verification¶

The repository provides development commands:

npm test

npm run check

Links¶

GitHub:

https://github.com/dongsheng123132/dsh-cache-stabilizer

Community directory page (provided by the community, not confirmed by README/package.json):

https://www.skillhub.cn/plugins/dongsheng123132/dsh-cache-stabilizer

The community directory is an independent site and is not an official app store for DeepSeek or Huanfang.