In the DeepSeek Harness ecosystem, model routing through external Sleev context optimization gateways needs to be observed. The dsh-sleev plugin provides this capability, recording routing at the Host level, exposing telemetry data, and making configuration options available in the Web UI.

Plugin Purpose

This plugin is maintained by xarleyn and aims to address observability challenges when integrating Sleev gateways with DSH. The plugin has been archived and migrated to the xarleyn/dsh-plugins monorepo.

Core Features

As a Host-side observer, the plugin provides the following capabilities:

  1. Routing and prefix observation: observe exact routes and/or prefixes (the default prefix is sleev-).
  2. Call classification: classify agent, compaction, session-title, and one-shot calls.
  3. Streaming output: pass through each stream chunk unchanged, without modifying the data content.
  4. Usage statistics: record provider usage and effective input token capacity.
  5. In-memory history: maintain a bounded, secret-free in-memory history.
  6. Structured logging: record one structured completion record for each observed call.
  7. Web UI configuration: expose observer matching, retention, and logging settings in the Web UI.

Installation and Enablement

Install it using the official command:

dsh plugin --profile web add dsh-sleev

To remove it, run:

dsh plugin --profile web remove dsh-sleev

After installation, restart the DeepSeek Harness Host to apply the configuration.

Configuration and Usage

1. Configure Sleev Routing

Merge provider routing under llm-pi-ai.providers in the $DSH_HOME/settings.yaml file. DSH resolves credential references; do not place API keys directly in the routing configuration.

Example configuration:

llm-pi-ai:
  providers:
    sleev-neuraldeep:
      displayName: Sleev / neuraldeep
      apiKeyEnv: NEURALDEEP_API_KEY
      api: openai-completions
      baseURL: http://127.0.0.1:17321/v1
      headers:
        sleev-base-url: https://api.neuraldeep.ru/v1
        sleev-harness: pi
      models:
        - id: gpt-oss-20b
          name: GPT OSS 20B via Sleev

2. Plugin Settings Card

In the Web UI, navigate to Settings → Plugins → Plugin configuration → Sleev.

The following can be adjusted:
* Exact provider aliases to observe
* Provider name prefixes to observe
* Number of recent call records retained in memory
* Structured telemetry log level (off, info, debug)

Notes and Limitations

  • Repository status: The original repository has been archived as read-only; current development is focused on the xarleyn/dsh-plugins monorepo.
  • Observer capabilities: The current M1 observer does not rewrite prompts, implement compaction, or route traffic autonomously. Routing must be configured normally in DSH’s llm-pi-ai.providers.
  • Privacy protection: The observer does not store prompts, request headers, credentials, or secret values.
  • Observation scope: Direct routes without a configured Sleev alias remain unobserved.