dsh-hermes-bridge

Native adaptive intelligence for DeepSeek Harness — no Hermes or Python dependency.

dsh-hermes-bridge reimplements Hermes's adaptive intelligence layer as a
standalone DSH plugin in JavaScript. It provides what Hermes provides on top
of the base agent loop — persistent memory, reusable skills, skill authoring,
background review, and curator — without requiring a separate Hermes installation.

What appears in DSH

DSH already provides the agent loop, tool calling, providers, subagents,
sessions, approvals, file/bash/web tools, and model routing. This plugin adds
the adaptive layer that DSH lacks natively — including the RL trajectory
collection and compression pipeline for generating fine-tuning data.

Differences from v0.1.0 (bridge)

v0.1.0 was a bridge that connected to a real Hermes Python installation.
v0.2.0 is a native reimplementation — no Hermes, no Python, no subprocess.
Memory and skills are pure JavaScript with standard fs operations.

Install

dsh plugin --profile web add github:aalvsz/dsh-hermes-bridge

Configure

Override the package row in your profile's cordis.patch.yml:

- id: dsh/hermes-bridge
  config:
    enabled: true
    namespace: hermes
    backgroundReview: false
    curator: false
    memoryCharLimit: 2200
    userCharLimit: 1375
    memoryNudgeInterval: 10
    skillNudgeInterval: 10
    saveTrajectories: false
    model: null
    trajectoryTargetMaxTokens: 15250
    trajectorySummaryTargetTokens: 750

Defaults

• backgroundReview, curator, and saveTrajectories are off until explicitly enabled.
• Memory and skill files use mode 0600; directories use 0700.
• All tools are namespaced hermes_* to avoid collisions with native DSH tools.

RL trajectory pipeline

When saveTrajectories: true, every completed conversation is converted to
ShareGPT trajectory format ({from, value} with <execute>, <result>, and
<think> XML tags) and appended to trajectory_samples.jsonl.

Use hermes_trajectory_compress to compress trajectories within a token budget:

1. Protected head turns (system, human, first gpt+tool) are preserved
2. Protected tail turns (last N turns) are preserved
3. Middle turns are accumulated until enough savings are achieved
4. Compressed turns are replaced with a single summary message
5. Boundary snapping prevents splitting gpt/tool pairs

This mirrors Hermes's trajectory_compressor.py for generating SFT/DPO-ready data.

Development

npm install
npm test
npm run verify

License

MIT.