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
| Capability | DSH surface |
|---|---|
| Persistent memory (MEMORY.md / USER.md) | hermes_memory + frozen system-prompt snapshot |
| Reusable skills (SKILL.md catalog) | hermes_skills_list, hermes_skill_view, hermes_skill_manage |
| Skill authoring | /hermes-learn |
| Background memory/skill review | optional DSH subagent fork |
| Curator | optional due-checked integration |
| RL trajectory capture | hermes_trajectory_save + automatic turn/end capture |
| Trajectory compression | hermes_trajectory_compress (protected regions + LLM summarization) |
| Capability diagnostics | hermes_status |
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, andsaveTrajectoriesare off until explicitly enabled.- Memory and skill files use mode
0600; directories use0700. - 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 ,, and
` XML tags) and appended totrajectory_samples.jsonl`.
Use hermes_trajectory_compress to compress trajectories within a token budget:
- Protected head turns (system, human, first gpt+tool) are preserved
- Protected tail turns (last N turns) are preserved
- Middle turns are accumulated until enough savings are achieved
- Compressed turns are replaced with a single summary message
- 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.
