HydraDB DeepSeek Harness Plugin
Persistent, cross-session memory for the DeepSeek Harness
(dsh), powered by HydraDB.
It recalls relevant context before each turn, syncs your workspace docs into
HydraDB, and captures conversations as durable memory, over the same shared
engine (scripts/plugin.mjs) as the HydraDB Claude Code / Codex / OpenClaw plugins.
What's dsh-specific here
- Entry module:
index.js, an ordinary Cordis
plugin (export const name + export function apply(ctx)). dsh mounts it by
package name and calls apply with the plugin ctx; it registers listeners
on the harness interception seams over the shared engine.
- Env prefix: per the shared client contract, this client reads the canonical
HYDRADB_* names plus the HYDRA_DEEPSEEK_* prefix (see CONTRACT.md).
- Skills: the shared HydraDB skills under
skills/ for manual recall/ingest.
Seams
| dsh extension point | Engine action |
|---|
agent/created | sync workspace docs into HydraDB when a session opens |
agent/pre-step (turn's first step) | recall relevant memory/knowledge and prepend it to the step messages |
tools/post-execute (file-writing tools) | incremental workspace sync |
session/event → turn/end | capture the completed turn |
Prompt-time auto-recall (agent/pre-step) and workspace sync work out of the box.
The agent/pre-step waterfall is cooperative: the plugin awaits next() and only
prepends a user-role context message to the downstream enter decision, so it
composes with other pre-step listeners instead of overriding them.
Prerequisites
- Node.js >= 18
- A HydraDB account: API key + database (hydradb.com)
- DeepSeek Harness (
dsh)
Install
This package ships as a dsh bundle (dsh.bundle + cordis.patch.yml), so
dsh plugin add installs and wires it as a profile layer in one step:
dsh plugin --profile <name> add dsh-hydradb
dsh appends it to the profile's dsh.profile.bundles and inserts the hydradb
plugin row automatically - no manual cordis.patch.yml editing. To pass config
inline instead of via the environment, add a config: block to the bundle's
cordis.patch.yml (see dsh.config.example.yml for the --patch overlay form).
Credentials come from the plugin config (api_key, database, optional
collection/base_url) or the environment:
export HYDRADB_API_KEY="your-api-key"
export HYDRADB_DATABASE="your-database"
# dsh-prefixed names also work: HYDRA_DEEPSEEK_API_KEY / HYDRA_DEEPSEEK_TENANT_ID
The default search mode is memory; set HYDRADB_SEARCH_MODE=both if you also
want synced workspace knowledge to auto-recall.
Usage
- Automatic recall runs before each turn's first step via
agent/pre-step
and prepends relevant memory/knowledge to the step messages.
- The shared skills provide manual
query / ingest / doctor / setup /
last-recall actions; the contract's canonical slash form is /hydradb-query,
/hydradb-ingest, /hydradb-list, /hydradb-inspect, /hydradb-delete.
Configuration
Config keys, environment overrides, and capture/search/ingest modes are identical
to the other HydraDB plugins - see config.example.json and CONTRACT.md.
Verified
Tested against DeepSeek Harness (dsh) 0.2.0-rc.2: the bundle installs via
dsh plugin add, registers as a profile layer, and its apply seams activate
cleanly. A live dsh headless turn (OpenRouter provider) answered a question
from HydraDB memory recalled and injected through agent/pre-step.
License
Apache 2.0 - Copyright (c) 2026 HydraDB
