SkillOpt-Sleep — DeepSeek Harness (dsh) integration

Give your DeepSeek Harness agent a nightly sleep cycle: it reviews past sessions offline, replays your recurring tasks on your own API budget, and consolidates what it learns into validated skills behind a held-out gate. Same engine as the Claude Code / Codex / Cursor integrations (skillopt_sleep), wired into dsh's plugin system as native tools plus a bundled skill.

DeepSeek Harness is the "everything is a plugin" agent framework (deepseek-ai/deepseek-harness). Plugins are TypeScript modules exporting an apply(ctx) function that register capabilities (tools, services, events, settings) on the Cordis context.

Why a sleep cycle?

Skills improve when an agent re-examines its own sessions — not just when you hand-tune a prompt. The sleep cycle turns your everyday usage into a nightly, validation-gated skill-improvement loop: harvest what you actually did, replay the recurring tasks on your own API budget, and keep only the changes that pass a held-out gate. It's the same engine as the Claude Code / Codex / Cursor integrations, wired into dsh as native tools plus a bundled skill.

What this integration adds

Tools

Prerequisites

• DeepSeek Harness (dsh) installed
• Python 3.10+ with the SkillOpt-Sleep engine:

pip install skillopt          # or use this source checkout

Install

As a bundle in a profile

Add dsh-skillopt to the profile's bundles, or in the profile cordis.patch.yml:

- insert:
    - id: skillopt
      name: './src/index.js'
      config:
        backend: mock          # or codex / claude / cursor / pi / opencode / handoff …
        project: /path/to/project
        preferences: 'Always use async/await'

Local patch overlay (dev)

Run from a DeepSeek Harness source checkout (the official dev workflow, pnpm resolves the workspace dsh bin):

pnpm dsh web --patch ./plugins/dsh/cordis.patch.yml

If dsh is installed globally (npm install -g), use it directly:

dsh web --patch ./plugins/dsh/cordis.patch.yml

Either way the patch inserts the skillopt plugin row into the profile; then ask the agent: "Use skillopt_status to check the sleep cycle state."

Config keys

Advanced engine keys (gate_mode, gate_metric, gate_no_regression, dream_rollouts, recall_k, evolve_memory/evolve_skill) go in ~/.skillopt-sleep/config.json — the same file shared by all integrations.

Data boundary

• Harvest is read-only; mock/handoff make no network calls.
• run stages proposals; adopt is the normal live-change boundary and backs up first.
• Real backends send truncated transcript excerpts and derived tasks to the selected provider. For sensitive sessions, export tasks first (skillopt_harvest with output=), redact, set "reviewed": true, then replay — real backends refuse unreviewed task files.
• Outbound prompts are not guaranteed secret-free; review source & provider policy.

Validate (no API spend)

python -m skillopt_sleep.experiments.run_experiment --persona researcher --assert-improves

See the SkillOpt-Sleep documentation for recorded results, limitations, and the supported integration surface.