dsh-ai-enhancements
Two plugins for DeepSeek Harness (dsh):
- dsh-self-improving — cross-session learning layer: the agent learns from historical experience.
- dsh-rule-enforcement — user rule injection: edit a markdown file to control agent behavior.
The two plugins can be used separately or together.
中文: README.zh-CN.md
Prerequisites
- Node.js >= 22 (check with
node -v)
- dsh installed and configured with an LLM provider
dsh-self-improving — Cross-Session Learning
After each task, the agent scores the turn, stores it in an experience database, and injects relevant experiences into the next task.
Install into dsh
Option 1: Tarball install (recommended for production)
Suitable for normal use. After installation the plugin appears in WebUI → Settings → Plugins.
# 1. Clone and build
git clone https://github.com/aminuoshi-378/dsh-self-improving.git
cd dsh-self-improving
pnpm install
pnpm run build
# 2. Pack into a tarball
pnpm pack # generates dsh-self-improving-0.1.0.tgz
# 3. Install into dsh
cd /path/to/deepseek-harness
dsh plugin --profile web add /absolute/path/to/dsh-self-improving-0.1.0.tgz
After the repo is updated you must re-run pnpm run build && pnpm pack and reinstall the tarball.
Option 2: Link install (recommended for development)
Suitable for development. After changing code you only need to rebuild and reinstall.
# 1. Clone and build
git clone https://github.com/aminuoshi-378/dsh-self-improving.git
cd dsh-self-improving
pnpm install
pnpm run build
# 2. Edit ~/.dsh/profiles/web/package.json and switch the dependency from tarball to link:
# Find the "dsh-self-improving" line and change it to:
# "dsh-self-improving": "link:/absolute/path/to/dsh-self-improving"
# 3. Reinstall dependencies (allow lockfile update)
cd ~/.dsh/profiles/web
CI=true pnpm install --no-frozen-lockfile
Link-installed plugins also appear in the WebUI plugin list.
Development workflow: change code → pnpm run build → cd ~/.dsh/profiles/web && CI=true pnpm install --no-frozen-lockfile → restart dsh web
Verify
cd /path/to/deepseek-harness
# Run the first task (accumulate experience)
dsh --profile web "create a file called hello.js"
stderr output:
[self-improving] plugin loaded
[self-improving] tool/result — write OK
[self-improving] turn 1 scored — score=0.78 | goal=advanced tools=1 successRate=1.00 steps=1 efficiency=1.00 difficulty=low
Run a second task and you will see experience injection (P0: once per turn):
[self-improving] agent/pre-step — turn=2 step=1 (injecting)
[self-improving] injecting 1 past experiences into pre-step (best score 0.78)
Configuration
The default configuration works out of the box. To override, edit ~/.dsh/profiles/web/cordis.patch.yml:
- id: self-improving
config:
dbPath: '~/.dsh/experiences.db' # Path to the experience DB; must be a file path for cross-session persistence
metaCognitionEnabled: true # Enable reflection / lesson extraction
behaviorAdapterEnabled: true # Enable experience injection
minInjectionScore: 0.3 # Only inject experiences with score >= 0.3
# Experience injection budget
maxInjectionChars: 8000 # Max characters of lesson text injected per turn
# Meta-cognition queue
maxPendingReflections: 100 # Max pending reflections before dropping oldest
# Experience store generational GC limits
youngGenMax: 200 # Max records in young generation
oldGenMax: 800 # Max records in old generation
lessonMergeThreshold: 20 # Merge similar lessons when unmerged count reaches this
experienceTtlDays: 30 # Old-gen experiences not reused in N days are downgraded
forgetScoreThreshold: 0.3 # Active-forgetting score threshold
forgetConfidenceThreshold: 0.2 # Active-forgetting confidence threshold
# Phase 6 adaptive strategies (optional, disabled by default)
adaptiveModelEnabled: false # Switch model based on historical task-type success rate
strongModel: '' # Strong reasoning model id (e.g. deepseek-reasoner); disabled if empty
standardModel: '' # Standard model id (e.g. deepseek-chat)
adaptiveToolGuardEnabled: false # Deny tools that repeatedly appear in failed experiences
failedToolDenyThreshold: 3 # Deny a tool that appears in N failed experiences
Built-in behaviors (no configuration required)
- Experience deduplication by tool sequence.
- One injection per turn.
- Step-efficiency and task-difficulty scoring.
- Implicit negative feedback detection (abort / correction / task restatement).
- Automatic lesson merging and generational GC.
- Task-type classification for better matching.
- Optional adaptive model switching and tool guard (see Phase 6 config above).
For implementation details, see docs/design.md.
Install the WebUI experience panel (optional)
cd dsh-self-improving/gui
pnpm install
pnpm run build
pnpm run bundle
pnpm pack # generates dsh-self-improving-gui-0.1.0.tgz
dsh plugin --profile web add /absolute/path/to/dsh-self-improving-gui-0.1.0.tgz
Restart dsh web, then view stats and import/export in Settings → Plugins → Experiences.
Link install: add "dsh-self-improving-gui": "link:/absolute/path/to/dsh-self-improving/gui" to ~/.dsh/profiles/web/package.json, then cd ~/.dsh/profiles/web && CI=true pnpm install --no-frozen-lockfile.
Local development and testing
cd dsh-self-improving
pnpm install # install dependencies
pnpm test # run 119 unit tests (8 test files)
pnpm run benchmark # run simulated A/B benchmark, generates benchmark-report.html
Run individual test files:
pnpm run test:store # experience store (14 tests)
pnpm run test:evaluator # outcome evaluator (8 tests)
pnpm run test:adapter # behavior adapter (11 tests)
pnpm run test:meta # meta-cognition engine (11 tests)
pnpm run test:memory # memory benchmark (15 tests)
pnpm run test:advanced # advanced features A1-B2 + T4 (30 tests)
pnpm run test:adaptive # adaptive strategies Phase 6 (11 tests)
pnpm run test:event # event parsing compatibility (9 tests)
dsh-rule-enforcement — User Rule Injection
Edit ~/.dsh/rules.md to add rules; they are automatically injected into the agent system prompt. Changes take effect immediately without restart.
Install into dsh
Option 1: Tarball install (recommended for production)
# 1. Clone and build
git clone https://github.com/aminuoshi-378/dsh-self-improving.git
cd dsh-self-improving/dsh-rule-enforcement
pnpm install
pnpm run build
pnpm pack # generates dsh-rule-enforcement-0.1.4.tgz
# 2. Install into dsh
cd /path/to/deepseek-harness
dsh plugin --profile web add /absolute/path/to/dsh-rule-enforcement-0.1.4.tgz
Option 2: Link install (recommended for development)
# 1. Clone and build
git clone https://github.com/aminuoshi-378/dsh-self-improving.git
cd dsh-self-improving/dsh-rule-enforcement
pnpm install
pnpm run build
# 2. Edit ~/.dsh/profiles/web/package.json and change to:
# "dsh-rule-enforcement": "link:/absolute/path/to/dsh-self-improving/dsh-rule-enforcement"
# 3. Reinstall dependencies (allow lockfile update)
cd ~/.dsh/profiles/web
CI=true pnpm install --no-frozen-lockfile
Install the WebUI rule editor (optional)
cd dsh-rule-enforcement/src/gui
pnpm install
pnpm run build
pnpm run bundle
pnpm pack # generates dsh-rule-enforcement-gui-0.1.3.tgz
dsh plugin --profile web add /absolute/path/to/dsh-rule-enforcement-gui-0.1.3.tgz
Restart dsh web, then edit rules in Settings → Plugins → Rules.
Edit rules
Edit the file ~/.dsh/rules.md directly:
# Project rules
- Reply in Chinese when the user asks in Chinese
- Update CHANGELOG before committing
- New files should use TypeScript
- Run security checks before git push
Or edit via the WebUI. Save and the rules take effect on the next agent request.
Local development and testing
cd dsh-rule-enforcement
pnpm install
pnpm run typecheck # type check
pnpm test # run 10 tests
Install all together
Tarball
cd /path/to/deepseek-harness
dsh plugin --profile web add /absolute/path/to/dsh-self-improving-0.1.0.tgz
dsh plugin --profile web add /absolute/path/to/dsh-self-improving-gui-0.1.0.tgz
dsh plugin --profile web add /absolute/path/to/dsh-rule-enforcement-0.1.4.tgz
dsh plugin --profile web add /absolute/path/to/dsh-rule-enforcement-gui-0.1.3.tgz
dsh --profile web
Link
Edit ~/.dsh/profiles/web/package.json:
{
"dependencies": {
"dsh-self-improving": "link:/path/to/dsh-self-improving",
"dsh-self-improving-gui": "link:/path/to/dsh-self-improving/gui",
"dsh-rule-enforcement": "link:/path/to/dsh-self-improving/dsh-rule-enforcement",
"dsh-rule-enforcement-gui": "link:/path/to/dsh-self-improving/dsh-rule-enforcement/src/gui"
}
}
cd ~/.dsh/profiles/web
CI=true pnpm install --no-frozen-lockfile
dsh --profile web
FAQ
better-sqlite3 error during install?
Add allowBuilds: { better-sqlite3: true } to ~/.dsh/profiles/web/pnpm-workspace.yaml, then reinstall.
Plugin not showing in WebUI plugin list?
Make sure the plugin is declared in ~/.dsh/profiles/web/package.json dependencies (link or tarball), registered in dsh.profile.bundles, and restart dsh web.
dbPath says directory does not exist?
The plugin automatically expands ~ and creates directories. If it still fails, run mkdir -p ~/.dsh manually.
Tarball install shows type stripping error?
You must run pnpm run build to compile .ts → .js before packing, and make sure the dist/ directory exists.
duplicate loader entry id?
The same plugin is being loaded twice. Make sure there is no plugin with the same name in the dsh repo packages/ directory, or its package.json has no dsh field.
Link install changes not taking effect?
After changing code, rebuild: pnpm run build, then reinstall in the profile directory: cd ~/.dsh/profiles/web && CI=true pnpm install --no-frozen-lockfile.
More documentation
- Architecture design — four-layer architecture, P0-P5 implementation details, safety boundaries, roadmap
- TODO — P0-P5 completion status and future plans
- Changelog
