harness-all-in-dsh
English | 中文
Reproductions of published Agent Harnesses built as native DeepSeek Harness plugins and Agent presets.
The project currently provides 13 fixed Harnesses. Original papers and official implementations are the primary references; each research note records the inspected revision and intentional differences.
Why this project
- Run every Harness through supported
dsh profiles instead of a separate launcher.
- Preserve model-visible plans, summaries, memory transitions, child results, and aggregation results in DSH Session data.
- Reuse the DSH Agent loop, subagent providers, scoped tools, and surface replacement instead of embedding another general-purpose Agent runtime.
- Compare Harnesses under the same model, task, tool, and budget settings.
- Separate implemented mechanism fidelity from future benchmark claims.
Included Harnesses
| Harness | Core mechanism | DSH implementation |
|---|
| ReAct | Interleaved reasoning, tool calls, and observations | Native Agent loop with full history |
| Plan-and-Execute | Linear plan followed by execution | Durable plan events and pre-step policy |
| ReSum | Context-pressure summarization and restart | Session surface compaction |
| Flash-Searcher | Dependency DAG and ready-node scheduling | Validated DAG projection and bounded parallel calls |
| GAM | Searchable memory pages and research integration | Durable page projection and retrieval tools |
| MemoBrain | Dependency-aware reasoning graph with flush and fold | Graph projection and surface replacement |
| AggAgent | Independent rollouts and evidence-first aggregation | Isolated child Sessions and trajectory tools |
| OAgent | Heterogeneous experts and critic selection | Child-Agent ensemble and logged critic result |
| AgentFold | Model-directed folding of trajectory ranges | Fold tool and replayable summaries |
| HiAgent | Hierarchical, value-sampled working memory | Full Session history with sampled model surface |
| DeepAgent | Three-tier memory folding and tool search | Memory tools and catalogue search |
| ROMA | Recursive atomize-plan-execute-aggregate control | Recursive child Sessions and dependency DAGs |
| AOrchestra | Runtime-selected instruction, context, tools, and model | Configurable child-Agent controller |
All 13 Harnesses are implemented at mechanism level. Benchmark comparisons require a matched evaluation matrix and are not claimed by this repository.
Use the Harnesses in DSH Web
Install this repository as a bundle in the web profile, then restart DSH Web:
dsh plugin --profile web add /absolute/path/to/harness-all-in-dsh
dsh web
Create a new Session and choose one of the modes whose name starts with Harness ·. A Session that already contains messages cannot change its Agent preset.
The presets share Shell, filesystem, search, Skill, and user-question tools. Planning, memory, folding, compaction, and child-Agent behavior come only from the selected Harness. The Web process owns the deployment-wide parallel tool-call limit, so a preset does not override that setting.
Run a Harness from the command line
Create a profile from headless, install the bundle, and apply the selected Harness overlay:
dsh --profile harness-react --from-default-profile headless --dump-config
dsh plugin --profile harness-react add /absolute/path/to/harness-all-in-dsh
dsh --profile harness-react --patch /absolute/path/to/harness-all-in-dsh/profiles/react.cordis.patch.yml "your task"
Replace react with another file name from profiles/ to select a different Harness. The headless overlays carry Harness-specific parallel-call limits for reproducible experiments.
Installing from source runs the package prepare script. Follow the DSH prompt before allowing the build script for a GitHub installation.
Documentation
- Architecture explains the DSH-native implementation rules and fidelity levels.
- Research sources links every Harness to its paper, official code, inspected revision, and implementation note.
presets/ contains the Web-selectable Agent compositions.profiles/ contains headless experiment overlays.
Development
pnpm install
pnpm run check
pnpm run check runs TypeScript checking, unit tests, and the production build. Real-model experiments require a configured DSH model provider and are separate from the keyless test suite.
License
This project is licensed under the MIT License. Upstream projects retain their own licenses; the research notes identify the license observed for each inspected revision.
