dsh-sisyphus
DiscoveredSisyphus orchestration agent preset for DeepSeek Harness: intent gate, decompose-and-delegate to specialist subagents, parallel execution, evidence loop.
dsh-sisyphus
English | 中文
Sisyphus orchestration agent preset for DeepSeek Harness (dsh).
An agent running this preset orchestrates instead of editing directly: it classifies every request first (intent gate), decomposes work into atomic units, delegates each unit to a specialist subagent, runs independent delegations in parallel, and verifies results before calling them done.
Inspiration
The Sisyphus working style — intent-gated routing, decompose-and-delegate to specialists, parallel execution, and the evidence loop — comes from OhMyOpenCode's (ohmy openagent) Sisyphus agent as used on top of opencode. This preset migrates that orchestration style to DeepSeek Harness, reusing dsh's own subagent capability seam instead of changing its main loop.
What it gives you
| Specialist subagent | Domain | Recursion |
|---|---|---|
explore | codebase search — where is X, how does Y work | leaf (maxDepth: 1) |
librarian | external docs, library best practices, OSS examples | leaf |
oracle | hard reasoning, architecture decisions, debugging | leaf |
plan_reviewer | critiques work plans for gaps and ambiguity | leaf |
implementer | concrete edits following existing patterns | may delegate (3) |
subagent | general delegated work | 3 |
subagent_fork | delegated work with the parent's completed history | — |
All specialists run in continuable background mode: the model starts independent delegations together and keeps working while they run; each returns a durable child id the model can continue via send_message, with list_agents for status. The leaf specialists accept exactly one call each and cannot delegate further.
The persona also encodes the Sisyphus working discipline: intent gate, decompose-and-delegate, parallel execution, session continuity, an evidence loop ("no evidence = not done"), and a six-part delegation prompt (TASK / EXPECTED OUTCOME / REQUIRED TOOLS / MUST DO / MUST NOT DO / CONTEXT).
Install
This repo is the preset itself. Installing copies the preset directory into the dsh agent-presets root.
One command
git clone https://github.com/wyhgoodjob/dsh-sisyphus.git
cd dsh-sisyphus
./install.sh
git clone https://github.com/wyhgoodjob/dsh-sisyphus.git
cd dsh-sisyphus
.\install.ps1
Manual
$DSH_HOME defaults to ~/.dsh (Windows: $env:USERPROFILE\.dsh). Copy the preset directory itself, flat — dsh scans only one level, so the target must be .agent-presets/sisyphus, never .agent-presets/presets/sisyphus:
cp -r presets/sisyphus "$DSH_HOME/.agent-presets/sisyphus"
Then restart dsh. A new session can pick Sisyphus in the UI, or set it as the default in the user settings document:
agent-presets:
default: sisyphus
From a git dependency (alternative)
dsh plugin --profile web add github:wyhgoodjob/dsh-sisyphus
and point the roster at the installed package:
- name: '@deepseek-ai/dsh-agent-presets'
config:
default: sisyphus
roots:
- path: <profile-node_modules>/dsh-sisyphus/presets
trust: user
Requirements
- A DeepSeek Harness whose persona row uses the
prefixfield (@deepseek-ai/dsh-persona≥0.1.3-alpha.2; the current0.1.5-rc/0.1.6-alphalines). On older harnesses that still taketext, the preset fails to mount with$.prefix missing required value. - DeepSeek Harness with the
spawnandforksubagent backends in its host composition (both ship with the standard app). - The subagent providers must advertise
depthLimit,persona,toolFilter, andprepareContinuablecapability — the in-processspawn/forkbackends do.
Model experience
The preset adds no schema of its own; token and KV-cache effects come from the plugins it composes. Every continuable specialist instance adds one tool:<toolName> prompt section (parallel-delegation guidance) and one schema, paid per parent request. The persona text replaces the deployment persona.
Known limitations and deferred work
- Specialists are soft-constrained by persona, not hard tool filters. The leaf specialists (
explore,librarian,oracle,plan_reviewer) are instructed not to edit, but they still carry the shell and filesystem tools; atoolFilterallow-list per specialist is planned and omitted only to avoid failing preset load on a host whose tool names differ. - Continuable fork invalidates some inherited KV-cache reuse (upstream
dshissue #2124); thesubagent_forkinstance accepts that cost. - Model selection is not enabled on the
subagent/implementerinstances (modelSelectionSettings: false), so children inherit the parent route.
License
MIT
Comments
Loading…
Similar plugins
by YYTbit
Multi-agent orchestration for DeepSeek Harness
★ 4
MIT
TypeScript
Aug 14, 2026
dsh plugin --profile web add oh-my-deepseek-harnessby Tim-ReJet
Durable subagent execution for DeepSeek Harness: a `durable` provider and a model-facing tool that drive pi-durable-subagents, so a crashed delegation converges on its first outcome instead of running twice.
★ 0
MIT
TypeScript
Oct 14, 2026
dsh plugin --profile web add dsh-plugin-durable-subagentby euuuuuuzer
A loop dock for DeepSeek Harness: one harness, multiple agent loops.
★ 6
MIT
JavaScript
Aug 16, 2026
dsh plugin --profile web add dsh-loop-dockDeepSeek Harness plugin: on task send, pick a different model for subagents and an independent reviewer that validates each subagent's work before it reaches the main agent
★ 0
MIT
TypeScript
dsh plugin --profile web add dsh-orquestratorby Liu-Bot24
DeepSeek Harness(DSH)只读执行轨迹复盘插件,支持规则分析、独立模型解读、证据定位、任务概览与运行对比。
★ 5
MIT
JavaScript
Oct 21, 2026
dsh plugin --profile web add dsh-plugin-trace-insightby elementor-i
agentmemory for DeepSeek Harness (dsh): full memory_* tools, capture hooks, and context injection over the local REST server
★ 4
↓ 31/wk
MIT
TypeScript
Aug 17, 2026
dsh plugin --profile web add @elementor-i/dsh-agentmemory