科研、数据与金融开源项目

Research Mode

通用研究 Agent 预设

以 Agent Preset 形式分发的轻量 DSH 原生研究模式。

项目 README(英文原文)

在 GitHub 中查看

DSH CDC Minimal Research Preset

An experimental DeepSeek Harness (DSH) agent preset for mathematical research. The goal of this project is to explore the boundaries of what DSH can do for long-horizon mathematical work: a small, deliberately minimal tool surface, an aggressive subagent-driven search protocol, and strict return discipline inspired by the OpenAI CDC prompt.

This is a user-authored preset, not an official DeepSeek product.

Design

Root agent: research-loop owner

The root agent thinks actively at the high level — possible routes, strategy, and priorities — while delegating tedious calculation, verification, coding, and numerical experiments to subagents. It maintains the research loop and enforces the rules:

  • Concrete agent scheduling follows the OpenAI CDC prompt: a diverse dynamic portfolio of subagents, an approach-family registry, adversarial audit of every candidate, and literature-grounded research before new ideas are invented from scratch.
  • The root repeatedly re-reads the user's prompt, tracks the high-level progress map, proposes batches of intermediate conjectures, and delegates detailed work. It normally consumes subagent summaries but may inspect any proof, code, data, or note directly whenever that helps, or ask a subagent for a focused explanation.
  • Numerical experiments are evidence, not truth. A computational result may redirect the search only after independent checks or reproduction; a buggy subagent script must never steer the portfolio.
  • Conjecture screening is a fallback: use it when the loop runs out of genuinely distinct ideas, and treat survivors only as hints for a fresh theoretical direction.
  • Return discipline: a complete proof or fully verified counterexample that survives two independent adversarial audits is returned immediately. Otherwise the root keeps working for at least 8 hours of effective effort and may then return only the strongest rigorous derivation plus its exact remaining gap, clearly labeled as not a resolution.

Subagents

Subagents do the thinking, coding, careful numerical experiments, auditing, and note writing. Delegation depth is capped at 2.

  • subagentdeepseek-v4-pro, for proofs, reductions, adversarial audit, final verification, and everything else the root delegates.
  • Calls are continuable by default: they return a durable subagent id and the parent receives a settlement notice when a child finishes.

Tool inventory

Model-facing tools:

ToolPurpose
bashPersistent shell for housekeeping, reading, and verification
str_replace_editorView / create / edit files
web_searchLiterature grounding; cite sources; verify before use in a proof
subagentContinuable pro-tier research worker
send_messageQueue a follow-up turn for a direct child
interrupt_agentStop a stuck child or descendant's current turn
list_agentsSnapshot status of children / descendants

Continuable children also receive the host-provided report tool, so they can send a child-to-parent message outside the normal settlement path.

Non-tool machinery mounted by the preset:

  • dsh-time-context — injects current time and browser timezone readings.
  • dsh-compaction-basic — automatic and manual context compaction.
  • prompt-recheck.js — local plugin that injects a durable reminder to re-read .dsh-prompt-path after every compaction and every 20 minutes.

Installation

The preset id is the directory name, so clone it to exactly this path:

git clone <your-repository-url> ~/.dsh/.agent-presets/minimal-web-subagent

Then refresh the DSH web UI and start a new session with CDC Minimal Research Mode.

Tested with @deepseek-ai/dsh 0.1.0-rc.6.

Optional: quieter child reports

The child-side report tool is host configuration and cannot be mounted by an agent preset. To inject child reports into the parent's next request instead of waking/queuing a separate turn, copy the entry from profile-patch.example.yml into ~/.dsh/profiles/web/cordis.patch.yml, then restart dsh web.

Usage

  • Put the research prompt in a local file and give its path to the agent. The root agent records the path in the workspace file .dsh-prompt-path.
  • Intermediate scripts, data, logs, plots, and notes all go directly into the flat workspace directory .dsh-generated/ with short descriptive names.
  • Let the root agent manage the search. Do not expect it to write code or prove lemmas itself; that work belongs to subagents.

Configuration

Edit agent.cordis.yml to adjust:

  • time-context: fallback timeZone and refresh interval
  • prompt-recheck: intervalMs (default 1200000, 20 minutes) and the prompt-path registry filename
  • tool-subagent: model tier, maxDepth, worker persona, and delegation policy
  • tool-subagent-control / tool-subagent-list-agents: follow-up, interrupt, and status tools

Note: DSH detects preset changes from agent.cordis.yml. If you edit only prompt-recheck.js, touch agent.cordis.yml (or change a comment) so new sessions mount the updated plugin. Sessions that have already started keep the generation they were created with.

Repository contents

FilePurpose
agent.cordis.ymlThe preset composition
preset.ymlDisplay metadata for the mode picker
prompt-recheck.jsLocal plugin that injects durable prompt re-read reminders
profile-patch.example.ymlOptional host-profile patch for quiet child reports
package.jsonMarks the preset directory as an ES module
LICENSEMIT license

License

MIT