dsh-chat-local

中文 · English

Multi-agent collaboration in DeepSeek Harness, with persistent identities, user-defined personas, and memories grounded in each agent's own experience.

Build a team, start separate conversations for different topics, and let members hand off work, submit results, and request review. An agent can recall material it has actually seen across conversations and execution sessions. Other members' private experiences do not automatically become shared memory.

Features

Getting started

The interface currently uses Chinese labels. The steps below retain those labels so you can find the corresponding controls.

1. Install

The verified DSH version is 0.1.7-rc.2; the declared host range is >=0.1.7-rc.2 <0.1.8-0. Install dsh-bridge first; it handles waking local sessions and delivering messages.

dsh plugin --profile web add github:ALPSIYH/dsh-chat-local

Restart dsh web after installation, then select 群聊 (Group chat) in the sidebar. Before upgrading an existing installation, back up the complete data directory as described in Operations.

2. Configure agents

Open 群聊 → Agent to create or edit members. Names, roles, and models belong to their work configuration. The 人格 (Persona) editor describes preferences that persist across tasks.

The initial persona is an inactive template awaiting configuration. For example:

## Working habits
Start with one testable approach, state the key assumptions, then implement it.
Revise conclusions when new evidence arrives and explain which premise changed.

## Communication preferences
Lead with the conclusion. Distinguish confirmed facts, inferences, and open questions.

Saved settings apply to subsequent turns. Personas are not rewritten automatically and do not grant additional execution permissions.

3. Create a group and start a conversation

Create a group, choose its members, and start a conversation for a topic. Use the team defaults or adjust the participants, working directory, and permissions for that conversation.

New conversations without a selected strategy default to work routing; reused group templates keep their saved strategy. Work routing sends ordinary requests to the responsible member, falling back to a coordinator before work is assigned. Choose a message purpose, recipient, and work item in the composer. Names mentioned in prose do not dispatch work. The overview separates pending requests, execution budgets, and acceptance. Existing conversations retain legacy routing until explicitly switched while idle. Saving member or team settings alone does not start a model call.

See the User guide for the full workflow.

How memory follows an agent

Persistent agent identity
├── Persona: user-maintained PERSONA.md
├── Experiences: material actually received, read, or authored
├── Beliefs: revisable interpretations with scope, validity, and counter-evidence
├── Methods: sourced learning candidates with explicit adoption
└── Appraisals: the agent's sourced judgments about other agents

Group deliveries receive a length-limited memory summary. Agents can use chat_recall to retrieve more. Automatic context in a native session also depends on its DSH preset: minimal suppresses runtime contexts, so persona and personal memory are not injected there. The participant panel shows the capability observed during the last prompt assembly; open the native session and choose a preset that enables runtime context if needed. Room-level counts of messages, deliveries, and reviews are maintained separately to help inspect collaboration.

Automatic consolidation currently means deduplication that preserves source text and provenance. Forgetting affects retrieval through decay or reversible suppression; the original logs remain. Model-generated semantic consolidation and automatic personality evolution are not implemented, and stable personality in real models has not been demonstrated. See Memory lifecycle.

Before you use it

• Permissions affect real sessions. Workspace-write and full-access settings may remain active after leaving group chat. Members sharing a working directory also operate on the same files. See Permissions and shared sessions.
• Data is stored locally. The default directory is ~/.dsh/dsh-chat-local/, and the current state format is v18. Rolling back to an incompatible older version also requires restoring data compatible with that version.
• Memory has visibility and capacity limits. Unreceived material, native history from before the plugin was mounted, and non-text content do not automatically become memories. Capacity limits can prevent new writes; unlimited history is not guaranteed.
• The local management API does not authenticate callers. Private-memory isolation in agent tools does not protect against programs with access to local files or the management API. Logs and persona files are not encrypted.
• Traceability does not establish truth. Logs support provenance and integrity checks, but models can still reason incorrectly. Passing tests or producing consistent short-term replies does not demonstrate a complete simulation of a person.

Documentation

Chinese remains the default documentation language. The detailed guides linked below are currently in Chinese.

Development

npm ci
npm run check

Run npm run build after changing the shared text or work protocols. check validates generated content, syntax, and regression tests. To test native DSH integration, set DSH_MODULES_DIR to the node_modules directory inside the DSH installation. See the Maintainer guide for details.

Use scripts/verify-event-log.mjs for offline log verification. Follow Operations and work on a data copy taken after writes have stopped.

MIT License