Find your next challenge.

The problem

Allow the harness to recognize declarations from the exact @deepseek-ai/dsh-typert-protocol package in independently published external plugins, so their typed Remote methods are discovered and Remote artifacts are generated without accepting unrelated lookalike packages.

Possibly addressed by:Writing Remote · Plugin Template · Plugin Design

The problem

Provide shared ethical and safety constraints that guide agent behavior and prevent unauthorized actions such as escaping containers, accessing host credentials, or making unsafe external changes.

The problem

Enable DeepSeek Harness to connect to and run Codex and Claude Code as integrated coding agents, including Claude Code access where the current Command Goat subscription setup does not support it.

Possibly addressed by:Claude Bridge · Relay Dsh Plugin Claude · Claude To Dsh

The problem

Provide an ACP subcommand or server/RPC mode with a persistent endpoint for external tools and supervisors to discover, control, and maintain stateful DSH or coding-agent sessions.

Possibly addressed by:Acp · Acp · Acp Server

The problem

Create a polished, community-maintained Oh My DSH-style bundle that makes DSH easy to set up and use through compatible, well-integrated plugins.

Possibly addressed by:Custom Plugin · Plugin Kit · Plugin Security Review

The problem

Allow agents to load and run locally hosted or specialized models directly on the device, without depending on Vercel-hosted inference or requiring a Vercel account.

Possibly addressed by:Local Ai · Local Models · Llm Qwen Local

The problem

Persist append-only, inspectable records of every agent run, including assembled prompts, model inputs and outputs, tool activity, scheduling, context injections, and thought traces. Support searching, inspection, resumption, forking, and replay so users can examine the agent’s reasoning process.

Possibly addressed by:Agent Run Logger · Trace Collapse · Replay

The problem

Make telemetry disabled by default and collect or transmit usage data only after explicit user consent, while supporting telemetry for agent improvement.

The problem

Allow users to enter, persist, and use their own API keys for model providers, including custom or externally hosted models. Ensure DSH correctly reads manually configured keys.

Possibly addressed by:Key Manager · Providers · Model Sync

The problem

Provide crop and zoom tools that let vision agents inspect specified image regions, including across multi-turn interactions, for focused visual analysis and OCR validation.

Possibly addressed by:Tool Vision · Auto Vision · Vision Exp Tile

The problem

Use automated model-driven engineering to inspect execution traces, identify bottlenecks, and recommend or apply prompt and tool improvements. Support CDLM-based automation that pilots the underlying LLM to develop and optimize the harness.

Possibly addressed by:Llm Discovery · Llm Latency · Llm As A Verifier

The problem

Have the agent explain its intended approach before implementation, especially for speculative or over-engineered work, and invite user guidance or approval before proceeding.

Possibly addressed by:Generative Ideas · Agent Selector

The problem

Associate commits with the agent session that produced them, including the relevant conversation, reasoning, and provenance. This lets humans and future agents review and verify how the changes were made.

Possibly addressed by:Agent Messaging · Agent Message · Agentmemory

The problem

Automatically retrieve and maintain relevant codebase context, including Git history, team-maintained skills and rules, and a navigable record of merge changelogs and major logic changes. Clean up stale context with human review so agents receive reliable, codebase-specific background proactively.

Possibly addressed by:Repo Context · Context Engineering · History Rewind

The problem

Provide stable per-agent isolation for workspaces, session state, credentials, and project directories, using routing keys based on channel and requester. Include scoped permissions and per-thread evidence trails to prevent cross-agent contamination, with routing that remains reliable before symlink configuration.

Possibly addressed by:Session Workspace · Background Agents · Agent Bus

The problem

Provide a small, fast, low-memory coding agent distributed as a natively compiled binary with no JavaScript runtime or other heavy runtime dependencies.

Possibly addressed by:Agent Jit · Managed Agents · Agentmemory