Find your next challenge.

The problem

Support deferred and selective tool loading so tool schemas are injected only on demand or when the available tool set changes. This reduces repeated prompt and caching overhead while allowing agents to access tools as needed.

Possibly addressed by:Tool Visibility · Tool Visibility · Tool Turbo

The problem

Add a persisted composer preference that lets users choose whether Enter or Cmd/Ctrl+Enter sends a message, with the other action inserting a newline. Preserve IME composition, default behavior, and Queue/Steer semantics during execution.

Possibly addressed by:Enter Send · Ctrl Enter Submit · Queue First Enter

The problem

Make sandbox-denied commands fail immediately with visible, streamed errors that explain the permission or namespace cause and guide users toward the correct escalation, retry, or writable-path configuration.

Possibly addressed by:Sandbox Escalation Guard · Strip Sandbox Permissions · Sandbox Permissions Not Strictly Wider Justification Empty Fix

The problem

Allow each model in a provider or catalog route to declare its wire protocol and use that declaration for discovery and request routing across Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses. Support adding protocols through the UI and provide a clear diagnostic when one cannot be inferred.

Possibly addressed by:Live Model Catalog · Provider Hub · Model Catalog Refresh

The problem

Provide an officially supported extension directory, lifecycle, loading mechanism, and documented stable plugin contract that remains compatible across refactors and harness versions. The API should cover service injection, route and remote method registration, prompt and tool integration, schema handling, reload behavior, public exports, and persistent third-party plugins.

Possibly addressed by:Plugin Kit · Plugin Dev Manager · Plugin Market

The problem

Expose runtime provider and model selection for subagent and subagent_fork tools, allowing users to inspect and switch running subagents independently while preserving parent-model inheritance by default.

Possibly addressed by:Subagent Model Picker · Subagent Model · Submodel Change

The problem

Add a toolChoice option that lets agents require a specified tool and propagate that requirement through adapters, wire requests, durable headers, and resumed sessions. This supports integrations that depend on stable tool invocation.

Possibly addressed by:Tool Call Limit · Toolcall Compat · Tool Manager

The problem

Enable bidirectional communication so Codex can send information directly to DeepSeek Harness, eliminating manual copy-and-paste between the two systems.

Possibly addressed by:Codex · Codex Community · Embedded Codex

The problem

Preserve complete conversation, memory, tools, prompts, workspace, and unfinished-work state when switching channels, agents, models, or workflows. Use provider-independent transcripts and centralized project configuration so sessions can move without manual copying or silent rebuilding.

Possibly addressed by:Background Agents · Agent Mode Switcher · Subagent Model Picker

The problem

Allow DeepSeek Harness to add and run locally hosted models such as Qwen, reducing reliance on remote services and accommodating lower-resource setups.

Possibly addressed by:Local Ai · Plugin Local Model · Local Llm

The problem

Add native prompt/prefix caching that saves prompt-processing slots, optionally in a RAM disk, and reuses them across forked conversations, chat windows, subagents, and parallel requests to avoid duplicate prefill work and reduce latency.

Possibly addressed by:Granular Prompt · Prompt Injector · Prompt Stash

The problem

Add an automatic approval mode that evaluates proposed agent actions using configurable policies and, where useful, a classifier. Support soft and hard denials while retaining explicit approval gates for risky, destructive, costly, or external operations.

Possibly addressed by:Auto Approve · Auto Approver · Auto Pass

The problem

Provide a standardized place to run, publish, and compare agent-harness benchmarks using consistent prompts and models so users can evaluate DSH against alternatives.

Possibly addressed by:Harness Audit · Harness All In Dsh · Harness Self Evolution

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

Support arbitrary models and providers, including local and frontier models, and allow workflows to delegate across them or switch between them without restarting the harness.

Possibly addressed by:Better Model Provider · Model Sync · Model Provider

The problem

Automatically load important instructions from persistent files and retain user preferences such as language and output style across conversations, long contexts, and sessions.

Possibly addressed by:Agent Instructions As Prompt · Session Prompt · Prompt Polish

The problem

Provide built-in SSH support for remote development and operations through native SSH workspace integration, similar to zcode.

Possibly addressed by:Remote Ssh · Ssh Workspaces · Ssh Remote

The problem

Let users configure a dedicated provider and model for child agents through a clear subagent control UI, while preserving explicit per-agent overrides and inheriting the parent model when no dedicated configuration is set.

Possibly addressed by:Subagent Model Picker · Submodel Change · Subagent Model

The problem

Add automated verification for assistant claims by checking relevant code and cited source text, requiring supporting evidence before claims enter deliverables. Flag high-risk claims for mandatory verification and route cases that cannot be automated to human review.

Possibly addressed by:Harness Audit · Verify Claims · Plugin Verify

The problem

Expose an agent’s current progress while it is running, including the number of steps taken, so the main agent can monitor long-running subagents and inspect their progress during or after execution.

Possibly addressed by:Subagent Progress · Subagent Steer · Background Agents