DeepSeek Harness 0.2.2: Evolving from AI Coding Tool to Agent Orchestration Platform

Version Overview

DeepSeek Harness (DSH) released version 0.2.2-alpha.2, compared against 0.2.1-rc.4 (both published 2026-10-28). The release remains an Alpha preview; new capabilities are experimental.

Version note: This article uses 0.2.1-rc.4 and 0.2.2-alpha.2 as baselines. Both versions were released on October 28, 2026. 0.2.2-alpha.2 is still an Alpha preview; some new features are experimental.

Agent Task Delegation & Git Development Workflow

External CLI Delegation

The update adds an experimental subagent_oar tool plugin that lets DSH delegate independent tasks to local Codex CLI or Claude CLI.

In 0.2.1, DSH already supported internal sub-agents and Agent Teams for distributing tasks among different agents. 0.2.2-alpha.2 extends delegation scope to external AI coding tools.

Problems Solved

Previously, developers using multiple AI coding tools had to manually switch terminals, copy context, assign tasks, and merge results. For example:

Use one agent to analyze requirements.

Use Codex CLI to write code.

Use Claude CLI to review implementation.

Manually consolidate changes.

With external CLI delegation, such workflows can be managed uniformly within DSH, reducing manual tool-switching.

The official envisioned scenario: the main DSH agent breaks down requirements, delegates code implementation to Codex CLI, delegates independent code review to Claude CLI, then aggregates results and decides next steps.

However, subagent_oar is still experimental; stability, context sharing, permission management, and complex collaboration need further validation.

The author notes a cost consideration: typically developers use Codex or Claude for planning and cheaper models (like DeepSeek) for sub-agent coding tasks. The new delegation reverses this pattern, so actual adoption may be limited.

Nevertheless, the feature provides a new extensibility direction: a unified task entry coordinating different AI coding tools for relatively independent development tasks. If task state management and result integration improve, DSH could orchestrate more complex multi-agent workflows.

Git Worktree & PR Integration Enhancements

Git Worktree is well-suited for multi-agent collaboration, allowing multiple isolated working directories in the same repository to reduce conflicts.

0.2.1 introduced an experimental Git Worktrees plugin for agents to create and enter independent checkouts. 0.2.2-alpha.2 adds an experimental "Change Request & Worktree" feature.

A bug fix addresses an issue where relative file links in chat still pointed to the original project directory after an agent entered a Worktree.

As AI agents handle more complex tasks, managing isolation, diffs, and merge processes becomes essential. The ideal flow: Task assignment → Isolated work directory → Code implementation → Diff check → PR review → Merge .

Git Worktree and PR capabilities provide a foundation for this flow. When multiple agents develop simultaneously, directory isolation lowers overwrite risk, while diffs and PR status help track progress.

If combined with external agent delegation, DSH could support a more complete parallel AI development workflow, with agents executing in isolated directories and results managed via a unified Git workflow.

These features remain experimental; they are not yet a mature automated PR management system, so code changes and merges still require manual verification.

Developer Statistics, SDK & PTC Upgrades

Developer Statistics Plugin

A new "Developer Tools - Statistics" plugin analyzes token consumption. It breaks down context and cache info, model calls, and tool call latency per conversation turn and work step, enabling performance troubleshooting for long-running agent tasks.

Cache Optimizations

The update fixes a cache invalidation issue when system prompts haven't changed, and optimizes history cache reuse for consecutive tool calls in Auto-approval mode. These improvements reduce unnecessary context retransmission in long, multi-turn agent tasks, potentially lowering token usage and improving efficiency. No official token-saving percentages or performance gains are provided yet.

SDK & Session API Enhancements

SDK interfaces for secondary developers are expanded: enhanced file delivery, Q&A, scheduling, skills, and context capabilities; added Q&A callbacks and operation approval support.

Session API now allows independent reading of session snapshots, subscribing to event logs, and querying run status. These provide finer-grained session management and runtime state access for embedding DSH or building custom agent applications.

PTC (Programmatic Tool Calling) Enhancements

Node PTC gains a VM Script mode supporting top-level await, dynamic imports, expression results, and optional cross-call state retention. This enables more flexible execution scenarios, especially for tasks combining multiple operations or preserving partial state.

The core pi-ai dependency upgrades to 1.3.0, with updated model catalog, endpoints, and inference replay rules.

Other UX Improvements

Several daily-use enhancements:

Chat history & context reference: Chat content can be inserted as references into the input box; full history sessions can be copied as Markdown.

Message sending reliability: Messages now display and clear the input draft only after host acknowledgment. On send failure, original text, references, and attachments are preserved.

Image & file handling: Images start uploading immediately when added to the input box, no longer waiting for message send. Single-image size limit raised to 50 MiB.

Background task interaction: Sending a new message while waiting for a background task immediately cancels the wait; the background task continues. Completion notifications have reduced duplicate information.

Upgrade Compatibility Notes

Minimum Node.js version raised to 24.9 . Check compatibility before upgrading.

Old tools.mode and DSH_TOOLS_MODE configurations must migrate to the new mechanism.

SSH protocol upgraded from v2 to v3. Self-deployed PTC bootstrap or SSH helper components must be updated in sync.

Session projection interfaces, agent model routing, and some tool plugin interfaces changed. Plugins should gradually use agent.model for model selection instead of the old routing method.

If using many custom plugins, PTC, or SDK integrations, validate Node.js version, legacy configs, protocols, and APIs in a test environment before production rollout.

Summary

0.2.1 already provided plugin management, sub-agents, Agent Teams, model integration, and basic Git Worktree. 0.2.2-alpha.2 adds external CLI delegation, developer statistics, Session API, and Git workflow enhancements. Though spread across modules, these all relate to complex agent task execution and management.

Chaining these capabilities enables a more clearly divided development process: main agent handles requirement understanding and task orchestration; execution agents work in isolated environments; unified session, tool, and Git workflows manage results.

The combination of external CLI delegation and Git workflow is the most promising direction. If these mature in future versions, DSH's applicability in complex AI programming projects could expand significantly.