DeepSeek Harness Two Updates in Two Days: Code Review, Mobile UX, 92T Tokens Daily

Overview

On November 3, DeepSeek experienced a partial service outage from 16:58 to 18:48, with users reporting "server busy" errors. Simultaneously, DeepSeek Harness (DSH) shipped two versions in two days: v0.2.2-alpha.3 on November 3 and rc.1 on November 4. The updates focus on code review workflow, mobile usability, session management, long-task context handling, and plugin extensibility. Official metrics show 1.54 million daily active users and approximately 92 trillion tokens consumed in a single day (November 4).

Code Review Enhancements

The experimental "Change Requests and Worktree" feature unifies code review and Git worktree entry points. Users can browse a searchable file tree, navigate per-file diffs, and track review status across local changes and pull requests. After spotting an issue, the selected code and comment can be handed directly to the Agent for remediation. For example, if an Agent modifies multiple files and a function needs adjustment, the user can continue the conversation around that snippet without manually copying code, file paths, or re-explaining context. File preview now includes a persistent directory tree and file location indicator for easier multi-file switching.

Execution traces ( Trajectory) support independent folding of sub-calls while preserving parent calls and summaries; search or cross-view navigation expands the relevant record.

Mobile Adaptation

Alpha.3 optimizes the mobile sidebar navigation and narrow-screen layout so that model selection, plugin discovery, and other panels remain fully visible. Menus and popovers follow browser zoom, pan, and soft-keyboard changes, staying in the visible viewport. The bottom input box reduces jitter during typing, menu expansion, and statistics updates. These changes directly address: buttons hidden by the keyboard, menu options pushed off-screen, and page shifts while typing. Rc.1 continues refining phone and tablet adaptation with the same focus on soft-keyboard interaction, navigation, and narrow-screen usability.

Session Management Improvements

Chat messages can be inserted as references into the input box; sent references show a compact preview.

Full conversation history can be copied as Markdown, including raw log paths when available.

Collapsed work steps now offer two display modes: keep the assistant's full reply or show only the final answer.

Long tool calls can be folded while preserving intermediate explanations and conclusions.

Message send reliability: the UI now confirms host receipt before clearing the draft; on failure, drafts, references, and attachments are retained.

Images upload immediately upon insertion and are reused on send; default per-image limit raised to 50 MiB.

Long-Task and Context Handling

When switching to a model with a smaller context window, DSH first compresses history using the original model; if compression fails, the switch is aborted. This prevents submitting requests that would exceed the new model's capacity. Cache reuse is fixed: unchanged system prompts allow the next model request to reuse the existing context prefix cache, avoiding unnecessary invalidation. Global and project-level AGENTS.md files now consume separate read budgets, preventing project instructions from crowding out global ones. History recovery fixes address sessions containing legacy sub-agent descriptions that failed to open, and cases where loading older compressed history halted further chat updates.

Plugin System Updates

Experimental subagent_oar plugin delegates independent tasks to local Codex or Claude CLI (requires local CLI environment).

Statistics plugin provides per-turn and per-step views of context, cache, model latency, and tool latency, helping developers pinpoint where time is spent.

Plugin management page supports rich-text details, third-party plugin package updates, and automatic rollback on hot-update failure.

Usage Statistics

After a month of desktop availability, November 4 recorded 1.54 million daily active users and ~92 trillion tokens processed in a single day, indicating rapid adoption.

Conclusion

These two releases tighten the loop around Agent task execution: post-completion review, actionable feedback, long-context continuity, and observability of tool calls and timing. For daily use, each eliminated draft loss, unopenable session, or manual code-copy round trip reduces a work interruption.