DeepSeek Harness 0.1.7: Shifting from Capability Expansion to Sustained Agent Operation

Version Focus Shift: 0.1.6 vs 0.1.7

The article frames the transition clearly: 0.1.6 expanded what agents can do by adding tools like Terminal, MCP Resources, SSH Remote Workspace, Browser Use, Computer Use, Plugin Manager, and Office Preview. 0.1.7 tackles how agents work long-term and stably , addressing session management, background execution, multi-turn conversation performance, thinking/tool-call output noise, code review speed, plugin/config maintenance, and session format evolution.

Session Upgraded to V4

Session format moves from V3 (used since 0.1.5) to V4, representing the most fundamental underlying change. The session increasingly resembles persistent agent working state . Official migration tooling handles batch conversion and tolerates some V3 sessions missing turn-end records. The author advises backing up ~/.dsh before upgrading, especially if you have accumulated history sessions, custom profiles, or plugin configs.

Session Sidebar Management

0.1.7 adds complete session management: pin, archive, filter, restore archived, and restore directly from search results. Archiving a running session no longer kills it bluntly; DSH checks for active Turn, Subagent, Job, or Reminder, shows which tasks would be affected, and requires a "stop and archive" confirmation. This reflects a shift from the early assumption of "open session → do task → close" to "maintain dozens of long-lived agent sessions simultaneously," approaching IDE project/workspace management logic.

Background Jobs with Auto-Wake

Long-running commands and workflows can now move to background execution. A background job panel shows real-time output; when the job finishes, the owning session can be re-awakened automatically. Example: previously pnpm run build (5 minutes) blocked the workflow — wait, wait, wait, build done, continue. Now: start build → exceed wait threshold → move to background → agent continues editing other files → build completes → session re-awakened → handle result. The author notes 0.1.7-alpha.2 fixed a bug where sessions stalled waiting for user input after consecutive background commands or one-off subagents completed. This Background Job + auto-wake pair is called the most meaningful runtime change, enabling an agent runtime that can schedule long-term tasks.

File Review Enhancements

Side-by-side diff becomes default with syntax highlighting, synchronized scrolling, line numbers, and hover preview on changed file cards. Quick confirmation via hover; deep review via full side-by-side diff. The author argues easier review expands the acceptable scope of autonomous agent edits. Excel Viewer now supports XLSX, XLS, CSV, TSV with worksheet switching, cell selection/copy, and formula view. Image, PDF, and Office preview zoom interactions are unified, preserving per-tab zoom levels.

Messages API Only — Chat Completions Removed

DeepSeek Adapter already favored Messages Protocol in 0.1.6. In 0.1.7, Chat Completions is formally removed ; the official adapter uses only Messages API. Custom DeepSeek endpoints must be Messages-compatible.

Plugin System: Ecosystem Compatibility

Plugins can declare multi-language titles, descriptions, and icons. Plugin Manager offers registry choices: official npm, npmmirror, or custom registry. Install/launch now checks plugin compatibility with the current DSH version and explains incompatibility reasons. This addresses frequent plugin API breakage due to rapid Harness releases.

Settings and Agent Preset Unification

Previously many settings lived in settings.yaml. In 0.1.7 they move into the current Profile's Plugin Configuration . Old settings.yaml is imported once. Agent Presets no longer rely primarily on traditional directory management. Profile, Plugin, Plugin Config, Agent Preset, and Settings converge into a single configuration system — beneficial long-term but potentially painful during this upgrade.

Agent Team Changes

Team Panel now shows real-time member and task status; conversation header allows viewing/switching members. Team tools increasingly target members by name. Task board becomes read-only ; task creation/update is done by agents via tools, not manual user edits. The role clarifies: not a user manually managing multiple bots, but a Lead Agent managing an agent team, with the user observing and intervening.

Upgrade Checklist

Key breaking changes to verify: Session V3→V4; DeepSeek Messages API only; settings.yaml → Profile Plugin Config; legacy directory-type Agent Preset → Plugin Bundle; Workspace read API → readBytes; spill-policy; maxInlineBytes → maxInlineTokens; PTC Runtime / Plugin API changes. Source-build upgrade steps: backup ~/.dsh, git fetch --tags, git checkout dsh-v0.1.7-rc.1, pnpm run clean, pnpm install --frozen-lockfile, pnpm run build, pnpm dsh web. Test on a copied environment first.

Three-Phase Summary

0.1.5: Give agents a working environment.

0.1.6: Give agents more tools.

0.1.7: Let agents work continuously in that environment.