DeepSeek Harness alpha adds worktrees, local-agent delegation and a long list of practical fixes

DeepSeek Harness v0.2.2-alpha.2 adds tools for running sessions in separate Git worktrees and delegating independent tasks to local Codex or Claude CLIs. The release also tackles a more everyday frustration: losing drafts or useful context when a task, connection or tool call goes wrong.

DeepSeek Watch analysis

What happened

The project’s release notes list new features, interface improvements and fixes across its AI-agent harness. Several changes matter directly to people building or using agent workflows, rather than merely adding another number to the version string.

Our top picks

• Run sessions in separate worktrees
  An experimental feature lets people work in an isolated Git worktree and inspect linked pull-request status and code changes.
• Delegate independent tasks locally
  An experimental plugin can hand separate tasks to locally installed Codex or Claude command-line tools.
• Keep drafts when sending fails
  The composer now waits for the host to accept a message and preserves drafts, quotes and attachments after a failed send.
• Recover more reliably from interruptions
  The minimal SDK saves session recovery records before model requests or tool execution, and stops if saving fails.
• Make tool access conditional
  Tools can be enabled step by step, with model declarations, bindings and execution admission kept aligned.
• Try scripts in Node PTC
  A new mode supports top-level await, dynamic imports and expression results; retaining state across calls is optional.
• Inspect agent work in more detail
  Experimental statistics show context, cache, model and tool timings by turn and step.
• Cut repeat downloads and heavy history transfers
  Web resources support ETags, while remote connections enable WebSocket compression by default.

Why it matters

Agent tooling is only as useful as the work around the model: context, permissions, recovery and the ability to inspect what happened. This release takes aim at several of those practical seams, from safer message handling to clearer control over tools and separate workspaces for coding tasks.

Some headline features remain experimental, and the notes specify migration work for custom plugins and older tool settings. That is relevant for maintainers, but it does not cancel out the user-facing improvements.

Our read

This is a substantial, useful alpha rather than a ceremonial version bump. The strongest changes make agent work easier to isolate, recover and inspect. If you build with DeepSeek Harness, read the migration notes before upgrading; if you use it, the worktree and draft-preservation changes are especially worth a look. Alpha means there may still be rough edges, a fact software has yet to find a way to make charming.

What to watch

• Whether the experimental worktree and local-delegation features become stable.
• How custom plugins must change for the new tool and prompt-assembly interfaces.
• Whether the project documents any limits or failure cases for session recovery and script-mode state.

Discussion spark: For coding agents, should developers prioritise isolated workspaces and recoverable sessions, or are those safeguards worth the extra complexity only after the agent itself becomes more capable?

not affiliated with or endorsed by DeepSeek