DeepSeek Harness v0.1.7-alpha.2: Sessions, Teams & Truncation

DeepSeek released dsh-v0.1.7-alpha.2 on September 22, 2026. It follows alpha.1 with reliability fixes for the new architecture: Sessions could stall after successive background commands or one-shot Sub-agent tasks, Web could appear connected after a restart while requests could not proceed, and long tool-content truncation could discard images or important trailing output.

The release also updates Agent Team initial tasks, model naming, PowerShell waits, and registry selection during the first plugin install. It is a separate corrective release, not a merged recap of alpha.1.

Release and installation

Pin the release:

npx @deepseek-ai/[email protected] web

Users already evaluating alpha.1 can compare exact versions. Environments following latest do not move to this alpha automatically.

Stalled Sessions: completion wakeups are unlimited by default

Alpha.2 fixes two ways a Session could fail to continue: the parent could miss the correct wakeup after several background commands, and a one-shot Sub-agent could finish without returning control to the remaining workflow.

Completion-event wakeups are unlimited by default while remaining configurable. This does not give an Agent unlimited task execution. It prevents an already completed job from leaving the Session stuck because its notification allowance was exhausted.

A useful regression test is:

1. start two short background commands;
2. delegate one one-shot Sub-agent task;
3. wait for all three to finish;
4. confirm the parent Session receives completion events and continues;
5. reload the page and verify events and final states are not duplicated.

Teams that explicitly configured a completion-wakeup limit should reassess that value instead of carrying the old constraint forward unchanged.

Connection state after a Web restart

Alpha.2 fixes the Web interface appearing connected after a service restart while requests could not reach the backend. The client now reconnects when the service is actually ready.

This matters for remote workspaces, reverse proxies, and development hot restarts. Test the status indicator, Session refresh, and a real model request together; a green icon alone is not an end-to-end health check.

Tool content uses a shared token budget

Tool text and images now share an estimated token budget. When content is too long, DSH preserves both the beginning and end and provides a path to the omitted material. This also fixes MCP images disappearing when preceding text caused truncation.

Custom spill policies have one configuration migration:

maxInlineBytes  ->  maxInlineTokens

Rename the field and choose a threshold using token semantics. Byte and token counts are not interchangeable, so copying the old number does not preserve the old behavior.

Test long text, mixed text and images, and output whose critical conclusion appears at the end. The model should retain the opening context, the trailing result, and a route to the full omitted content.

Clearer Agent Team initial tasks

New Agent Team members receive initial-task guidance about teammates and the Lead. This reduces cases where a member understands its local task but not where to report or how to hand work back.

The structured task board and tools still define collaboration. Prompt guidance does not replace task state, dependencies, or completion criteria. Team templates should check for duplicated instructions after the upgrade.

Responsive registry selection for first plugin install

When no npm registry is configured, the first plugin install selects a responsive registry. This reduces long waits against an unreachable source in constrained networks.

Enterprise and reproducible environments should still pin a registry. Automatic selection improves first use; it does not replace package-source and supply-chain policy.

Interaction, model, and Office fixes

Alpha.2 also improves:

• scrolling, history loading, turn navigation, and message stability;
• code-block styling, copy, wrapping, diffs, and line numbers;
• preservation of line breaks in queued messages;
• long filenames and review-popover layout;
• Excel special content and resize state;
• persistent PowerShell waits;
• speech-recognition setting persistence;
• readable names for discovered models while retaining the real model ID.

Record the underlying model ID during diagnosis. A readable display name helps selection, but Provider requests and migrations still depend on the ID.

Narrower vendor-package auto-updates

Vendor-package automatic updates are restricted to patch releases within the same minor version. This reduces unintended compatibility jumps. Pre-releases and custom Bundles should still pin exact versions and test load, unload, configuration migration, and Session recovery.

Upgrade checklist for alpha.1 users

Alpha.2 is the better baseline for teams already testing the 0.1.7 alpha line. Profiles staying on the rc line can use it as the current reference when evaluating Session V4 and the new plugin architecture.

Related reading

• DeepSeek Harness v0.1.7-alpha.1 update
• DSH Sub-agent usage and billing analysis
• DeepSeek Harness Plugin Manager
• Why did a simple task take 50 turns?

FAQ

Is alpha.2 stable?

No. GitHub marks it as a Pre-release, and npm latest remains on an rc release.

Do unlimited completion wakeups create infinite Agent loops?

No. They address completion events that failed to resume a Session. Agent turns, tasks, and permission boundaries remain separate controls.

Must every custom plugin change its configuration?

Plugins with a custom spill policy that still uses maxInlineBytes must migrate to maxInlineTokens. Other plugins should still test long tool results and image content.