DeepSeek Harness 0.2: Why Bundle Architecture & Sandbox Recovery Matter More Than Desktop

1. Schedule Moves from Default Capability to Optional Bundle

In v0.1.7, DeepSeek Harness (DSH) began using a Bundle/Profile mechanism to split its own capabilities. Version 0.2.0-rc.1 takes this further by converting Automation Tasks into an official Optional Bundle rather than a built-in feature.

DSH's composition model: a Profile's package.json lists Bundles in load order; each Bundle patches the final Cordis composition by inserting or overriding plugin configuration. Schedule is now removed from the default composition and becomes an explicitly selectable capability layer.

0.1.7
Web Profile
└── Schedule / Automation capability

0.2
Web Profile
├── Default Bundle
├── User Bundle
└── Optional Schedule Bundle  ← opt-in

This separates installation from activation : an official capability ships with DSH but does not enter every user's runtime unless chosen.

Important nuance: Schedule is not a system-level cron or background notification service. It provides only three tools:

schedule_create
schedule_list
schedule_delete

Supporting delayed execution, absolute timestamps, and fixed-interval reminders. State persists in the session event log and survives restarts. However, its delivery mode is session-local — reminders arrive as follow-up messages inside the original session. If the session has no live root agent, overdue reminders stay pending until the session resumes. They are not emailed, SMSed, or pushed system-wide.

DSH did not turn cron into a plugin. Instead, it extracted a complete session-local Schedule capability from the default composition and made it an explicitly selectable official capability layer.

This marks the Bundle architecture being used more systematically by the core team.

2. Windows Sandbox: Permission Diagnosis, Backup & Repair

Version 0.2.0-rc.1 adds a permission-diagnosis skill to the built-in Windows Sandbox. It identifies the root cause of Access Denied errors (ACL, Integrity Level, inherited permissions, target directory state) within a supported scope, then performs a backed-up, recoverable repair. RC.2 merges diagnosis and repair into a single authorized operation while preserving pre-change backups and a restore command.

The problem: a bare Access Denied gives the model insufficient information, leading to blind command retries that never fix the underlying permission state. The new flow:

Access Denied
    ↓
Identify permission issue
    ↓
Determine if within repairable scope
    ↓
Request user authorization
    ↓
Backup original permissions
    ↓
Execute repair
    ↓
Retain recovery path

Goal: eliminate stalls in complex local dev environments caused by permission anomalies, enabling legitimate workloads to be diagnosed and recovered automatically.

3. Async Question Mode: Decoupling "Wait for User" from "Stop Execution"

RC.2 introduces an opt-in experimental Async Question Mode . In the traditional agent loop, ask-user-question is a synchronous block:

Execute task
  ↓
Need user info
  ↓
Ask
  ↓
Pause entire agent
  ↓
Receive answer
  ↓
Continue

This simplicity hurts complex tasks. Example: agent needs deployment confirmation but could simultaneously run tests, type-check, or analyze unrelated code. If the question locks the whole flow, independent work halts.

Async Question Mode lets the agent continue independent work after a timeout while the user answers later — similar to ChatGPT's follow-up question behavior.

4. Model & Search Integration Changes

DeepSeek account web search : sessions using DeepSeek models can now invoke web search without configuring an extra API key, removing one integration layer.

pi-ai upgrade to 0.87.1 : third-party model catalog and compatibility integration updated. Official notice: some legacy Model IDs were removed, so persisted model selections may need reselection. If a saved model disappears after upgrade, check catalog/ID changes first — it's a compatibility issue.

Fuzzy search & keyboard navigation added for long model lists, targeting aggregator APIs like OpenRouter.

5. DeepSeek Harness Desktop

Official desktop builds now exist for Windows and macOS; Linux is explicitly unsupported. DeepSeek account login works, suggesting a subscription tier may follow. Pricing remains unknown.

Summary

In one sentence: DeepSeek Harness 0.2 continues refining a composable, recoverable Agent Runtime with explicit permission boundaries and execution-state semantics.

If you want to understand how models, tools, state, permissions, execution environments, and plugin lifecycles fit on a single architecture diagram, this release provides concrete steps in that direction.