Chapter 10 — Troubleshooting, Safety, and Next Steps

3 min read · 685 words · 9 sections

What You Will Learn

• Locate common problems with a table;
• Know where to look for diagnostics;
• Build your own safety checklist;
• Control cost;
• Plan your next practice steps.

10.1 Common Problems

10.2 Reading Diagnostics

Start with the UI

• Red errors;
• Failed tool cards;
• Plugin Manager warnings;
• Model connection status.

Then check context

• Is the workspace correct?
• Is the mode correct?
• Is the plugin enabled?
• Is the Session too long?

Then check logs

Desktop/CLI usually has a log or diagnostics entry. CLI users can run:

dsh --profile web --dump-config

Do not edit profile files until you understand the layer order.

A general troubleshooting order

1. Minimal reproduction
2. New Session
3. Standard mode
4. Disable the most recently installed plugin
5. Check model/network
6. Check workspace/permissions
7. Read logs/diagnostics
8. Reinstall/roll back only after that

10.3 Safety Checklist

Before an important task:

[ ] Does the workspace contain only what is needed?
[ ] Are important files backed up or committed?
[ ] Can I start read-only?
[ ] Will it delete, overwrite, push, or upload?
[ ] Am I about to expose keys, passwords, or private data?
[ ] Is the plugin source trustworthy?
[ ] Do I really need full access?
[ ] How do I roll back?

Three bottom lines

1. Secrets never go into chat;
2. Back up before destructive actions;
3. Do not give untrusted plugins high permissions.

10.4 Cost and Usage

Agent cost is mostly tokens:[2]

• Longer context costs more;
• Large files and long pages cost more;
• Subagents and workflows add model calls;
• Retries add cost.

Habits that save money

• Read only what is needed;
• Read long documents in sections;
• Summarize to files and start a new Session;
• Do not use subagents when one Agent is enough;
• Choose the right model for the task.

10.5 Next Steps

Week 1: Everyday use

• Complete three real tasks in Standard;
• Practice goal + input + constraints + acceptance;
• Learn to read tool cards and approvals.

Week 2: Workspace and context

• Write an AGENTS.md;
• Use PROJECT.md and PROGRESS.md;
• Practice resuming in a new Session.

Week 3: Tools and Skills

• Use todos for a multi-step task;
• Try one subagent;
• Read a SKILL.md;
• Deliver a file with present.

Week 4: Extension

• Read Chapter 8 and understand Plugin/Bundle/Profile;
• Build one small tool in Creator;
• Install, verify, and uninstall;
• Write your own plugin notes.

10.6 One-Page Cheat Sheet

Agent = Model + Harness

Everyday work  → Standard
Extending DSH  → Creator
Batch tools    → PTC
Small tasks    → Minimal

Workspace = folder + permission boundary
Session   = one conversation
Context   = working memory; it fills up
Skill     = knowledge/procedure
Tool      = action
Plugin    = runtime capability
Bundle    = installer package
Profile   = installation environment
Preset    = capability composition for one Agent

Good task = goal + input + constraints + acceptance
Read-only first, then write
Plan first, then execute
Back up before destructive actions
Secrets never go into chat

Summary

• Troubleshoot from minimal reproduction + new Session + Standard + disable new plugins.
• Three safety rules: secrets, backups, permissions.
• Cost comes from context and model calls.
• Learning path: daily use → workspace → tools/Skills → plugins.
• Use first; extend later. Do not start with a huge plugin.

References

1. DeepSeek Harness quickstart
2. DeepSeek API token usage
3. Git documentation
4. npm publish documentation

All links are public sources; product behavior may change between versions.