FAQ — Common Questions from Real-World Scenarios

Q1: Do business users need to type CLI commands?

No.

The Lovrabet CLI works more like the execution layer behind the Agent. What business users should do instead:

• State the business question clearly
• Say whether it's read-only or writes are allowed
• Include key identifiers when available — order numbers, phone numbers, customer IDs

The actual commands are usually decided by the Agent and skills in the background.

---

Q2: Why is there no sql list or bff list?

That's intentional.

SQL and BFF entries map to real business logic, and a general-purpose Agent shouldn't be free to scan and execute them at will. The preferred pattern:

• Maintainers state explicitly in the skill "which SQL / which BFF this scenario calls"
• Business users describe only the business goal
• The Agent follows the skill

---

Q3: What does --global actually mean?

It controls where a config write goes.

Simply put:

• Global config: works everywhere on this machine
• Project config: applies only inside the current project directory

Typical behavior:

• auth login leans global by default, so multiple projects can share the same AK
• app use and config set usually write project config first when run inside a project

If you only use the Agent from the business side, you normally never manage this detail yourself.

---

Q4: What if installation reports PromptScript does not support global skill installation?

TIP

This is a short-term upstream compatibility issue with skills@1.5.10 — not a Lovrabet CLI installation failure. Business users don't need to figure out --agent, --all, or Agent directories themselves: check the status first, then let the Agent or a technical colleague handle it in a local terminal.

First, separate two states:

• If lovrabet --version prints a version number, the Lovrabet CLI itself is installed fine.
• A failed Skill install only affects whether the AI Agent understands the Lovrabet CLI correctly — it doesn't mean the CLI is unusable.

Recommended handling:

• Don't retry the same install command repeatedly inside IM Agents such as Feishu Aily or DingTalk Wukong — you'll be misled by 30-second timeouts, sandbox paths, or app-private SkillHubs.
• Have the Agent or a technical colleague run the install in a real local terminal, and confirm the result with the local skills CLI output.
• As a temporary workaround, pin the previous installer version:

npx skills@1.5.9 add lovrabet/lovrabet-cli -g -y

• --agent <agent> is only for technical users who know exactly which tool they target — say Codex, Claude Code, or Cursor. Business users shouldn't fill it in themselves, and shouldn't let the Agent guess.
• If the log shows an installed success message followed by Failed to install 1, verify with the real local output of npx skills list -g --json — don't trust the natural-language summary alone.

Technical reference: upstream Issue #1352 and the fix PR #1362. https://github.com/vercel-labs/skills/issues/1362 redirects to the same PR.

---

Q5: Why is there no app add?

Because the runtime CLI no longer treats "maintaining a local list of app profiles" as its main model.

The current recommended approach:

• Use app list to see the apps visible to the current account
• Use app use to pick the default app
• Let the remote app catalog plus the local cache be the source of truth

---

Q6: Do I have to know appcode, dataset codes, or sqlcodes?

Not necessarily.

Ideally:

• Business users don't need to know them
• Skill maintainers ideally should
• The Agent can run a discovery step when needed

If you can provide these identifiers, execution is faster. If you don't know them, let the Agent scout for you first.

---

Q7: How do I phrase a request so the Agent gets it right?

Try to state all of these in one go:

• Which business system or business domain
• Which business object
• The time range or filter conditions
• What result you ultimately want
• Whether this is read-only or writes are allowed

For example:

Find orders from the last 7 days that were paid but not shipped, summarize by warehouse, read-only.

That works far better than "look up some orders for me".

---

Q8: How do I phrase write requests safely?

Just add one line to the request:

• preview first, then execute
• read-only, don't write anything

This makes it much easier for the Agent to land on the right safety path automatically.

---

Q9: What is --dry-run for?

It shows you "what this run would do" without writing anything.

It's especially useful when:

• Batch-correcting statuses
• Backfilling data
• Giving a final check before deletion
• You're not fully sure the parameters are right

---

Q10: Is Cookie auth still supported?

The current Lovrabet Runtime CLI uses AccessKey as its primary auth path. If older docs still mention Cookie, read them in terms of the new AccessKey model.

---

Q11: What if the Agent gets it wrong?

Sort the problem into one of two buckets first:

• The commands never ran: check login, default app, doctor, and the logs first
• The commands ran but the results are wrong: check the skill, the business prompt, and the business-flow assumptions first

Very often the problem isn't the CLI at all — it's that "the business flow bound in the skill is outdated".