When the user says Holix is wrong or asks it to check itself («проверь себя», «почему ты делаешь не так», «ты отвечаешь неправильно», check yourself), spawn the session_doctor sub-agent with a clean context.
Installs into .claude/skills of the current project.
Are you the author of Holix Self Diagnose?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/javded-itres-holix-self-diagnose)
---
name: holix-self-diagnose
description: >
When the user says Holix is wrong or asks it to check itself
(«проверь себя», «почему ты делаешь не так», «ты отвечаешь неправильно»,
check yourself), spawn the session_doctor sub-agent with a clean context.
tags:
- self-diagnose
- session
- honesty
- holix
- required
required: true
user-invocable: true
---
## When to use
The user is criticizing **this agent**, not asking to debug their own code:
- «проверь себя», «проверь свою работу», «самодиагностика»
- «почему ты делаешь не так», «ты отвечаешь неправильно»
- «check yourself», «you're answering incorrectly»
## Procedure
### Main agent
1. Spawn a **clean-context** doctor (do not fork parent turns):
`delegate_to_subagent(agent_type="session_doctor", fork=false, task=…)`.
2. In `task` include this `conversation_id` and the user's complaint.
3. `wait_subagent_result` and show the doctor's briefing.
4. Do **not** call `self_diagnose` yourself (this chat is polluted).
5. Do **not** change system settings, skills, or extensions.
### session_doctor (you already are the child)
1. Call `self_diagnose` immediately (parent session is autopsied automatically).
2. Read `findings`, `plan`, `session.timeline`, `session.last_real_ask`, `llm`, `settings`.
3. Explain what went wrong in **this session** (tools vs claims vs user prompts vs settings).
4. Coach the user: how to write the next prompt and how to use Holix so the mistake does not repeat.
5. Do **not** write files, patch skills, run the terminal, or change system settings.
6. If an operator must change model / jail / extensions or read logs: call `request_admin_support`.
That tool **always asks the user to confirm**. Deny = send nothing. Do not retry after Deny.
## Pitfalls
- Do not answer the complaint from memory. The autopsy is in `self_diagnose`.
- `read_file` / `cat` is not delivering a file. The report will flag that.
- Do not auto-spawn any type other than `session_doctor` for this request.
- Never claim a support ticket was sent unless `request_admin_support` returned `ok: true`.
## Verification
- Main: this turn spawned `session_doctor` (`fork=false`) and waited.
- Doctor: this turn has a `self_diagnose` result and quotes finding codes.