Module 09 — Practice: your own toolset

Time: 160 minutes · Artifact: labs/09-custom/

Task 1. Command (30 min)

Create a /check-fix command in opencode.jsonc:

{
  "commands": {
    "check-fix": {
      "description": "Run the checks and fix the causes of failures",
      "template": "Run ./scripts/check.sh. If it fails — fix the cause, don't disable the checks. Repeat at most 3 times. At the end, show the diff and explain it in 2 lines."
    }
  }
}

Test the /check-fix invocation on a broken test.

Acceptance: the command works, the test is green again, the diff has been reviewed.

Task 2. Skill (40 min)

Create the skill .opencode/skills/migrations/SKILL.md:

---
name: migrations
description: Rules for creating and applying DB migrations in this project
---

1. Migration file: `migrations/YYYYMMDD_description.sql`
2. Always a down migration
3. Forbidden: DROP COLUMN without a backup
4. After creating: `alembic upgrade head` in dev, `pytest -q`

Verify: in a new session, ask "add a migration" — the agent should reference the rules.

Acceptance: the skill is picked up; the report includes a quote from the agent's response.

Task 3. Reviewer subagent (40 min)

{
  "agents": {
    "reviewer": {
      "description": "Strict diff review, no edits",
      "mode": "subagent",
      "permissions": [
        { "action": "edit", "resource": "*", "effect": "deny" },
        { "action": "shell", "resource": "git diff *", "effect": "allow" },
        { "action": "shell", "resource": "*", "effect": "deny" }
      ]
    }
  }
}

Give it the diff from module 07.

Acceptance: the subagent returned a list of problems, the files are unchanged (git status is clean).

Task 4. A model for the job (25 min)

Configure two setups:

{
  "agents": {
    "fast": { "mode": "primary", "model": "<cheap>" },
    "think": { "mode": "primary", "model": "<reasoning>" }
  }
}

Compare them on the same task.

Acceptance: a "task / agent / cost / quality" table.

Task 5. MCP or a reasoned refusal (25 min)

Check the list of available MCP servers:

opencode mcp list

If you need an external tool — connect it read-only and describe the risk. If you don't need one — write 3 reasons why not in the report (extra attack surface, doesn't solve the task, there's a simpler path).

Acceptance: either a working read-only MCP or an argued refusal.

Submission criteria

• The /check-fix command works.
• The skill is picked up by the agent.
• The reviewer subagent has no edit permissions.
• There's a model comparison by agent.
• MCP is connected or reasonably rejected.