Choose a subagent for the composer and it stays selected until you remove or replace it. Each one is built for a different kind of work, and only general carries out changes end to end.
SettingsAgents
Pick a subagent and it stays selected until you swap it. Only general carries out changes end to end.
@explore · Where is the refresh token rotated?
› Searched refreshToken · 14 matches
› Read src/auth/session.ts, src/auth/rotate.ts
› Answered · 0 files changed
@exploreAsk ML.ai... @ to delegate
Developers who want finding, planning and editing to be separate acts, each handled by a subagent scoped to that job.
Ask @explore how something works and get an answer with nothing on disk changed.
Let architect name the critical files, trade-offs and steps before any edit.
Hand general a multi-step edit or refactor that it carries out inside the project.
Type @ in the composer to pick a subagent. It stays selected until you remove or replace it, and each one can touch only what its job needs.
Searched getUser9 matches
Edited 4 files+12 −12
Ran pnpm test48 passed
Inside the project only · back to you after one turn
Edits and commands, inside the project
For a task you can describe precisely, such as a multi-step edit or a refactor. It edits files and runs commands inside the project only, makes safe, reversible decisions on its own, and hands control back after one turn.
Grep process.env.DATABASE_URL7 matches
Read src/config/env.ts, src/db/client.ts
Read scripts/seed.ts
0 files changed · no permission prompt
Read-only search, nothing changed
A read-only search specialist. It reads files, greps, globs and fetches, and changes nothing while doing it. You can tell it how hard to look, from quick to very thorough.
Critical files
src/billing/webhook.ts · src/billing/retry.ts
Trade-off
Retry inline or move retries to a queue
Steps
1. add a retry table 2. queue failed events 3. test the backoff
Files, trade-offs and steps, no edits
For a change big or unfamiliar enough that the order of work is not obvious. It reads the code and returns the critical files, the trade-offs and a step-by-step plan, without editing or writing anything.
## Billing webhook retry
- [ ] add a retry table
- [ ] queue failed events
- [ ] test: billing/retry.test.ts
Can write this file and nothing else
The plan saved to a file for review
The same investigation as architect, with the finished plan written to a markdown file for you to review and approve. It can change nothing except that plan file.
· generalWhere is the refresh token rotated?
Searched refreshToken14 matches
Read src/auth/session.ts, src/auth/rotate.ts
Tools in reach: Read · Edit · Bash
· @explore@explore Where is the refresh token rotated?
Searched refreshToken14 matches
Read src/auth/session.ts, src/auth/rotate.ts
Answered · 0 files changed
general can answer a look-up, but it can also edit files and run commands, so even a question carries the reach of a tool that could act.
The same search and the same answer, from an agent that cannot edit or write. Nothing on disk changes, and there is no permission prompt because there is nothing to approve.
Finding out and changing things stay two separate acts. Here is one question answered read-only, then the edit handed off.
@exploreFind every place this repo reads process.env.DATABASE_URL and summarize the pattern.
@explore · Read · Grep · Glob, no edit or write tools
@explore Find every place this repo reads process.env.DATABASE_URL and summarize the pattern.
Routed to @exploreread-only, whatever the mode
@explore Find every place this repo reads process.env.DATABASE_URL and summarize the pattern.
Routed to @exploreread-only, whatever the mode
Grep process.env.DATABASE_URL7 matches
Read src/config/env.ts, src/db/client.ts, scripts/seed.ts
0 files changed · no permission prompt
@generalRoute scripts/seed.ts through getDatabaseUrl()
The edit goes to general in Build mode, as its own step
@exploreFind every place this repo reads process.env.DATABASE_URL and summarize the pattern.
@explore · Read · Grep · Glob, no edit or write tools
What trips people up when they first pick a subagent.
What goes wrong
How handles it
general cannot ask questions mid-task, so a task that still needs clarifying comes back built on its best guess.
How handles it
Scope it first with @explore or @architect, then hand general a task that is already well defined.
A custom subagent you just created does not appear in the @ picker.
How handles it
Agent definitions are read at startup: run ML.ai: Restart Engine and it appears in the picker and in Settings → Agents.
You select general inside a Plan-mode session and wait for it to change files.
How handles it
general’s project-only scope sits under the session’s mode, not above it, so Plan still denies edits. Switch the session to Build to apply changes.
Fix the registration validation bug and run the relevant tests.
plan.md
Fix registration validation
Ask ML.ai...
PLAN.md · image uploads
## Step 1: add an uploads queue
- create src/jobs/resizeImage.ts
- enqueue from POST /uploads
- test: jobs/resizeImage.test.ts
The plan, written as steps to review
Edit and write tools switched off for the turn, and selecting general does not turn them back on.
Explain how auth refresh works
refresh() in src/auth/session.ts rotates the token pair 60 seconds before expiry.
session.tslines 41-42
41if (expiresAt - now() < 60_000)42 return rotate(tokens)
Find and fix the failing test
Background tasks
RunningAdd input validation to the signup formgeneral · write allowed
Read SignupForm.tsx, validate.ts
Delegated from the composer
Delegate a job to a subagent in its own session, with its write capability stated up front.
Yes, but you describe it rather than write the definition by hand. Ask ML.ai in plain language, e.g. "Create a subagent that reviews pull requests for security issues," and its bundled create-subagent skill composes a full definition, covering scope, instructions, and allowed tools, written to .opencode/agent/ in your project. You can open and edit that file afterward if it got something wrong.
Agent definitions are only read at engine startup. Run ML.ai: Restart Engine from the Command Palette after creating one, and it will appear in both the @ picker and Settings → Agents.
No. explore is read-only by construction: it searches, greps, globs, and fetches, and cannot edit or write, so there is never an operation for a permission prompt to gate.
No. general’s own scope (project-only, but able to edit and run commands) sits underneath the session’s standing mode, not above it. A Plan-mode session with general selected still has the edit and write tools denied at the mode layer.
Look with @explore, plan with architect, and let general make the change, each scoped to what the job needs.
