concepts
Plan the work, then execute.
Long jobs are broken into a task list. Status survives interruption, so you can quit, restart, and the agent picks up where it left off.
The shape of a task
~/.anthracode/tasks/0042-refactor-auth.json
{
"id": "0042",
"title": "Refactor auth middleware",
"status": "in_progress",
"created_at": "2026-05-08T09:14:00Z",
"steps": [
{ "id": "s1", "text": "read auth.ts and 2 dependents", "status": "completed" },
{ "id": "s2", "text": "extract verifyToken into lib/jwt.ts", "status": "completed" },
{ "id": "s3", "text": "update callers, run tests", "status": "in_progress" }
]
}Resumability
On startup the agent looks for any in_progress task in the current project and offers to resume it. You can also list and pick manually:
~/terminal
anthracode tasks ls~/terminal
anthracode tasks resume 0042When the agent creates a task
- Multi-step jobs that need 3+ distinct edits.
- Anything that crosses files or modules.
- Anything you might want to interrupt and resume.
Single edits stay in the conversation — the overhead of a task file isn't worth it.
Tasks vs memory
Tasks are working state — they describe what's happening right now. Memory is durable knowledge — what to remember across sessions. A task ends when its steps complete; the lessons learned during it may end up in a memory file.