Sessions, checkpoints & long-running tasks

1 · Sessions persist and replay

Every conversation, tool activity block and task list is persisted with its session. Reopen a task and you get the full history back — tool calls replay as collapsed activity, task lists restore with their exact progress, and Build/Plan modes restore from the latest assistant turn. Every new chat starts fresh in Build mode.

2 · Live task lists

Multi-step work gets a visible todo card: one in_progress item at a time, status icons as steps verify, progress bar and priority badges. Updates stream over SSE as they happen and survive restarts. The model carries the full todo policy in its system prompt, so the list stays honest — items are marked complete the moment they're verified, not before.

3 · Long conversations get compacted, not truncated

When a provider conversation grows too large, Sonderr compacts it automatically: the original request, recent activity and the latest checkpoint carry forward; bulky old tool outputs and image payloads drop. Compactions are untrusted hints — the agent re-checks current files and results instead of trusting a summary.

4 · Checkpoints: the resume point

Substantial Build work writes durable checkpoints — goal, current milestone, verified evidence, decisions made, and one exact next action. If the browser closes, the local process can continue checkpointed work in bounded provider chunks, and reconnecting shows live progress. If Sonderr itself exits, the next launch marks the task interrupted and offers Resume interrupted task.

5 · Work budgets, stated plainly

• S-tasks (S1–S4): 10–60 seconds of active work.
• H-tasks (H1–H4): 5–20 minutes.
• U-tasks (U1–U10): 6–30 hours, across up to 260 provider chunks and 1,500 tool calls.

Idle time and downtime never count. Provider chunks are capped (24 rounds / 80 tool calls each). These are workload targets, not guarantees — the agent should stop rather than invent work when the goal is met. Permission and confirmation gates stay in force for the entire run.

6 · Private task notes

For exceptionally long jobs the agent can keep concise private notes — up to 16 notes of 12 KB each, stored locally with restrictive permissions, never shown in chat cards, removed when the task verifies complete or after 30 days. This is the only "memory" beyond your sessions, and it never leaves your machine.

7 · Collaboration rooms (bonus memory)

Rooms keep a lead agent and up to three read-only workers in one chat with shared findings, peer review and saved-room recovery. Worker personas are labeled fictional AI; their output is unverified evidence until the lead (or you) verifies it.