💡 Goal Mode: Checkpoint-Based Background Workflow Execution with Avatar Notifications

[github-actions[bot]]

Summary

Users assign complex, multi-step tasks to TalkTerm and the avatar manages execution in the background, saving state at each workflow checkpoint. Users can step away and return later — the avatar resumes from the last checkpoint and summarizes progress. System tray notifications alert users when decisions are needed or workflows complete.

Market Signal

OpenAI Codex launched Goal mode enabling tasks that run for hours or days with remote monitoring via Codex Mobile. Manus (Meta) offers remote orchestration from any device. Claude Managed Agents added self-hosted sandboxes for enterprise-grade execution control. Background, long-running agent workflows are becoming a competitive differentiator in desktop AI agents.

User Signal

TalkTerm's PRD requires durable session state persistence (FR32) on close, crash, or unexpected termination. Session resume (FR31) is already planned. The gap is in managing actively-running workflows when the user isn't actively watching — the current design assumes synchronous user attention throughout workflows.

Technical Opportunity

The Claude Agent SDK's session persistence via session ID (stored in SQLite) already enables resume across app restarts. The confirm-plan pattern (FR20) provides natural checkpoint boundaries. Electron supports system tray presence for background notifications. The architectural foundation for checkpoint-based execution exists — it needs explicit workflow state tracking and notification integration.

Assessment

Dimension	Score	Rationale
Feasibility	med	Checkpoint-based execution is feasible with current architecture; full cloud offload needs Phase 2
Impact	high	Enables "delegate and walk away" — the ultimate non-technical user experience
Urgency	med	Codex Goal mode sets competitive expectation; TalkTerm should have an answer

Adversarial Review

Strongest objection: The architecture requires Claude Agent SDK in Electron's main process. Long-running background tasks would either block the UI or need cloud infrastructure (Managed Agents sandboxes), adding significant complexity and cost to a BYOK model.

Rebuttal: Scope to checkpoint-based execution first: the SDK executes workflow steps sequentially, saving state at each confirm-plan checkpoint. If the user steps away, session state persists (FR32). When they return, the avatar resumes and summarizes. The Electron main process can continue SDK execution while the renderer shows a "working in background" state. True hours-long cloud-offloaded execution via Managed Agents sandboxes is a Phase 2 enhancement. The MVP value is compelling: "Start this research task, I'll be back in 30 minutes."

Suggested Next Step

Define the checkpoint state machine (workflow states, persistence format, resume protocol). Design system tray notification UX for decision-needed and workflow-complete events. Evaluate Electron main process constraints for background SDK execution.