💡 Voice-Defined Delegation Rules: Verbal Permission Profiles to Reduce Confirmation Fatigue

[github-actions[bot]]

Summary

Users verbally create persistent approval rules ("always save brainstorming docs without asking", "never modify files outside my project folder") that reduce FR20 confirmation prompts for trusted, repetitive actions. Rules are scoped per workspace, auditable via FR27-29, and include immutable safety floors that cannot be overridden for destructive operations. This bridges the gap between "approve everything" (unsafe) and "confirm every action" (fatiguing), giving non-technical users granular control through natural conversation.

Market Signal

ChatGPT's new voice agentic features (Sep 23, 2026) offer only binary approve-all or approve-each — no middle ground for permission management. Enterprise AI adoption research (Writer 2026) identifies confirmation fatigue as a top friction point preventing daily AI agent use. No competitor — ChatGPT, Fazm, Manus, or any desktop AI agent — offers voice-defined, granular permission profiles. This is a genuine UX innovation in the AI agent trust space.

User Signal

The PRD targets daily active usage (≥1 session/day within 30 days of first use). FR20 requires pre-action confirmation for every file create/modify/delete and API call. For power users performing 10+ agent actions per session, this creates significant friction that threatens the Day-30 retention target. The onboarding initiative (#333-338) gets users started, but permission fatigue is the silent adoption killer for sustained daily use. Related: existing discussions on progressive trust (#320) and conversational permission approvals (#70) address aspects of this problem but neither offers user-defined delegation rules.

Technical Opportunity

Leverages the existing confirm-plan pattern (FR20), audit trail (FR27-29), and electron-store for persistent rule storage. Rules are simple JSON predicates ({action_type, workspace_path, auto_approve: true}) evaluated as a pre-check before showing the confirmation dialog. Voice input parsing uses the existing STT pipeline — "always approve file saves in my docs folder" → {action: "file:write", scope: "/Users/sarah/docs/**", auto_approve: true}. Implementation touches the security bounded context (main/security/) for rule evaluation and IPC handlers (main/ipc/) for confirmation flow. Architecture supports this via the AgentBackend interface and IPC channel segregation (agent:confirm).

Assessment

Dimension	Score	Rationale
Feasibility	high	Simple rule engine on existing architecture — JSON predicates in electron-store, pre-check before confirmation dialog
Impact	high	Directly addresses the #1 friction point for daily use retention; unique in market
Urgency	high	Needed before DAU target deadline; confirmation fatigue compounds with every session

Adversarial Review

Strongest objection: Users could create overly permissive rules and blame TalkTerm when something goes wrong ("I said approve everything" → deletes wrong file).

Rebuttal: Rules include immutable safety floors — destructive operations (file delete, overwrite existing files, external API calls) always require confirmation regardless of user-defined rules. Rule creation itself goes through the confirm-plan pattern (FR20), so the avatar explains exactly what the rule will allow. All rule applications are logged in the audit trail (FR27-29), providing accountability. Start with 3 simple templates (not a freeform rule engine) via progressive disclosure: "Auto-save new files", "Auto-approve read-only operations", "Trust this workspace for all file writes."

Suggested Next Step

Design the delegation rule schema (action types, workspace scopes, safety floors) and define the 3 starter templates. Create a tech spec covering rule storage in electron-store, the evaluation pipeline in main/security/, and the avatar UX for rule creation and review via voice commands.