Agent Mode: Add plan approval gate before execution

[CarGDev]

Problem

During agent evaluation, models executed plans without waiting for user confirmation. This is a critical safety issue - users should review and approve plans before the agent modifies their filesystem.

Observed Behavior

• GPT-5-mini: Showed plan, immediately started executing without waiting
• Claude Sonnet 4.5: Showed plan, didn't ask for approval, created files directly
• Gemini: Created plan but didn't wait for review

Expected Behavior

1. Agent generates plan
2. Plan displayed to user with summary of changes
3. User prompted: "Approve this plan? [y/n/edit]"
4. Only after approval → execution begins
5. User can abort at any step

Implementation Notes

Partial implementation exists:

• src/services/plan-mode/
• src/tools/plan-approval/
• src/types/plan-mode.ts

References

• Claude Code uses EnterPlanMode / ExitPlanMode tools
• opencode has similar plan-then-execute flow

Priority

🔴 Critical - Blocks production readiness for agent completion

---

Generated from model evaluation test

[CarGDev]

Implemented in commit e877818. Added plan approval gate:

Changes:

• Registered plan_approval tool in tools index
• Added plan approval flow in message handler
• Created plan approval modal TUI component
• Updated system prompts for balanced/thorough tiers to require plan_approval

Workflow:

1. Agent uses plan_approval action="analyze_task" to check complexity
2. Agent creates plan with steps, context, risks using the tool
3. Agent submits plan with plan_approval action="submit"
4. User approves (yes/proceed) or rejects (no/stop)
5. Only after approval does agent execute file modifications

User Responses:

• Approve: "yes", "proceed", "go ahead", "approve", "looks good"
• Reject: "no", "stop", "cancel", "don't"
• Feedback: Any other message modifies the plan