HITL: Human-in-the-loop approval workflow for autonomous fixes

[michaeloboyle]

Description

Formalize the human-in-the-loop approval workflow that connects the Fix Review panel (#10), pattern library (#8), and self-healing mode (#7) into a cohesive HITL system.

Current State (Delivered)

• Fix Review panel: Before/after screenshots, Approve/Partial/Reject buttons
• Pattern library: Stores successful fixes, matches on new issues
• Telemetry: resolve() closes the feedback loop, metrics() shows dashboard

HITL Requirements

1. Approval Queue

• Persistent queue of pending fixes (survives page refresh)
• Queue items show: description, issue class, before/after screenshots, code diff
• Batch approve/reject for same-class fixes
• Queue badge count in the visualizer header bar

2. Code Diff View

• Show the actual file edits made by Claude Code (parse session JSONL for Edit/Write tool calls)
• Syntax-highlighted diff alongside the visual before/after
• One-click revert if rejected (git checkout the changed files)

3. Approval Graduation (ADR-0007)

• Track approval rate per issue class
• When class hits 95%+ approval over 10+ fixes, propose auto-apply
• "Graduate" button promotes class from HITL → autonomous
• Graduated classes still log telemetry but skip the review queue
• Regression detection: if a graduated fix fails, revoke auto-apply

4. Confidence Scoring

• Display confidence score based on pattern match quality
• High-confidence fixes (known class, strong pattern match) show green badge
• Low-confidence fixes (novel class, no prior patterns) show orange badge
• Confidence threshold configurable (default: auto-apply above 0.9)

5. Notification Integration

• Desktop notification when fix is ready for review
• Optional sound (via TTS feedback skill)
• Slack/webhook notification for remote review

Architecture

__claude.fix() → Before screenshot → Claude Code session → After screenshot
                                                                    ↓
                                              Fix Review Panel (HITL)
                                              ├── Before/After images
                                              ├── Code diff
                                              ├── Confidence score
                                              └── [Approve] [Partial] [Reject]
                                                      ↓
                                              Pattern stored (success)
                                              → Metrics updated
                                              → Graduation check

Depends On

• ADR-0006: DevTools Bridge integration #6 (ADR-0006: DevTools Bridge) — Done
• Fix pattern library: store and reuse successful fix patterns #8 (Fix pattern library) — Done
• Before/after visual diff for fix approval workflow #10 (Before/after visual diff) — Done
• Self-healing UI: proactive issue detection and auto-fix queue #7 (Self-healing UI) — Open (provides the fix source)
• Context pruning: reduce context size using telemetry signal-to-noise data #9 (Context pruning) — Open (optimizes what gets sent)

References

• PKM ADR-0007: Always-On Agent Fleet (operational autonomy)
• PKM ADR-0013 Benchmarks: Self-Improvement Loop 3 (Approval Lane Graduation)
• PKM ADR-0013 Benchmarks: North Star principles