Issue Garbage Collector

Quick Start

# Phase 1: Run Python scanner (fast, free)
python3 .github/scripts/issue-scanner.py

# Phase 1: Get suspects only (for Phase 2 input)
python3 .github/scripts/issue-scanner.py --suspects-only

# Phase 1: JSON output (for automation)
python3 .github/scripts/issue-scanner.py --json

# Phase 1: Validation check (CI integration, exit 1 if errors)
python3 .github/scripts/issue-scanner.py --check

Harness Integration

• Repo-defined entry: docs/harness/automations.yml contains issue-gc-review
• Harness surface: settings/harness → Cleanup & Correction
• Data source: the Harness automation view reads suspect data from python3 .github/scripts/issue-scanner.py --suspects-only
• Intended usage: review pending duplicate / stale / open-check suspects in Harness first, then decide whether to run the cleanup workflow below

Two-Phase Strategy (Cost Optimization)

Problem: Running deep AI analysis on every issue is expensive.

Solution: Two-phase approach:

1. Phase 1 (Fast/Free) — Python script for pattern matching
2. Phase 2 (Deep/Expensive) — claude -p only on suspects

┌─────────────────────────────────────────────────────────┐
│  All Issues (N files)                                   │
│  ┌───────────────────────────────────────────────────┐  │
│  │ Phase 1: Python Scanner (.github/scripts/issue-scanner.py)│ │
│  │ - Filename keyword extraction                     │  │
│  │ - YAML front-matter validation                    │  │
│  │ - Same area + keyword overlap detection           │  │
│  │ - Age-based staleness check                       │  │
│  │ → Output: Suspect list (M files, M << N)          │  │
│  └───────────────────────────────────────────────────┘  │
│                         ↓                               │
│  ┌───────────────────────────────────────────────────┐  │
│  │ Phase 2: Deep Analysis (claude -p, only M files)  │  │
│  │ - Content similarity                              │  │
│  │ - Semantic duplicate detection                    │  │
│  │ - Merge recommendations                           │  │
│  └───────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────┘

Phase 1: Python Scanner

Run python3 .github/scripts/issue-scanner.py to get:

1.1 Formatted Table View

====================================================================================================
📋 ISSUE SCANNER REPORT
====================================================================================================

📊 ISSUE TABLE:
----------------------------------------------------------------------------------------------------
Status       Sev  Date         Area               Title
----------------------------------------------------------------------------------------------------
✅ resolv     🟠    2026-03-02   background-worker  HMR 导致 sessionToTask 内存 Map 丢失
🔴 open       🟡    2026-03-04   ui                 Task Execute button disabled
...
----------------------------------------------------------------------------------------------------
Total: 12 issues

📈 SUMMARY BY STATUS:
  🔴 open: 5
  ✅ resolved: 7

1.2 Validation Errors

If any issue has malformed front-matter, the scanner reports:

❌ VALIDATION ERRORS (need AI fix):
------------------------------------------------------------
  2026-03-08-broken-issue.md:
    - Missing required field: area
    - Invalid status: pending (valid: ['open', 'investigating', 'resolved', 'wontfix', 'duplicate'])

Action: Ask AI to fix the file:

claude -p "Fix the front-matter in docs/issues/2026-03-08-broken-issue.md. Add missing 'area' field and change status to a valid value."

1.3 Suspect Detection

The scanner automatically detects:

Output:

⚠️  SUSPECTS (need Phase 2 deep analysis):
------------------------------------------------------------

  🔗 Potential Duplicates:
    - 2026-03-02-hmr-resets-session-to-task-map.md
      ↔ 2026-03-08-background-task-hmr-recovery.md
      Reason: Same area 'background-worker', keywords: {'task', 'hmr'}

  ⏰ Stale Issues:
    - 2026-02-01-old-bug.md: Open for 35 days (>30)

1.4 JSON Output for Automation

# Get suspects as JSON for scripting
python3 .github/scripts/issue-scanner.py --suspects-only

Output:

[
  {
    "file_a": "2026-03-02-hmr-resets-session-to-task-map.md",
    "file_b": "2026-03-08-background-task-hmr-recovery.md",
    "reason": "Same area 'background-worker', keywords: {'task', 'hmr'}",
    "type": "duplicate"
  }
]

Phase 2: Automatic Deep Analysis

IMPORTANT: After Phase 1, proceed automatically to Phase 2 without asking. Do NOT ask "Would you like me to proceed?" — just do it.

Execution Flow

1. Run python3 .github/scripts/issue-scanner.py
2. For each suspect found, automatically perform deep analysis
3. For each action needed, execute immediately (update status, merge, etc.)
4. Only ask for confirmation on destructive actions (delete, merge)
5. Report final summary when done

For Each Suspect Type

Duplicates — Read both files, compare content:

• If same root cause → Merge (keep newer, add context from older)
• If related but different → Add related_issues cross-reference
• If distinct → Skip (false positive)

Open Issues — Check if resolved:

• Read the issue, check Relevant Files in codebase
• If code shows fix → python3 .github/scripts/issue-scanner.py --resolve <file>
• If still broken → Leave as open
• If unclear → Leave as open, add comment in issue

Stale Issues (open > 30 days):

• Check if code still exists
• If fixed → Resolve
• If code removed → Close with --close
• If still relevant → Create GitHub issue for tracking

Quick Update Commands

Use the scanner's update commands for fast changes:

# Resolve issues (status: open → resolved)
python3 .github/scripts/issue-scanner.py --resolve file1.md file2.md

# Close issues (status: open → wontfix)
python3 .github/scripts/issue-scanner.py --close file.md

# Generic field update
python3 .github/scripts/issue-scanner.py --set severity high --files file.md

Safety Rules

1. Never delete _template.md
2. Never delete issues with status: investigating — active work
3. Ask for confirmation only for: delete, merge
4. Auto-execute for: status updates, adding cross-references
5. Preserve knowledge — resolved issues are valuable

Periodic Maintenance

Cost Optimization

Savings: ~90% cost reduction by filtering in Phase 1.