Thread Detection: Option A vs B Deep Dive

Output

🧵 Thread Detection: Option A vs B Deep DiveOption A: Thread as First-Class Entity
✅ Pros


Rich Metadata
Title, trigger reason, resolution notes, status, metrics all live on the thread

Easy Queries
SELECT * FROM threads WHERE flowId = X AND status = 'resolved' - show all rabbit holes

Aggregation
"This week: 3 threads, 15 turns, 2.1M tokens on detours" - easy to compute

UI Grouping
Timeline can collapse thread kontasks under a single "Mobile Padding Bug" header

Learning
Can tag threads with root cause ("CSS cascade", "missing auth") for pattern recognition

Status Tracking
Thread can be "ongoing", "resolved", "abandoned" - clear lifecycle


❌ Cons


New Entity
Another table, more schema, more CRUD endpoints, more MCP tools

Creation Friction
Who creates the thread? When? Automatic detection needs to be good

Boundary Ambiguity
When does a thread end? Resolving the bug? Or when you get back to main work?

Over-Engineering?
Might be solving a visibility problem with schema when UI grouping could suffice

Option B: Edge-Based Threading
✅ Pros


No New Entity
Uses existing graph infrastructure - just add edge type continues or spawned_by

Simpler Start
Can ship quickly, iterate on detection logic without schema migrations

Flexible
Threads emerge from edge traversal - can recompute groupings differently later

Gradual
Can promote to full entity later if edges prove insufficient


❌ Cons


No Thread Identity
Can't name the thread, can't add "trigger" or "resolution" notes

Harder Queries
"Show me all threads" requires graph traversal to find connected components

Metrics Scattered
Total turns/tokens for a thread = sum across all kontasks in chain (computed, not stored)

No Status
Can't mark a thread as "resolved" - it's just a chain of kontasks

🔍 Auto-Detection Strategies
Signal 1: Keyword Detection in Prompts/Titles
// Detect "still broken", "didn't work", "regression", "try again"
const THREAD_TRIGGER_PATTERNS = [
  /still (broken|not working|failing)/i,
  /didn't (work|fix|help)/i,
  /regression/i,
  /try (again|different|another)/i,
  /same (issue|problem|error)/i,
  /why is.*(still|again)/i,
];

// If prompt matches AND references recent kontask topic → thread continuation


Signal 2: Same File/Component Within Time Window
// If last 3 kontasks all touched the same file(s)
function detectFileCluster(recentKontasks) {
  const files = recentKontasks.flatMap(k => k.filesModified);
  const counts = countOccurrences(files);
  
  // If same file appears 3+ times in last 5 kontasks → likely a thread
  return Object.entries(counts)
    .filter(([file, count]) => count >= 3);
}

// Example: layout.ts modified in 4 consecutive kontasks → thread


Signal 3: Topic Stability + Bug Tags
// Kontask tags: ["bug", "regression", "fix", "broken"]
// If 2+ consecutive kontasks have bug-related tags → thread

const BUG_TAGS = ['bug', 'regression', 'fix', 'broken', 'issue'];

function isBugSequence(kontasks) {
  return kontasks.every(k => 
    k.tags.some(t => BUG_TAGS.includes(t))
  );
}


Signal 4: Temporal Clustering
// If kontasks are < 30 min apart AND seem related → thread
// "Related" = same topic OR same domain OR same files

Kontask A: 08:40 - Input Area Fix
Kontask B: 08:52 - Still broken      ← 12 min gap, same topic
Kontask C: 08:52 - CSS cascade       ← 0 min gap, same files
Kontask D: 09:02 - Finally fixed     ← 10 min gap, same files

→ All 4 form a thread

🖥️ Where Detection Happens


Location
When
How

Claude (in-turn)
During turn, before creating kontask
Claude sees recent kontasks in context, recognizes "I'm continuing the same problem" → adds threadId or continues edge

Konui (post-turn)
After kontask created, background job
Analyzes recent kontasks, detects clusters using signals above, creates edges or threads

User (manual)
In Flow Timeline UI
"Group these 4 kontasks as a thread" - manual override for edge cases

🎯 UI Surfacing
Option A (Entity) - Clean Grouping
▼ 🧵 Mobile Padding Bug (4 turns • 12 min • 450K tokens)
  │ Trigger: Expected quick CSS fix
  │ ✓ Resolved: CSS cascade order
  │
  ├─ 09:02 CSS Cascade Fix
  ├─ 08:52 Input Area Padding Fix  
  ├─ 08:52 Still broken...
  └─ 08:40 Full-Width Input Fix


Option B (Edges) - Inferred Grouping
▼ Linked kontasks (detected) (4 items)
  │ ← No title, no trigger, no resolution
  │
  ├─ 09:02 CSS Cascade Fix         continues ↓
  ├─ 08:52 Input Area Padding Fix  continues ↓
  ├─ 08:52 Still broken...         continues ↓
  └─ 08:40 Full-Width Input Fix

💡 RecommendationStart with Option B (edge-based) to validate detection signals, then promote to Option A (entity) once you know what metadata matters.
Phase 1: Add continues edge type, have Claude add it when detecting "still broken" patterns
Phase 2: Build UI that traverses continues edges to group kontasks
Phase 3: If you find yourself wanting to name threads or track resolution → graduate to gvThread entity
This lets you iterate on detection logic without committing to schema. The edge data becomes the migration path if you need entities later.

Location	When	How
Claude (in-turn)	During turn, before creating kontask	Claude sees recent kontasks in context, recognizes "I'm continuing the same problem" → adds `threadId` or `continues` edge
Konui (post-turn)	After kontask created, background job	Analyzes recent kontasks, detects clusters using signals above, creates edges or threads
User (manual)	In Flow Timeline UI	"Group these 4 kontasks as a thread" - manual override for edge cases

Quick Actions

▶ View Turn

Original Request

No layout configured

Details

Type General

Status Completed

Scope greatvibe

Tags designflowsschemathreadingarchitecture

Created 11 Jan 2026, 9:39 am

Updated 11 Jan 2026, 9:39 am

Timing

⏱️ 80.6s

0 tools

● TTFE 0% ● Thinking 100% ● Tools 0%

Created By claude

📋 Claude's Context

Session ses_40728090...

Raw Data

{
  "id": "850e4ea8-a906-4f33-a17c-96635d6495b2",
  "type": "general",
  "status": "completed",
  "title": "Thread Detection: Option A vs B Deep Dive",
  "description": "Detailed comparison of entity-based vs edge-based threading with auto-detection strategies",
  "context": {
    "output": "<div style=\"background:#1e1e1e;color:#e0e0e0;padding:20px;border-radius:8px;font-family:'Consolas','Monaco',monospace;\">\n<h2 style=\"font-family:sans-serif;color:#fff;margin:0 0 16px 0;\">🧵 Thread Detection: Option A vs B Deep Dive</h2>\n\n<h3 style=\"color:#22c55e;margin:16px 0 8px 0;\">Option A: Thread as First-Class Entity</h3>\n\n<div style=\"background:#2d2d2d;padding:16px;border-radius:6px;margin-bottom:16px;\">\n<h4 style=\"color:#98c379;margin:0 0 12px 0;\">✅ Pros</h4>\n<table style=\"width:100%;border-collapse:collapse;\">\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;width:25%;\">Rich Metadata</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Title, trigger reason, resolution notes, status, metrics all live on the thread</td>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;\">Easy Queries</td>\n<td style=\"padding:8px;color:#e0e0e0;\"><code>SELECT * FROM threads WHERE flowId = X AND status = 'resolved'</code> - show all rabbit holes</td>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;\">Aggregation</td>\n<td style=\"padding:8px;color:#e0e0e0;\">\"This week: 3 threads, 15 turns, 2.1M tokens on detours\" - easy to compute</td>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;\">UI Grouping</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Timeline can collapse thread kontasks under a single \"Mobile Padding Bug\" header</td>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;\">Learning</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Can tag threads with root cause (\"CSS cascade\", \"missing auth\") for pattern recognition</td>\n</tr>\n<tr>\n<td style=\"padding:8px;color:#e5c07b;\">Status Tracking</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Thread can be \"ongoing\", \"resolved\", \"abandoned\" - clear lifecycle</td>\n</tr>\n</table>\n\n<h4 style=\"color:#f87171;margin:16px 0 12px 0;\">❌ Cons</h4>\n<table style=\"width:100%;border-collapse:collapse;\">\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;width:25%;\">New Entity</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Another table, more schema, more CRUD endpoints, more MCP tools</td>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;\">Creation Friction</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Who creates the thread? When? Automatic detection needs to be good</td>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;\">Boundary Ambiguity</td>\n<td style=\"padding:8px;color:#e0e0e0;\">When does a thread end? Resolving the bug? Or when you get back to main work?</td>\n</tr>\n<tr>\n<td style=\"padding:8px;color:#e5c07b;\">Over-Engineering?</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Might be solving a visibility problem with schema when UI grouping could suffice</td>\n</tr>\n</table>\n</div>\n\n<h3 style=\"color:#3b82f6;margin:16px 0 8px 0;\">Option B: Edge-Based Threading</h3>\n\n<div style=\"background:#2d2d2d;padding:16px;border-radius:6px;margin-bottom:16px;\">\n<h4 style=\"color:#98c379;margin:0 0 12px 0;\">✅ Pros</h4>\n<table style=\"width:100%;border-collapse:collapse;\">\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;width:25%;\">No New Entity</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Uses existing graph infrastructure - just add edge type <code>continues</code> or <code>spawned_by</code></td>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;\">Simpler Start</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Can ship quickly, iterate on detection logic without schema migrations</td>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;\">Flexible</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Threads emerge from edge traversal - can recompute groupings differently later</td>\n</tr>\n<tr>\n<td style=\"padding:8px;color:#e5c07b;\">Gradual</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Can promote to full entity later if edges prove insufficient</td>\n</tr>\n</table>\n\n<h4 style=\"color:#f87171;margin:16px 0 12px 0;\">❌ Cons</h4>\n<table style=\"width:100%;border-collapse:collapse;\">\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;width:25%;\">No Thread Identity</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Can't name the thread, can't add \"trigger\" or \"resolution\" notes</td>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;\">Harder Queries</td>\n<td style=\"padding:8px;color:#e0e0e0;\">\"Show me all threads\" requires graph traversal to find connected components</td>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:8px;color:#e5c07b;\">Metrics Scattered</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Total turns/tokens for a thread = sum across all kontasks in chain (computed, not stored)</td>\n</tr>\n<tr>\n<td style=\"padding:8px;color:#e5c07b;\">No Status</td>\n<td style=\"padding:8px;color:#e0e0e0;\">Can't mark a thread as \"resolved\" - it's just a chain of kontasks</td>\n</tr>\n</table>\n</div>\n\n<h3 style=\"color:#f97316;margin:16px 0 8px 0;\">🔍 Auto-Detection Strategies</h3>\n\n<div style=\"background:#2d2d2d;padding:16px;border-radius:6px;margin-bottom:16px;\">\n<h4 style=\"color:#e5c07b;margin:0 0 12px 0;\">Signal 1: Keyword Detection in Prompts/Titles</h4>\n<pre style=\"background:#252526;padding:12px;border-radius:4px;margin:8px 0;line-height:1.6;overflow-x:auto;\">\n<span style=\"color:#7f848e;\">// Detect \"still broken\", \"didn't work\", \"regression\", \"try again\"</span>\n<span style=\"color:#c678dd;\">const</span> THREAD_TRIGGER_PATTERNS = [\n  <span style=\"color:#98c379;\">/still (broken|not working|failing)/i</span>,\n  <span style=\"color:#98c379;\">/didn't (work|fix|help)/i</span>,\n  <span style=\"color:#98c379;\">/regression/i</span>,\n  <span style=\"color:#98c379;\">/try (again|different|another)/i</span>,\n  <span style=\"color:#98c379;\">/same (issue|problem|error)/i</span>,\n  <span style=\"color:#98c379;\">/why is.*(still|again)/i</span>,\n];\n\n<span style=\"color:#7f848e;\">// If prompt matches AND references recent kontask topic → thread continuation</span>\n</pre>\n</div>\n\n<div style=\"background:#2d2d2d;padding:16px;border-radius:6px;margin-bottom:16px;\">\n<h4 style=\"color:#e5c07b;margin:0 0 12px 0;\">Signal 2: Same File/Component Within Time Window</h4>\n<pre style=\"background:#252526;padding:12px;border-radius:4px;margin:8px 0;line-height:1.6;overflow-x:auto;\">\n<span style=\"color:#7f848e;\">// If last 3 kontasks all touched the same file(s)</span>\n<span style=\"color:#c678dd;\">function</span> detectFileCluster(recentKontasks) {\n  <span style=\"color:#c678dd;\">const</span> files = recentKontasks.flatMap(k => k.filesModified);\n  <span style=\"color:#c678dd;\">const</span> counts = countOccurrences(files);\n  \n  <span style=\"color:#7f848e;\">// If same file appears 3+ times in last 5 kontasks → likely a thread</span>\n  <span style=\"color:#c678dd;\">return</span> Object.entries(counts)\n    .filter(([file, count]) => count >= 3);\n}\n\n<span style=\"color:#7f848e;\">// Example: layout.ts modified in 4 consecutive kontasks → thread</span>\n</pre>\n</div>\n\n<div style=\"background:#2d2d2d;padding:16px;border-radius:6px;margin-bottom:16px;\">\n<h4 style=\"color:#e5c07b;margin:0 0 12px 0;\">Signal 3: Topic Stability + Bug Tags</h4>\n<pre style=\"background:#252526;padding:12px;border-radius:4px;margin:8px 0;line-height:1.6;overflow-x:auto;\">\n<span style=\"color:#7f848e;\">// Kontask tags: [\"bug\", \"regression\", \"fix\", \"broken\"]</span>\n<span style=\"color:#7f848e;\">// If 2+ consecutive kontasks have bug-related tags → thread</span>\n\n<span style=\"color:#c678dd;\">const</span> BUG_TAGS = [<span style=\"color:#98c379;\">'bug'</span>, <span style=\"color:#98c379;\">'regression'</span>, <span style=\"color:#98c379;\">'fix'</span>, <span style=\"color:#98c379;\">'broken'</span>, <span style=\"color:#98c379;\">'issue'</span>];\n\n<span style=\"color:#c678dd;\">function</span> isBugSequence(kontasks) {\n  <span style=\"color:#c678dd;\">return</span> kontasks.every(k => \n    k.tags.some(t => BUG_TAGS.includes(t))\n  );\n}\n</pre>\n</div>\n\n<div style=\"background:#2d2d2d;padding:16px;border-radius:6px;margin-bottom:16px;\">\n<h4 style=\"color:#e5c07b;margin:0 0 12px 0;\">Signal 4: Temporal Clustering</h4>\n<pre style=\"background:#252526;padding:12px;border-radius:4px;margin:8px 0;line-height:1.6;overflow-x:auto;\">\n<span style=\"color:#7f848e;\">// If kontasks are < 30 min apart AND seem related → thread</span>\n<span style=\"color:#7f848e;\">// \"Related\" = same topic OR same domain OR same files</span>\n\nKontask A: 08:40 - Input Area Fix\nKontask B: 08:52 - Still broken      <span style=\"color:#f97316;\">← 12 min gap, same topic</span>\nKontask C: 08:52 - CSS cascade       <span style=\"color:#f97316;\">← 0 min gap, same files</span>\nKontask D: 09:02 - Finally fixed     <span style=\"color:#f97316;\">← 10 min gap, same files</span>\n\n<span style=\"color:#22c55e;\">→ All 4 form a thread</span>\n</pre>\n</div>\n\n<h3 style=\"color:#c678dd;margin:16px 0 8px 0;\">🖥️ Where Detection Happens</h3>\n\n<div style=\"background:#2d2d2d;padding:16px;border-radius:6px;margin-bottom:16px;\">\n<table style=\"width:100%;border-collapse:collapse;\">\n<tr style=\"border-bottom:1px solid #3d3d3d;background:#252526;\">\n<th style=\"padding:10px;color:#e5c07b;text-align:left;\">Location</th>\n<th style=\"padding:10px;color:#e5c07b;text-align:left;\">When</th>\n<th style=\"padding:10px;color:#e5c07b;text-align:left;\">How</th>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:10px;color:#3b82f6;\">Claude (in-turn)</td>\n<td style=\"padding:10px;color:#e0e0e0;\">During turn, before creating kontask</td>\n<td style=\"padding:10px;color:#e0e0e0;\">Claude sees recent kontasks in context, recognizes \"I'm continuing the same problem\" → adds <code>threadId</code> or <code>continues</code> edge</td>\n</tr>\n<tr style=\"border-bottom:1px solid #3d3d3d;\">\n<td style=\"padding:10px;color:#22c55e;\">Konui (post-turn)</td>\n<td style=\"padding:10px;color:#e0e0e0;\">After kontask created, background job</td>\n<td style=\"padding:10px;color:#e0e0e0;\">Analyzes recent kontasks, detects clusters using signals above, creates edges or threads</td>\n</tr>\n<tr>\n<td style=\"padding:10px;color:#f97316;\">User (manual)</td>\n<td style=\"padding:10px;color:#e0e0e0;\">In Flow Timeline UI</td>\n<td style=\"padding:10px;color:#e0e0e0;\">\"Group these 4 kontasks as a thread\" - manual override for edge cases</td>\n</tr>\n</table>\n</div>\n\n<h3 style=\"color:#a855f7;margin:16px 0 8px 0;\">🎯 UI Surfacing</h3>\n\n<div style=\"background:#2d2d2d;padding:16px;border-radius:6px;margin-bottom:16px;\">\n<h4 style=\"color:#e5c07b;margin:0 0 8px 0;\">Option A (Entity) - Clean Grouping</h4>\n<pre style=\"background:#252526;padding:12px;border-radius:4px;margin:8px 0;line-height:1.6;overflow-x:auto;\">\n<span style=\"color:#f97316;\">▼ 🧵 Mobile Padding Bug</span> <span style=\"color:#7f848e;\">(4 turns • 12 min • 450K tokens)</span>\n  │ <span style=\"color:#7f848e;\">Trigger: Expected quick CSS fix</span>\n  │ <span style=\"color:#22c55e;\">✓ Resolved: CSS cascade order</span>\n  │\n  ├─ 09:02 CSS Cascade Fix\n  ├─ 08:52 Input Area Padding Fix  \n  ├─ 08:52 Still broken...\n  └─ 08:40 Full-Width Input Fix\n</pre>\n\n<h4 style=\"color:#e5c07b;margin:16px 0 8px 0;\">Option B (Edges) - Inferred Grouping</h4>\n<pre style=\"background:#252526;padding:12px;border-radius:4px;margin:8px 0;line-height:1.6;overflow-x:auto;\">\n<span style=\"color:#3b82f6;\">▼ Linked kontasks (detected)</span> <span style=\"color:#7f848e;\">(4 items)</span>\n  │ <span style=\"color:#7f848e;\">← No title, no trigger, no resolution</span>\n  │\n  ├─ 09:02 CSS Cascade Fix         <span style=\"color:#7f848e;\">continues ↓</span>\n  ├─ 08:52 Input Area Padding Fix  <span style=\"color:#7f848e;\">continues ↓</span>\n  ├─ 08:52 Still broken...         <span style=\"color:#7f848e;\">continues ↓</span>\n  └─ 08:40 Full-Width Input Fix\n</pre>\n</div>\n\n<h3 style=\"color:#eab308;margin:16px 0 8px 0;\">💡 Recommendation</h3>\n\n<div style=\"background:#1a2a1a;border:1px solid #22c55e;padding:16px;border-radius:6px;\">\n<p style=\"color:#e0e0e0;margin:0;line-height:1.6;\"><strong style=\"color:#22c55e;\">Start with Option B</strong> (edge-based) to validate detection signals, then <strong style=\"color:#22c55e;\">promote to Option A</strong> (entity) once you know what metadata matters.</p>\n\n<ol style=\"margin:12px 0 0 0;padding-left:20px;line-height:2;\">\n<li><strong>Phase 1:</strong> Add <code>continues</code> edge type, have Claude add it when detecting \"still broken\" patterns</li>\n<li><strong>Phase 2:</strong> Build UI that traverses <code>continues</code> edges to group kontasks</li>\n<li><strong>Phase 3:</strong> If you find yourself wanting to name threads or track resolution → graduate to gvThread entity</li>\n</ol>\n\n<p style=\"color:#7f848e;margin:12px 0 0 0;font-size:0.9rem;\">This lets you iterate on detection logic without committing to schema. The edge data becomes the migration path if you need entities later.</p>\n</div>\n</div>",
    "requestedAt": "2026-01-11T00:45:00.000Z",
    "requestId": "a84d6070-b3c0-4050-9ae4-b184935d4f99",
    "choices": [
      {
        "label": "Start with edges",
        "value": "Add 'continues' edge type and implement detection signals in Claude/konui - validate before adding entity",
        "primary": true
      },
      {
        "label": "Go straight to entity",
        "value": "Design full gvThread schema now - we know we want rich metadata like trigger, resolution, and aggregate metrics"
      },
      {
        "label": "Claude detection first",
        "value": "Start by having Claude detect and add threadId to kontasks based on prompt analysis - simplest starting point"
      },
      {
        "label": "Defer to backlog",
        "value": "This is good exploration but let's finish current flow/topic work first before adding threading"
      }
    ],
    "turnTiming": {
      "totalMs": 80613,
      "ttfeMs": 50,
      "thinkingMs": 80563,
      "toolExecutionMs": 0,
      "toolCallCount": 0,
      "thinkingPct": 100,
      "toolsPct": 0,
      "ttfePct": 0
    }
  },
  "createdBy": "claude",
  "createdAt": "2026-01-10T23:39:02.309Z",
  "updatedAt": "2026-01-10T23:39:07.522Z",
  "requestId": "a84d6070-b3c0-4050-9ae4-b184935d4f99",
  "scope": "greatvibe",
  "tags": [
    "design",
    "flows",
    "schema",
    "threading",
    "architecture"
  ],
  "sessionId": "ses_40728090-b86"
}

Rich Metadata	Title, trigger reason, resolution notes, status, metrics all live on the thread
Easy Queries	`SELECT * FROM threads WHERE flowId = X AND status = 'resolved'` - show all rabbit holes
Aggregation	"This week: 3 threads, 15 turns, 2.1M tokens on detours" - easy to compute
UI Grouping	Timeline can collapse thread kontasks under a single "Mobile Padding Bug" header
Learning	Can tag threads with root cause ("CSS cascade", "missing auth") for pattern recognition
Status Tracking	Thread can be "ongoing", "resolved", "abandoned" - clear lifecycle

New Entity	Another table, more schema, more CRUD endpoints, more MCP tools
Creation Friction	Who creates the thread? When? Automatic detection needs to be good
Boundary Ambiguity	When does a thread end? Resolving the bug? Or when you get back to main work?
Over-Engineering?	Might be solving a visibility problem with schema when UI grouping could suffice

No New Entity	Uses existing graph infrastructure - just add edge type `continues` or `spawned_by`
Simpler Start	Can ship quickly, iterate on detection logic without schema migrations
Flexible	Threads emerge from edge traversal - can recompute groupings differently later
Gradual	Can promote to full entity later if edges prove insufficient

No Thread Identity	Can't name the thread, can't add "trigger" or "resolution" notes
Harder Queries	"Show me all threads" requires graph traversal to find connected components
Metrics Scattered	Total turns/tokens for a thread = sum across all kontasks in chain (computed, not stored)
No Status	Can't mark a thread as "resolved" - it's just a chain of kontasks