Card Queue

Automatic staging buffer that captures AI responses and synthesizes high-quality knowledge card proposals.

Overview

The Card Queue solves a fundamental problem: most valuable insights from AI conversations are never captured because the developer is focused on the task at hand, not on documentation.

ContextManager silently queues every AI response that meets a minimum length threshold. When you’re ready, a single “Distill” action synthesizes all queued responses into knowledge card proposals - surfacing cross-cutting patterns that no single response contains clearly enough on its own.

Cards extracted by the Auto-Capture / Auto-Learn service are also staged here rather than being silently created. Every auto-detected knowledge card and architecture note lands in the queue for your review first.

How It Works 

graph TD
    A[AI Response over 300 chars] --> B[Automatically queued]
    B --> C[Tool call evidence captured]
    C --> D[User clicks Distill into Cards]
    D --> E[Single LLM call synthesizes all]
    E --> F[Card proposals with scores]
    F --> G[One-click Add Card or Approve All]

    style A fill:#1f6feb,stroke:#388bfd,color:#fff
    style B fill:#1158c7,stroke:#388bfd,color:#fff
    style D fill:#9e6a03,stroke:#d29922,color:#fff
    style E fill:#9e6a03,stroke:#d29922,color:#fff
    style G fill:#238636,stroke:#3fb950,color:#fff

Zero-Cost Capture

Unlike previous designs, the Card Queue does not run an LLM call per response to assess worthiness. Every response over the minimum length is silently queued at zero cost. Quality judgement happens once, at distill time.

Cross-Response Synthesis

The distillQueue() pipeline reads the entire card queue in one LLM call and:

  • Identifies cross-cutting themes across multiple responses
  • Generates card proposals with titles, categories, and content
  • Assigns confidence scores (0–1)
  • Maps each proposal to sourceIndices[] showing which responses contributed
  • Surfaces insights that span multiple conversations

Tool Call Evidence

When responses are captured via the agent hook system, tool calls are captured alongside:

Field Description
Tool Name Which tool was used (e.g., read_file, grep_search)
Input Truncated input arguments (≤200 chars per call)
Count Total tool calls per turn (capped at 10)

This evidence helps the AI feel confident that the knowledge was already investigated - reducing redundant codebase exploration when distilling.

Dashboard UI

The Card Queue lives in the Knowledge tab:

  • 📬 Card Queue header with pending count
  • Distill into Cards button - runs the synthesis pipeline
  • Clear Queue - remove all queued items
  • Raw items accordion - collapsible list showing each queued response:
    • Participant source and prompt preview
    • Response character count
    • 🔧 Tool call badge (when present)
    • Rendered response content (collapsible)
    • Tool call details (collapsible monospace list)
    • Per-item ✕ removal

Distill Results

After distilling, results appear inline in the Knowledge tab:

  • Category badge and confidence percentage
  • Content preview
  • Tags (2–5 keywords auto-generated by the LLM)
  • Reasoning explanation
  • Source indices linking to original responses
  • + Add Card button per proposal
  • Approve All for batch creation

Tags are preserved through the entire approval chain — whether approving individual cards or batch-approving all, the LLM-generated tags are stored on the resulting knowledge card.

Smart-Merge on Approval

When you approve any card (from distill results or the raw queue), ContextManager checks existing cards for similarity using Jaccard word-set overlap. If a similar card is found (≥ 30% overlap), a QuickPick appears:

  • Create new card — proceed as normal
  • Merge into: [existing title] (XX% similar) — append the new card’s content to the existing card with a dated separator

This prevents duplicate cards from accumulating as auto-capture and distill run over time.

Settings

Setting Default Description
cardQueue.enabled true Enable automatic response queuing
cardQueue.minResponseLength 300 Minimum response length to queue (50–5000)

Queue Limits

  • Maximum 30 items - oldest items evicted when cap is exceeded (FIFO)
  • Tool calls capped at 10 per response - to control token budget
  • Tool inputs truncated to 200 chars - aggressive truncation for budget control

Pipeline Architecture 

graph TD
    A[capture.ps1] --> B[hook-queue.jsonl]
    B --> C[HookWatcher.processEntry]
    C --> D[queueCardCandidate]
    D --> E[createQueuedCard]
    E --> F[ProjectManager.addToCardQueue]
    F --> G[Dashboard - Intelligence tab]

    style A fill:#9e6a03,stroke:#d29922,color:#fff
    style C fill:#1158c7,stroke:#388bfd,color:#fff
    style F fill:#1158c7,stroke:#388bfd,color:#fff
    style G fill:#238636,stroke:#3fb950,color:#fff

Note: the Auto-Capture service (`autoCapture.ts`) feeds the same queue — auto-detected knowledge cards enter at **`ProjectManager.addToCardQueue`**, never bypassing user review.

Next Steps

Dashboard →