extract（提取）

Extract Skill

Typically runs automatically via SessionStart hook.

Process pending learning extractions from previous sessions.

How It Works

The SessionStart hook runs:

ao extract

This checks for queued extractions and outputs prompts for Claude to process.

Manual Execution

Given /extract:

Step 1: Check for Pending Extractions

ao extract 2>/dev/null

Or check the pending queue:

cat .agents/ao/pending.jsonl 2>/dev/null | head -5

Step 1.5: Without ao CLI — Manual Extraction

If ao CLI is not available, process the pending queue manually:

if ! command -v ao &>/dev/null; then
  echo "ao CLI not available — running manual extraction"

  # Check for pending queue
  if [ -f .agents/ao/pending.jsonl ] && [ -s .agents/ao/pending.jsonl ]; then
    echo "Found pending extractions:"
    cat .agents/ao/pending.jsonl

    # For each pending entry, check for corresponding forge output
    # Forge outputs live in .agents/forge/
    for forge_file in .agents/forge/*.md; do
      [ -f "$forge_file" ] || continue
      echo "Processing: $forge_file"
    done
  else
    echo "No pending extractions found."
  fi

  # After processing, check .agents/forge/ for unprocessed candidates
  FORGE_COUNT=$(ls .agents/forge/*.md 2>/dev/null | wc -l | tr -d ' ')
  if [ "$FORGE_COUNT" -gt 0 ]; then
    echo "$FORGE_COUNT forge candidates found — review and extract learnings manually."
    echo "For each candidate in .agents/forge/:"
    echo "  1. Read the candidate file"
    echo "  2. Extract actionable learnings using the template in Step 3"
    echo "  3. Write to .agents/learnings/YYYY-MM-DD-<topic>.md"
    echo "  4. High-confidence items (>= 0.7) can be promoted directly"
  fi
fi

For each forge candidate, extract learnings using the same template format defined in Step 3 of this skill. Write results to .agents/learnings/. After processing, clear the pending queue:

# Clear processed entries
> .agents/ao/pending.jsonl
echo "Pending queue cleared"

Step 2: Process Each Pending Item

For each queued session:

1. Read the session summary
2. Extract actionable learnings
3. Write to .agents/learnings/

Step 3: Write Learnings

Write to: .agents/learnings/YYYY-MM-DD-<session-id>.md

# Learning: <Short Title>

**ID**: L1
**Category**: <debugging|architecture|process|testing|security>
**Confidence**: <high|medium|low>

## What We Learned

<1-2 sentences describing the insight>

## Why It Matters

<1 sentence on impact/value>

## Source

Session: <session-id>

Step 3.5: Validate Learnings

After writing learning files, validate each has required fields:

1. Scan newly written files:

ls -t .agents/learnings/YYYY-MM-DD-*.md 2>/dev/null | head -5

1. For each file, check required fields:

- Heading: File must start with # Learning: <title> (non-empty title) - Category: Must contain **Category**: <value> where value is one of: debugging, architecture, process, testing, security - Confidence: Must contain **Confidence**: <value> where value is one of: high, medium, low - Content: Must contain a ## What We Learned section with at least one non-empty line after the heading

1. Report validation results:

- For each valid learning: "✓: valid" - For each invalid learning: "⚠: missing " (list each missing field)

1. Do NOT delete or retry invalid learnings. Log the warning and proceed. Invalid learnings are still better than no learnings — the warning helps identify extraction quality issues over time.

Step 4: Clear the Queue

ao extract --clear 2>/dev/null

Step 5: Report Completion

Tell the user:

• Number of learnings extracted
• Key insights
• Location of learning files

The Knowledge Loop

Session N ends:
  → ao forge --last-session --queue
  → Session queued in pending.jsonl

Session N+1 starts:
  → ao extract (this skill)
  → Claude processes the queue
  → Writes to .agents/learnings/
  → Validates required fields
  → Loop closed

Key Rules

• Runs automatically - usually via hook
• Process the queue - don't leave extractions pending
• Be specific - actionable learnings, not vague observations
• Close the loop - extraction completes the knowledge cycle

Examples

SessionStart Hook Invocation

Hook triggers: session-start.sh runs at session start

What happens:

1. Hook calls ao extract 2>/dev/null
2. CLI outputs queued session IDs and prompts
3. Agent processes each pending extraction
4. Agent writes learnings to .agents/learnings/<date>-<session>.md
5. Agent validates required fields and reports results
6. Hook calls ao extract --clear to empty queue

Result: Prior session knowledge automatically extracted at session start without user action.

Manual Extraction Trigger

User says: /extract or "extract learnings from last session"

What happens:

1. Agent checks pending queue with ao extract
2. Agent reads session summaries from queue
3. Agent extracts decisions, learnings, failures
4. Agent writes to .agents/learnings/ with proper structure
5. Agent validates fields (category, confidence, content)
6. Agent clears queue and reports completion

Result: Pending extractions processed manually, queue cleared, learnings indexed.

Troubleshooting