Analyze Transcripts

Review captured worker session transcripts to identify:

• Skills/agents that were or weren't used
• Patterns of confusion or getting stuck
• Follow-up work that was missed
• Workflow improvements

Usage

# Analyze all transcripts
/conductor:analyze-transcripts

# Analyze specific issue(s)
/conductor:analyze-transcripts TabzChrome-abc TabzChrome-def

# Analyze recent transcripts (last N)
/conductor:analyze-transcripts --recent 5

Transcript Location

Transcripts are saved to .beads/transcripts/<issue-id>.txt by wave-done before sessions are killed.

Analysis Steps

1. Find Transcripts

echo "=== Finding Transcripts ==="

TRANSCRIPT_DIR=".beads/transcripts"
if [ ! -d "$TRANSCRIPT_DIR" ]; then
  echo "No transcripts found at $TRANSCRIPT_DIR"
  exit 0
fi

# List available transcripts
ls -la "$TRANSCRIPT_DIR"/*.txt 2>/dev/null | head -20

2. Read and Analyze

For each transcript, analyze:

a) Skill Usage

# Check if skills were invoked
grep -E "^/[a-z]+-[a-z]+|Skill\(" "$TRANSCRIPT_FILE" || echo "No skill invocations found"

Look for:

• /skill-name:skill-name invocations
• Skill(...) tool calls
• References to skill content being loaded

b) Confusion Patterns

# Look for signs of confusion or retries
grep -iE "error|failed|retry|stuck|confused|unclear|not sure" "$TRANSCRIPT_FILE" | head -10

Patterns indicating issues:

• Multiple retries of same action
• "I'm not sure how to..."
• Build/test failures requiring multiple fixes
• Long pauses before action

c) Missed Follow-ups

# Look for TODOs, FIXMEs mentioned but not tracked
grep -E "TODO|FIXME|should also|need to|later|follow.?up" "$TRANSCRIPT_FILE" | head -10

Check if worker:

• Mentioned future work but didn't create issues
• Found edge cases but didn't track them
• Noted improvements but moved on

d) Tool Usage

# Check which tools were used
grep -E "Read|Write|Edit|Bash|Glob|Grep" "$TRANSCRIPT_FILE" | wc -l

3. Generate Report

## Transcript Analysis: ISSUE-ID

### Skill Usage
- [ ] Skills invoked at start: (list)
- [ ] Skills used during work: (list)
- [ ] Skills that should have been used: (list)

### Worker Effectiveness
- Time to first action: ~X minutes
- Build/test retries: N
- Confusion points: (describe)

### Missed Opportunities
- Follow-ups not created: (list)
- TODOs left untracked: (list)

### Recommendations
1. ...
2. ...

Batch Analysis

For analyzing patterns across multiple transcripts:

echo "=== Batch Analysis ==="

# Aggregate token usage and costs
echo "=== Token & Cost Summary ==="
TOTAL_COST=0
TOTAL_IN=0
TOTAL_OUT=0
TOTAL_TOOLS=0

for f in .beads/transcripts/*.txt; do
  [ -f "$f" ] || continue
  COST=$(grep -oE 'Cost: \$[0-9.]+' "$f" 2>/dev/null | head -1 | grep -oE '[0-9.]+' || echo "0")
  IN=$(grep -oE 'Tokens In: [0-9]+' "$f" 2>/dev/null | head -1 | grep -oE '[0-9]+' || echo "0")
  OUT=$(grep -oE 'Tokens Out: [0-9]+' "$f" 2>/dev/null | head -1 | grep -oE '[0-9]+' || echo "0")
  TOOLS=$(grep -oE 'Tool Calls: [0-9]+' "$f" 2>/dev/null | head -1 | grep -oE '[0-9]+' || echo "0")

  TOTAL_COST=$(echo "$TOTAL_COST + $COST" | bc 2>/dev/null || echo "$TOTAL_COST")
  TOTAL_IN=$((TOTAL_IN + IN))
  TOTAL_OUT=$((TOTAL_OUT + OUT))
  TOTAL_TOOLS=$((TOTAL_TOOLS + TOOLS))

  echo "$(basename "$f" .txt): \$${COST}, ${IN}/${OUT} tokens, ${TOOLS} tools"
done

echo ""
echo "TOTALS: \$${TOTAL_COST}, ${TOTAL_IN} in / ${TOTAL_OUT} out, ${TOTAL_TOOLS} tool calls"
echo ""

# Count skill usage across all transcripts
echo "=== Skill Invocation Frequency ==="
grep -rh "^/[a-z]" .beads/transcripts/ 2>/dev/null | sort | uniq -c | sort -rn | head -10

# Common errors
echo ""
echo "=== Common Errors ==="
grep -rhi "error\|failed" .beads/transcripts/ 2>/dev/null | sort | uniq -c | sort -rn | head -10

# Workers that didn't use skills
echo ""
echo "=== Transcripts With No Skill Invocations ==="
for f in .beads/transcripts/*.txt; do
  if ! grep -qE "^/[a-z]+-[a-z]+" "$f" 2>/dev/null; then
    basename "$f" .txt
  fi
done

# Most expensive issues
echo ""
echo "=== Most Expensive Issues ==="
for f in .beads/transcripts/*.txt; do
  [ -f "$f" ] || continue
  COST=$(grep -oE 'Cost: \$[0-9.]+' "$f" 2>/dev/null | head -1 | grep -oE '[0-9.]+' || echo "0")
  echo "$COST $(basename "$f" .txt)"
done | sort -rn | head -5

AI-Assisted Analysis

For deeper analysis, use Codex:

# Analyze transcript with Codex
TRANSCRIPT=$(cat .beads/transcripts/ISSUE-ID.txt)

mcp-cli call codex/review "$(jq -n --arg content "$TRANSCRIPT" '{
  prompt: "Analyze this Claude worker session transcript. Identify: 1) Were skills properly used? 2) What caused confusion? 3) What follow-up work was missed? 4) How could the prompt have been better?",
  content: $content
}')"

Improvement Actions

Based on analysis, update:

Finding	Action
Skills not invoked	Update prompt template to be more explicit
Same errors repeated	Add to worker-done error handling
Follow-ups missed	Strengthen create-followups step
Confusion on X	Add guidance to bd-swarm prompts

Related

• /conductor:wave-done - Captures transcripts before killing sessions
• /conductor:create-followups - Track discovered work
• .beads/transcripts/ - Transcript storage location