session-search Skill

Search your complete conversation history stored in job session JSONL files. Use this when you need to find how similar problems were solved in past jobs, reference previous code patterns, or answer questions about past agent work.

Trigger

Use this skill when:

• User asks about "what was done before" or "how was X solved"
• You need to reference solutions from past autonomous jobs
• Searching for code patterns, tools used, or approaches taken
• Reviewing cost/token usage across jobs

Location

Session logs live at: /job/logs/<JOB_ID>/*.jsonl

Each job folder contains:

• job.md - The original task prompt
• *.jsonl - Full conversation transcripts (multiple sessions per job)
• job.config.json - Job configuration

Structure

Each .jsonl file contains messages with:

• type: "session" (metadata), "message", "thinking_level_change", or "model_change"
• timestamp: ISO timestamp
• message.role: "user", "assistant", or "toolResult"
• message.content[]: Text, thinking, or tool calls (filter type=="text" for human-readable content)
• message.usage.cost.total: Cost per response

Commands

List all jobs by date and size

for dir in /job/logs/*/; do
  if [ -d "$dir" ]; then
    job_id=$(basename "$dir")
    first_log=$(ls -t "$dir"/*.jsonl 2>/dev/null | head -1)
    if [ -n "$first_log" ]; then
      date=$(head -1 "$first_log" | jq -r '.timestamp' | cut -dT -f1)
      size=$(du -sh "$dir" | cut -f1)
      echo "$date $size $job_id"
    fi
  fi
done | sort -r

Find jobs from a specific day

for dir in /job/logs/*/; do
  ls "$dir"/*.jsonl 2>/dev/null | while read f; do
    head -1 "$f" | jq -r '.timestamp' 2>/dev/null | grep -q "2026-02-25" && echo "$dir"
  done
done | sort -u

Extract user prompts from a session

jq -r 'select(.message.role == "user") | .message.content[]? | select(.type == "text") | .text' <session>.jsonl

Search for keyword in assistant responses

jq -r 'select(.message.role == "assistant") | .message.content[]? | select(.type == "text") | .text' <session>.jsonl | rg -i "keyword"

Get total cost for a job

cat /job/logs/<JOB_ID>/*.jsonl | jq -s '[.[] | .message.usage.cost.total // 0] | add'

Search across ALL sessions for a phrase

rg -l "phrase" /job/logs/*/*.jsonl

Find jobs where a specific tool was used

for dir in /job/logs/*/; do
  if grep -q '"toolName":"bash"' "$dir"/*.jsonl 2>/dev/null; then
    echo "$(basename "$dir")"
  fi
done

Get tool usage breakdown for a job

cat /job/logs/<JOB_ID>/*.jsonl | jq -r '.message.content[]? | select(.type == "toolCall") | .name' | sort | uniq -c | sort -rn

Find the job prompt for a job

cat /job/logs/<JOB_ID>/job.md

Search job prompts for a topic

rg -l "topic" /job/logs/*/job.md

Find most recent job on a topic

for dir in $(ls -t /job/logs/*/); do
  if rg -q "topic" "$dir"/*.jsonl 2>/dev/null; then
    echo "$dir"
    break
  fi
done

Tips

• Sessions are append-only JSONL (one JSON object per line)
• Large jobs can have many session files (each represents a model switch or thinking level change)
• Use head -1 to sample the first message of a session for metadata
• The job.md file contains the original task prompt - always check this first for context
• Costs accumulate across all session files in a job folder

Fast text-only hint (low noise)

cat /job/logs/<JOB_ID>/*.jsonl | jq -r 'select(.type=="message") | .message.content[]? | select(.type=="text") | .text' | rg 'keyword'

Example: Find how a similar task was solved

# 1. Find jobs about "email" or "send"
rg -l "email|send" /job/logs/*/job.md

# 2. Look at the most recent result
head -50 /job/logs/<found_job_id>/job.md

# 3. Search the session for the solution pattern
cat /job/logs/<found_job_id>/*.jsonl | jq -r 'select(.message.role == "assistant") | .message.content[]? | select(.type == "text") | .text' | rg -A 10 -B 2 "send_email"