compact

Compact - Context Compaction

Overview

Long engagements burn context fast. This skill compresses completed phase outputs into a single engagement_summary.md without losing any certified intelligence.

Rule: session.json is always the authoritative source of truth. After compaction, the AI re-reads session.json instead of relying on conversation history.

---

When to Trigger

Auto-trigger at these moments:

• After completing any full phase (recon, secrets, exploit, zerodayhunt)
• When session.json signals[] exceeds 30 entries
• When the user asks explicitly
• Before spawning a new fork (give the fork a clean context budget)

---

Steps

1. Get target:

   TARGET=$1
   SESSION=~/pentest-toolkit/results/$TARGET/session.json
   RESULTS=~/pentest-toolkit/results/$TARGET
   

2. Identify completed phases:

   ls $RESULTS/interesting_*.md 2>/dev/null
   

3. For each completed phase - compress to 5 bullet signals:

   Read each interesting_<phase>.md and distill to essential intelligence only:

   RECON SIGNALS (from interesting_recon.md):
   - Live hosts: <N> discovered, key: <most interesting>
   - Tech stack: <comma-separated confirmed tech>
   - WAF: <vendor or null>
   - Key endpoints: <top 3>
   - Hypothesis calibration: H1 <N>%, H2 <N>%
   
   SECRETS SIGNALS (from interesting_secrets.md):
   - Credentials found: <Y/N>, type: <aws_key|jwt|password>
   - Key files: <source locations>
   - Cloud provider confirmed: <Y/N>
   
   EXPLOIT SIGNALS (from interesting_exploit.md):
   - Confirmed findings: <N>, classes: <list>
   - SSRF vectors: <Y/N>, endpoints: <list>
   - Potential findings: <N>
   

4. Write engagement_summary.md:

   cat > $RESULTS/engagement_summary.md << EOF
   # Engagement Summary: $TARGET
   Compacted: $(date +%Y-%m-%d\ %H:%M)
   Source of truth: session.json
   
   ## Phase Intelligence (Compressed)
   <5-bullet summaries per completed phase>
   
   ## Active Hypotheses
   <from session.json hypotheses[] - all active ones with current probability>
   
   ## Confirmed Findings
   <from session.json report_draft.findings[] - title + severity + status per finding>
   
   ## Open Discovery Queue
   <from session.json discovery_queue[] - surface + priority per item>
   
   ## Signal Summary
   Total signals: <N>
   Critical signals: <list VULN_CONFIRMED + CRED_FOUND entries>
   
   ## Next Recommended Action
   <based on engagement_state and top hypothesis probability>
   EOF