VibeGuard: Learn

Core Concept

• Agent's mistakes are not the end, but a signal to improve the defense system.
• When the Agent discovers non-obvious solutions, it should be extracted into reusable knowledge
• Goal: Mistakes will no longer be repeated + Experience will no longer be forgotten

Pattern routing

Automatically select a mode based on parameters and context:

input	mode
User describes error/bug/guard failure	Mode A: error analysis → output guard/hook/rule
User said "extract" / "Extract experience" / "What was learned"	Mode B: Experience extraction → Output SKILL.md
Stop hook automatically triggered (no parameters)	Evaluate first → select A or B as needed
Session with both bug fixes and non-obvious findings	A + B both executed

---

Mode A: Error analysis (defense-oriented)

Learn from mistakes and generate guard rules or hooks to prevent similar mistakes from happening again.

Trigger scene

• Agent created duplicate files/types (L1 failed)
• Agent hard-coded path/port/URL (L4 failure)
• Agent introduces data splitting (multiple entries are inconsistent)
• Agent is unnecessarily over-designed (L5 failure)
• The guard script has false positives or false negatives
• Any recurring Agent error patterns

Guardrails

• No direct fixes to the business code — only improvements to VibeGuard itself
• New rules must be verifiable (can write scripts to detect or test assertions)
• Don’t overgeneralize — one error pattern corresponds to one precise rule

Steps

1. Automatic pattern recognition (extracted from events.jsonl + learn-digest.jsonl)
   • Read ~/.vibeguard/projects/<hash>/events.jsonl and analyze recent event records
   • Read ~/.vibeguard/learn-digest.jsonl to obtain the cross-session signals recognized by GC regular learning (repeated_warn / chronic_block / hot_files / slow_sessions / warn_escalation)
   • For a bounded current-project preview, resolve the VibeGuard checkout before running the analyzer:

     VIBEGUARD_REPO_DIR="${VIBEGUARD_REPO_DIR:-$(cat "$HOME/.vibeguard/repo-path" 2>/dev/null)}"
     test -f "$VIBEGUARD_REPO_DIR/scripts/gc/learn_digest.py" || { echo "VibeGuard repo path missing; rerun setup.sh from a VibeGuard checkout" >&2; exit 1; }
     python3 "$VIBEGUARD_REPO_DIR/scripts/gc/learn_digest.py" --scope current --project-root "$PWD" --dry-run --format json --no-code-scan
     

   • Preview mode reads only the resolved current project log directory and does not append learn-digest.jsonl or update .learn-watermark
   • Extract high-frequency warn patterns: similar problems that have been warned many times but still reoccur
   • Extract similar operations that are repeatedly blocked: identify the operation patterns in which the agent repeatedly hits the wall
   • Group by hook + reason to output top 5 high-frequency problems
   • Use recognized patterns as input for subsequent analysis (no longer relying solely on user descriptions)