Delivery Gate

A dual-layer mechanical gate for Claude Code Stop hooks. Two Python scripts, zero dependencies, deterministic checks only.

config-health（Process Monitor）→ tracks rule execution → soft feedback (never blocks)
         ↓
quality-gate（Output Enforcer）→ checks output completeness → hard block (≥3 stale → exit 2)
         ↓
       Delivery

When to Activate

• Any project where you want Claude to consistently capture learning across sessions
• Teams that want mechanical enforcement of output quality (not just trust the AI's word)
• Solo developers who want the system to remind them, not nag them

Installation

# Copy both scripts into your Claude Code scripts directory
cp config-health.py ~/.claude/scripts/
cp quality-gate.py ~/.claude/scripts/

Add to ~/.claude/settings.json (order matters — config-health first):

{
  "hooks": {
    "Stop": [
      {
        "type": "command",
        "command": "python3 ~/.claude/scripts/config-health.py --hook",
        "timeout": 5000
      },
      {
        "type": "command",
        "command": "python3 ~/.claude/scripts/quality-gate.py",
        "timeout": 5000
      }
    ]
  }
}

Manual health dashboard:

python3 ~/.claude/scripts/config-health.py --check

How It Works

config-health.py — Process Monitor (soft, never blocks)

quality-gate.py — Output Enforcer (hard, blocks when stale)

The Dual-Layer Boundary

The boundary between soft and hard is not "importance" — it's "can this be fixed later?"

• Missed rule markers → can retroactively mark and count → soft
• Missed output records → lost forever → hard

Customization

config-health.py

• RULES — rules to track with regex markers
• VERIFICATION_WINDOW — sessions needed for a rule to "pass"
• MIN_TOOL_CALLS_FOR_CHECK — minimum tools to consider session complex
• MIN_EDITS_FOR_DELIVERY — minimum edits to expect DELIVERY marker

quality-gate.py

• RATIONALIZE — regex patterns for rationalization detection
• LIBS — files/dirs to check for today's updates
• COMPLEX_THRESHOLD — Edit/Write calls to classify as complex
• DISK_CRIT_GB — block below this

Examples

Normal path — silence

Session: 2 edits, growth-log updated, rules followed
→ config-health: 85% marker rate, wrote to JSONL (exit 0, no stdout)
→ quality-gate: 1/5 stale (exit 0, no stdout)
→ Delivery. Zero tokens consumed.

Rule drift — surfaced, not blocked

Session: rules followed but THINK marker rate dropped to 40%
→ config-health: wrote "THINK: 40%" to pending-verifications.md (exit 0)
→ quality-gate: all libs fresh (exit 0)
→ Delivery. Next session: BODY.md surfaces pending item → AI pays attention.

Output incomplete — BLOCKED

Session: 5 edits, nothing written to memory
→ config-health: normal (exit 0)
→ quality-gate: 5/5 libs stale → exit 2
  stderr: "BLOCK: 5 libraries stale (threshold: ≥3)"
→ Claude CANNOT finish. Must update libraries first.

Limitations

The mechanical gate enforces habits and completeness, not content quality. It checks that you captured learning, not whether the learning is correct. For output consistency checks, pair with self-audit.

Compatibility

• Python 3.8+
• Windows, macOS, Linux
• Zero dependencies beyond stdlib

License

MIT