feat(templates): add governance and approval gate templates (Paperclip pattern)

Session 4 step 17 — 5 autonomy levels (0-4), PreToolUse approval-gate
hook polls approval-responses.jsonl with 60s timeout, blocks on no-response.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Kjell Tore Guttormsen 2026-04-12 06:55:33 +02:00
commit 912689f3c5
3 changed files with 277 additions and 0 deletions

View file

@ -0,0 +1,40 @@
# Governance: {{PROJECT_NAME}}
## Autonomy Levels
- Level 0: Full manual approval (all tool calls require human OK)
- Level 1: Auto-approve safe operations (Read, Glob, Grep)
- Level 2: Auto-approve file operations (+ Write, Edit within project)
- Level 3: Auto-approve all except destructive (+ Bash non-destructive)
- Level 4: Full autonomy with hooks as guardrails
Current level: {{AUTONOMY_LEVEL}}
## Approval Gates
Gates are checkpoints where the agent MUST pause and request human approval.
- {{GATE_1_NAME}}: {{GATE_1_CONDITION}}
Action: {{GATE_1_ACTION}}
- {{GATE_2_NAME}}: {{GATE_2_CONDITION}}
Action: {{GATE_2_ACTION}}
## Escalation Rules
- Budget exceeded: pause agent, notify via {{NOTIFICATION_METHOD}}
- Error threshold: after {{ERROR_THRESHOLD}} consecutive errors, pause agent
- Unknown tool call: block and log
- Scope violation: block and notify
## Audit Requirements
- All tool calls logged to audit.log
- Budget events logged to cost-events.jsonl
- Approval decisions logged to approvals.log
- Retention: {{LOG_RETENTION_DAYS}} days
## Philosophy
Autonomy is a privilege you grant. Start at Level 0 and increase only
when the agent has demonstrated reliable behavior at the current level.
Each level adds capability while hooks maintain the guardrails.

View file

@ -0,0 +1,75 @@
# Governance and Approval Gates
PreToolUse hook implementing human oversight with configurable autonomy levels.
Inspired by Paperclip's "autonomy is a privilege you grant" philosophy.
## Autonomy model
Five levels (04) control how much the agent auto-approves:
| Level | Description | Auto-approves |
|-------|-------------|---------------|
| 0 | Full manual | Nothing — every tool requires human OK |
| 1 | Safe ops | Read, Glob, Grep |
| 2 | File ops | + Write, Edit within project |
| 3 | Non-destructive | + Bash (non-destructive commands) |
| 4 | Full autonomy | Everything — hooks act only as guardrails |
Set `Current level: N` in `GOVERNANCE.md` to change autonomy.
## Approval gates
Gates are conditional checkpoints where the agent pauses and waits for
human approval regardless of autonomy level. Configure them in `GOVERNANCE.md`:
```
- deploy-gate: after all tests pass AND reviewer approved
Action: pause
```
## How approval flow works
1. `approval-gate.sh` runs as a PreToolUse hook
2. If auto-approved (based on level): exits 0, tool proceeds
3. If gated: writes request to `governance/pending-approvals.jsonl`
4. Polls `governance/approval-responses.jsonl` for matching response (by ID)
5. Response `approve` → allow (exit 0), log to `approvals.log`
6. Response `deny` → block (exit 2), log denial
7. No response within timeout (default 60s) → block with timeout message
## Responding to approval requests
When an agent is waiting for approval, add to `governance/approval-responses.jsonl`:
```json
{"id": "<request-id>", "decision": "approve"}
```
or
```json
{"id": "<request-id>", "decision": "deny"}
```
## Audit trail
All tool calls are logged to `governance/audit.log`.
All approval decisions are logged to `governance/approvals.log`.
## Integration
Add to `.claude/settings.json`:
```json
{
"hooks": {
"PreToolUse": [{
"matcher": "*",
"hooks": [{"type": "command", "command": "bash governance/approval-gate.sh"}]
}]
}
}
```
## Paperclip comparison
Paperclip stores approvals in a database table with async notification.
This implementation uses file-based polling — simpler, no service dependency,
suitable for single-machine deployments. For multi-agent systems that need
concurrent approval routing, consider upgrading to a queue or webhook approach.

View file

@ -0,0 +1,162 @@
#!/bin/bash
# PreToolUse hook: Implement approval gates based on GOVERNANCE.md.
# Bash 3.2 compatible. Uses python3 for JSON/MD parsing.
#
# Follows Paperclip's approval mechanism:
# 1. Read GOVERNANCE.md for current autonomy level and gate definitions
# 2. Auto-approve or require human approval based on level
# 3. Write pending approval requests; check for responses
# 4. Timeout with no response → block
#
# Placeholders:
# {{WORKING_DIR}} - absolute path to project directory
WORKING_DIR="{{WORKING_DIR}}"
GOVERNANCE_DIR="$WORKING_DIR/governance"
GOVERNANCE_FILE="$WORKING_DIR/GOVERNANCE.md"
PENDING_FILE="$GOVERNANCE_DIR/pending-approvals.jsonl"
RESPONSES_FILE="$GOVERNANCE_DIR/approval-responses.jsonl"
AUDIT_LOG="$GOVERNANCE_DIR/audit.log"
APPROVALS_LOG="$GOVERNANCE_DIR/approvals.log"
APPROVAL_TIMEOUT=60
mkdir -p "$GOVERNANCE_DIR"
# Read hook input
INPUT=$(cat)
# Auto-approve tools based on autonomy level
DECISION=$(GOVERNANCE_FILE="$GOVERNANCE_FILE" python3 -c "
import re, json, sys, os
governance_file = os.environ.get('GOVERNANCE_FILE', '')
try:
data = json.loads('''$INPUT''')
except:
print('approve')
sys.exit(0)
tool_name = data.get('tool_name', '')
# Read governance policy
if not os.path.exists(governance_file):
print('approve')
sys.exit(0)
content = open(governance_file).read()
level_m = re.search(r'Current level:\s*(\d+)', content)
level = int(level_m.group(1)) if level_m else 0
# Safe read-only tools (always approved at level 1+)
read_tools = ['Read', 'Glob', 'Grep', 'LS']
# File operation tools (approved at level 2+)
file_tools = ['Write', 'Edit', 'MultiEdit']
# Non-destructive bash (approved at level 3+)
# Level 4: everything auto-approved
if level >= 4:
print('approve')
elif level >= 3 and tool_name not in ['Bash']:
print('approve')
elif level >= 2 and tool_name in read_tools + file_tools:
print('approve')
elif level >= 1 and tool_name in read_tools:
print('approve')
else:
print('gate')
" 2>/dev/null)
# Log every tool call to audit log
python3 -c "
import json, time, os
try:
data = json.loads('''$INPUT''')
tool_name = data.get('tool_name', 'unknown')
entry = time.strftime('%Y-%m-%dT%H:%M:%SZ', time.gmtime()) + ' TOOL ' + tool_name + ' decision=$DECISION'
with open('$AUDIT_LOG', 'a') as f:
f.write(entry + '\n')
except:
pass
" 2>/dev/null
if [ "$DECISION" = "approve" ]; then
exit 0
fi
# Gate: write pending approval request
REQUEST_ID=$(python3 -c "import time; print(str(int(time.time())))" 2>/dev/null)
python3 -c "
import json, time, os
try:
data = json.loads('''$INPUT''')
tool_name = data.get('tool_name', 'unknown')
tool_input = data.get('tool_input', {})
req = {
'id': '$REQUEST_ID',
'timestamp': time.strftime('%Y-%m-%dT%H:%M:%SZ', time.gmtime()),
'tool_name': tool_name,
'tool_input_summary': str(tool_input)[:200],
'status': 'pending'
}
with open('$PENDING_FILE', 'a') as f:
f.write(json.dumps(req) + '\n')
print('Approval required for: ' + tool_name)
print('Request ID: $REQUEST_ID')
print('Add response to: $RESPONSES_FILE')
print('Format: {\"id\": \"$REQUEST_ID\", \"decision\": \"approve\"}')
except:
pass
" >&2
# Poll for response
SECONDS_WAITED=0
while [ "$SECONDS_WAITED" -lt "$APPROVAL_TIMEOUT" ]; do
if [ -f "$RESPONSES_FILE" ]; then
RESPONSE=$(python3 -c "
import json, os
req_id = '$REQUEST_ID'
responses_file = '$RESPONSES_FILE'
try:
with open(responses_file) as f:
for line in f:
line = line.strip()
if line:
try:
r = json.loads(line)
if r.get('id') == req_id:
print(r.get('decision', 'deny'))
exit(0)
except:
pass
except:
pass
print('pending')
" 2>/dev/null)
if [ "$RESPONSE" = "approve" ]; then
python3 -c "
import json, time
entry = {'id': '$REQUEST_ID', 'timestamp': '$(date -u +%Y-%m-%dT%H:%M:%SZ)', 'decision': 'approved'}
with open('$APPROVALS_LOG', 'a') as f:
f.write(json.dumps(entry) + '\n')
" 2>/dev/null
exit 0
elif [ "$RESPONSE" = "deny" ]; then
python3 -c "
import json, time
entry = {'id': '$REQUEST_ID', 'timestamp': '$(date -u +%Y-%m-%dT%H:%M:%SZ)', 'decision': 'denied'}
with open('$APPROVALS_LOG', 'a') as f:
f.write(json.dumps(entry) + '\n')
" 2>/dev/null
echo '{"decision": "block", "reason": "Approval denied by operator."}'
exit 2
fi
fi
sleep 5
SECONDS_WAITED=$(( SECONDS_WAITED + 5 ))
done
# Timeout — block
echo "Approval timeout after ${APPROVAL_TIMEOUT}s — blocking tool call." >&2
echo '{"decision": "block", "reason": "Approval timeout. Check governance/pending-approvals.jsonl."}'
exit 2