◈ Observify

{
  "questions": [
    {
      "header": "Arm guard",
      "multiSelect": false,
      "options": [
        {
          "description": "Add the PreToolUse block to settings.json. From then on Article I is mechanical, not remembered — I physically cannot author code again. Evidence for safety: 11/11 selftest, 4-way proof including a delegate-write ALLOW, and it fails closed on unknown paths. Undo is deleting the block; the hook logs every decision to ~/.claude/hooks/article1-conductor-guard.log.",
          "label": "Arm it now (Recommended)"
        },
        {
          "description": "If the guard supports a warn/observe mode, run it for a day logging what it WOULD have blocked, without blocking anything. Zero risk of halting work, and it produces evidence about false positives before it has teeth. Costs a day of it not actually preventing anything — and tonight showed the discipline fails without teeth.",
          "label": "Arm it in log-only mode first"
        },
        {
          "description": "Keep it installed and proven but inactive; arm it when you have time to watch the first hours. Nothing changes today — including that I'm still relying on the same discipline that failed twice tonight.",
          "label": "Leave it unarmed for now"
        }
      ],
      "question": "The Article I guard is installed and passes 11/11 selftests, but it is INERT until you add it to ~/.claude/settings.json. Armed, it intercepts every Write/Edit/NotebookEdit on this machine and denies conductor writes outside the permitted set (dispatch packets, directives, audit records, memory). It correctly allows delegate writes, so it should not halt the fleet — but a bad matcher on Write|Edit is a machine-wide blast radius, which is why codex left it to you."
    }
  ]
}

{
  "file_path": "/Users/wolf/Projects/observify/_work/spend-meter/.packet.md",
  "new_string": "## DELIVERABLE 0 — ARM THE ARTICLE I GUARD (do this FIRST, it is small)\n\nThe Founder ratified arming it at 03:05. `~/.claude/hooks/article1-pretool-guard.py` is\ninstalled and passes 11/11 selftests but is inert. Add its `PreToolUse` block to\n`~/.claude/settings.json` — the exact block is in\n`/Users/wolf/Projects/observify/_work/article1-enforcement/settings-snippet.json`, but\n**repoint it at the installed path** `~/.claude/hooks/article1-pretool-guard.py` (the snippet\nstill names the scratch `_work/` copy, which will be cleaned).\n- Back up `settings.json` first; it is the Founder's file.\n- Validate the JSON parses before and after (`python3 -m json.tool`); a malformed settings.json\n breaks every session on this machine.\n- Merge into any existing `hooks` block — do NOT clobber the `PreCompact`/`SessionStart` hooks\n already configured there.\n- After arming, verify: pipe a conductor-style write at the hook and confirm `deny`, and pipe a\n permitted path and confirm `allow`. Report both.\n\n## THE INCIDENT (established from disk, 2026-08-08)",
  "old_string": "## THE INCIDENT (established from disk, 2026-08-08)",
  "replace_all": false
}

cd /Users/wolf/Projects/observify sed -i '' 's|Do not arm anything in$|(Deliverable 0 is the ONE authorized settings.json edit.) Do not otherwise touch|' _work/spend-meter/.packet.md 2>/dev/null VENDOR_CODEX_SANDBOX=danger-full-access VENDOR_MAX_S=2700 \ bash /Users/wolf/Projects/Kit/think-like-fable/tools/dispatch.sh start codex spend-meter \ /Users/wolf/Projects/observify/_work/spend-meter gpt-5.6-sol 2>&1 | tail -2

SD=$HOME/.kit-vendor-runs/vendor-runs/spend-meter for i in $(seq 1 90); do if [ -f "$SD/exit" ]; then echo "[$(date +%H:%M:%S)] FINISHED exit=$(cat $SD/exit)"; break; fi kill -0 $(cat $SD/pid 2>/dev/null) 2>/dev/null || { echo "[$(date +%H:%M:%S)] DIED no exit"; break; } echo "[$(date +%H:%M:%S)] running $(stat -f%z $SD/raw 2>/dev/null)b" sleep 30 done

tail -2 /private/tmp/claude-501/-Users-wolf-Projects-observify/c971db0d-fe85-4f5f-bb76-c8f5aa940ea3/tasks/bu127adlh.output echo "=== guard armed? ==="; python3 -c " import json;s=json.load(open('/Users/wolf/.claude/settings.json')) h=s.get('hooks',{});print(' hook types:',list(h.keys())) pt=h.get('PreToolUse',[]);print(' PreToolUse entries:',len(pt)) for e in pt: print(' matcher:',e.get('matcher'),'->',(e.get('hooks') or [{}])[0].get('command','')[:70])" echo "=== spend meter ==="; ls -l ~/.observer-watch/*spend* ~/.observer-watch/*cost* 2>/dev/null | awk '{print " ",$9,$5"b"}' || echo " none" echo "=== canon ==="; grep -c '^ - id: L' /Users/wolf/Projects/Kit/think-like-fable/kernel/lessons.yaml

ls ~/.observer-watch/ | grep -iE "spend|budget|cost" ; echo "--- meter armed? ---"; for f in ~/.observer-watch/*spend*.pid; do p=$(cat $f 2>/dev/null); ps -p $p -o pid,ppid,etime 2>/dev/null|tail -1; done 2>/dev/null echo "--- guard log (is it firing?) ---"; tail -3 ~/.claude/hooks/article1-conductor-guard.log 2>/dev/null || echo " no decisions logged yet" echo "--- L218 ---"; grep -A1 "id: L218" /Users/wolf/Projects/Kit/think-like-fable/kernel/lessons.yaml | head -2

{
  "content": "# Deliberate Article I violation attempt — conductor authoring code.\n# If this file exists, the armed guard did NOT intercept a real Write tool call\n# and is dark despite passing its synthetic tests. Delete on sight.\nprint(\"guard did not fire\")\n",
  "file_path": "/Users/wolf/Projects/observify/_guard-live-test.py"
}