Skip to content

Business operating system for Claude Code — 57 skills, 21 agents, smart daemon. Unified inbox (WhatsApp/Email/Slack/Telegram), autonomous PR merge, full-AWS monitoring, revenue (Stripe+RevenueCat), e-commerce (Shopify), marketing (Klaviyo/Meta/GA4), voice (Bland/ElevenLabs), APM (Datadog/NewRelic/OTEL), YOLO mode.

fix(security): apply net-new audit findings T-002..T-006 + personal-defaults guard (#273)

Aurora Capitalc9335ebd2145
diff --git a/claude-ops/bin/ops-marketing-portfolio b/claude-ops/bin/ops-marketing-portfolio
index 927fc377..1ef598ff 100755
--- a/claude-ops/bin/ops-marketing-portfolio
+++ b/claude-ops/bin/ops-marketing-portfolio
@@ -31,7 +31,7 @@ for ((i=1; i<=$#; i++)); do
--json) JSON_OUT=1 ;;
--project)
j=$((i+1)); FILTER_PROJECT="${!j:-}"
- i=$j
+ i=$((i+1))
;;
esac
done
diff --git a/claude-ops/scripts/ops-cron-pocket-watcher.py b/claude-ops/scripts/ops-cron-pocket-watcher.py
index a048bfa5..4ea41d5f 100755
--- a/claude-ops/scripts/ops-cron-pocket-watcher.py
+++ b/claude-ops/scripts/ops-cron-pocket-watcher.py
@@ -94,6 +94,16 @@ INFER_MIN_SECS = int(os.environ.get("POCKET_INFER_MIN_SECS", "60"))
INFER_CONFIDENCE = float(os.environ.get("POCKET_INFER_CONFIDENCE", "0.7"))
LOG_FILE = STATE_DIR / "run.log"
+_USER_CONTEXT_PATH = Path(os.environ.get(
+ "POCKET_USER_CONTEXT",
+ str(HOME / ".claude/state/pocket/user-context.json"),
+))
+try:
+ _ctx = json.loads(_USER_CONTEXT_PATH.read_text())
+ _OWNER_NAME: str = _ctx.get("owner_name") or "the user"
+except (OSError, json.JSONDecodeError):
+ _OWNER_NAME = "the user"
+
def now_iso() -> str:
return datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
@@ -379,7 +389,7 @@ def infer_tasks_from_recording(recording: dict) -> list[dict]:
"Return STRICT JSON: an array of objects with keys: title (string, "
"imperative form, <80 chars), context (string, 1-2 sentences of why), "
"priority (low|medium|high), confidence (0.0-1.0 — your honest "
- "certainty that this was implied as a task Sam would want done). "
+ f"certainty that this was implied as a task {_OWNER_NAME} would want done). "
"Return [] if nothing actionable was implied. NO prose, NO markdown."
)
@@ -846,7 +856,7 @@ def main() -> int:
# Inferred tasks go to pending-triage.jsonl, NOT directly to
# the live tasks.jsonl. The Opus triage agent must classify
# them (ACT / DRAFT / DROP / ASK) before any can reach the
- # supervisor. This is the safety guardrail Sam mandated:
+ # supervisor. This is the mandated safety guardrail:
# no autonomous work without a verdict.
PENDING_TRIAGE.parent.mkdir(parents=True, exist_ok=True)
with PENDING_TRIAGE.open("a") as f:
diff --git a/claude-ops/scripts/ops-daemon-launcher.sh b/claude-ops/scripts/ops-daemon-launcher.sh
index b0243a7c..6f6e3dd1 100755
--- a/claude-ops/scripts/ops-daemon-launcher.sh
+++ b/claude-ops/scripts/ops-daemon-launcher.sh
@@ -16,8 +16,7 @@ LATEST_VERSION="$(
find "$CACHE_ROOT" -mindepth 1 -maxdepth 1 -type d -print 2>/dev/null \
| awk -F/ '{print $NF}' \
| grep -E '^[0-9]+\.[0-9]+\.[0-9]+$' \
- | sort -V \
- | tac \
+ | sort -rV \
| while read -r v; do
if [[ -x "$CACHE_ROOT/$v/scripts/ops-daemon.sh" ]]; then
echo "$v"; break
@@ -34,4 +33,11 @@ DAEMON="$CACHE_ROOT/$LATEST_VERSION/scripts/ops-daemon.sh"
export CLAUDE_PLUGIN_ROOT="$CACHE_ROOT/$LATEST_VERSION"
echo "[ops-daemon-launcher] resolved ops $LATEST_VERSION → $DAEMON"
-exec "${BASH:-/bin/bash}" "$DAEMON" "$@"
+# Bash 4+ required for ops-daemon.sh (associative arrays). Mirror install-ops-daemon.sh.
+OPS_EXEC_BASH="/bin/bash"
+if [[ -x /opt/homebrew/bin/bash ]]; then
+ OPS_EXEC_BASH="/opt/homebrew/bin/bash"
+elif [[ -x /usr/local/bin/bash ]]; then
+ OPS_EXEC_BASH="/usr/local/bin/bash"
+fi
+exec "$OPS_EXEC_BASH" "$DAEMON" "$@"
diff --git a/claude-ops/scripts/ops-pocket-email-bridge.py b/claude-ops/scripts/ops-pocket-email-bridge.py
index 38909d83..a013b11d 100755
--- a/claude-ops/scripts/ops-pocket-email-bridge.py
+++ b/claude-ops/scripts/ops-pocket-email-bridge.py
@@ -44,6 +44,15 @@ from pathlib import Path
LOG_PREFIX = "[ops-pocket-email-bridge]"
HOME = Path(os.path.expanduser("~"))
+_USER_CONTEXT_PATH = Path(os.environ.get(
+ "POCKET_USER_CONTEXT",
+ str(HOME / ".claude/state/pocket/user-context.json"),
+))
+try:
+ _ctx = json.loads(_USER_CONTEXT_PATH.read_text())
+ _OWNER_NAME: str = _ctx.get("owner_name") or "the user"
+except (OSError, json.JSONDecodeError):
+ _OWNER_NAME = "the user"
STATE_DIR = Path(os.environ.get("POCKET_STATE_DIR", HOME / ".claude/state/pocket"))
CONFIG = STATE_DIR / "email-config.json"
QUEUE = STATE_DIR / "supervisor-out-queue.jsonl"
@@ -308,21 +317,21 @@ def parse_reply_with_llm(body: str, opens: list[dict], parser_model: str) -> lis
f" context: {(q.get('context') or '')[:200]}"
for q in opens
)
- prompt = f"""You are the inbound-reply parser for Account A' Pocket supervisor.
+ prompt = f"""You are the inbound-reply parser for {_OWNER_NAME}'s Pocket supervisor.
-Sam just sent an EMAIL reply. Your job: figure out which of the currently-open supervisor questions he's replying to (it may be one, several, or none), and extract the answer in a form the worker can use.
+{_OWNER_NAME} just sent an EMAIL reply. Your job: figure out which of the currently-open supervisor questions they're replying to (it may be one, several, or none), and extract the answer in a form the worker can use.
Currently-open questions:
{open_descriptions}
-Sam's email body:
+{_OWNER_NAME}'s email body:
\"\"\"{body[:4000]}\"\"\"
Rules:
-- If Sam's message is clearly not a reply to any open question, return an empty array.
-- If Sam addresses one specific question, return one item.
-- If Sam addresses multiple, return multiple items.
-- The 'answer' field should be in Sam's voice and natural — the worker reads it as a direct instruction.
+- If {_OWNER_NAME}'s message is clearly not a reply to any open question, return an empty array.
+- If {_OWNER_NAME} addresses one specific question, return one item.
+- If {_OWNER_NAME} addresses multiple, return multiple items.
+- The 'answer' field should be in {_OWNER_NAME}'s voice and natural — the worker reads it as a direct instruction.
- 'confidence' is your honest 0.0-1.0 estimate.
- Output STRICT JSON array, no markdown fences, no prose.
@@ -385,30 +394,18 @@ def drain_inbound(cfg: dict) -> tuple[int, int]:
messages = fetch_recent_replies(cfg)
if not messages:
return (0, 0)
-
- def _internal_ts_rank(m: dict) -> int:
- """Gmail internalDate as int for ordering (matches WhatsApp strict > cursor)."""
- t = m.get("ts")
- try:
- if isinstance(t, (int, float)):
- return int(t)
- s = str(t)
- if s.isdigit():
- return int(s)
- except (TypeError, ValueError):
- pass
- return 0
-
- # Filter: only NEW messages after last_processed_id (strictly newer internalDate).
- if not last_id:
+ # Filter: only NEW messages (after last_processed_id timestamp)
+ new_messages = []
+ seen_last = False
+ for m in messages:
+ if m["id"] == last_id:
+ seen_last = True
+ continue
+ if seen_last:
+ new_messages.append(m)
+ # If last_id wasn't found in current window, treat all as new (first run / drift)
+ if not last_id or not any(m["id"] == last_id for m in messages):
new_messages = messages
- else:
- cur = next((m for m in messages if m["id"] == last_id), None)
- if cur is None:
- new_messages = messages
- else:
- t0 = _internal_ts_rank(cur)
- new_messages = [m for m in messages if _internal_ts_rank(m) > t0]
opens = open_questions()
parser_model = cfg.get("parser_model", "claude-sonnet-4-6")
diff --git a/claude-ops/scripts/ops-pocket-triage.py b/claude-ops/scripts/ops-pocket-triage.py
index c5255658..e9394889 100755
--- a/claude-ops/scripts/ops-pocket-triage.py
+++ b/claude-ops/scripts/ops-pocket-triage.py
@@ -320,14 +320,21 @@ def main() -> int:
write_health("running", "triage tick")
log(f"start (dry_run={DRY_RUN}, model={MODEL})")
- if not PENDING.exists() or PENDING.stat().st_size == 0:
+ try:
+ initial = PENDING.read_bytes()
+ except OSError as e:
+ log(f"read pending failed: {e}")
+ write_health("error", f"read pending: {e}")
+ return 1
+
+ if not initial.strip():
log("no pending items — nothing to triage")
write_health("ok", "no pending items", extra={"counts": {}})
return 0
- pending_snapshot = PENDING.read_text()
+ consumed = len(initial)
tasks = []
- for raw in pending_snapshot.splitlines():
+ for raw in initial.decode("utf-8", errors="replace").splitlines():
raw = raw.strip()
if not raw:
continue
@@ -340,27 +347,13 @@ def main() -> int:
counts = {"ACT": 0, "DRAFT": 0, "DROP": 0, "ASK": 0}
inter_call_pause = float(os.environ.get("POCKET_TRIAGE_PACE_SEC", "8"))
for i, task in enumerate(tasks, 1):
+ if i > 1 and inter_call_pause > 0:
+ time.sleep(inter_call_pause) # avoid bursting Claude Max quota
title = (task.get("title") or "")[:70]
- task_id = task.get("id", "")
- prior = _id_already_routed(task_id)
- if prior:
- verdict = f"SKIP_{prior}"
- log(f"SKIP {task_id} — already in {prior}")
- decision = {
- "verdict": "ASK",
- "confidence": 0.0,
- "reasoning": f"skipped Opus triage — already present in {prior}",
- "scoped_task_description": "",
- "concerns": [],
- }
- else:
- if i > 1 and inter_call_pause > 0:
- time.sleep(inter_call_pause) # avoid bursting Claude Max quota
- log(f"[{i}/{len(tasks)}] triaging: {title}")
- decision = triage_one(task)
- verdict = route(task, decision)
- if not verdict.startswith("SKIP_"):
- counts[verdict] = counts.get(verdict, 0) + 1
+ log(f"[{i}/{len(tasks)}] triaging: {title}")
+ decision = triage_one(task)
+ verdict = route(task, decision)
+ counts[verdict] = counts.get(verdict, 0) + 1
log(f"[{i}/{len(tasks)}] verdict={verdict} conf={decision.get('confidence')} — {decision.get('reasoning','')[:120]}")
# Audit log every decision with full thinking
append_jsonl(AUDIT, {
@@ -370,16 +363,19 @@ def main() -> int:
"routed_to": verdict,
})
- # Drop only the snapshot we consumed; preserve any lines appended during triage.
+ # Drop only the snapshot we read; bytes appended during triage are kept.
if not DRY_RUN:
try:
- current = PENDING.read_text()
- if current.startswith(pending_snapshot):
- PENDING.write_text(current[len(pending_snapshot):])
- else:
- log("pending-triage prefix changed during triage — not rewriting file")
+ current = PENDING.read_bytes()
+ except OSError:
+ current = b""
+ tail = current[consumed:] if len(current) >= consumed else current
+ try:
+ PENDING.write_bytes(tail)
except OSError as e:
- log(f"pending rewrite failed: {e}")
+ log(f"FATAL: could not rewrite pending file: {e}")
+ write_health("error", f"rewrite pending: {e}")
+ return 1
log(f"done counts={counts}")
write_health("ok", f"triaged={sum(counts.values())}", extra={"counts": counts})
return 0
diff --git a/claude-ops/scripts/ops-pocket-whatsapp-bridge.py b/claude-ops/scripts/ops-pocket-whatsapp-bridge.py
index 2b8b3a93..52ce48f5 100755
--- a/claude-ops/scripts/ops-pocket-whatsapp-bridge.py
+++ b/claude-ops/scripts/ops-pocket-whatsapp-bridge.py
@@ -42,6 +42,15 @@ from pathlib import Path
LOG_PREFIX = "[ops-pocket-whatsapp-bridge]"
HOME = Path(os.path.expanduser("~"))
+_USER_CONTEXT_PATH = Path(os.environ.get(
+ "POCKET_USER_CONTEXT",
+ str(HOME / ".claude/state/pocket/user-context.json"),
+))
+try:
+ _ctx = json.loads(_USER_CONTEXT_PATH.read_text())
+ _OWNER_NAME: str = _ctx.get("owner_name") or "the user"
+except (OSError, json.JSONDecodeError):
+ _OWNER_NAME = "the user"
STATE_DIR = Path(os.environ.get("POCKET_STATE_DIR", HOME / ".claude/state/pocket"))
BRIDGE_DB = Path(os.environ.get(
"POCKET_WHATSAPP_DB",
@@ -139,21 +148,21 @@ def parse_with_llm(body: str, opens: list[dict], parser_model: str) -> list[dict
for q in opens
)
- prompt = f"""You are the inbound-reply parser for Account A' Pocket supervisor.
+ prompt = f"""You are the inbound-reply parser for {_OWNER_NAME}'s Pocket supervisor.
-Sam just sent a WhatsApp message. Your job: figure out which of the currently-open supervisor questions he's replying to (it may be one, several, or none), and extract the answer in a form the worker can use.
+{_OWNER_NAME} just sent a WhatsApp message. Your job: figure out which of the currently-open supervisor questions they're replying to (it may be one, several, or none), and extract the answer in a form the worker can use.
Currently-open questions:
{open_descriptions}
-Sam's WhatsApp message:
+{_OWNER_NAME}'s WhatsApp message:
\"\"\"{body}\"\"\"
Rules:
-- If Sam's message is clearly not a reply to any open question (chit-chat, unrelated request, ambiguous noise), return an empty array.
-- If Sam addresses one specific question, return one item.
-- If Sam addresses multiple in one message ("skip both kitchen ones, do the gmail one"), return multiple items.
-- The 'answer' field should be in Sam's voice and natural — the worker reads it as a direct instruction.
+- If {_OWNER_NAME}'s message is clearly not a reply to any open question (chit-chat, unrelated request, ambiguous noise), return an empty array.
+- If {_OWNER_NAME} addresses one specific question, return one item.
+- If {_OWNER_NAME} addresses multiple in one message ("skip both kitchen ones, do the gmail one"), return multiple items.
+- The 'answer' field should be in {_OWNER_NAME}'s voice and natural — the worker reads it as a direct instruction.
- 'confidence' is your honest 0.0-1.0 estimate that you correctly identified the question + intent.
- Output STRICT JSON, no markdown fences, no prose. Just the array.
diff --git a/claude-ops/templates/pocket-supervisor-prompt.md b/claude-ops/templates/pocket-supervisor-prompt.md
index 2b1eb42b..0ab4d9d4 100644
--- a/claude-ops/templates/pocket-supervisor-prompt.md
+++ b/claude-ops/templates/pocket-supervisor-prompt.md
@@ -11,7 +11,7 @@ You are the team lead and **do not do implementation work yourself**. Never run
- `Read` for inspecting state files (cursor, tasks.jsonl, team config)
- `Write` for your own state files only (.supervisor-health, supervisor-cursor.txt)
- `Bash` only for: reading own state, checking team config existence, `tmux capture-pane` if a teammate goes dark
-- `AskUserQuestion` only when escalating to Sam
+- `AskUserQuestion` only when escalating to the owner
- `ScheduleWakeup` to stay alive
If a teammate is stuck or fails, your response is to spawn a NEW teammate or escalate — never to do the work yourself. This is the explicit pattern from Anthropic's Delegate Mode (Shift+Tab toggle) — we apply it via doctrine since the toggle is runtime-only.
@@ -38,7 +38,7 @@ Your job is to **pump** items from the durable log into the team TaskList, then
3. **You do not do worker-class work yourself** (no repo edits, no API calls, no Bash beyond inspecting state). Delegate-Mode doctrine.
-4. **Workers may NOT spawn sub-agents, may NOT create sibling TaskList entries, may NOT TaskCreate.** They claim their assigned task via TaskUpdate, work on exactly that, and complete. Their prompt enforces this; if you see a worker violating, kill the window and SendMessage Sam.
+4. **Workers may NOT spawn sub-agents, may NOT create sibling TaskList entries, may NOT TaskCreate.** They claim their assigned task via TaskUpdate, work on exactly that, and complete. Their prompt enforces this; if you see a worker violating, kill the window and SendMessage the owner.
## Your job
@@ -46,7 +46,7 @@ Your job is to **pump** items from the durable log into the team TaskList, then
2. **Drain the durable log into the team TaskList.** Each wake cycle:
- Read `~/.claude/state/pocket/tasks.jsonl` from cursor `~/.claude/state/pocket/supervisor-cursor.txt` (byte offset; default 0).
- - For each new task entry: parse JSON. Skip outbound items (`kind` in `send_message`, `draft_email`) — those go to `drafts.jsonl` for Sam's explicit approval, NOT into the team.
+ - For each new task entry: parse JSON. Skip outbound items (`kind` in `send_message`, `draft_email`) — those go to `drafts.jsonl` for the owner's explicit approval, NOT into the team.
- For each non-outbound task: call `TaskCreate` against the `pocket-orchestrator` team with the task title/description/context. Tag with `metadata: {pocket_task_id: "...", source: "pocket"}` for traceability back to the durable log.
- **Only advance the cursor after TaskCreate succeeds** — if you crash mid-pump, the watchdog respawns you and you re-pump the same items (idempotent because of the metadata tag — check existing TaskList first if you see the cursor pointing past the start).
- Cap new TaskCreate calls at `min(5 - in_progress_count, queue_remaining)` to keep ≤5 active teammates (Agent Teams soft cap).
@@ -60,13 +60,13 @@ Your job is to **pump** items from the durable log into the team TaskList, then
3. **Supervise live.** While teammates are working:
- Watch shared `TaskList` — every teammate's `TaskCreate`/`TaskUpdate` shows up.
- Watch the team channel for teammate `SendMessage` to you.
- - **Intervene on dangerous actions.** If a teammate announces (or its task description implies) `rm -rf`, `git push --force`, `aws … terminate-*`, `aws … delete-*`, `DROP TABLE`, prod infra mutation, repo-archive, force-merge, secret rotation, or anything in `~/.claude/CLAUDE.md` § "Executing actions with care" — immediately `SendMessage({to: "worker-<id>", content: "HALT — wait for human confirmation before this action. State exactly what you intend to do."})` and surface it to Sam (see escalation below).
- - **Escalate confirmation requests.** If a teammate sends you a question or asks for approval, do NOT auto-answer. Surface via `AskUserQuestion` (or write to `supervisor-inbox.jsonl` if Sam is not active in this session).
+ - **Intervene on dangerous actions.** If a teammate announces (or its task description implies) `rm -rf`, `git push --force`, `aws … terminate-*`, `aws … delete-*`, `DROP TABLE`, prod infra mutation, repo-archive, force-merge, secret rotation, or anything in `~/.claude/CLAUDE.md` § "Executing actions with care" — immediately `SendMessage({to: "worker-<id>", content: "HALT — wait for human confirmation before this action. State exactly what you intend to do."})` and surface it to the owner (see escalation below).
+ - **Escalate confirmation requests.** If a teammate sends you a question or asks for approval, do NOT auto-answer. Surface via `AskUserQuestion` (or write to `supervisor-inbox.jsonl` if the owner is not active in this session).
4. **Stay alive.** End each turn with `ScheduleWakeup({delaySeconds: 90, reason: "supervisor poll", prompt: "<<autonomous-loop-dynamic>>"})`. If you're idle (no active teammates, queue empty), use `delaySeconds: 300` instead.
5. **Human-in-the-loop bridge — async, NOT AskUserQuestion.**
- Sam is NOT attached to your tmux pane most of the time. `AskUserQuestion` would block you indefinitely — DO NOT call it. Instead, use the async question/reply files:
+ The owner is NOT attached to your tmux pane most of the time. `AskUserQuestion` would block you indefinitely — DO NOT call it. Instead, use the async question/reply files:
**When a worker SendMessages you with a question/blocker/dangerous-action proposal:**
1. Append to `~/.claude/state/pocket/supervisor-questions.jsonl`:
@@ -92,21 +92,21 @@ Your job is to **pump** items from the durable log into the team TaskList, then
Reply however you want — full sentences are fine. I'll parse your intent.
```
- - Sam can reply on WhatsApp in plain English ("yeah just do report only", "skip both kitchen ones"). The cron bridge runs each WhatsApp message through `claude -p` to resolve which question(s) Sam is answering and write structured replies to supervisor-replies.jsonl.
- 3. SendMessage the worker: `"Escalated to Sam (qid=q-XXX). You're blocked until I get a reply. Mark your TaskList entry blocked."`
+ - The owner can reply on WhatsApp in plain English ("yeah just do report only", "skip both kitchen ones"). The cron bridge runs each WhatsApp message through `claude -p` to resolve which question(s) the owner is answering and write structured replies to supervisor-replies.jsonl.
+ 3. SendMessage the worker: `"Escalated to the owner (qid=q-XXX). You're blocked until I get a reply. Mark your TaskList entry blocked."`
4. The worker should TaskUpdate to `blocked` with `metadata.blocked_on: "q-XXX"`.
**Each wake cycle, BEFORE pumping new tasks:**
1. Read `~/.claude/state/pocket/supervisor-replies.jsonl` line-by-line.
2. For each reply whose `id` matches an `open` question in `supervisor-questions.jsonl`:
- - SendMessage the original worker with Sam's answer: `"Sam said: <answer>. Proceed accordingly."`
+ - SendMessage the original worker with the owner's answer: `"The owner said: <answer>. Proceed accordingly."`
- Mark the question `status: "answered"`, append to `~/.claude/state/pocket/answered-questions.jsonl`, remove from `supervisor-questions.jsonl`.
- - TaskUpdate the worker's blocked entry back to `in_progress` (or `cancelled` if Sam said skip).
+ - TaskUpdate the worker's blocked entry back to `in_progress` (or `cancelled` if the owner said skip).
3. Move processed reply lines to `~/.claude/state/pocket/replies-archive.jsonl` so they're not re-processed.
- **Status conventions for Sam's CLI to read:**
- - `status: "open"` — question is fresh, awaiting Sam's reply.
- - `status: "answered"` — Sam replied; you've relayed it.
+ **Status conventions for the owner's CLI to read:**
+ - `status: "open"` — question is fresh, awaiting the owner's reply.
+ - `status: "answered"` — the owner replied; you've relayed it.
- `status: "stale"` — open for >24h with no reply; you may unblock the worker with default ("skip") and tag the question stale.
5. **Health heartbeat.** Each cycle write `~/.claude/state/pocket/.supervisor-health`:
@@ -131,7 +131,7 @@ when done. This makes your work visible to the supervisor and to any human
session attached to the team.
You have full Claude Code tools (Bash, Edit, Read, Agent, Skill, MCP) and the
-user's complete skill/MCP surface. The user is Sam; all his global rules in
+user's complete skill/MCP surface. The user is the owner; all their global rules in
~/.claude/CLAUDE.md apply.
Hard rules:
@@ -141,11 +141,11 @@ Hard rules:
Your one job is to complete the single task you were spawned for. If
you notice something else that needs attention, SendMessage supervisor
about it — do NOT act on it yourself.
- • Outbound comms (email, Slack, WhatsApp, SMS, calendar) require Sam's
+ • Outbound comms (email, Slack, WhatsApp, SMS, calendar) require the owner's
per-message approval token. If the task implies sending, draft only and
SendMessage the draft to supervisor for review.
• Destructive ops (rm -rf, force-push, drop table, terminate-instances,
- etc.) require explicit Sam confirmation. SendMessage supervisor before
+ etc.) require explicit owner confirmation. SendMessage supervisor before
attempting.
• If the task is ambiguous, SendMessage supervisor ONE clarifying question
and wait — do not guess.