feat(mr-number): OCR extraction mode + lookup script + installer

mr_lookup.py: lazy vision-SDK import + --extract ocr (tesseract, no SDK) so the tool
runs on the redroid cloud box. extract_via_ocr parses both the 'Recent reports'
summary and 'User reports' detail layouts → reports/red_flags/verdict, with escort-
relevant negative keywords (ghosted, pressured, time waster, FT, lowball, flake).
lookup.sh: one-command cloud preview (SSH → drive app + OCR on the box → print
verdict). install.sh: idempotent setup of droplet (tesseract+tool+services), plum
(deps+MCP SDK+mrlookup on PATH), and the Claude Desktop MCP (run while Desktop quit
so it sticks past the clobber). Verified end-to-end: real lookup → denied + red flags.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

users/transquinnftw/tools/mr-number-lookup/install.sh

@@ -0,0 +1,95 @@
+#!/usr/bin/env bash
+#
+# install.sh — set up the Mr. Number lookup feature end to end.
+#
+# Idempotent. Run from plum. RUN IT WHILE CLAUDE DESKTOP IS QUIT — the desktop
+# config is rewritten on quit, so adding the MCP while Desktop runs gets clobbered;
+# adding it while quit makes it stick.
+#
+#   1. DROPLET (redroid box): tesseract + the tool + the OCR/keyboard services.
+#   2. PLUM: local deps (tesseract/adb), the MCP SDK, and the `mrlookup` command.
+#   3. CLAUDE DESKTOP: register the quinn-mr-number MCP (durably, while quit).
+#
+set -euo pipefail
+
+HOST="${MR_NUMBER_HOST:-root@45.55.191.82}"
+KEY="${MR_NUMBER_KEY:-$HOME/.ssh/id_ed25519_1984}"
+TOOL_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REMOTE_DIR="/opt/mr-number"
+DESKTOP_CFG="$HOME/Library/Application Support/Claude/claude_desktop_config.json"
+BUN="$(command -v bun || echo "$HOME/.bun/bin/bun")"
+
+say() { printf '\n\033[1;36m== %s\033[0m\n' "$*"; }
+ok()  { printf '  \033[32m✓\033[0m %s\n' "$*"; }
+warn() { printf '  \033[33m!\033[0m %s\n' "$*"; }
+
+ssh_box() { ssh -i "$KEY" -o ConnectTimeout=15 "$HOST" "$@"; }
+scp_box() { scp -i "$KEY" -o ConnectTimeout=15 "$1" "$HOST:$2"; }
+
+# ── 1. DROPLET ──────────────────────────────────────────────────────────────
+say "1/3  Droplet ($HOST)"
+if ! ssh_box 'echo ok' >/dev/null 2>&1; then
+  warn "cannot reach $HOST with $KEY — skipping droplet setup"
+else
+  ssh_box 'command -v tesseract >/dev/null || { export DEBIAN_FRONTEND=noninteractive; apt-get update -qq && apt-get install -y -qq tesseract-ocr; }'
+  ok "tesseract present: $(ssh_box 'tesseract --version 2>&1 | head -1')"
+
+  ssh_box "mkdir -p $REMOTE_DIR /opt/mrnumber-ocr /opt/adb-web"
+  scp_box "$TOOL_DIR/mr_lookup.py" "$REMOTE_DIR/mr_lookup.py"
+  scp_box "$TOOL_DIR/cloud-setup/ocr-service/server.py" "/opt/mrnumber-ocr/server.py"
+  scp_box "$TOOL_DIR/cloud-setup/ocr-service/mrnumber-ocr.service" "/etc/systemd/system/mrnumber-ocr.service"
+  scp_box "$TOOL_DIR/cloud-setup/adb-keyboard/server.py" "/opt/adb-web/server.py"
+  ok "tool + services copied to droplet"
+
+  ssh_box 'systemctl daemon-reload \
+    && systemctl enable --now mrnumber-ocr >/dev/null 2>&1 \
+    && systemctl restart adb-web 2>/dev/null || true'
+  ssh_box 'for s in mrnumber-ocr adb-web; do printf "    %s: " "$s"; systemctl is-active "$s"; done'
+  ok "services enabled (mrnumber-ocr :8003, adb-web :8001)"
+fi
+
+# ── 2. PLUM ─────────────────────────────────────────────────────────────────
+say "2/3  Plum (local)"
+if command -v brew >/dev/null; then
+  command -v tesseract >/dev/null || brew install tesseract >/dev/null && ok "tesseract" || warn "tesseract install skipped"
+  command -v adb >/dev/null || brew install android-platform-tools >/dev/null && ok "adb" || warn "adb install skipped"
+else
+  warn "no brew — ensure tesseract + adb are installed for the MCP path"
+fi
+
+# MCP SDK (public npm; bun auto-resolves, but materialize for Desktop's minimal PATH)
+if [ -x "$BUN" ]; then
+  ( cd "$TOOL_DIR/mcp" && "$BUN" install >/dev/null 2>&1 ) && ok "MCP SDK installed (mcp/)" || warn "bun install failed in mcp/"
+else
+  warn "bun not found — the quinn-mr-number MCP needs it on PATH"
+fi
+
+# `mrlookup` on PATH
+chmod +x "$TOOL_DIR/lookup.sh"
+for BINDIR in "$HOME/Code/@scripts/session-tools" "$HOME/.local/bin" "/usr/local/bin"; do
+  if [ -d "$BINDIR" ] && [ -w "$BINDIR" ]; then
+    ln -sf "$TOOL_DIR/lookup.sh" "$BINDIR/mrlookup" && ok "mrlookup → $BINDIR/mrlookup" && break
+  fi
+done
+
+# ── 3. CLAUDE DESKTOP MCP (run while Desktop is QUIT) ───────────────────────
+say "3/3  Claude Desktop MCP"
+if pgrep -f "Claude.app/Contents/MacOS/Claude" >/dev/null 2>&1; then
+  warn "Claude Desktop is RUNNING — quit it and re-run, or this entry will be clobbered on next quit."
+fi
+INDEX="$TOOL_DIR/mcp/index.ts"
+python3 - "$DESKTOP_CFG" "$INDEX" "$BUN" <<'PY'
+import json, os, sys
+cfg, index, bun = sys.argv[1], sys.argv[2], sys.argv[3]
+if not os.path.exists(cfg):
+    print("  ! desktop config not found:", cfg); sys.exit(0)
+d = json.load(open(cfg))
+servers = d.setdefault("mcpServers", {})
+servers["quinn-mr-number"] = {"command": bun, "args": ["run", index]}
+json.dump(d, open(cfg, "w"), indent=2); open(cfg, "a").write("\n")
+print("  \033[32m✓\033[0m quinn-mr-number registered (servers:", ", ".join(servers) + ")")
+PY
+
+say "Done."
+echo "  Quick preview:   mrlookup +15551234567"
+echo "  Record a verdict: use the quinn-mr-number MCP (restart Claude Desktop first)."

users/transquinnftw/tools/mr-number-lookup/lookup.sh

@@ -0,0 +1,55 @@
+#!/usr/bin/env bash
+#
+# lookup — screen a phone number against Mr. Number via the cloud redroid box.
+#
+# Runs the lookup ENTIRELY on the DO droplet (45.55.191.82): adb drives the app,
+# tesseract OCRs the reports — no vision SDK, no tunnel, one SSH round-trip. Prints
+# the community reports + a suggested screening verdict.
+#
+# Preview by default (does NOT record). Recording into the screening gate goes
+# through the quinn-mr-number MCP (token stays on plum) — see README.
+#
+#   lookup +15551234567
+#   lookup "+1 555 123 4567"        # any format; cleaned before input
+#
+set -euo pipefail
+
+HOST="${MR_NUMBER_HOST:-root@45.55.191.82}"
+KEY="${MR_NUMBER_KEY:-$HOME/.ssh/id_ed25519_1984}"
+REMOTE_DIR="${MR_NUMBER_REMOTE_DIR:-/opt/mr-number}"
+SERIAL="${MR_NUMBER_REDROID_SERIAL:-localhost:5555}"
+
+PHONE="${1:-}"
+if [ -z "$PHONE" ]; then
+  echo "usage: lookup <phone>        e.g. lookup +15551234567" >&2
+  exit 1
+fi
+
+# Single SSH round-trip: drive the app + OCR on the box, emit one JSON object.
+JSON="$(ssh -i "$KEY" -o ConnectTimeout=15 "$HOST" \
+  "cd '$REMOTE_DIR' && python3 mr_lookup.py --json --extract ocr --dry-run --device '$SERIAL' --phone '$PHONE'")"
+
+# Pretty-print the result for a human (comma-style prints — no nested quotes so it
+# stays valid inside the bash single-quoted heredoc).
+printf '%s' "$JSON" | python3 -c '
+import sys, json
+d = json.load(sys.stdin)
+if d.get("error"):
+    print("ERROR:", d.get("message") or d["error"]); sys.exit(1)
+ex = d.get("extracted") or {}
+print()
+print("  phone       ", d.get("phone"))
+print("  verdict     ", d.get("result"), "  (suggested:", ex.get("suggested_result"), ")")
+if ex.get("classification"): print("  caller      ", ex.get("classification"))
+if ex.get("report_count") is not None: print("  reports     ", ex.get("report_count"))
+flags = ex.get("red_flags") or []
+if flags: print("  red flags   ", "; ".join(flags))
+reps = ex.get("reports") or []
+if reps:
+    print("  recent:")
+    for r in reps[:6]: print("    -", r)
+print()
+print("  screenshot  ", d.get("screenshot"), "(on the box)")
+print("  (preview only - not recorded; use the quinn-mr-number MCP to record)")
+print()
+'

users/transquinnftw/tools/mr-number-lookup/mr_lookup.py

@@ -35,19 +35,24 @@ from typing import Any
 # requests is only needed for the final recording step (guarded import so unit tests
 # can run in environments without it; the emulator path itself is fully testable).
 
-# --- Vision SDK (exact same pattern as codebase/@features/ad-watch/scripts/classify_photos.py)
-_SDK_SRC = os.environ.get(
-    "CLAUDE_CODE_BATCH_SDK_PATH",
-    str(Path.home() / "Code/@applications/@ml/@packages/@py/claude-code-batch-sdk/src"),
-)
-if _SDK_SRC not in sys.path:
-    sys.path.insert(0, _SDK_SRC)
-
-try:
-    from claude_code_batch_sdk import ClaudeClient, parse_json_response  # noqa: E402
-except ImportError:
-    print("ERROR: claude-code-batch-sdk not found. Set CLAUDE_CODE_BATCH_SDK_PATH or clone it to the expected location.", file=sys.stderr)
-    sys.exit(1)
+# --- Vision SDK loaded lazily (only the --extract vision path needs it). The OCR
+# path (--extract ocr) runs with stdlib + tesseract only, so the tool works on the
+# redroid cloud box where the batch SDK is absent.
+def _load_vision_sdk():
+    """Import claude-code-batch-sdk on demand; exit clearly if it's missing."""
+    sdk_src = os.environ.get(
+        "CLAUDE_CODE_BATCH_SDK_PATH",
+        str(Path.home() / "Code/@applications/@ml/@packages/@py/claude-code-batch-sdk/src"),
+    )
+    if sdk_src not in sys.path:
+        sys.path.insert(0, sdk_src)
+    try:
+        from claude_code_batch_sdk import ClaudeClient, parse_json_response  # noqa: E402
+    except ImportError:
+        print("ERROR: claude-code-batch-sdk not found (needed for --extract vision). "
+              "Set CLAUDE_CODE_BATCH_SDK_PATH, or use --extract ocr.", file=sys.stderr)
+        sys.exit(1)
+    return ClaudeClient, parse_json_response
 
 # --- Config / env
 QUINN_MY_URL = (os.environ.get("QUINN_MY_URL") or "https://my.transquinnftw.com").rstrip("/")
@@ -90,6 +95,7 @@ def _build_vision_prompt(screenshot_path: str, phone: str) -> str:
     )
 
 async def _extract_from_screenshot(screenshot_path: str, phone: str) -> dict[str, Any]:
+    ClaudeClient, parse_json_response = _load_vision_sdk()
     client = ClaudeClient(model="haiku", max_concurrent=1)  # haiku is fast and sufficient for text extraction
     prompt = _build_vision_prompt(str(screenshot_path), phone)
 
@@ -108,6 +114,86 @@ async def _extract_from_screenshot(screenshot_path: str, phone: str) -> dict[str
 
     return parsed
 
+# --- OCR extraction (no vision SDK) — for the redroid cloud box where tesseract is
+# present but the batch SDK is not. Produces the SAME dict shape as the vision path.
+_OCR_NEG_KW = ("no show", "noshow", "no-show", "rude", "cop", "police", "scam", "spam",
+               "fraud", "timewaste", "time waster", "time-waster", "ghost", "pressured",
+               "harass", "boundary", "unsafe", "aggressive", "fake", "flake", "lowball",
+               "haggle", "facetime", "wants to ft", "creep", "rob", "danger")
+_OCR_DATE_RE = re.compile(
+    r"^[a-z]{0,2}\s*(jan|feb|mar|apr|may|jun|jul|aug|sep|oct|nov|dec)[a-z]*\.?\s+\d", re.I)
+# Header that begins the reports list — the summary screen says "Recent reports",
+# the detail screen says "User reports", some builds "Community reports".
+_OCR_REPORTS_HEADER = re.compile(r"\b(recent|user|community)\s+reports\b", re.I)
+_OCR_SKIP = ("report caller", "recent reports", "user reports", "community reports",
+             "view all", "look up", "search", "block", "unblock", "add contact",
+             "call", "message", "text", "report number", "is this you")
+
+
+def _tesseract(image_path: str, psm: int = 6) -> str:
+    """OCR an image with tesseract. Configurable via TESSERACT_BIN."""
+    binary = os.environ.get("TESSERACT_BIN", "tesseract")
+    proc = subprocess.run([binary, image_path, "stdout", "--psm", str(psm)],
+                          capture_output=True, text=True, timeout=60)
+    if proc.returncode != 0:
+        raise RuntimeError(f"tesseract failed (rc={proc.returncode}): {proc.stderr[:300]}")
+    return proc.stdout
+
+
+def extract_via_ocr(image_path: str, phone: str) -> dict[str, Any]:
+    """Parse the Mr. Number reports screen from OCR text into the extract dict."""
+    text = _tesseract(image_path)
+    lines = [ln.strip() for ln in text.splitlines() if ln.strip()]
+
+    count = None
+    classification = None
+    reports: list[str] = []
+    capturing = False
+    for ln in lines:
+        low = ln.lower()
+        m = re.search(r"view all (\d+)\s+report", low)
+        if m:
+            count = int(m.group(1))
+        if classification is None and re.search(r"spam|scam|telemarket|fraud", low):
+            classification = ln
+        if _OCR_REPORTS_HEADER.search(low):
+            capturing = True
+            continue
+        if "view all" in low:  # summary screen: reports end at the "View all N" link
+            capturing = False
+            continue
+        if not capturing:
+            continue
+        if any(s in low for s in _OCR_SKIP) or _OCR_DATE_RE.match(ln):
+            continue
+        # Drop short OCR noise (icon glyphs come through as 1-2 chars like "q", "CQ)").
+        cleaned = re.sub(r"^[^a-zA-Z0-9]+", "", ln).strip()
+        if len(cleaned) >= 4:
+            reports.append(cleaned)
+
+    red_flags = [r for r in reports if any(k in r.lower() for k in _OCR_NEG_KW)]
+    if classification and re.search(r"spam|scam|fraud", classification.lower()):
+        red_flags.append(classification)
+    summary = "; ".join(reports[:3]) or (classification or "no reports parsed")
+    if red_flags:
+        suggested = "denied"
+    elif count or reports:
+        suggested = "approved"
+    else:
+        suggested = "not_found"
+    return {
+        "phone": phone,
+        "report_count": count,
+        "reports": reports,
+        "classification": classification,
+        "red_flags": red_flags,
+        "summary": summary,
+        "suggested_result": suggested,
+        "extractor": "ocr",
+        "ocr_text": text,
+    }
+
+
 # --- adb helpers (refactored into class for testability of the emulator method)
 class MrNumberEmulator:
     """Encapsulates adb interactions with the Mr. Number app in an emulator.
@@ -346,7 +432,8 @@ def record_screening(client_id: int, phone: str, result: str, raw: str) -> dict[
     resp.raise_for_status()
     return resp.json()
 
-async def main_async(phone: str, client_id: int | None, dry_run: bool, dump_ui: bool = False) -> dict[str, Any]:
+async def main_async(phone: str, client_id: int | None, dry_run: bool, dump_ui: bool = False,
+                     extract: str = "vision") -> dict[str, Any]:
     log(f"[mr-number] Starting lookup for {phone} on {DEVICE} (client_id={client_id}, dry_run={dry_run})")
 
     input_phone = clean_phone(phone)
@@ -394,10 +481,14 @@ async def main_async(phone: str, client_id: int | None, dry_run: bool, dump_ui:
     shot = take_screenshot(input_phone)
     log(f"[mr-number] Screenshot saved to {shot}")
 
-    # 3. Vision extraction
-    log("[mr-number] Running vision extraction on screenshot...")
-    extracted = await _extract_from_screenshot(str(shot), phone)
-    log("[mr-number] Extraction:", json.dumps(extracted, indent=2)[:800])
+    # 3. Extraction — OCR (tesseract, no SDK) or vision (claude batch SDK)
+    if extract == "ocr":
+        log("[mr-number] Running OCR extraction (tesseract) on screenshot...")
+        extracted = extract_via_ocr(str(shot), phone)
+    else:
+        log("[mr-number] Running vision extraction on screenshot...")
+        extracted = await _extract_from_screenshot(str(shot), phone)
+    log("[mr-number] Extraction:", json.dumps({k: v for k, v in extracted.items() if k != 'ocr_text'}, indent=2)[:800])
 
     # 4. Decide result + build raw
     result = decide_result(extracted)
@@ -451,6 +542,8 @@ def main() -> None:
     parser.add_argument("--device", default=DEVICE, help="adb serial (default emulator-5554)")
     parser.add_argument("--dump-ui", action="store_true", help="Always dump the current UI hierarchy before actions (for calibration/troubleshooting)")
     parser.add_argument("--json", action="store_true", help="Emit one JSON result object on stdout (progress to stderr). Used by the mr-number MCP.")
+    parser.add_argument("--extract", choices=["vision", "ocr"], default="vision",
+                        help="Extraction backend: 'vision' (claude batch SDK) or 'ocr' (tesseract, no SDK — for the redroid cloud box).")
     args = parser.parse_args()
 
     DEVICE = args.device
@@ -471,7 +564,7 @@ def main() -> None:
         log(get_ui_dump()[:2000] + "\n... (truncated)")
 
     import asyncio
-    result = asyncio.run(main_async(args.phone, args.client_id, args.dry_run, args.dump_ui))
+    result = asyncio.run(main_async(args.phone, args.client_id, args.dry_run, args.dump_ui, args.extract))
     if JSON_MODE:
         print(json.dumps(result))