diff --git a/benchmarks/dev/bulk_read.py b/benchmarks/dev/bulk_read.py
new file mode 100644
index 0000000..04c3157
--- /dev/null
+++ b/benchmarks/dev/bulk_read.py
@@ -0,0 +1,62 @@
+"""Bulk reader: answer a question about one or more files with the cheap chat model, citing line numbers.
+
+The Portal pattern (Spotify, 2026-09): reading five files to answer one question is I/O, not reasoning, so a
+cheap model reads and the expensive one gets the answer. Files are sent in numbered chunks; the answer must cite
+`path:line`, which is what a summary otherwise lacks. Reasoning stays with the caller: this reports what the
+files say, it does not decide anything.
+
+Usage: .venv/bin/python benchmarks/dev/bulk_read.py "<question>" FILE [FILE ...] [--max-usd 0.05]
+"""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+CHUNK_LINES = 400
+PROMPT = (
+    "Answer the question using only the files below (numbered lines). Be specific and brief; cite every claim as "
+    "path:line. If the files do not answer it, say so.\n\nQuestion: {question}\n\n{files}"
+)
+
+
+def numbered(path: Path) -> list[str]:
+    lines = path.read_text(errors="ignore").splitlines()
+    chunks = []
+    for start in range(0, max(len(lines), 1), CHUNK_LINES):
+        body = "\n".join(f"{n + 1:5d}| {line}" for n, line in enumerate(lines[start : start + CHUNK_LINES], start=start))
+        chunks.append(f"--- {path} (lines {start + 1}-{min(start + CHUNK_LINES, len(lines))})\n{body}")
+    return chunks
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    parser.add_argument("question")
+    parser.add_argument("files", nargs="+", type=Path)
+    parser.add_argument("--max-usd", type=float, default=0.05)
+    args = parser.parse_args()
+    from openultrasast.config import load_config
+    from openultrasast.model.endpoint import DEFAULT_DETECTOR_MODEL
+    from openultrasast import tool_hunter
+
+    load_config()
+    client = tool_hunter.resolve_hunter_client()
+    if client is None:
+        print("no chat client: set DEEPSEEK_API_KEY", file=sys.stderr)
+        return 2
+    chunks = [chunk for path in args.files for chunk in numbered(path)]
+    answers = []
+    for chunk in chunks:
+        if float(client.cost_usd()) >= args.max_usd:
+            answers.append(f"[stopped at the ${args.max_usd} ceiling before: {chunk.splitlines()[0]}]")
+            break
+        response = client.complete(model=DEFAULT_DETECTOR_MODEL, messages=[{"role": "user", "content": PROMPT.format(question=args.question, files=chunk)}], tools=[], timeout_seconds=90)
+        answers.append((response.content or "").strip())
+    print("\n\n".join(a for a in answers if a))
+    print(f"\n[bulk_read: {len(chunks)} chunk(s), ${float(client.cost_usd()):.4f}]", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/benchmarks/dev/read_guard.py b/benchmarks/dev/read_guard.py
new file mode 100644
index 0000000..c6fa3cc
--- /dev/null
+++ b/benchmarks/dev/read_guard.py
@@ -0,0 +1,45 @@
+#!/usr/bin/env python3
+"""PreToolUse hook: a whole-file Read of more than LIMIT lines is denied and redirected.
+
+Measured 2026-09-29: 90% of a development session's tokens were tool I/O -- inline scripts and what they
+printed -- not reasoning. Whole-file reads are the largest single item that a summary could replace. A Read with
+`offset`/`limit` is a deliberate range and passes; a file at or under LIMIT lines passes.
+
+Wired in .claude/settings.local.json (PreToolUse, matcher Read). Stdin: the hook JSON; stdout: a permission
+decision only when denying. Never fails the tool: any error here exits 0 silently.
+"""
+
+import json
+import sys
+
+LIMIT = 350
+
+
+def main() -> int:
+    try:
+        event = json.load(sys.stdin)
+    except (json.JSONDecodeError, OSError):
+        return 0
+    if event.get("tool_name") != "Read":
+        return 0
+    tool_input = event.get("tool_input") or {}
+    path = tool_input.get("file_path")
+    if not path or tool_input.get("limit") or tool_input.get("offset"):
+        return 0
+    try:
+        with open(path, "rb") as handle:
+            lines = sum(1 for _ in handle)
+    except OSError:
+        return 0
+    if lines <= LIMIT:
+        return 0
+    reason = (
+        f"{path} has {lines} lines (limit {LIMIT} for a whole-file read). Read a range with offset/limit, or ask "
+        f"the bulk reader for what you need: .venv/bin/python benchmarks/dev/bulk_read.py '<question>' {path}"
+    )
+    print(json.dumps({"hookSpecificOutput": {"hookEventName": "PreToolUse", "permissionDecision": "deny", "permissionDecisionReason": reason}}))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/benchmarks/dev/token_report.py b/benchmarks/dev/token_report.py
new file mode 100644
index 0000000..51bf6a2
--- /dev/null
+++ b/benchmarks/dev/token_report.py
@@ -0,0 +1,59 @@
+"""Where a Claude Code session's tokens go: tool input vs tool output vs text, from the transcript.
+
+The 2026-09-29 measurement of this project's main session: tool_use input 47.5% (inline scripts), tool results
+42%, assistant text 6%, thinking 2%, user 2%. Re-run after a workflow change to see whether it moved.
+
+Usage: python benchmarks/dev/token_report.py [TRANSCRIPT.jsonl]   (default: the newest under ~/.claude/projects)
+"""
+
+from __future__ import annotations
+
+import collections
+import json
+import sys
+from pathlib import Path
+
+
+def newest() -> Path:
+    candidates = sorted(Path.home().glob(".claude/projects/*/*.jsonl"), key=lambda p: p.stat().st_mtime)
+    if not candidates:
+        raise SystemExit("no transcript found")
+    return candidates[-1]
+
+
+def main() -> int:
+    path = Path(sys.argv[1]) if len(sys.argv) > 1 else newest()
+    chars: collections.Counter[str] = collections.Counter()
+    tools: collections.Counter[str] = collections.Counter()
+    for line in path.open():
+        try:
+            entry = json.loads(line)
+        except json.JSONDecodeError:
+            continue
+        content = (entry.get("message") or {}).get("content")
+        kind = entry.get("type", "?")
+        if isinstance(content, str):
+            chars[kind] += len(content)
+        elif isinstance(content, list):
+            for block in content:
+                block_type = block.get("type")
+                if block_type == "text":
+                    chars[f"{kind}:text"] += len(block.get("text", ""))
+                elif block_type == "tool_use":
+                    tools[str(block.get("name"))] += 1
+                    chars["tool_use_input"] += len(json.dumps(block.get("input")))
+                elif block_type == "tool_result":
+                    result = block.get("content")
+                    chars["tool_result"] += len(result if isinstance(result, str) else json.dumps(result))
+                elif block_type == "thinking":
+                    chars["thinking"] += len(block.get("thinking", ""))
+    total = sum(chars.values()) or 1
+    print(f"{path.name}: {total / 1e6:.1f}M chars unique (~{total / 4e6:.2f}M tokens once; re-sent every turn)")
+    for key, value in chars.most_common():
+        print(f"  {key:16s} {value / 1e6:6.2f}M  {value / total:5.1%}")
+    print("tool calls:", sum(tools.values()), dict(tools.most_common(5)))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())