@wendellresearch-ai/benchwarmers 0.1.1 → 0.2.1

package/gateway.py

@@ -1,712 +1,755 @@
-"""The bench method as an OpenAI-compatible endpoint, at http://127.0.0.1:8123/v1.
-
-    python gateway.py
-
-Ask for a COLLECTION by name — `bench/core`, `bench/reviewers` — and the prompt
-goes in parallel to every member of it plus explorer seats `route.py` fills by
-staleness, so the roster keeps getting evidence even when the same collection
-is called every day:
-
-    explorers = min(max_explorers, max_calls - members)     # defaults 2 and 5
-
-Collections and those limits live in `config/collections.json` and
-`config/gateway.json` (see `metrics/bench_collections.py`), edited from the
-dashboard. There is no default: a request that names no collection, or one
-that does not exist, is refused with every collection's description and tags,
-so the caller can choose — `GET /v1/models` carries the same.
-
-The reply is every answer, each labelled by seat and model with a call id,
-including seats that failed or came back empty: who answered is evidence about
-the roster, and a summary naming only the good answer hides it.
-
-Unless the reply is `merged` (a collection's `reply`, or per request a
-`:merged` suffix on the model or `"bench": {"reply": "merged"}`; `:all` the
-other way): then, after the panel, one more model writes a single answer
-saying where the seats agreed, where they disagreed and who said what, for a
-client that can show only one. It has its own call id, so a verdict can judge
-the merge, and the `bench` field still carries every seat. If no merger has
-the context for the merge, or it fails, the reply is every answer with a
-one-line note: the panel the caller waited for is never lost.
-
-A collection's `shape` (review, design, …) decides which models may fill the
-explorer seats. Left empty, a decision model labels the prompt when one that
-takes choice questions is onboarded (not ad hoc); otherwise, or on any doubt,
-the shape is `general`. A request body may add `"with": ["<alias>"]` to seat a
-model in an explorer chair, as `route.py --with` does.
-
-THE LOOP
---------
-Every seat is recorded in `metrics/offloads.jsonl` as an unjudged offload the
-moment it answers, so the explorer seats rotate even when nobody judges. The
-caller then checks the answers against the real code — agreement between seats
-is not a verdict, they fail together — and records one per call id:
-
-    python metrics/offload.py verdict <call id> --useful --note "..." \\
-        --real "<a finding it got right>" --false "<a claim it made that is untrue>"
-    POST /bench/verdict {"call": "<call id>", "useful": true, "note": "...",
-                         "real": ["..."], "false": ["..."]}
-
-The real/false claims are what `suite/mine.py` turns into suite tasks.
-
-The verdict replaces the unjudged entry, and route.py's evidence — which
-explorers are due, which models earn a place in a collection — reads it.
-
-IMAGES
-------
-A request whose messages carry `image_url` parts (http(s) or data: URLs) goes
-to each seat with its messages as sent, through the built-in client whatever
-BENCH_CLIENT says (`llm -m` takes a prompt). Only models the catalog says take
-images sit (`cannot_see`): a member that cannot, or that the catalog knows
-nothing about, is skipped and named in `bench.skipped` with why; explorers
-are drawn from models that can, under the shape `vision` unless the
-collection sets one; with no seat that can see, the request is refused (400)
-rather than answered blind. Each image counts IMAGE_TOKENS in the size check.
-Wherever the request is written down (the call log, offloads, learned
-limits, error text, the dashboard) an image is a placeholder, never its
-bytes. The merger gets the request as text with those placeholders, plus
-the seats' answers: its job is to reconcile what the seats said, not to look
-again, so any model can merge and a disagreement about the picture is
-reported rather than settled.
-
-WHAT IT IS NOT
---------------
-A single-model proxy. Tools such as Roo Code call Open WebUI directly; this
-exists only for the multi-model work. Seats are called with `llm -m`, exactly
-as offloads always were, or the built-in client (BENCH_CLIENT), so endpoints,
-headers and per-model options all come from the same registry. Full prompts
-and answers go to `runs/gateway/`, gitignored, with the name of the gateway
-key that asked. Bound to 127.0.0.1, like the dashboard (BENCH_BIND and
-BENCH_GATEWAY_PORT change that; beyond this machine it asks for a gateway key,
-metrics/guard.py, which also covers keys on this machine, scopes, caps and
-CORS). The dashboard answers these same routes on its own port.
-"""
-
-from __future__ import annotations
-
-import datetime as dt
-import json
-import os
-import pathlib
-import re
-import secrets
-import subprocess
-import sys
-import time
-import urllib.parse
-from concurrent.futures import ThreadPoolExecutor
-from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
-
-ROOT = pathlib.Path(__file__).resolve().parent
-sys.path.insert(0, str(ROOT / "metrics"))
-from paths import DATA  # noqa: E402  -- the install's data folder (metrics/paths.py)
-PORT = int(os.environ.get("BENCH_GATEWAY_PORT") or 8123)
-LOG_DIR = DATA / "runs" / "gateway"
-# The suite's own timeout. Above the slowest real answer seen (383 s); at 900
-# one stuck explorer held a whole request for fifteen minutes.
-SEAT_TIMEOUT = 420
-sys.path.insert(0, str(ROOT / "metrics"))
-sys.path.insert(0, str(ROOT / "suite"))
-import bench_collections  # noqa: E402
-import client  # noqa: E402
-import decide  # noqa: E402
-import guard  # noqa: E402
-import offload  # noqa: E402
-import route  # noqa: E402
-
-# What the shape picker is asked to choose between. Each option carries a
-# description because bare names mislead it: Mercury Decide called "list every
-# place serve.py writes a file" `code` at 0.93 when given names alone.
-SHAPES = {
-    "general": "A general question or task that fits none of the other kinds.",
-    "design": "Architecture, process or approach decisions: weighing options, an ADR, a plan.",
-    "prose": "Writing or editing documentation or other text.",
-    "review": "Judging whether given code is correct: finding bugs, or where a guarantee breaks.",
-    "code": "Writing new code or changing existing code.",
-    "enumeration": "Listing every place something happens in given code or files.",
-    "quick": "A short, bounded task: classify, extract or summarise a small input.",
-}
-CONFIDENT = 0.6
-LARGE = 8000          # characters; route.py keeps models that fail big prompts off big tasks
-# Bytes. A million-token prompt is about 4 MB, and a photo sent as a data URL
-# is 4/3 of its file. 20 MB is the tightest whole-request limit among the big
-# providers (Gemini's for inline data), so what passes here they take too.
-MAX_BODY = 20_000_000
-TOO_BIG = (" Send images as http(s) URLs, or smaller ones: models scale a photo down to about"
-           " 1,500-2,000 pixels on its long side anyway.")
-IMAGE_TOKENS = 2_500   # per image, in the size check (estimate_tokens says why)
-IMAGE_URL = re.compile(r"(https?://[^\s/?#]+|data:image/[\w.+-]+;base64,)", re.I)
-DATA_URLS = re.compile(r"data:[\w/.+-]*;base64,[A-Za-z0-9+/=_\\-]*", re.I)
-URLS = re.compile(r"https?://[^\s\"'<>\\]+", re.I)
-BASE64 = re.compile(r"[A-Za-z0-9+/]{100,}={0,2}")
-HEADER_NAMES = re.compile(r"[A-Za-z0-9-]{1,64}(\s*,\s*[A-Za-z0-9-]{1,64}){0,31}")
-SEAT_KEYS = ("call", "seat", "alias", "status", "seconds", "answer", "error")
-# A merge that reads as consensus hides the dissent the bench runs on, so the
-# merger is told to keep who said what.
-MERGE_SYSTEM = ("Several models answered the same request. Write one answer to it from theirs: say where"
-                " they agreed, where they disagreed, and who said what, naming each model as labelled."
-                " Keep every dissent, and do not present a claim as settled because several made it.")
-
-
-def flatten(messages: list[dict]) -> tuple[str, str]:
-    """(system, prompt) for `llm`. A panel question is one prompt; a longer
-    conversation is kept as labelled turns rather than silently truncated.
-    An image is its placeholder: this text is what gets written down."""
-    def text(m):
-        c = m.get("content") or ""
-        return c if isinstance(c, str) else "\n".join(
-            placeholder(_url(p)) if p.get("type") == "image_url" else str(p.get("text") or "")
-            for p in c if isinstance(p, dict))
-    # Newer OpenAI SDKs send "developer" where older ones send "system".
-    lead = ("system", "developer")
-    system = "\n\n".join(text(m) for m in messages if m.get("role") in lead)
-    turns = [m for m in messages if m.get("role") not in lead]
-    if len(turns) == 1:
-        return system, text(turns[0])
-    return system, "\n\n".join(f"{m.get('role', 'user')}: {text(m)}" for m in turns)
-
-
-def _url(part: dict) -> str:
-    """An image part's URL: {"image_url": {"url": ...}} as OpenAI sends it, or a bare string."""
-    u = part.get("image_url")
-    u = u.get("url") if isinstance(u, dict) else u
-    return u if isinstance(u, str) else ""
-
-
-def images(messages: list[dict]) -> list[str]:
-    """Every image part's URL, in order. ValueError for one no seat could be
-    sent: the gateway never fetches an image, the upstream does, so only
-    http(s) URLs and inline base64 images pass; and an image in another API's
-    shape is refused rather than dropped, which would have a seat answer blind."""
-    urls = []
-    for m in messages:
-        parts = m.get("content")
-        for p in parts if isinstance(parts, list) else ():
-            kind = p.get("type") if isinstance(p, dict) else None
-            if kind in ("image", "input_image"):
-                raise ValueError(f'a "{kind}" part is another API\'s shape; send each image as'
-                                 ' {"type": "image_url", "image_url": {"url": "<http(s) or data: URL>"}}')
-            if kind == "image_url":
-                url = _url(p)
-                if not IMAGE_URL.match(url):
-                    raise ValueError("an image_url must be an http(s) URL or a data:image/...;base64, URL,"
-                                     f" not {placeholder(url) if url else 'empty'}")
-                urls.append(url)
-    return urls
-
-
-def placeholder(url: str) -> str:
-    """How an image is written down (the call log, offloads, the dashboard,
-    error text, the merger's copy of the request), never its bytes: a data
-    URL by its size, an http URL without its query (a signed URL's signature
-    is a credential) or a user:password."""
-    if url[:5].lower() == "data:":
-        return f"[image: data URL, {(len(url) * 3 // 4 + 999) // 1000:,} KB]"
-    u = urllib.parse.urlsplit(url)
-    return f"[image: {u.scheme}://{u.netloc.rpartition('@')[2]}{u.path[:200]}{'?...' if u.query else ''}]"
-
-
-def redact(text: str) -> str:
-    """`text` with each inline image as its placeholder, every http(s) URL
-    without its query or user:password, and any other long base64 run as its
-    size: an upstream may echo the request in an error, in whatever form, and
-    errors are logged, shown and learned from."""
-    def bare(m):
-        try:
-            u = urllib.parse.urlsplit(m.group())
-        except ValueError:                             # an unparseable netloc, such as "[::1"
-            return "[url]"
-        return f"{u.scheme}://{u.netloc.rpartition('@')[2]}{u.path}{'?...' if u.query else ''}"
-    text = URLS.sub(bare, DATA_URLS.sub(lambda m: placeholder(m.group()), text))
-    return BASE64.sub(lambda m: f"[base64, {(len(m.group()) * 3 // 4 + 999) // 1000:,} KB]", text)
-
-
-def cannot_see(alias: str) -> str:
-    """Why an image cannot be put to `alias`, or "" when it can. The catalog
-    decides (metrics/catalog.py: the model's input types, your Vision / OCR
-    override in its model popup, a capability seen working). A model the
-    catalog knows nothing about does not sit either: one that cannot see
-    answers anyway, describing a photo it never got, and the override is one
-    click."""
-    import catalog
-    import endpoints
-    if endpoints.entry_of(alias).get("api") == "responses":
-        return "reached through the Responses API, which the gateway sends no images to yet"
-    rec = catalog.info(alias)
-    if rec.get("vision"):
-        return ""
-    if rec.get("vision") is None:
-        return "the catalog does not say whether it takes images; tick Vision / OCR in its model popup if it does"
-    return f"does not take images (says {rec['sources'].get('vision', 'the catalog')})"
-
-
-def pick_shape(prompt: str) -> tuple[str, str]:
-    """(shape, how it was chosen). Falls back to general on any doubt."""
-    try:
-        held = set(json.loads((DATA / "config" / "model-roles.json").read_text(encoding="utf-8"))
-                   .get("ad_hoc", []))
-    except (OSError, ValueError, AttributeError):
-        return "general", "config/model-roles.json unreadable"
-    picker = next((a for a, d in decide.load().items()
-                   if "choice" in d["question_types"] and a not in held), None)
-    if not picker:
-        return "general", "no decision model onboarded"
-    question = {"shape": {"type": "choice",
-                          "instructions": "What kind of task does this prompt ask for?",
-                          "criteria": SHAPES}}
-    try:
-        answer = decide.ask(picker, prompt[:20000], question)["answers"]["shape"]
-    except Exception as exc:                           # noqa: BLE001
-        return "general", f"{picker} failed ({decide.error_text(exc)[:80]})"
-    conf = answer.get("confidence", 0)
-    if answer.get("choice") not in SHAPES or conf < CONFIDENT:
-        return "general", f"{picker} unsure ({answer.get('choice')} at {conf:.2f})"
-    return answer["choice"], f"{picker}, {conf:.2f}"
-
-
-def retiring() -> set[str]:
-    """Aliases not to spend an explorer seat on: those on a retiring alert
-    (failing their probes; 2026-10-16 both explorers of a bench/core call were
-    and both failed), and those cooling down after a rate limit, until the
-    wait the provider stated, or 30 minutes (metrics/learn.py)."""
-    out: set[str] = set()
-    try:
-        import alerts
-        out |= {a.get("alias", "") for a in alerts.active().values() if a.get("kind") == "retiring"}
-    except Exception:                                  # noqa: BLE001
-        pass
-    try:
-        import learn
-        out |= set(learn.cooldowns())
-    except Exception:                                  # noqa: BLE001
-        pass
-    return out
-
-
-def ask_seat(alias: str, system: str, prompt: str, messages: list[dict] | None = None) -> dict:
-    """One seat's answer. `messages`: the request as sent, when it carries
-    images; it goes through the built-in client whatever BENCH_CLIENT says,
-    since `llm -m` takes a prompt, not inline images."""
-    cmd = ["llm", "-m", alias] + (["-s", system] if system else [])
-    started = time.monotonic()
-    if messages is not None or client.backend(alias) == "builtin":
-        # No subprocess, and no shared llm log for parallel seats to lock.
-        res = client.chat(alias, prompt, system, SEAT_TIMEOUT, messages=messages)
-        answer = res["text"]
-        error = res["error"] or (f"no answer in {SEAT_TIMEOUT}s" if res["timed_out"] else "")
-    else:
-        answer, error = _ask_llm(cmd, prompt)
-    # A quota or credit notice delivered as the reply's text, with a success
-    # exit, is not an answer (suite/run.py PROVIDER_NOTICES).
-    from run import provider_notice
-    if not error and (notice := provider_notice(answer)):
-        error, answer = f"provider notice in a success reply ({notice}): {answer[:200]}", ""
-    if messages is not None:
-        # Before the error is cut, logged or learned from. ponytail: the answer
-        # is not redacted; a seat gets an image as pixels, not as its URL or bytes.
-        error = redact(error)
-    if error:
-        # A stated context window, rate limit or wait is kept (metrics/learn.py);
-        # a rate limit also cools the model, so explorer seats skip it.
-        try:
-            import learn
-            learn.record(alias, error, "gateway")
-        except Exception:                              # noqa: BLE001
-            pass
-    status = "error" if error else ("empty" if not answer else "ok")
-    return {"status": status, "answer": answer, "error": error[-500:],
-            "seconds": round(time.monotonic() - started, 1)}
-
-
-def _ask_llm(cmd: list[str], prompt: str) -> tuple[str, str]:
-    """(answer, error) from the `llm` CLI."""
-    try:
-        done = subprocess.run(cmd, input=prompt, capture_output=True, text=True,
-                              encoding="utf-8", errors="replace", timeout=SEAT_TIMEOUT,
-                              env={**os.environ, "PYTHONIOENCODING": "utf-8"})
-        answer = done.stdout.strip()
-        # A non-zero exit is an error even when stderr is silent.
-        error = (done.stderr.strip() or f"llm exited {done.returncode}") if done.returncode else ""
-    except subprocess.TimeoutExpired:
-        answer, error = "", f"no answer in {SEAT_TIMEOUT}s"
-    except Exception as exc:                           # noqa: BLE001
-        # One seat failing to start must not lose the other seats' records.
-        answer, error = "", f"{type(exc).__name__}: {exc}"
-    return answer, error
-
-
-class TooLarge(ValueError):
-    """The prompt does not fit a member's context window (HTTP 413)."""
-
-
-def estimate_tokens(text: str) -> int:
-    """Characters / 4 plus 10%: tokenizers differ by family, so this errs high.
-    The windows it is compared with are corrected by learned 'maximum context
-    length' errors (metrics/learn.py). An image adds IMAGE_TOKENS whatever its
-    size, which errs high too: OpenAI's, Anthropic's and Gemini's models spend
-    up to about 1,600 on one at their usual detail. Qwen and Llama vision
-    models can spend more on a very large image; a learned error corrects that."""
-    return int(len(text) / 4 * 1.1)
-
-
-def too_small(needed: int, aliases) -> dict[str, int]:
-    """Aliases whose KNOWN context window is under `needed`, with the window."""
-    import catalog
-    out = {}
-    for a in aliases:
-        cw = catalog.info(a).get("context_window")
-        if cw and cw < needed:
-            out[a] = cw
-    return out
-
-
-def record_seat(task: str, call: str, label: str, alias: str, r: dict) -> dict:
-    """Log one answer as an unjudged offload, the moment it is in, so explorer
-    rotation counts it even if nobody judges; returns the seat as logged."""
-    offload.add(alias, task, None, f"{label}; {r['status']} in {r['seconds']}s"
-                + (f": {r['error'][:120]}" if r["error"] else ""), call)
-    return {"call": call, "seat": label, "alias": alias, **r}
-
-
-def merger_candidates(collection: dict, members: list[str], shape: str, large: bool):
-    """Who may merge, best first. The collection's `merger` alone when set;
-    else the members with verdicts on this shape, most useful first, then the
-    router's primary, then the other members (a fresh install has neither
-    verdicts nor suite scores). A generator, so the router is asked only if
-    no member with a record fits."""
-    if collection.get("merger"):
-        yield collection["merger"]
-        return
-    record: dict[str, list[int]] = {}
-    for e in offload.entries():
-        # The shape ends each seat's task as run_collection writes it: "(bench/<name>, <shape>)".
-        if e["model"] in members and e.get("useful") is not None and e.get("task", "").endswith(f", {shape})"):
-            r = record.setdefault(e["model"], [0, 0])
-            r[0], r[1] = r[0] + bool(e["useful"]), r[1] + 1
-    # Useful verdicts counted, not a ratio, as route.py counts findings: one
-    # lucky verdict must not outrank a long good record.
-    yield from sorted(record, key=lambda a: (-record[a][0], -record[a][0] / record[a][1]))
-    try:
-        primary = route.panel(shape, large=large)["primary"]
-    except Exception:                                  # noqa: BLE001 -- the members can still merge
-        primary = None
-    if primary:
-        yield primary["model"]
-    yield from members
-
-
-def merge(collection: dict, members: list[str], shape: str, whole: str, seats: list[dict],
-          task: str, panel_id: str) -> tuple[dict | None, str]:
-    """The merger's seat, and why the reply falls back to every answer ("" if
-    it does not). The merger reads the request and every answer, so it gets
-    the members' size check; a failed merger is still logged like a seat."""
-    answered = [s for s in seats if s["status"] == "ok"]
-    if not answered:
-        return None, "no seat answered"
-    text = f"THE REQUEST\n\n{whole}\n\nTHE ANSWERS\n\n" + "\n\n".join(
-        f"## {s['seat']} — {s['alias']}\n\n{s['answer']}" for s in answered)
-    needed = estimate_tokens(f"{MERGE_SYSTEM}\n\n{text}")
-    alias = next((a for a in merger_candidates(collection, members, shape, len(text) > LARGE)
-                  if not too_small(needed, [a])), None)
-    if alias is None:
-        return None, f"no merger has the context window for the merge (about {needed:,} tokens)"
-    seat = record_seat(task, f"{panel_id}.merger", "merger", alias, ask_seat(alias, MERGE_SYSTEM, text))
-    if seat["status"] != "ok":
-        return seat, f"the merger {alias} came back {seat['status']}" + (f": {seat['error'][:120]}" if seat["error"] else "")
-    return seat, ""
-
-
-def run_collection(name: str, messages: list[dict], nominated: list[str] | None = None,
-                   reply: str = "", who: str = "local") -> tuple[str, dict]:
-    """The panel's reply as markdown, and the same seats as data for the
-    response's `bench` field. `reply`: the request's mode, "" for the
-    collection's own. `who`: the gateway key that asked, for the call log."""
-    pictures = images(messages)
-    system, prompt = flatten(messages)
-    if not prompt.strip():
-        raise ValueError("no user message to put to the collection")
-    collection = bench_collections.load()[name]
-    s = bench_collections.settings()
-    raw = collection.get("members") or []
-    # The editor dedupes and caps; a hand-edited file may not have.
-    members = list(dict.fromkeys(m for m in raw if isinstance(m, str) and m))[:s["max_calls"]]
-    n_explorers = bench_collections.explorers_for(len(members), s)
-    # A member named in `with` is already seated; it must not use up the slot.
-    nominated = [a for a in dict.fromkeys(nominated or []) if a not in members]
-    # AN IMAGE GOES ONLY TO A MODEL THAT CAN SEE IT: one that cannot would
-    # answer anyway, about a picture it never got. Skipped, and named.
-    skipped = [{"seat": seat, "alias": a, "why": why} for seat, group in (("member", members), ("with", nominated))
-               for a in group if pictures and (why := cannot_see(a))]
-    unseen = {k["alias"] for k in skipped}
-    sitting = [a for a in members if a not in unseen]
-    nominated = [a for a in nominated if a not in unseen]
-    # The system text goes to every seat too, so it counts for shape and size.
-    whole = f"{system}\n\n{prompt}".strip()
-    # REFUSE WHAT A MEMBER CANNOT READ, so the caller can trim the prompt or
-    # pick another collection instead of collecting a seat's context error.
-    needed = estimate_tokens(whole) + IMAGE_TOKENS * len(pictures)
-    small = too_small(needed, sitting)
-    if small:
-        windows = ", ".join(f"{a} {cw:,}" for a, cw in sorted(small.items(), key=lambda kv: kv[1]))
-        counted = f" (each image counted as {IMAGE_TOKENS:,})" if pictures else ""
-        raise TooLarge(f"prompt is about {needed:,} tokens{counted}; too large for {windows} (context"
-                       f" windows) in bench/{name}. Trim it under {min(small.values()):,} tokens,"
-                       " or choose a collection whose members have larger windows"
-                       " (GET /v1/models lists each collection's context_window).")
-    if pictures and not collection.get("shape"):
-        shape, how = "vision", "the request carries images"
-    elif not n_explorers:
-        shape, how = collection.get("shape") or "general", "no explorer seats to fill"
-    elif collection.get("shape"):
-        shape, how = collection["shape"], "set on the collection"
-    else:
-        shape, how = pick_shape(whole)
-    seats = [(f"member{i}", "member", alias) for i, alias in enumerate(sitting, 1)]
-    if n_explorers:
-        # `with` in the request body is route.py's --with: an explorer seat.
-        from run import _registered_names
-        avoid = retiring() | set(too_small(needed, _registered_names()))   # explorers must fit too
-        if pictures:
-            avoid |= {a for a in _registered_names() if cannot_see(a)}    # and see
-        p = route.panel(shape, large=len(whole) > LARGE, nominated=nominated or None,
-                        members=sitting, explorers=n_explorers, avoid=avoid)
-        seats += [(key, label, p[key]["model"]) for key, label in
-                  (("explorer", "explorer"), ("explorer2", "explorer 2")) if p.get(key)]
-    if not seats and pictures:
-        raise ValueError(f"no seat in bench/{name} can see images, so none is asked rather than answer"
-                         " blind: " + ("; ".join(f"{k['alias']} {k['why']}" for k in skipped) or "no members")
-                         + ", and no model that takes images was free for an explorer seat. Add one to the"
-                         " collection, or call one whose together.capabilities on GET /v1/models lists vision.")
-    if not seats:
-        raise LookupError(f"collection {name!r} has no members and no explorer could be seated")
-    panel_id = f"p{dt.datetime.now():%Y%m%d-%H%M%S}-{secrets.token_hex(2)}"
-    model = bench_collections.PREFIX + name
-    with ThreadPoolExecutor(max_workers=len(seats)) as pool:
-        results = list(pool.map(lambda s: ask_seat(s[2], system, prompt, messages if pictures else None), seats))
-
-    task = " ".join(prompt.split())[:90] + f" ({model}, {shape})"
-    LOG_DIR.mkdir(parents=True, exist_ok=True)
-    record = {"panel": panel_id, "at": dt.datetime.now().astimezone().isoformat(timespec="seconds"),
-              "model": model, "shape": shape, "shape_by": how, "key": who, "system": system,
-              "prompt": prompt, "seats": [], **({"skipped": skipped} if skipped else {})}
-    out = [f"Bench {model} {panel_id} · {len(sitting)} members + {len(seats) - len(sitting)}"
-           f" explorers · explorer shape {shape} ({how})\n"]
-    if skipped:
-        out.append("Not asked, as they cannot be sent the images: "
-                   + "; ".join(f"{k['alias']} ({k['why']})" for k in skipped) + "\n")
-    for (key, label, alias), r in zip(seats, results):
-        call = f"{panel_id}.{key}"
-        record["seats"].append(record_seat(task, call, label, alias, r))
-        body = r["answer"] or (f"_error: {r['error'][:300]}_" if r["error"] else
-                               "_empty reply — usually a token budget, not incapacity_")
-        out.append(f"## {label} — {alias}\n`{call}` · {r['status']} · {r['seconds']}s\n\n{body}\n")
-    out.append("---\nRecord a verdict per seat once checked against the code:\n"
-               "`python metrics/offload.py verdict <call id> --useful|--not-useful --note \"...\"`")
-    content = "\n".join(out)
-    # Beyond the OpenAI fields, so a caller reads seats without parsing the
-    # markdown. Clients that do not know it ignore it.
-    bench = {"panel": panel_id, "collection": model, "shape": shape, "shape_by": how,
-             "seats": [{k: s[k] for k in SEAT_KEYS} for s in record["seats"]],
-             "verdict": f"POST http://127.0.0.1:{PORT}/bench/verdict"
-                        ' {"call", "useful": true|false, "note", "real": [...], "false": [...]}',
-             **({"skipped": skipped} if skipped else {})}
-    if (reply or collection.get("reply") or "all") == "merged":
-        try:
-            merger, note = merge(collection, members, shape, whole, record["seats"], task, panel_id)
-        except Exception as exc:                       # noqa: BLE001 -- never lose the panel to the merge
-            merger, note = None, f"the merge failed ({type(exc).__name__}: {exc})"
-        if merger:
-            bench["merger"] = {k: merger[k] for k in SEAT_KEYS}
-            record["seats"].append(merger)             # the dashboard shows it, verdict and all
-        if note:
-            bench.update(reply="all", note=note)
-            content = f"_Merged reply unavailable: {note}. Every answer follows._\n\n{content}"
-        else:
-            who = ", ".join(f"{s['alias']} {s['status']}" for s in bench["seats"])
-            bench["reply"] = "merged"
-            content = (f"{merger['answer']}\n\n---\nMerged by {merger['alias']} (`{merger['call']}`) from:"
-                       f" {who}. Every seat's answer and call id is in the response's `bench` field.")
-    with (LOG_DIR / "calls.jsonl").open("a", encoding="utf-8") as fh:
-        fh.write(json.dumps(record, ensure_ascii=False) + "\n")
-    return content, bench
-
-
-def models() -> list[str]:
-    return [bench_collections.PREFIX + n for n in bench_collections.load()]
-
-
-class Handler(BaseHTTPRequestHandler):
-    """The gateway's routes. serve.py's Handler subclasses this to answer them
-    on the dashboard's port too, so the checks use the port actually served."""
-    who = "-"
-
-    def _json(self, code: int, data: dict, headers: dict | None = None) -> None:
-        body = json.dumps(data).encode()
-        self.send_response(code)
-        self.send_header("Content-Type", "application/json")
-        self.send_header("Content-Length", str(len(body)))
-        for k, v in (headers or {}).items():
-            self.send_header(k, v)
-        self.end_headers()
-        self.wfile.write(body)
-
-    def _error(self, code: int, message: str, headers: dict | None = None) -> None:
-        self._json(code, {"error": {"message": message, "type": "bench_gateway", "code": code}}, headers=headers)
-
-    def end_headers(self):
-        # CORS for a browser app on an allowed origin, on the gateway's routes
-        # only (metrics/guard.py); Vary, so a cache never hands one origin's
-        # answer to another. A request line too broken to parse has no headers.
-        headers = getattr(self, "headers", None)
-        if headers is not None and guard.gateway_route((getattr(self, "path", "") or "").split("?", 1)[0]) \
-                and (origin := guard.cors_origin(headers, self.server.server_port)):
-            self.send_header("Access-Control-Allow-Origin", origin)
-            self.send_header("Vary", "Origin")
-        super().end_headers()
-
-    def do_OPTIONS(self):                              # noqa: N802
-        """A browser's CORS preflight: answered on the gateway's routes, for an
-        allowed origin only. The dashboard's routes answer none, as before."""
-        if not guard.gateway_route(self.path.split("?", 1)[0]):
-            return self.send_error(501, f"Unsupported method ({self.command!r})")
-        try:
-            guard.check(self.headers, self.server.server_port, post=False, max_body=0)
-        except guard.Refused as r:
-            return self._error(r.status, str(r))
-        if not guard.cors_origin(self.headers, self.server.server_port):
-            return self._error(403, f"Origin {self.headers.get('Origin') or '(none)'!r} is not allowed;"
-                                    " set BENCH_ALLOWED_ORIGINS or BENCHWARMERS_PUBLIC_URL to add one")
-        # The headers it asks for as well, when they are plain names: the
-        # OpenAI JS SDK sends X-Stainless-* on every call, and a preflight that
-        # refused them would block it. The origin is the boundary, not these.
-        asked = (self.headers.get("Access-Control-Request-Headers") or "").strip()
-        self.send_response(204)
-        self.send_header("Access-Control-Allow-Methods", "GET, POST, OPTIONS")
-        self.send_header("Access-Control-Allow-Headers", "Authorization, Content-Type"
-                         + (f", {asked}" if HEADER_NAMES.fullmatch(asked) else ""))
-        self.send_header("Access-Control-Max-Age", "600")
-        self.end_headers()
-
-    def do_GET(self):                                  # noqa: N802
-        try:
-            guard.check(self.headers, self.server.server_port, post=False, max_body=0)
-            self.who = guard.gateway_caller(self.headers)
-        except guard.Refused as r:
-            return self._error(r.status, str(r), r.headers)
-        if self.path.split("?", 1)[0].rstrip("/") in ("/v1/models", "/models"):
-            # Beyond the OpenAI fields: what each collection is for, so an
-            # orchestrator can choose one by task. Clients ignore extra keys.
-            # Each collection twice: as set, and in the other reply mode, so a
-            # client that offers only listed models can still pick either.
-            s, data = bench_collections.settings(), bench_collections.load()
-            rows = []
-            for n, c in data.items():
-                if not guard.may_call(self.who, n):    # a key limited to some collections sees those
-                    continue
-                d = bench_collections.described(n, c, s)
-                other = "all" if d["reply"] == "merged" else "merged"
-                rows += [{"id": bench_collections.PREFIX + n, "object": "model", "owned_by": "benchwarmers", **d},
-                         {"id": f"{bench_collections.PREFIX}{n}:{other}", "object": "model",
-                          "owned_by": "benchwarmers", **d, "reply": other}]
-            self._json(200, {"object": "list", "data": rows})
-        else:
-            self._error(404, "GET /v1/models is the only GET here")
-
-    def do_POST(self):                                 # noqa: N802
-        # Without this any web page could spend model calls, seat ad hoc
-        # models through `with`, or write verdicts (metrics/guard.py).
-        try:
-            length = guard.check(self.headers, self.server.server_port, post=True, max_body=MAX_BODY, cors=True)
-            self.who = guard.gateway_caller(self.headers)
-        except guard.Refused as r:
-            return self._error(r.status, str(r) + (TOO_BIG if r.status == 413 else ""), r.headers)
-        try:
-            body = json.loads(self.rfile.read(length) or b"{}")
-        except ValueError:
-            return self._error(400, "body is not JSON")
-        if not isinstance(body, dict):
-            return self._error(400, "body must be a JSON object")
-        path = self.path.split("?", 1)[0].rstrip("/")
-        if path == "/bench/verdict":
-            if not isinstance(body.get("useful"), bool):
-                return self._error(400, '"useful" must be true or false')
-            lists = {k: body.get(k) or [] for k in ("real", "false")}
-            if not all(isinstance(v, list) and all(isinstance(s, str) for s in v) for v in lists.values()):
-                return self._error(400, '"real" and "false" must be lists of one-sentence claims')
-            try:
-                e = offload.verdict(str(body.get("call", "")), body["useful"], str(body.get("note") or ""),
-                                    lists["real"], lists["false"])
-            except KeyError as exc:
-                return self._error(404, exc.args[0])
-            return self._json(200, {"ok": True, "recorded": e})
-        if path not in ("/v1/chat/completions", "/chat/completions"):
-            return self._error(404, "POST /v1/chat/completions or /bench/verdict")
-        model = body.get("model") or ""
-        if not isinstance(model, str):
-            return self._error(400, '"model" must be a collection name such as bench/core')
-        model, suffix = bench_collections.split_reply(model)
-        name = bench_collections.resolve(model)
-        if name is None:
-            return self._error(400 if not model else 404, bench_collections.choose_message(model))
-        model = bench_collections.PREFIX + name
-        messages, nominated, opts = body.get("messages"), body.get("with") or [], body.get("bench") or {}
-        if not (isinstance(messages, list) and all(isinstance(m, dict) for m in messages)):
-            return self._error(400, '"messages" must be a list of message objects')
-        if not (isinstance(nominated, list) and all(isinstance(a, str) for a in nominated)):
-            return self._error(400, '"with" must be a list of model aliases')
-        if not isinstance(opts, dict) or opts.get("reply", "all") not in bench_collections.REPLIES:
-            return self._error(400, '"bench" must be an object whose "reply" is "all" or "merged"')
-        if not guard.may_call(self.who, name):
-            allowed = ", ".join(bench_collections.PREFIX + c for c in guard.scope(self.who) or []) or "none"
-            return self._error(403, f"the gateway key {self.who} may not call {model}; it may call {allowed}")
-        try:
-            guard.spend(self.who)                      # counted once it names a collection it may call
-        except guard.Refused as r:
-            return self._error(r.status, str(r), r.headers)
-        try:
-            # The body is the most specific, then the suffix, then the collection.
-            content, bench = run_collection(name, messages, nominated, opts.get("reply") or suffix, self.who)
-        except TooLarge as exc:
-            return self._error(413, str(exc))
-        except ValueError as exc:
-            return self._error(400, str(exc))
-        except LookupError as exc:
-            return self._error(503, str(exc))
-        reply_id, created = f"chatcmpl-{secrets.token_hex(8)}", int(time.time())
-        if not body.get("stream"):
-            return self._json(200, {
-                "id": reply_id, "object": "chat.completion", "created": created, "model": model,
-                "choices": [{"index": 0, "finish_reason": "stop",
-                             "message": {"role": "assistant", "content": content}}], "bench": bench})
-        # A panel finishes all at once, so a streaming client gets it as one
-        # chunk, merged or not. ponytail: nothing is sent until the panel and
-        # the merge are done, so a client or proxy with a shorter idle timeout
-        # drops it. Keep-alives mean sending the 200 before the panel runs, so
-        # the 413/503 refusals move into the stream; streaming the merge as it
-        # is written needs a streaming client.chat (RELEASE-PLAN acceptance 11).
-        self.send_response(200)
-        self.send_header("Content-Type", "text/event-stream")
-        self.end_headers()
-        for delta, finish in (({"role": "assistant", "content": content}, None), ({}, "stop")):
-            chunk = {"id": reply_id, "object": "chat.completion.chunk", "created": created,
-                     "model": model, "choices": [{"index": 0, "delta": delta, "finish_reason": finish}],
-                     **({"bench": bench} if finish is None else {})}
-            self.wfile.write(f"data: {json.dumps(chunk)}\n\n".encode())
-        self.wfile.write(b"data: [DONE]\n\n")
-
-    def log_request(self, code="-", size="-"):         # route, status, who; never a query or a key
-        guard.access(self, code)
-
-    def log_message(self, fmt, *args):                 # quieter than the default
-        sys.stderr.write(f"{self.log_date_time_string()} {fmt % args}\n")
-
-
-def main() -> None:
-    sys.stdout.reconfigure(encoding="utf-8", errors="replace")
-    address = guard.bind()
-    guard.startup(address)
-    print(f"bench gateway on http://{address}:{PORT}/v1 — models: {', '.join(models())}")
-    ThreadingHTTPServer((address, PORT), Handler).serve_forever()
-
-
-if __name__ == "__main__":
-    main()
+"""The bench method as an OpenAI-compatible endpoint, at http://127.0.0.1:8123/v1.
+
+    python gateway.py
+
+Ask for a COLLECTION by name — `bench/core`, `bench/reviewers` — and the prompt
+goes in parallel to every member of it plus explorer seats `route.py` fills by
+staleness, so the roster keeps getting evidence even when the same collection
+is called every day:
+
+    explorers = min(max_explorers, max_calls - members)     # defaults 2 and 5
+
+Collections and those limits live in `config/collections.json` and
+`config/gateway.json` (see `metrics/bench_collections.py`), edited from the
+dashboard. There is no default: a request that names no collection, or one
+that does not exist, is refused with every collection's description and tags,
+so the caller can choose — `GET /v1/models` carries the same.
+
+The reply is every answer, each labelled by seat and model with a call id,
+including seats that failed or came back empty: who answered is evidence about
+the roster, and a summary naming only the good answer hides it.
+
+Unless the reply is `merged` (a collection's `reply`, or per request a
+`:merged` suffix on the model or `"bench": {"reply": "merged"}`; `:all` the
+other way): then, after the panel, one more model writes a single answer
+saying where the seats agreed, where they disagreed and who said what, for a
+client that can show only one. It has its own call id, so a verdict can judge
+the merge, and the `bench` field still carries every seat. If no merger has
+the context for the merge, or it fails, the reply is every answer with a
+one-line note: the panel the caller waited for is never lost.
+
+A collection's `shape` (review, design, …) decides which models may fill the
+explorer seats. Left empty, a decision model labels the prompt when one that
+takes choice questions is onboarded (not ad hoc); otherwise, or on any doubt,
+the shape is `general`. A request body may add `"with": ["<alias>"]` to seat a
+model in an explorer chair, as `route.py --with` does.
+
+A retired model (`config/retired.json`) is never called: retiring takes it out
+of every collection, and one a collection still names — as a member, `with`
+or its merger — is skipped and named in `bench.skipped`, its seat going to an
+explorer and the merge to a merger chosen by verdicts.
+
+THE LOOP
+--------
+Every seat is recorded in `metrics/offloads.jsonl` as an unjudged offload the
+moment it answers, so the explorer seats rotate even when nobody judges. The
+caller then checks the answers against the real code — agreement between seats
+is not a verdict, they fail together — and records one per call id:
+
+    python metrics/offload.py verdict <call id> --useful --note "..." \\
+        --real "<a finding it got right>" --false "<a claim it made that is untrue>"
+    POST /bench/verdict {"call": "<call id>", "useful": true, "note": "...",
+                         "real": ["..."], "false": ["..."]}
+
+The real/false claims are what `suite/mine.py` turns into suite tasks.
+
+The verdict replaces the unjudged entry, and route.py's evidence — which
+explorers are due, which models earn a place in a collection — reads it.
+
+IMAGES
+------
+A request whose messages carry `image_url` parts (http(s) or data: URLs) goes
+to each seat with its messages as sent, through the built-in client whatever
+BENCH_CLIENT says (`llm -m` takes a prompt). Only models the catalog says take
+images sit (`cannot_see`): a member that cannot, or that the catalog knows
+nothing about, is skipped and named in `bench.skipped` with why; explorers
+are drawn from models that can, under the shape `vision` unless the
+collection sets one; with no seat that can see, the request is refused (400)
+rather than answered blind. Each image counts IMAGE_TOKENS in the size check.
+Wherever the request is written down (the call log, offloads, learned
+limits, error text, the dashboard) an image is a placeholder, never its
+bytes. The merger gets the request as text with those placeholders, plus
+the seats' answers: its job is to reconcile what the seats said, not to look
+again, so any model can merge and a disagreement about the picture is
+reported rather than settled.
+
+WHAT IT IS NOT
+--------------
+A single-model proxy. Tools such as Roo Code call Open WebUI directly; this
+exists only for the multi-model work. Seats are called with `llm -m`, exactly
+as offloads always were, or the built-in client (BENCH_CLIENT), so endpoints,
+headers and per-model options all come from the same registry. Full prompts
+and answers go to `runs/gateway/`, gitignored, with the name of the gateway
+key that asked. Bound to 127.0.0.1, like the dashboard (BENCH_BIND and
+BENCH_GATEWAY_PORT change that; beyond this machine it asks for a gateway key,
+metrics/guard.py, which also covers keys on this machine, scopes, caps and
+CORS). The dashboard answers these same routes on its own port.
+"""
+
+from __future__ import annotations
+
+import datetime as dt
+import json
+import os
+import pathlib
+import re
+import secrets
+import subprocess
+import sys
+import time
+import urllib.parse
+from concurrent.futures import ThreadPoolExecutor
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+ROOT = pathlib.Path(__file__).resolve().parent
+sys.path.insert(0, str(ROOT / "metrics"))
+from paths import DATA  # noqa: E402  -- the install's data folder (metrics/paths.py)
+PORT = int(os.environ.get("BENCH_GATEWAY_PORT") or 8123)
+LOG_DIR = DATA / "runs" / "gateway"
+# The suite's own timeout. Above the slowest real answer seen (383 s); at 900
+# one stuck explorer held a whole request for fifteen minutes.
+SEAT_TIMEOUT = 420
+sys.path.insert(0, str(ROOT / "metrics"))
+sys.path.insert(0, str(ROOT / "suite"))
+import bench_collections  # noqa: E402
+import client  # noqa: E402
+import decide  # noqa: E402
+import guard  # noqa: E402
+import offload  # noqa: E402
+import route  # noqa: E402
+
+# What the shape picker is asked to choose between. Each option carries a
+# description because bare names mislead it: Mercury Decide called "list every
+# place serve.py writes a file" `code` at 0.93 when given names alone.
+SHAPES = {
+    "general": "A general question or task that fits none of the other kinds.",
+    "design": "Architecture, process or approach decisions: weighing options, an ADR, a plan.",
+    "prose": "Writing or editing documentation or other text.",
+    "review": "Judging whether given code is correct: finding bugs, or where a guarantee breaks.",
+    "code": "Writing new code or changing existing code.",
+    "enumeration": "Listing every place something happens in given code or files.",
+    "quick": "A short, bounded task: classify, extract or summarise a small input.",
+}
+CONFIDENT = 0.6
+LARGE = 8000          # characters; route.py keeps models that fail big prompts off big tasks
+# Bytes. A million-token prompt is about 4 MB, and a photo sent as a data URL
+# is 4/3 of its file. 20 MB is the tightest whole-request limit among the big
+# providers (Gemini's for inline data), so what passes here they take too.
+MAX_BODY = 20_000_000
+TOO_BIG = (" Send images as http(s) URLs, or smaller ones: models scale a photo down to about"
+           " 1,500-2,000 pixels on its long side anyway.")
+IMAGE_TOKENS = 2_500   # per image, in the size check (estimate_tokens says why)
+IMAGE_URL = re.compile(r"(https?://[^\s/?#]+|data:image/[\w.+-]+;base64,)", re.I)
+DATA_URLS = re.compile(r"data:[\w/.+-]*;base64,[A-Za-z0-9+/=_\\-]*", re.I)
+URLS = re.compile(r"https?://[^\s\"'<>\\]+", re.I)
+BASE64 = re.compile(r"[A-Za-z0-9+/]{100,}={0,2}")
+HEADER_NAMES = re.compile(r"[A-Za-z0-9-]{1,64}(\s*,\s*[A-Za-z0-9-]{1,64}){0,31}")
+SEAT_KEYS = ("call", "seat", "alias", "status", "seconds", "answer", "error")
+# A merge that reads as consensus hides the dissent the bench runs on, so the
+# merger is told to keep who said what.
+MERGE_SYSTEM = ("Several models answered the same request. Write one answer to it from theirs: say where"
+                " they agreed, where they disagreed, and who said what, naming each model as labelled."
+                " Keep every dissent, and do not present a claim as settled because several made it.")
+
+
+def flatten(messages: list[dict]) -> tuple[str, str]:
+    """(system, prompt) for `llm`. A panel question is one prompt; a longer
+    conversation is kept as labelled turns rather than silently truncated.
+    An image is its placeholder: this text is what gets written down."""
+    def text(m):
+        c = m.get("content") or ""
+        return c if isinstance(c, str) else "\n".join(
+            placeholder(_url(p)) if p.get("type") == "image_url" else str(p.get("text") or "")
+            for p in c if isinstance(p, dict))
+    # Newer OpenAI SDKs send "developer" where older ones send "system".
+    lead = ("system", "developer")
+    system = "\n\n".join(text(m) for m in messages if m.get("role") in lead)
+    turns = [m for m in messages if m.get("role") not in lead]
+    if len(turns) == 1:
+        return system, text(turns[0])
+    return system, "\n\n".join(f"{m.get('role', 'user')}: {text(m)}" for m in turns)
+
+
+def _url(part: dict) -> str:
+    """An image part's URL: {"image_url": {"url": ...}} as OpenAI sends it, or a bare string."""
+    u = part.get("image_url")
+    u = u.get("url") if isinstance(u, dict) else u
+    return u if isinstance(u, str) else ""
+
+
+def images(messages: list[dict]) -> list[str]:
+    """Every image part's URL, in order. ValueError for one no seat could be
+    sent: the gateway never fetches an image, the upstream does, so only
+    http(s) URLs and inline base64 images pass; and an image in another API's
+    shape is refused rather than dropped, which would have a seat answer blind."""
+    urls = []
+    for m in messages:
+        parts = m.get("content")
+        for p in parts if isinstance(parts, list) else ():
+            kind = p.get("type") if isinstance(p, dict) else None
+            if kind in ("image", "input_image"):
+                raise ValueError(f'a "{kind}" part is another API\'s shape; send each image as'
+                                 ' {"type": "image_url", "image_url": {"url": "<http(s) or data: URL>"}}')
+            if kind == "image_url":
+                url = _url(p)
+                if not IMAGE_URL.match(url):
+                    raise ValueError("an image_url must be an http(s) URL or a data:image/...;base64, URL,"
+                                     f" not {placeholder(url) if url else 'empty'}")
+                urls.append(url)
+    return urls
+
+
+def placeholder(url: str) -> str:
+    """How an image is written down (the call log, offloads, the dashboard,
+    error text, the merger's copy of the request), never its bytes: a data
+    URL by its size, an http URL without its query (a signed URL's signature
+    is a credential) or a user:password."""
+    if url[:5].lower() == "data:":
+        return f"[image: data URL, {(len(url) * 3 // 4 + 999) // 1000:,} KB]"
+    u = urllib.parse.urlsplit(url)
+    return f"[image: {u.scheme}://{u.netloc.rpartition('@')[2]}{u.path[:200]}{'?...' if u.query else ''}]"
+
+
+def redact(text: str) -> str:
+    """`text` with each inline image as its placeholder, every http(s) URL
+    without its query or user:password, and any other long base64 run as its
+    size: an upstream may echo the request in an error, in whatever form, and
+    errors are logged, shown and learned from."""
+    def bare(m):
+        try:
+            u = urllib.parse.urlsplit(m.group())
+        except ValueError:                             # an unparseable netloc, such as "[::1"
+            return "[url]"
+        return f"{u.scheme}://{u.netloc.rpartition('@')[2]}{u.path}{'?...' if u.query else ''}"
+    text = URLS.sub(bare, DATA_URLS.sub(lambda m: placeholder(m.group()), text))
+    return BASE64.sub(lambda m: f"[base64, {(len(m.group()) * 3 // 4 + 999) // 1000:,} KB]", text)
+
+
+def cannot_see(alias: str) -> str:
+    """Why an image cannot be put to `alias`, or "" when it can. The catalog
+    decides (metrics/catalog.py: the model's input types, your Vision / OCR
+    override in its model popup, a capability seen working). A model the
+    catalog knows nothing about does not sit either: one that cannot see
+    answers anyway, describing a photo it never got, and the override is one
+    click."""
+    import catalog
+    import endpoints
+    if endpoints.entry_of(alias).get("api") == "responses":
+        return "reached through the Responses API, which the gateway sends no images to yet"
+    rec = catalog.info(alias)
+    if rec.get("vision"):
+        return ""
+    if rec.get("vision") is None:
+        return "the catalog does not say whether it takes images; tick Vision / OCR in its model popup if it does"
+    return f"does not take images (says {rec['sources'].get('vision', 'the catalog')})"
+
+
+def pick_shape(prompt: str) -> tuple[str, str]:
+    """(shape, how it was chosen). Falls back to general on any doubt."""
+    try:
+        held = set(json.loads((DATA / "config" / "model-roles.json").read_text(encoding="utf-8"))
+                   .get("ad_hoc", []))
+    except (OSError, ValueError, AttributeError):
+        return "general", "config/model-roles.json unreadable"
+    picker = next((a for a, d in decide.load().items()
+                   if "choice" in d["question_types"] and a not in held), None)
+    if not picker:
+        return "general", "no decision model onboarded"
+    question = {"shape": {"type": "choice",
+                          "instructions": "What kind of task does this prompt ask for?",
+                          "criteria": SHAPES}}
+    try:
+        answer = decide.ask(picker, prompt[:20000], question)["answers"]["shape"]
+    except Exception as exc:                           # noqa: BLE001
+        return "general", f"{picker} failed ({decide.error_text(exc)[:80]})"
+    conf = answer.get("confidence", 0)
+    if answer.get("choice") not in SHAPES or conf < CONFIDENT:
+        return "general", f"{picker} unsure ({answer.get('choice')} at {conf:.2f})"
+    return answer["choice"], f"{picker}, {conf:.2f}"
+
+
+def retired() -> dict[str, dict]:
+    """config/retired.json: alias -> its record (when, why). Retiring a model
+    takes it out of every collection (serve.remove_model); this is for a
+    collection that still names one -- a hand edit, or one saved before that
+    -- so the seat is skipped with a reason instead of failing on the call."""
+    try:
+        data = json.loads((DATA / "config" / "retired.json").read_text(encoding="utf-8"))
+    except (OSError, ValueError):
+        return {}
+    return data if isinstance(data, dict) else {}
+
+
+def retired_why(entry: dict) -> str:
+    when = str(entry.get("at", ""))[:10] if isinstance(entry, dict) else ""
+    return (f"retired from the bench{' on ' + when if when else ''}; take it out of the collection"
+            " (Restore in its model popup brings it back)")
+
+
+def retiring() -> set[str]:
+    """Aliases not to spend an explorer seat on: those on a retiring alert
+    (failing their probes; 2026-10-16 both explorers of a bench/core call were
+    and both failed), and those cooling down after a rate limit, until the
+    wait the provider stated, or 30 minutes (metrics/learn.py)."""
+    out: set[str] = set()
+    try:
+        import alerts
+        out |= {a.get("alias", "") for a in alerts.active().values() if a.get("kind") == "retiring"}
+    except Exception:                                  # noqa: BLE001
+        pass
+    try:
+        import learn
+        out |= set(learn.cooldowns())
+    except Exception:                                  # noqa: BLE001
+        pass
+    return out
+
+
+def ask_seat(alias: str, system: str, prompt: str, messages: list[dict] | None = None) -> dict:
+    """One seat's answer. `messages`: the request as sent, when it carries
+    images; it goes through the built-in client whatever BENCH_CLIENT says,
+    since `llm -m` takes a prompt, not inline images."""
+    cmd = ["llm", "-m", alias] + (["-s", system] if system else [])
+    started = time.monotonic()
+    if messages is not None or client.backend(alias) == "builtin":
+        # No subprocess, and no shared llm log for parallel seats to lock.
+        res = client.chat(alias, prompt, system, SEAT_TIMEOUT, messages=messages)
+        answer = res["text"]
+        error = res["error"] or (f"no answer in {SEAT_TIMEOUT}s" if res["timed_out"] else "")
+    else:
+        answer, error = _ask_llm(cmd, prompt)
+    # A quota or credit notice delivered as the reply's text, with a success
+    # exit, is not an answer (suite/run.py PROVIDER_NOTICES).
+    from run import provider_notice
+    if not error and (notice := provider_notice(answer)):
+        error, answer = f"provider notice in a success reply ({notice}): {answer[:200]}", ""
+    if messages is not None:
+        # Before the error is cut, logged or learned from. ponytail: the answer
+        # is not redacted; a seat gets an image as pixels, not as its URL or bytes.
+        error = redact(error)
+    if error:
+        # A stated context window, rate limit or wait is kept (metrics/learn.py);
+        # a rate limit also cools the model, so explorer seats skip it.
+        try:
+            import learn
+            learn.record(alias, error, "gateway")
+        except Exception:                              # noqa: BLE001
+            pass
+    status = "error" if error else ("empty" if not answer else "ok")
+    return {"status": status, "answer": answer, "error": error[-500:],
+            "seconds": round(time.monotonic() - started, 1)}
+
+
+def _ask_llm(cmd: list[str], prompt: str) -> tuple[str, str]:
+    """(answer, error) from the `llm` CLI."""
+    try:
+        done = subprocess.run(cmd, input=prompt, capture_output=True, text=True,
+                              encoding="utf-8", errors="replace", timeout=SEAT_TIMEOUT,
+                              env={**os.environ, "PYTHONIOENCODING": "utf-8"})
+        answer = done.stdout.strip()
+        # A non-zero exit is an error even when stderr is silent, unless it is
+        # only llm failing to log an answer it printed (client.py).
+        failed = done.returncode and not client.llm_log_failed_only(done.stdout, done.stderr)
+        error = (done.stderr.strip() or f"llm exited {done.returncode}") if failed else ""
+    except subprocess.TimeoutExpired:
+        answer, error = "", f"no answer in {SEAT_TIMEOUT}s"
+    except Exception as exc:                           # noqa: BLE001
+        # One seat failing to start must not lose the other seats' records.
+        answer, error = "", f"{type(exc).__name__}: {exc}"
+    return answer, error
+
+
+class TooLarge(ValueError):
+    """The prompt does not fit a member's context window (HTTP 413)."""
+
+
+def estimate_tokens(text: str) -> int:
+    """Characters / 4 plus 10%: tokenizers differ by family, so this errs high.
+    The windows it is compared with are corrected by learned 'maximum context
+    length' errors (metrics/learn.py). An image adds IMAGE_TOKENS whatever its
+    size, which errs high too: OpenAI's, Anthropic's and Gemini's models spend
+    up to about 1,600 on one at their usual detail. Qwen and Llama vision
+    models can spend more on a very large image; a learned error corrects that."""
+    return int(len(text) / 4 * 1.1)
+
+
+def too_small(needed: int, aliases) -> dict[str, int]:
+    """Aliases whose KNOWN context window is under `needed`, with the window."""
+    import catalog
+    out = {}
+    for a in aliases:
+        cw = catalog.info(a).get("context_window")
+        if cw and cw < needed:
+            out[a] = cw
+    return out
+
+
+def record_seat(task: str, call: str, label: str, alias: str, r: dict) -> dict:
+    """Log one answer as an unjudged offload, the moment it is in, so explorer
+    rotation counts it even if nobody judges; returns the seat as logged."""
+    offload.add(alias, task, None, f"{label}; {r['status']} in {r['seconds']}s"
+                + (f": {r['error'][:120]}" if r["error"] else ""), call)
+    return {"call": call, "seat": label, "alias": alias, **r}
+
+
+def merger_candidates(collection: dict, members: list[str], shape: str, large: bool):
+    """Who may merge, best first. The collection's `merger` alone when set;
+    else the members with verdicts on this shape, most useful first, then the
+    router's primary, then the other members (a fresh install has neither
+    verdicts nor suite scores). A generator, so the router is asked only if
+    no member with a record fits."""
+    if collection.get("merger"):
+        yield collection["merger"]
+        return
+    record: dict[str, list[int]] = {}
+    for e in offload.entries():
+        # The shape ends each seat's task as run_collection writes it: "(bench/<name>, <shape>)".
+        if e["model"] in members and e.get("useful") is not None and e.get("task", "").endswith(f", {shape})"):
+            r = record.setdefault(e["model"], [0, 0])
+            r[0], r[1] = r[0] + bool(e["useful"]), r[1] + 1
+    # Useful verdicts counted, not a ratio, as route.py counts findings: one
+    # lucky verdict must not outrank a long good record.
+    yield from sorted(record, key=lambda a: (-record[a][0], -record[a][0] / record[a][1]))
+    try:
+        primary = route.panel(shape, large=large)["primary"]
+    except Exception:                                  # noqa: BLE001 -- the members can still merge
+        primary = None
+    if primary:
+        yield primary["model"]
+    yield from members
+
+
+def merge(collection: dict, members: list[str], shape: str, whole: str, seats: list[dict],
+          task: str, panel_id: str) -> tuple[dict | None, str]:
+    """The merger's seat, and why the reply falls back to every answer ("" if
+    it does not). The merger reads the request and every answer, so it gets
+    the members' size check; a failed merger is still logged like a seat."""
+    answered = [s for s in seats if s["status"] == "ok"]
+    if not answered:
+        return None, "no seat answered"
+    text = f"THE REQUEST\n\n{whole}\n\nTHE ANSWERS\n\n" + "\n\n".join(
+        f"## {s['seat']} — {s['alias']}\n\n{s['answer']}" for s in answered)
+    needed = estimate_tokens(f"{MERGE_SYSTEM}\n\n{text}")
+    alias = next((a for a in merger_candidates(collection, members, shape, len(text) > LARGE)
+                  if not too_small(needed, [a])), None)
+    if alias is None:
+        return None, f"no merger has the context window for the merge (about {needed:,} tokens)"
+    seat = record_seat(task, f"{panel_id}.merger", "merger", alias, ask_seat(alias, MERGE_SYSTEM, text))
+    if seat["status"] != "ok":
+        return seat, f"the merger {alias} came back {seat['status']}" + (f": {seat['error'][:120]}" if seat["error"] else "")
+    return seat, ""
+
+
+def run_collection(name: str, messages: list[dict], nominated: list[str] | None = None,
+                   reply: str = "", who: str = "local") -> tuple[str, dict]:
+    """The panel's reply as markdown, and the same seats as data for the
+    response's `bench` field. `reply`: the request's mode, "" for the
+    collection's own. `who`: the gateway key that asked, for the call log."""
+    pictures = images(messages)
+    system, prompt = flatten(messages)
+    if not prompt.strip():
+        raise ValueError("no user message to put to the collection")
+    collection = bench_collections.load()[name]
+    s = bench_collections.settings()
+    raw = collection.get("members") or []
+    # The editor dedupes and caps; a hand-edited file may not have.
+    members = list(dict.fromkeys(m for m in raw if isinstance(m, str) and m))
+    # A RETIRED MODEL IS NOT CALLED: it is no longer registered, so the call
+    # could only fail ("Unknown model"). Skipped, and named, before the cap so
+    # it frees its seat for the next member or an explorer. Retiring takes it
+    # out of the collection, so this is a collection that still names one.
+    gone = retired()
+    skipped = [{"seat": "member", "alias": a, "why": retired_why(gone[a])} for a in members if a in gone]
+    members = [a for a in members if a not in gone][:s["max_calls"]]
+    # A member named in `with` is already seated; it must not use up the slot.
+    nominated = [a for a in dict.fromkeys(nominated or []) if a not in members]
+    skipped += [{"seat": "with", "alias": a, "why": retired_why(gone[a])} for a in nominated
+                if a in gone and a not in {k["alias"] for k in skipped}]
+    nominated = [a for a in nominated if a not in gone]
+    merged = (reply or collection.get("reply") or "all") == "merged"
+    if collection.get("merger") in gone:
+        if merged:
+            skipped.append({"seat": "merger", "alias": collection["merger"],
+                            "why": retired_why(gone[collection["merger"]]) + "; the merger is chosen by verdicts"})
+        collection = {**collection, "merger": ""}
+    n_explorers = bench_collections.explorers_for(len(members), s)
+    # AN IMAGE GOES ONLY TO A MODEL THAT CAN SEE IT: one that cannot would
+    # answer anyway, about a picture it never got. Skipped, and named.
+    skipped += [{"seat": seat, "alias": a, "why": why} for seat, group in (("member", members), ("with", nominated))
+                for a in group if pictures and (why := cannot_see(a))]
+    unseen = {k["alias"] for k in skipped}
+    sitting = [a for a in members if a not in unseen]
+    nominated = [a for a in nominated if a not in unseen]
+    # The system text goes to every seat too, so it counts for shape and size.
+    whole = f"{system}\n\n{prompt}".strip()
+    # REFUSE WHAT A MEMBER CANNOT READ, so the caller can trim the prompt or
+    # pick another collection instead of collecting a seat's context error.
+    needed = estimate_tokens(whole) + IMAGE_TOKENS * len(pictures)
+    small = too_small(needed, sitting)
+    if small:
+        windows = ", ".join(f"{a} {cw:,}" for a, cw in sorted(small.items(), key=lambda kv: kv[1]))
+        counted = f" (each image counted as {IMAGE_TOKENS:,})" if pictures else ""
+        raise TooLarge(f"prompt is about {needed:,} tokens{counted}; too large for {windows} (context"
+                       f" windows) in bench/{name}. Trim it under {min(small.values()):,} tokens,"
+                       " or choose a collection whose members have larger windows"
+                       " (GET /v1/models lists each collection's context_window).")
+    if pictures and not collection.get("shape"):
+        shape, how = "vision", "the request carries images"
+    elif not n_explorers:
+        shape, how = collection.get("shape") or "general", "no explorer seats to fill"
+    elif collection.get("shape"):
+        shape, how = collection["shape"], "set on the collection"
+    else:
+        shape, how = pick_shape(whole)
+    seats = [(f"member{i}", "member", alias) for i, alias in enumerate(sitting, 1)]
+    if n_explorers:
+        # `with` in the request body is route.py's --with: an explorer seat.
+        from run import _registered_names
+        avoid = retiring() | set(too_small(needed, _registered_names()))   # explorers must fit too
+        if pictures:
+            avoid |= {a for a in _registered_names() if cannot_see(a)}    # and see
+        p = route.panel(shape, large=len(whole) > LARGE, nominated=nominated or None,
+                        members=sitting, explorers=n_explorers, avoid=avoid)
+        seats += [(key, label, p[key]["model"]) for key, label in
+                  (("explorer", "explorer"), ("explorer2", "explorer 2")) if p.get(key)]
+    if not seats and pictures:
+        raise ValueError(f"no seat in bench/{name} can see images, so none is asked rather than answer"
+                         " blind: " + ("; ".join(f"{k['alias']} {k['why']}" for k in skipped) or "no members")
+                         + ", and no model that takes images was free for an explorer seat. Add one to the"
+                         " collection, or call one whose together.capabilities on GET /v1/models lists vision.")
+    if not seats:
+        raise LookupError(f"collection {name!r} has no members and no explorer could be seated"
+                          + "".join(f"; {k['alias']} not asked: {k['why']}" for k in skipped))
+    panel_id = f"p{dt.datetime.now():%Y%m%d-%H%M%S}-{secrets.token_hex(2)}"
+    model = bench_collections.PREFIX + name
+    with ThreadPoolExecutor(max_workers=len(seats)) as pool:
+        results = list(pool.map(lambda s: ask_seat(s[2], system, prompt, messages if pictures else None), seats))
+
+    task = " ".join(prompt.split())[:90] + f" ({model}, {shape})"
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    record = {"panel": panel_id, "at": dt.datetime.now().astimezone().isoformat(timespec="seconds"),
+              "model": model, "shape": shape, "shape_by": how, "key": who, "system": system,
+              "prompt": prompt, "seats": [], **({"skipped": skipped} if skipped else {})}
+    out = [f"Bench {model} {panel_id} · {len(sitting)} members + {len(seats) - len(sitting)}"
+           f" explorers · explorer shape {shape} ({how})\n"]
+    if skipped:
+        out.append("Not asked: " + "; ".join(f"{k['alias']} ({k['why']})" for k in skipped) + "\n")
+    for (key, label, alias), r in zip(seats, results):
+        call = f"{panel_id}.{key}"
+        record["seats"].append(record_seat(task, call, label, alias, r))
+        body = r["answer"] or (f"_error: {r['error'][:300]}_" if r["error"] else
+                               "_empty reply — usually a token budget, not incapacity_")
+        out.append(f"## {label} — {alias}\n`{call}` · {r['status']} · {r['seconds']}s\n\n{body}\n")
+    out.append("---\nRecord a verdict per seat once checked against the code:\n"
+               "`POST /bench/verdict` on this same server (see `bench.verdict`), or"
+               " `python metrics/offload.py verdict <call id> --useful|--not-useful --note \"...\"`")
+    content = "\n".join(out)
+    # Beyond the OpenAI fields, so a caller reads seats without parsing the
+    # markdown. Clients that do not know it ignore it.
+    bench = {"panel": panel_id, "collection": model, "shape": shape, "shape_by": how,
+             "seats": [{k: s[k] for k in SEAT_KEYS} for s in record["seats"]],
+             # A path, not a URL: the same server answered, whichever port or
+             # proxy the caller came through (the dashboard serves these routes too).
+             "verdict": 'POST /bench/verdict {"call", "useful": true|false, "note", "real": [...], "false": [...]}',
+             **({"skipped": skipped} if skipped else {})}
+    if merged:
+        try:
+            merger, note = merge(collection, members, shape, whole, record["seats"], task, panel_id)
+        except Exception as exc:                       # noqa: BLE001 -- never lose the panel to the merge
+            merger, note = None, f"the merge failed ({type(exc).__name__}: {exc})"
+        if merger:
+            bench["merger"] = {k: merger[k] for k in SEAT_KEYS}
+            record["seats"].append(merger)             # the dashboard shows it, verdict and all
+        if note:
+            bench.update(reply="all", note=note)
+            content = f"_Merged reply unavailable: {note}. Every answer follows._\n\n{content}"
+        else:
+            who = ", ".join(f"{s['alias']} {s['status']}" for s in bench["seats"])
+            bench["reply"] = "merged"
+            content = (f"{merger['answer']}\n\n---\nMerged by {merger['alias']} (`{merger['call']}`) from:"
+                       f" {who}. Every seat's answer and call id is in the response's `bench` field.")
+    with (LOG_DIR / "calls.jsonl").open("a", encoding="utf-8") as fh:
+        fh.write(json.dumps(record, ensure_ascii=False) + "\n")
+    return content, bench
+
+
+def models() -> list[str]:
+    return [bench_collections.PREFIX + n for n in bench_collections.load()]
+
+
+class Handler(BaseHTTPRequestHandler):
+    """The gateway's routes. serve.py's Handler subclasses this to answer them
+    on the dashboard's port too, so the checks use the port actually served."""
+    who = "-"
+
+    def _json(self, code: int, data: dict, headers: dict | None = None) -> None:
+        body = json.dumps(data).encode()
+        self.send_response(code)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(body)))
+        for k, v in (headers or {}).items():
+            self.send_header(k, v)
+        self.end_headers()
+        self.wfile.write(body)
+
+    def _error(self, code: int, message: str, headers: dict | None = None) -> None:
+        self._json(code, {"error": {"message": message, "type": "bench_gateway", "code": code}}, headers=headers)
+
+    def end_headers(self):
+        # CORS for a browser app on an allowed origin, on the gateway's routes
+        # only (metrics/guard.py); Vary, so a cache never hands one origin's
+        # answer to another. A request line too broken to parse has no headers.
+        headers = getattr(self, "headers", None)
+        if headers is not None and guard.gateway_route((getattr(self, "path", "") or "").split("?", 1)[0]) \
+                and (origin := guard.cors_origin(headers, self.server.server_port)):
+            self.send_header("Access-Control-Allow-Origin", origin)
+            self.send_header("Vary", "Origin")
+        super().end_headers()
+
+    def do_OPTIONS(self):                              # noqa: N802
+        """A browser's CORS preflight: answered on the gateway's routes, for an
+        allowed origin only. The dashboard's routes answer none, as before."""
+        if not guard.gateway_route(self.path.split("?", 1)[0]):
+            return self.send_error(501, f"Unsupported method ({self.command!r})")
+        try:
+            guard.check(self.headers, self.server.server_port, post=False, max_body=0)
+        except guard.Refused as r:
+            return self._error(r.status, str(r))
+        if not guard.cors_origin(self.headers, self.server.server_port):
+            return self._error(403, f"Origin {self.headers.get('Origin') or '(none)'!r} is not allowed;"
+                                    " set BENCH_ALLOWED_ORIGINS or BENCHWARMERS_PUBLIC_URL to add one")
+        # The headers it asks for as well, when they are plain names: the
+        # OpenAI JS SDK sends X-Stainless-* on every call, and a preflight that
+        # refused them would block it. The origin is the boundary, not these.
+        asked = (self.headers.get("Access-Control-Request-Headers") or "").strip()
+        self.send_response(204)
+        self.send_header("Access-Control-Allow-Methods", "GET, POST, OPTIONS")
+        self.send_header("Access-Control-Allow-Headers", "Authorization, Content-Type"
+                         + (f", {asked}" if HEADER_NAMES.fullmatch(asked) else ""))
+        self.send_header("Access-Control-Max-Age", "600")
+        self.end_headers()
+
+    def do_GET(self):                                  # noqa: N802
+        try:
+            guard.check(self.headers, self.server.server_port, post=False, max_body=0)
+            self.who = guard.gateway_caller(self.headers)
+        except guard.Refused as r:
+            return self._error(r.status, str(r), r.headers)
+        if self.path.split("?", 1)[0].rstrip("/") in ("/v1/models", "/models"):
+            # Beyond the OpenAI fields: what each collection is for, so an
+            # orchestrator can choose one by task. Clients ignore extra keys.
+            # Each collection twice: as set, and in the other reply mode, so a
+            # client that offers only listed models can still pick either.
+            s, data = bench_collections.settings(), bench_collections.load()
+            rows = []
+            for n, c in data.items():
+                if not guard.may_call(self.who, n):    # a key limited to some collections sees those
+                    continue
+                d = bench_collections.described(n, c, s)
+                other = "all" if d["reply"] == "merged" else "merged"
+                rows += [{"id": bench_collections.PREFIX + n, "object": "model", "owned_by": "benchwarmers", **d},
+                         {"id": f"{bench_collections.PREFIX}{n}:{other}", "object": "model",
+                          "owned_by": "benchwarmers", **d, "reply": other}]
+            self._json(200, {"object": "list", "data": rows})
+        else:
+            self._error(404, "GET /v1/models is the only GET here")
+
+    def do_POST(self):                                 # noqa: N802
+        # Without this any web page could spend model calls, seat ad hoc
+        # models through `with`, or write verdicts (metrics/guard.py).
+        try:
+            length = guard.check(self.headers, self.server.server_port, post=True, max_body=MAX_BODY, cors=True)
+            self.who = guard.gateway_caller(self.headers)
+        except guard.Refused as r:
+            return self._error(r.status, str(r) + (TOO_BIG if r.status == 413 else ""), r.headers)
+        try:
+            body = json.loads(self.rfile.read(length) or b"{}")
+        except ValueError:
+            return self._error(400, "body is not JSON")
+        if not isinstance(body, dict):
+            return self._error(400, "body must be a JSON object")
+        path = self.path.split("?", 1)[0].rstrip("/")
+        if path == "/bench/verdict":
+            if not isinstance(body.get("useful"), bool):
+                return self._error(400, '"useful" must be true or false')
+            lists = {k: body.get(k) or [] for k in ("real", "false")}
+            if not all(isinstance(v, list) and all(isinstance(s, str) for s in v) for v in lists.values()):
+                return self._error(400, '"real" and "false" must be lists of one-sentence claims')
+            try:
+                e = offload.verdict(str(body.get("call", "")), body["useful"], str(body.get("note") or ""),
+                                    lists["real"], lists["false"])
+            except KeyError as exc:
+                return self._error(404, exc.args[0])
+            return self._json(200, {"ok": True, "recorded": e})
+        if path not in ("/v1/chat/completions", "/chat/completions"):
+            return self._error(404, "POST /v1/chat/completions or /bench/verdict")
+        model = body.get("model") or ""
+        if not isinstance(model, str):
+            return self._error(400, '"model" must be a collection name such as bench/core')
+        model, suffix = bench_collections.split_reply(model)
+        name = bench_collections.resolve(model)
+        if name is None:
+            return self._error(400 if not model else 404, bench_collections.choose_message(model))
+        model = bench_collections.PREFIX + name
+        messages, nominated, opts = body.get("messages"), body.get("with") or [], body.get("bench") or {}
+        if not (isinstance(messages, list) and all(isinstance(m, dict) for m in messages)):
+            return self._error(400, '"messages" must be a list of message objects')
+        if not (isinstance(nominated, list) and all(isinstance(a, str) for a in nominated)):
+            return self._error(400, '"with" must be a list of model aliases')
+        if not isinstance(opts, dict) or opts.get("reply", "all") not in bench_collections.REPLIES:
+            return self._error(400, '"bench" must be an object whose "reply" is "all" or "merged"')
+        if not guard.may_call(self.who, name):
+            allowed = ", ".join(bench_collections.PREFIX + c for c in guard.scope(self.who) or []) or "none"
+            return self._error(403, f"the gateway key {self.who} may not call {model}; it may call {allowed}")
+        try:
+            guard.spend(self.who)                      # counted once it names a collection it may call
+        except guard.Refused as r:
+            return self._error(r.status, str(r), r.headers)
+        try:
+            # The body is the most specific, then the suffix, then the collection.
+            content, bench = run_collection(name, messages, nominated, opts.get("reply") or suffix, self.who)
+        except TooLarge as exc:
+            return self._error(413, str(exc))
+        except ValueError as exc:
+            return self._error(400, str(exc))
+        except LookupError as exc:
+            return self._error(503, str(exc))
+        reply_id, created = f"chatcmpl-{secrets.token_hex(8)}", int(time.time())
+        if not body.get("stream"):
+            return self._json(200, {
+                "id": reply_id, "object": "chat.completion", "created": created, "model": model,
+                "choices": [{"index": 0, "finish_reason": "stop",
+                             "message": {"role": "assistant", "content": content}}], "bench": bench})
+        # A panel finishes all at once, so a streaming client gets it as one
+        # chunk, merged or not. ponytail: nothing is sent until the panel and
+        # the merge are done, so a client or proxy with a shorter idle timeout
+        # drops it. Keep-alives mean sending the 200 before the panel runs, so
+        # the 413/503 refusals move into the stream; streaming the merge as it
+        # is written needs a streaming client.chat (RELEASE-PLAN acceptance 11).
+        self.send_response(200)
+        self.send_header("Content-Type", "text/event-stream")
+        self.end_headers()
+        for delta, finish in (({"role": "assistant", "content": content}, None), ({}, "stop")):
+            chunk = {"id": reply_id, "object": "chat.completion.chunk", "created": created,
+                     "model": model, "choices": [{"index": 0, "delta": delta, "finish_reason": finish}],
+                     **({"bench": bench} if finish is None else {})}
+            self.wfile.write(f"data: {json.dumps(chunk)}\n\n".encode())
+        self.wfile.write(b"data: [DONE]\n\n")
+
+    def log_request(self, code="-", size="-"):         # route, status, who; never a query or a key
+        guard.access(self, code)
+
+    def log_message(self, fmt, *args):                 # quieter than the default
+        sys.stderr.write(f"{self.log_date_time_string()} {fmt % args}\n")
+
+
+def main() -> None:
+    sys.stdout.reconfigure(encoding="utf-8", errors="replace")
+    address = guard.bind()
+    guard.startup(address)
+    print(f"bench gateway on http://{address}:{PORT}/v1 — models: {', '.join(models())}")
+    ThreadingHTTPServer((address, PORT), Handler).serve_forever()
+
+
+if __name__ == "__main__":
+    main()